For users to load in one key.

BiRefNet_codes

@@ -1 +0,0 @@
-Subproject commit 6921f57da442c87e2020bf2aea2cee85527be482

BiRefNet_github/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 ZhengPeng
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

BiRefNet_github/README.md

@@ -0,0 +1,234 @@
+# <p align=center>`Bilateral Reference for High-Resolution Dichotomous Image Segmentation`</p>
+
+|            *DIS-Sample_1*        |             *DIS-Sample_2*        |
+| :------------------------------: | :-------------------------------: |
+| <img src="https://drive.google.com/thumbnail?id=1ItXaA26iYnE8XQ_GgNLy71MOWePoS2-g&sz=w400" /> |  <img src="https://drive.google.com/thumbnail?id=1Z-esCujQF_uEa_YJjkibc3NUrW4aR_d4&sz=w400" /> |
+
+This repo is the official implementation of "[**Bilateral Reference for High-Resolution Dichotomous Image Segmentation**](https://arxiv.org/pdf/2401.03407.pdf)" (___arXiv 2024___).
+
+> **Authors:**
+> [Peng Zheng](https://scholar.google.com/citations?user=TZRzWOsAAAAJ),
+> [Dehong Gao](https://scholar.google.com/citations?user=0uPb8MMAAAAJ),
+> [Deng-Ping Fan](https://scholar.google.com/citations?user=kakwJ5QAAAAJ),
+> [Li Liu](https://scholar.google.com/citations?user=9cMQrVsAAAAJ),
+> [Jorma Laaksonen](https://scholar.google.com/citations?user=qQP6WXIAAAAJ),
+> [Wanli Ouyang](https://scholar.google.com/citations?user=pw_0Z_UAAAAJ), &
+> [Nicu Sebe](https://scholar.google.com/citations?user=stFCYOAAAAAJ).
+
+[[**arXiv**](https://arxiv.org/abs/2401.03407)] [[**code**](https://github.com/ZhengPeng7/BiRefNet)] [[**stuff**](https://drive.google.com/drive/folders/1s2Xe0cjq-2ctnJBR24563yMSCOu4CcxM)]  [[**中文版**](https://drive.google.com/file/d/1aBnJ_R9lbnC2dm8dqD0-pzP2Cu-U1Xpt/view?usp=drive_link)] 
+
+Our BiRefNet has achieved SOTA on many similar HR tasks:
+
+**DIS**: [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/dichotomous-image-segmentation-on-dis-te1)](https://paperswithcode.com/sota/dichotomous-image-segmentation-on-dis-te1?p=bilateral-reference-for-high-resolution) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/dichotomous-image-segmentation-on-dis-te2)](https://paperswithcode.com/sota/dichotomous-image-segmentation-on-dis-te2?p=bilateral-reference-for-high-resolution) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/dichotomous-image-segmentation-on-dis-te3)](https://paperswithcode.com/sota/dichotomous-image-segmentation-on-dis-te3?p=bilateral-reference-for-high-resolution) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/dichotomous-image-segmentation-on-dis-te4)](https://paperswithcode.com/sota/dichotomous-image-segmentation-on-dis-te4?p=bilateral-reference-for-high-resolution) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/dichotomous-image-segmentation-on-dis-vd)](https://paperswithcode.com/sota/dichotomous-image-segmentation-on-dis-vd?p=bilateral-reference-for-high-resolution)
+
+<details><summary>Figure of Comparison on Papers with Codes (by the time of this work):</summary><p>
+<img src="https://drive.google.com/thumbnail?id=1DLt6CFXdT1QSWDj_6jRkyZINXZ4vmyRp&sz=w1620" />
+<img src="https://drive.google.com/thumbnail?id=1gn5GyKFlJbMIkre1JyEdHDSYcrFmcLD0&sz=w1620" />
+<img src="https://drive.google.com/thumbnail?id=16CVYYOtafEeZhHqv0am2Daku1n_exMP6&sz=w1620" />
+<img src="https://drive.google.com/thumbnail?id=10K45xwPXmaTG4Ex-29ss9payA9yBnyLn&sz=w1620" />
+<img src="https://drive.google.com/thumbnail?id=16EuyqKFJOqwMmagvfnbC9hUurL9pYLLB&sz=w1620" />
+</details>
+<br />
+
+**COD**:[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/camouflaged-object-segmentation-on-cod)](https://paperswithcode.com/sota/camouflaged-object-segmentation-on-cod?p=bilateral-reference-for-high-resolution) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/camouflaged-object-segmentation-on-nc4k)](https://paperswithcode.com/sota/camouflaged-object-segmentation-on-nc4k?p=bilateral-reference-for-high-resolution) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/camouflaged-object-segmentation-on-camo)](https://paperswithcode.com/sota/camouflaged-object-segmentation-on-camo?p=bilateral-reference-for-high-resolution) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/camouflaged-object-segmentation-on-chameleon)](https://paperswithcode.com/sota/camouflaged-object-segmentation-on-chameleon?p=bilateral-reference-for-high-resolution)
+
+<details><summary>Figure of Comparison on Papers with Codes (by the time of this work):</summary><p>
+<img src="https://drive.google.com/thumbnail?id=1DLt6CFXdT1QSWDj_6jRkyZINXZ4vmyRp&sz=w1620" />
+<img src="https://drive.google.com/thumbnail?id=1gn5GyKFlJbMIkre1JyEdHDSYcrFmcLD0&sz=w1620" />
+<img src="https://drive.google.com/thumbnail?id=16CVYYOtafEeZhHqv0am2Daku1n_exMP6&sz=w1620" />
+</details>
+<br />
+
+**HRSOD**: [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/rgb-salient-object-detection-on-davis-s)](https://paperswithcode.com/sota/rgb-salient-object-detection-on-davis-s?p=bilateral-reference-for-high-resolution) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/rgb-salient-object-detection-on-hrsod)](https://paperswithcode.com/sota/rgb-salient-object-detection-on-hrsod?p=bilateral-reference-for-high-resolution) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/rgb-salient-object-detection-on-uhrsd)](https://paperswithcode.com/sota/rgb-salient-object-detection-on-uhrsd?p=bilateral-reference-for-high-resolution) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/salient-object-detection-on-duts-te)](https://paperswithcode.com/sota/salient-object-detection-on-duts-te?p=bilateral-reference-for-high-resolution) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bilateral-reference-for-high-resolution/salient-object-detection-on-dut-omron)](https://paperswithcode.com/sota/salient-object-detection-on-dut-omron?p=bilateral-reference-for-high-resolution)
+
+<details><summary>Figure of Comparison on Papers with Codes (by the time of this work):</summary><p>
+<img src="https://drive.google.com/thumbnail?id=1hNfQtlTAHT4-AVbk_47852zyRp1NOFLs&sz=w1620" />
+<img src="https://drive.google.com/thumbnail?id=1bcVldUAxYkMI3OMTyaP_jNuOugDfYj-d&sz=w1620" />
+<img src="https://drive.google.com/thumbnail?id=1p1zgyVz27cGEqQMtOKzm_6zoYK3Sw_Zk&sz=w1620" />
+<img src="https://drive.google.com/thumbnail?id=1TubAvcoEbH_mHu3I-AxflnB71nkf35jJ&sz=w1620" />
+<img src="https://drive.google.com/thumbnail?id=1A3V9HjVtcMQdnGPwuy-DBVhwKuo0q2lT&sz=w1620" />
+</details>
+<br />
+
+#### Try our online demos for inference:
+
++ **Inference and evaluation** of your given weights: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1MaEiBfJ4xIaZZn0DqKrhydHB8X97hNXl#scrollTo=DJ4meUYjia6S)
++ **Online Inference with GUI** with adjustable resolutions: [![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/ZhengPeng7/BiRefNet_demo)  
++ Online **Single Image Inference** on Colab: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/14Dqg7oeBkFEtchaHLNpig2BcdkZEogba?usp=drive_link)
+<img src="https://drive.google.com/thumbnail?id=12XmDhKtO1o2fEvBu4OE4ULVB2BK0ecWi&sz=w1620" />
+
+
+
+## Model Zoo
+
+> For more general use of our BiRefNet, I managed to extend the original adademic one to more general ones for better application in real life.
+>
+> Datasets and datasets are suggested to download from official pages. But you can also download the packaged ones: [DIS](https://drive.google.com/drive/folders/1hZW6tAGPJwo9mPS7qGGGdpxuvuXiyoMJ?usp=drive_link), [HRSOD](https://drive.google.com/drive/folders/18_hAE3QM4cwAzEAKXuSNtKjmgFXTQXZN?usp=drive_link), [COD](https://drive.google.com/drive/folders/1EyHmKWsXfaCR9O0BiZEc3roZbRcs4ECO?usp=drive_link), [Backbones](https://drive.google.com/drive/folders/1cmce_emsS8A5ha5XT2c_CZiJzlLM81ms?usp=drive_link).
+>
+> Find performances (almost all metrics) of all models in the `exp-TASK_SETTINGS` folders in [[**stuff**](https://drive.google.com/drive/folders/1s2Xe0cjq-2ctnJBR24563yMSCOu4CcxM)].
+
+
+
+<details><summary>Models in the original paper, for <b>comparison on benchmarks</b>:</summary><p>
+
+| Task  |        Training Sets        |   Backbone    |                           Download                           |
+| :---: | :-------------------------: | :-----------: | :----------------------------------------------------------: |
+|  DIS  |          DIS5K-TR           | swin_v1_large | [google-drive](https://drive.google.com/file/d/1J90LucvDQaS3R_-9E7QUh1mgJ8eQvccb/view?usp=drive_link) |
+|  COD  |     COD10K-TR, CAMO-TR      | swin_v1_large | [google-drive](https://drive.google.com/file/d/1tM5M72k7a8aKF-dYy-QXaqvfEhbFaWkC/view?usp=drive_link) |
+| HRSOD |           DUTS-TR           | swin_v1_large | [google-drive](https://drive.google.com/file/d/1f7L0Pb1Y3RkOMbqLCW_zO31dik9AiUFa/view?usp=drive_link) |
+| HRSOD |          HRSOD-TR           | swin_v1_large |                         google-drive                         |
+| HRSOD |          UHRSD-TR           | swin_v1_large |                         google-drive                         |
+| HRSOD |      DUTS-TR, HRSOD-TR      | swin_v1_large | [google-drive](https://drive.google.com/file/d/1WJooyTkhoDLllaqwbpur_9Hle0XTHEs_/view?usp=drive_link) |
+| HRSOD |      DUTS-TR, UHRSD-TR      | swin_v1_large | [google-drive](https://drive.google.com/file/d/1Pu1mv3ORobJatIuUoEuZaWDl2ylP3Gw7/view?usp=drive_link) |
+| HRSOD |     HRSOD-TR, UHRSD-TR      | swin_v1_large | [google-drive](https://drive.google.com/file/d/1xEh7fsgWGaS5c3IffMswasv0_u-aVM9E/view?usp=drive_link) |
+| HRSOD | DUTS-TR, HRSOD-TR, UHRSD-TR | swin_v1_large | [google-drive](https://drive.google.com/file/d/13FaxyyOwyCddfZn2vZo1xG1KNZ3cZ-6B/view?usp=drive_link) |
+
+</details>
+
+
+
+<details><summary>Models trained with customed data (massive, portrait), for <b>general use in practical application</b>:</summary>
+
+|           Task            |                        Training Sets                         |   Backbone    | Test Set  | Metric (S, wF[, HCE]) |                           Download                           |
+| :-----------------------: | :----------------------------------------------------------: | :-----------: | :-------: | :-------------------: | :----------------------------------------------------------: |
+|      **general use**      | DIS5K-TR,DIS-TEs, DUTS-TR_TE,HRSOD-TR_TE,UHRSD-TR_TE, HRS10K-TR_TE | swin_v1_large |  DIS-VD   |  0.889, 0.840, 1152   | [google-drive](https://drive.google.com/file/d/1KRVE-U3OHrUuuFPY4FFdE4eYBeHJSA0H/view?usp=drive_link) |
+|      **general use**      | DIS5K-TR,DIS-TEs, DUTS-TR_TE,HRSOD-TR_TE,UHRSD-TR_TE, HRS10K-TR_TE | swin_v1_tiny  |  DIS-VD   |  0.867, 0.809, 1182   | [Google-drive](https://drive.google.com/file/d/16gDZISjNp7rKi5vsJm6_fbYF8ZBK8AoF/view?usp=drive_link) |
+|      **general use**      |                      DIS5K-TR, DIS-TEs                       | swin_v1_large |  DIS-VD   |  0.907, 0.865, 1059   | [google-drive](https://drive.google.com/file/d/1P6NJzG3Jf1sl7js2q1CPC3yqvBn_O8UJ/view?usp=drive_link) |
+| **portrait segmentation** |                           P3M-10k                            | swin_v1_large | P3M-500-P |     0.982, 0.990      | [google-drive](https://drive.google.com/file/d/1vrjPoOGj05iSxb4MMeznX5k67VlyfZX5/view?usp=drive_link) |
+
+</details>
+
+
+
+<details><summary>Segmentation with box <b>guidance</b>:</summary>
+
+​	*In progress...*
+
+</details>
+
+
+
+<details><summary>Model <b>efficiency</b>:</summary><p>
+
+> Screenshot from the original paper. All tests are conducted on a single A100 GPU.
+
+<img src="https://drive.google.com/thumbnail?id=1mTfSD_qt-rFO1t8DRQcyIa5cgWLf1w2-&sz=h300" />  <img src="https://drive.google.com/thumbnail?id=1F_OURIWILVe4u1rSz-aqt6ur__bAef25&sz=h300" />
+
+</details>
+
+
+
+## Third-Party Creations
+
+> Concerning edge devices with less computing power, we provide a lightweight version with `swin_v1_tiny` as the backbone, which is x4+ faster and x5+ smaller. The details can be found in [this issue](https://github.com/ZhengPeng7/BiRefNet/issues/11#issuecomment-2041033576) and links there.
+
+We found there've been some 3rd party applications based on our BiRefNet. Many thanks for their contribution to the community!  
+Choose the one you like to try with clicks instead of codes:  
+1. **Applications**:
+   + Thanks [**fal.ai/birefnet**](https://fal.ai/models/birefnet): this project on `fal.ai` encapsulates BiRefNet **online** with more useful options in **UI** and **API** to call the model.
+     <p align="center"><img src="https://drive.google.com/thumbnail?id=1rNk81YV_Pzb2GykrzfGvX6T7KBXR0wrA&sz=w1620" /></p>
+
+   + Thanks [**ZHO-ZHO-ZHO/ComfyUI-BiRefNet-ZHO**](https://github.com/ZHO-ZHO-ZHO/ComfyUI-BiRefNet-ZHO): this project further improves the **UI** for BiRefNet in ComfyUI, especially for **video data**.
+     <p align="center"><img src="https://drive.google.com/thumbnail?id=1GOqEreyS7ENzTPN0RqxEjaA76RpMlkYM&sz=w1620" /></p>
+     
+     <https://github.com/ZhengPeng7/BiRefNet/assets/25921713/3a1c7ab2-9847-4dac-8935-43a2d3cd2671>
+
+   + Thanks [**viperyl/ComfyUI-BiRefNet**](https://github.com/viperyl/ComfyUI-BiRefNet): this project packs BiRefNet as **ComfyUI nodes**, and makes this SOTA model easier use for everyone.
+     <p align="center"><img src="https://drive.google.com/thumbnail?id=1KfxCQUUa2y9T-aysEaeVVjCUt3Z0zSkL&sz=w1620" /></p>
+
+   + Thanks [**Rishabh**](https://github.com/rishabh063) for offerring a demo for the [easier single image inference on colab](https://colab.research.google.com/drive/14Dqg7oeBkFEtchaHLNpig2BcdkZEogba?usp=drive_link).
+
+2. **More Visual Comparisons**
+   + Thanks [**twitter.com/ZHOZHO672070**](https://twitter.com/ZHOZHO672070) for the comparison with more background-removal methods in images:
+
+     <img src="https://drive.google.com/thumbnail?id=1nvVIFt_Ezs-crPSQxUDqkUBz598fTe63&sz=w1620" />
+
+   + Thanks [**twitter.com/toyxyz3**](https://twitter.com/toyxyz3) for the comparison with more background-removal methods in videos:
+
+    <https://github.com/ZhengPeng7/BiRefNet/assets/25921713/40136198-01cc-4106-81f9-81c985f02e31>
+
+    <https://github.com/ZhengPeng7/BiRefNet/assets/25921713/1a32860c-0893-49dd-b557-c2e35a83c160>
+
+
+## Usage
+
+#### Environment Setup
+
+```shell
+# PyTorch==2.0.1 is used for faster training with compilation.
+conda create -n dis python=3.9 -y && conda activate dis
+pip install -r requirements.txt
+```
+
+#### Dataset Preparation
+
+Download combined training / test sets I have organized well from: [DIS](https://drive.google.com/drive/folders/1hZW6tAGPJwo9mPS7qGGGdpxuvuXiyoMJ)--[COD](https://drive.google.com/drive/folders/1EyHmKWsXfaCR9O0BiZEc3roZbRcs4ECO)--[HRSOD](https://drive.google.com/drive/folders/18_hAE3QM4cwAzEAKXuSNtKjmgFXTQXZN) or the single official ones in the `single_ones` folder, or their official pages. You can also find the same ones on my **BaiduDisk**: [DIS](https://pan.baidu.com/s/1O_pQIGAE4DKqL93xOxHpxw?pwd=PSWD)--[COD](https://pan.baidu.com/s/1RnxAzaHSTGBC1N6r_RfeqQ?pwd=PSWD)--[HRSOD](https://pan.baidu.com/s/1_Del53_0lBuG0DKJJAk4UA?pwd=PSWD).
+
+#### Weights Preparation
+
+Download backbone weights from [my google-drive folder](https://drive.google.com/drive/folders/1s2Xe0cjq-2ctnJBR24563yMSCOu4CcxM) or their official pages.
+
+#### Run
+
+```shell
+# Train & Test & Evaluation
+./train_test.sh RUN_NAME GPU_NUMBERS_FOR_TRAINING GPU_NUMBERS_FOR_TEST
+# See train.sh / test.sh for only training / test-evaluation.
+# After the evluation, run `gen_best_ep.py` to select the best ckpt from a specific metric (you choose it from Sm, wFm, HCE (DIS only)).
+```
+
+#### Well-trained weights:
+
+Download the `BiRefNet-{TASK}-{EPOCH}.pth` from [[**stuff**](https://drive.google.com/drive/folders/1s2Xe0cjq-2ctnJBR24563yMSCOu4CcxM)]. Info of the corresponding (predicted\_maps/performance/training\_log) weights can be also found in folders like `exp-BiRefNet-{TASK_SETTINGS}` in the same directory.
+
+You can also download the weights from the release of this repo.
+
+The results might be a bit different from those in the original paper, you can see them in the `eval_results-BiRefNet-{TASK_SETTINGS}` folder in each `exp-xx`, we will update them in the following days. Due to the very high cost I used (A100-80G x 8) which many people cannot afford to (including myself....),  I re-trained BiRefNet on a single A100-40G only and achieve the performance on the same level (even better). It means you can directly train the model on a single GPU with 36.5G+ memory. BTW, 5.5G GPU memory is needed for inference in 1024x1024. (I personally paid a lot for renting an A100-40G to re-train BiRefNet on the three tasks... T_T. Hope it can help you.)
+
+But if you have more and more powerful GPUs, you can set GPU IDs and increase the batch size in `config.py` to accelerate the training. We have made all this kind of things adaptive in scripts to seamlessly switch between single-card training and multi-card training. Enjoy it :)
+
+#### Some of my messages:
+
+This project was originally built for DIS only. But after the updates one by one, I made it larger and larger with many functions embedded together. Finally, you can **use it for any binary image segmentation tasks**, such as DIS/COD/SOD, medical image segmentation, anomaly segmentation, etc. You can eaily open/close below things (usually in `config.py`):
++ Multi-GPU training: open/close with one variable.
++ Backbone choices: Swin_v1, PVT_v2, ConvNets, ...
++ Weighted losses: BCE, IoU, SSIM, MAE, Reg, ...
++ Adversarial loss for binary segmentation (proposed in my previous work [MCCL](https://arxiv.org/pdf/2302.14485.pdf)).
++ Training tricks: multi-scale supervision, freezing backbone, multi-scale input...
++ Data collator: loading all in memory, smooth combination of different datasets for combined training and test.
++ ...
+I really hope you enjoy this project and use it in more works to achieve new SOTAs.
+
+
+### Quantitative Results
+
+<p align="center"><img src="https://drive.google.com/thumbnail?id=184e84BwLuNu1FytSAQ2EnANZ0RFHKPip&sz=w1620" /></p>
+
+<p align="center"><img src="https://drive.google.com/thumbnail?id=1W0mi0ZiYbqsaGuohNXU8Gh7Zj4M3neFg&sz=w1620" /></p>
+
+
+
+### Qualitative Results
+
+<p align="center"><img src="https://drive.google.com/thumbnail?id=1TYZF8pVZc2V0V6g3ik4iAr9iKvJ8BNrf&sz=w1620" /></p>
+
+<p align="center"><img src="https://drive.google.com/thumbnail?id=1ZGHC32CAdT9cwRloPzOCKWCrVQZvUAlJ&sz=w1620" /></p>
+
+
+
+### Citation
+
+```
+@article{zheng2024birefnet,
+  title={Bilateral Reference for High-Resolution Dichotomous Image Segmentation},
+  author={Zheng, Peng and Gao, Dehong and Fan, Deng-Ping and Liu, Li and Laaksonen, Jorma and Ouyang, Wanli and Sebe, Nicu},
+  journal={arXiv},
+  year={2024}
+}
+```
+
+
+
+## Contact
+
+Any question, discussion or even complaint, feel free to leave issues here or send me e-mails ([email protected]).

BiRefNet_github/config.py

@@ -0,0 +1,156 @@
+import os
+import math
+
+
+class Config():
+    def __init__(self) -> None:
+        # PATH settings
+        self.sys_home_dir = os.environ['HOME']     # Make up your file system as: SYS_HOME_DIR/codes/dis/BiRefNet, SYS_HOME_DIR/datasets/dis/xx, SYS_HOME_DIR/weights/xx
+
+        # TASK settings
+        self.task = ['DIS5K', 'COD', 'HRSOD', 'DIS5K+HRSOD+HRS10K', 'P3M-10k'][0]
+        self.training_set = {
+            'DIS5K': ['DIS-TR', 'DIS-TR+DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4'][0],
+            'COD': 'TR-COD10K+TR-CAMO',
+            'HRSOD': ['TR-DUTS', 'TR-HRSOD', 'TR-UHRSD', 'TR-DUTS+TR-HRSOD', 'TR-DUTS+TR-UHRSD', 'TR-HRSOD+TR-UHRSD', 'TR-DUTS+TR-HRSOD+TR-UHRSD'][5],
+            'DIS5K+HRSOD+HRS10K': 'DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4+DIS-TR+TE-HRS10K+TE-HRSOD+TE-UHRSD+TR-HRS10K+TR-HRSOD+TR-UHRSD',     # leave DIS-VD for evaluation.
+            'P3M-10k': 'TR-P3M-10k',
+        }[self.task]
+        self.prompt4loc = ['dense', 'sparse'][0]
+
+        # Faster-Training settings
+        self.load_all = True
+        self.compile = True     # 1. Trigger CPU memory leak in some extend, which is an inherent problem of PyTorch.
+                                #   Machines with > 70GB CPU memory can run the whole training on DIS5K with default setting.
+                                # 2. Higher PyTorch version may fix it: https://github.com/pytorch/pytorch/issues/119607.
+                                # 3. But compile in Pytorch > 2.0.1 seems to bring no acceleration for training.
+        self.precisionHigh = True
+
+        # MODEL settings
+        self.ms_supervision = True
+        self.out_ref = self.ms_supervision and True
+        self.dec_ipt = True
+        self.dec_ipt_split = True
+        self.cxt_num = [0, 3][1]    # multi-scale skip connections from encoder
+        self.mul_scl_ipt = ['', 'add', 'cat'][2]
+        self.dec_att = ['', 'ASPP', 'ASPPDeformable'][2]
+        self.squeeze_block = ['', 'BasicDecBlk_x1', 'ResBlk_x4', 'ASPP_x3', 'ASPPDeformable_x3'][1]
+        self.dec_blk = ['BasicDecBlk', 'ResBlk', 'HierarAttDecBlk'][0]
+
+        # TRAINING settings
+        self.batch_size = 4
+        self.IoU_finetune_last_epochs = [
+            0,
+            {
+                'DIS5K': -50,
+                'COD': -20,
+                'HRSOD': -20,
+                'DIS5K+HRSOD+HRS10K': -20,
+                'P3M-10k': -20,
+            }[self.task]
+        ][1]    # choose 0 to skip
+        self.lr = (1e-4 if 'DIS5K' in self.task else 1e-5) * math.sqrt(self.batch_size / 4)     # DIS needs high lr to converge faster. Adapt the lr linearly
+        self.size = 1024
+        self.num_workers = max(4, self.batch_size)          # will be decrease to min(it, batch_size) at the initialization of the data_loader
+
+        # Backbone settings
+        self.bb = [
+            'vgg16', 'vgg16bn', 'resnet50',         # 0, 1, 2
+            'swin_v1_t', 'swin_v1_s',               # 3, 4
+            'swin_v1_b', 'swin_v1_l',               # 5-bs9, 6-bs4
+            'pvt_v2_b0', 'pvt_v2_b1',               # 7, 8
+            'pvt_v2_b2', 'pvt_v2_b5',               # 9-bs10, 10-bs5
+        ][6]
+        self.lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+            'swin_v1_t': [768, 384, 192, 96], 'swin_v1_s': [768, 384, 192, 96],
+            'pvt_v2_b0': [256, 160, 64, 32], 'pvt_v2_b1': [512, 320, 128, 64],
+        }[self.bb]
+        if self.mul_scl_ipt == 'cat':
+            self.lateral_channels_in_collection = [channel * 2 for channel in self.lateral_channels_in_collection]
+        self.cxt = self.lateral_channels_in_collection[1:][::-1][-self.cxt_num:] if self.cxt_num else []
+
+        # MODEL settings - inactive
+        self.lat_blk = ['BasicLatBlk'][0]
+        self.dec_channels_inter = ['fixed', 'adap'][0]
+        self.refine = ['', 'itself', 'RefUNet', 'Refiner', 'RefinerPVTInChannels4'][0]
+        self.progressive_ref = self.refine and True
+        self.ender = self.progressive_ref and False
+        self.scale = self.progressive_ref and 2
+        self.auxiliary_classification = False       # Only for DIS5K, where class labels are saved in `dataset.py`.
+        self.refine_iteration = 1
+        self.freeze_bb = False
+        self.model = [
+            'BiRefNet',
+        ][0]
+        if self.dec_blk == 'HierarAttDecBlk':
+            self.batch_size = 2 ** [0, 1, 2, 3, 4][2]
+
+        # TRAINING settings - inactive
+        self.preproc_methods = ['flip', 'enhance', 'rotate', 'pepper', 'crop'][:4]
+        self.optimizer = ['Adam', 'AdamW'][1]
+        self.lr_decay_epochs = [1e5]    # Set to negative N to decay the lr in the last N-th epoch.
+        self.lr_decay_rate = 0.5
+        # Loss
+        self.lambdas_pix_last = {
+            # not 0 means opening this loss
+            # original rate -- 1 : 30 : 1.5 : 0.2, bce x 30
+            'bce': 30 * 1,          # high performance
+            'iou': 0.5 * 1,         # 0 / 255
+            'iou_patch': 0.5 * 0,   # 0 / 255, win_size = (64, 64)
+            'mse': 150 * 0,         # can smooth the saliency map
+            'triplet': 3 * 0,
+            'reg': 100 * 0,
+            'ssim': 10 * 1,          # help contours,
+            'cnt': 5 * 0,          # help contours
+            'structure': 5 * 0,    # structure loss from codes of MVANet. A little improvement on DIS-TE[1,2,3], a bit more decrease on DIS-TE4.
+        }
+        self.lambdas_cls = {
+            'ce': 5.0
+        }
+        # Adv
+        self.lambda_adv_g = 10. * 0        # turn to 0 to avoid adv training
+        self.lambda_adv_d = 3. * (self.lambda_adv_g > 0)
+
+        # PATH settings - inactive
+        self.data_root_dir = os.path.join(self.sys_home_dir, 'datasets/dis')
+        self.weights_root_dir = os.path.join(self.sys_home_dir, 'weights')
+        self.weights = {
+            'pvt_v2_b2': os.path.join(self.weights_root_dir, 'pvt_v2_b2.pth'),
+            'pvt_v2_b5': os.path.join(self.weights_root_dir, ['pvt_v2_b5.pth', 'pvt_v2_b5_22k.pth'][0]),
+            'swin_v1_b': os.path.join(self.weights_root_dir, ['swin_base_patch4_window12_384_22kto1k.pth', 'swin_base_patch4_window12_384_22k.pth'][0]),
+            'swin_v1_l': os.path.join(self.weights_root_dir, ['swin_large_patch4_window12_384_22kto1k.pth', 'swin_large_patch4_window12_384_22k.pth'][0]),
+            'swin_v1_t': os.path.join(self.weights_root_dir, ['swin_tiny_patch4_window7_224_22kto1k_finetune.pth'][0]),
+            'swin_v1_s': os.path.join(self.weights_root_dir, ['swin_small_patch4_window7_224_22kto1k_finetune.pth'][0]),
+            'pvt_v2_b0': os.path.join(self.weights_root_dir, ['pvt_v2_b0.pth'][0]),
+            'pvt_v2_b1': os.path.join(self.weights_root_dir, ['pvt_v2_b1.pth'][0]),
+        }
+
+        # Callbacks - inactive
+        self.verbose_eval = True
+        self.only_S_MAE = False
+        self.use_fp16 = False   # Bugs. It may cause nan in training.
+        self.SDPA_enabled = False    # Bugs. Slower and errors occur in multi-GPUs
+
+        # others
+        self.device = [0, 'cpu'][0]     # .to(0) == .to('cuda:0')
+
+        self.batch_size_valid = 1
+        self.rand_seed = 7
+        run_sh_file = [f for f in os.listdir('.') if 'train.sh' == f] + [os.path.join('..', f) for f in os.listdir('..') if 'train.sh' == f]
+        with open(run_sh_file[0], 'r') as f:
+            lines = f.readlines()
+            self.save_last = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'val_last=' in l][0].split('val_last=')[-1].split()[0])
+            self.save_step = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'step=' in l][0].split('step=')[-1].split()[0])
+        self.val_step = [0, self.save_step][0]
+
+    def print_task(self) -> None:
+        # Return task for choosing settings in shell scripts.
+        print(self.task)
+
+if __name__ == '__main__':
+    config = Config()
+    config.print_task()
+    

BiRefNet_github/dataset.py

@@ -0,0 +1,112 @@
+import os
+import cv2
+from tqdm import tqdm
+from PIL import Image
+from torch.utils import data
+from torchvision import transforms
+
+from preproc import preproc
+from config import Config
+from utils import path_to_image
+
+
+Image.MAX_IMAGE_PIXELS = None       # remove DecompressionBombWarning
+config = Config()
+_class_labels_TR_sorted = (
+    'Airplane, Ant, Antenna, Archery, Axe, BabyCarriage, Bag, BalanceBeam, Balcony, Balloon, Basket, BasketballHoop, Beatle, Bed, Bee, Bench, Bicycle, '
+    'BicycleFrame, BicycleStand, Boat, Bonsai, BoomLift, Bridge, BunkBed, Butterfly, Button, Cable, CableLift, Cage, Camcorder, Cannon, Canoe, Car, '
+    'CarParkDropArm, Carriage, Cart, Caterpillar, CeilingLamp, Centipede, Chair, Clip, Clock, Clothes, CoatHanger, Comb, ConcretePumpTruck, Crack, Crane, '
+    'Cup, DentalChair, Desk, DeskChair, Diagram, DishRack, DoorHandle, Dragonfish, Dragonfly, Drum, Earphone, Easel, ElectricIron, Excavator, Eyeglasses, '
+    'Fan, Fence, Fencing, FerrisWheel, FireExtinguisher, Fishing, Flag, FloorLamp, Forklift, GasStation, Gate, Gear, Goal, Golf, GymEquipment, Hammock, '
+    'Handcart, Handcraft, Handrail, HangGlider, Harp, Harvester, Headset, Helicopter, Helmet, Hook, HorizontalBar, Hydrovalve, IroningTable, Jewelry, Key, '
+    'KidsPlayground, Kitchenware, Kite, Knife, Ladder, LaundryRack, Lightning, Lobster, Locust, Machine, MachineGun, MagazineRack, Mantis, Medal, MemorialArchway, '
+    'Microphone, Missile, MobileHolder, Monitor, Mosquito, Motorcycle, MovingTrolley, Mower, MusicPlayer, MusicStand, ObservationTower, Octopus, OilWell, '
+    'OlympicLogo, OperatingTable, OutdoorFitnessEquipment, Parachute, Pavilion, Piano, Pipe, PlowHarrow, PoleVault, Punchbag, Rack, Racket, Rifle, Ring, Robot, '
+    'RockClimbing, Rope, Sailboat, Satellite, Scaffold, Scale, Scissor, Scooter, Sculpture, Seadragon, Seahorse, Seal, SewingMachine, Ship, Shoe, ShoppingCart, '
+    'ShoppingTrolley, Shower, Shrimp, Signboard, Skateboarding, Skeleton, Skiing, Spade, SpeedBoat, Spider, Spoon, Stair, Stand, Stationary, SteeringWheel, '
+    'Stethoscope, Stool, Stove, StreetLamp, SweetStand, Swing, Sword, TV, Table, TableChair, TableLamp, TableTennis, Tank, Tapeline, Teapot, Telescope, Tent, '
+    'TobaccoPipe, Toy, Tractor, TrafficLight, TrafficSign, Trampoline, TransmissionTower, Tree, Tricycle, TrimmerCover, Tripod, Trombone, Truck, Trumpet, Tuba, '
+    'UAV, Umbrella, UnevenBars, UtilityPole, VacuumCleaner, Violin, Wakesurfing, Watch, WaterTower, WateringPot, Well, WellLid, Wheel, Wheelchair, WindTurbine, Windmill, WineGlass, WireWhisk, Yacht'
+)
+class_labels_TR_sorted = _class_labels_TR_sorted.split(', ')
+
+
+class MyData(data.Dataset):
+    def __init__(self, datasets, image_size, is_train=True):
+        self.size_train = image_size
+        self.size_test = image_size
+        self.keep_size = not config.size
+        self.data_size = (config.size, config.size)
+        self.is_train = is_train
+        self.load_all = config.load_all
+        self.device = config.device
+        if self.is_train and config.auxiliary_classification:
+            self.cls_name2id = {_name: _id for _id, _name in enumerate(class_labels_TR_sorted)}
+        self.transform_image = transforms.Compose([
+            transforms.Resize(self.data_size),
+            transforms.ToTensor(),
+            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+        ][self.load_all or self.keep_size:])
+        self.transform_label = transforms.Compose([
+            transforms.Resize(self.data_size),
+            transforms.ToTensor(),
+        ][self.load_all or self.keep_size:])
+        dataset_root = os.path.join(config.data_root_dir, config.task)
+        # datasets can be a list of different datasets for training on combined sets.
+        self.image_paths = []
+        for dataset in datasets.split('+'):
+            image_root = os.path.join(dataset_root, dataset, 'im')
+            self.image_paths += [os.path.join(image_root, p) for p in os.listdir(image_root)]
+        self.label_paths = []
+        for p in self.image_paths:
+            for ext in ['.png', '.jpg', '.PNG', '.JPG', '.JPEG']:
+                ## 'im' and 'gt' may need modifying
+                p_gt = p.replace('/im/', '/gt/')[:-(len(p.split('.')[-1])+1)] + ext
+                file_exists = False
+                if os.path.exists(p_gt):
+                    self.label_paths.append(p_gt)
+                    file_exists = True
+                    break
+            if not file_exists:
+                print('Not exists:', p_gt)
+        if self.load_all:
+            self.images_loaded, self.labels_loaded = [], []
+            self.class_labels_loaded = []
+            # for image_path, label_path in zip(self.image_paths, self.label_paths):
+            for image_path, label_path in tqdm(zip(self.image_paths, self.label_paths), total=len(self.image_paths)):
+                _image = path_to_image(image_path, size=(config.size, config.size), color_type='rgb')
+                _label = path_to_image(label_path, size=(config.size, config.size), color_type='gray')
+                self.images_loaded.append(_image)
+                self.labels_loaded.append(_label)
+                self.class_labels_loaded.append(
+                    self.cls_name2id[label_path.split('/')[-1].split('#')[3]] if self.is_train and config.auxiliary_classification else -1
+                )
+
+    def __getitem__(self, index):
+
+        if self.load_all:
+            image = self.images_loaded[index]
+            label = self.labels_loaded[index]
+            class_label = self.class_labels_loaded[index] if self.is_train and config.auxiliary_classification else -1
+        else:
+            image = path_to_image(self.image_paths[index], size=(config.size, config.size), color_type='rgb')
+            label = path_to_image(self.label_paths[index], size=(config.size, config.size), color_type='gray')
+            class_label = self.cls_name2id[self.label_paths[index].split('/')[-1].split('#')[3]] if self.is_train and config.auxiliary_classification else -1
+
+        # loading image and label
+        if self.is_train:
+            image, label = preproc(image, label, preproc_methods=config.preproc_methods)
+        # else:
+        #     if _label.shape[0] > 2048 or _label.shape[1] > 2048:
+        #         _image = cv2.resize(_image, (2048, 2048), interpolation=cv2.INTER_LINEAR)
+        #         _label = cv2.resize(_label, (2048, 2048), interpolation=cv2.INTER_LINEAR)
+
+        image, label = self.transform_image(image), self.transform_label(label)
+
+        if self.is_train:
+            return image, label, class_label
+        else:
+            return image, label, self.label_paths[index]
+
+    def __len__(self):
+        return len(self.image_paths)

BiRefNet_github/eval_existingOnes.py

@@ -0,0 +1,139 @@
+import os
+import argparse
+from glob import glob
+import prettytable as pt
+
+from evaluation.evaluate import evaluator
+from config import Config
+
+
+config = Config()
+
+
+def do_eval(args):
+    # evaluation for whole dataset
+    # dataset first in evaluation
+    for _data_name in args.data_lst.split('+'):
+        pred_data_dir =  sorted(glob(os.path.join(args.pred_root, args.model_lst[0], _data_name)))
+        if not pred_data_dir:
+            print('Skip dataset {}.'.format(_data_name))
+            continue
+        gt_src = os.path.join(args.gt_root, _data_name)
+        gt_paths = sorted(glob(os.path.join(gt_src, 'gt', '*')))
+        print('#' * 20, _data_name, '#' * 20)
+        filename = os.path.join(args.save_dir, '{}_eval.txt'.format(_data_name))
+        tb = pt.PrettyTable()
+        tb.vertical_char = '&'
+        if config.task == 'DIS5K':
+            tb.field_names = ["Dataset", "Method", "maxFm", "wFmeasure", 'MAE', "Smeasure", "meanEm", "HCE", "maxEm", "meanFm", "adpEm", "adpFm"]
+        elif config.task == 'COD':
+            tb.field_names = ["Dataset", "Method", "Smeasure", "wFmeasure", "meanFm", "meanEm", "maxEm", 'MAE', "maxFm", "adpEm", "adpFm", "HCE"]
+        elif config.task == 'HRSOD':
+            tb.field_names = ["Dataset", "Method", "Smeasure", "maxFm", "meanEm", 'MAE', "maxEm", "meanFm", "wFmeasure", "adpEm", "adpFm", "HCE"]
+        elif config.task == 'DIS5K+HRSOD+HRS10K':
+            tb.field_names = ["Dataset", "Method", "maxFm", "wFmeasure", 'MAE', "Smeasure", "meanEm", "HCE", "maxEm", "meanFm", "adpEm", "adpFm"]
+        elif config.task == 'P3M-10k':
+            tb.field_names = ["Dataset", "Method", "Smeasure", "maxFm", "meanEm", 'MAE', "maxEm", "meanFm", "wFmeasure", "adpEm", "adpFm", "HCE"]
+        else:
+            tb.field_names = ["Dataset", "Method", "Smeasure", 'MAE', "maxEm", "meanEm", "maxFm", "meanFm", "wFmeasure", "adpEm", "adpFm", "HCE"]
+        for _model_name in args.model_lst[:]:
+            print('\t', 'Evaluating model: {}...'.format(_model_name))
+            pred_paths = [p.replace(args.gt_root, os.path.join(args.pred_root, _model_name)).replace('/gt/', '/') for p in gt_paths]
+            # print(pred_paths[:1], gt_paths[:1])
+            em, sm, fm, mae, wfm, hce = evaluator(
+                gt_paths=gt_paths,
+                pred_paths=pred_paths,
+                metrics=args.metrics.split('+'),
+                verbose=config.verbose_eval
+            )
+            if config.task == 'DIS5K':
+                scores = [
+                    fm['curve'].max().round(3), wfm.round(3), mae.round(3), sm.round(3), em['curve'].mean().round(3), int(hce.round()), 
+                    em['curve'].max().round(3), fm['curve'].mean().round(3), em['adp'].round(3), fm['adp'].round(3),
+                ]
+            elif config.task == 'COD':
+                scores = [
+                    sm.round(3), wfm.round(3), fm['curve'].mean().round(3), em['curve'].mean().round(3), em['curve'].max().round(3), mae.round(3),
+                    fm['curve'].max().round(3), em['adp'].round(3), fm['adp'].round(3), int(hce.round()),
+                ]
+            elif config.task == 'HRSOD':
+                scores = [
+                    sm.round(3), fm['curve'].max().round(3), em['curve'].mean().round(3), mae.round(3),
+                    em['curve'].max().round(3), fm['curve'].mean().round(3), wfm.round(3), em['adp'].round(3), fm['adp'].round(3), int(hce.round()),
+                ]
+            elif config.task == 'DIS5K+HRSOD+HRS10K':
+                scores = [
+                    fm['curve'].max().round(3), wfm.round(3), mae.round(3), sm.round(3), em['curve'].mean().round(3), int(hce.round()), 
+                    em['curve'].max().round(3), fm['curve'].mean().round(3), em['adp'].round(3), fm['adp'].round(3),
+                ]
+            elif config.task == 'P3M-10k':
+                scores = [
+                    sm.round(3), fm['curve'].max().round(3), em['curve'].mean().round(3), mae.round(3),
+                    em['curve'].max().round(3), fm['curve'].mean().round(3), wfm.round(3), em['adp'].round(3), fm['adp'].round(3), int(hce.round()),
+                ]
+            else:
+                scores = [
+                    sm.round(3), mae.round(3), em['curve'].max().round(3), em['curve'].mean().round(3),
+                    fm['curve'].max().round(3), fm['curve'].mean().round(3), wfm.round(3),
+                    em['adp'].round(3), fm['adp'].round(3), int(hce.round()),
+                ]
+            
+            for idx_score, score in enumerate(scores):
+                scores[idx_score] = '.' + format(score, '.3f').split('.')[-1] if score <= 1  else format(score, '<4')
+            records = [_data_name, _model_name] + scores
+            tb.add_row(records)
+            # Write results after every check.
+            with open(filename, 'w+') as file_to_write:
+                file_to_write.write(str(tb)+'\n')
+        print(tb)
+
+
+if __name__ == '__main__':
+    # set parameters
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        '--gt_root', type=str, help='ground-truth root',
+        default=os.path.join(config.data_root_dir, config.task))
+    parser.add_argument(
+        '--pred_root', type=str, help='prediction root',
+        default='./e_preds')
+    parser.add_argument(
+        '--data_lst', type=str, help='test dataset',
+        default={
+            'DIS5K': '+'.join(['DIS-VD', 'DIS-TE1', 'DIS-TE2', 'DIS-TE3', 'DIS-TE4'][:]),
+            'COD': '+'.join(['TE-COD10K', 'NC4K', 'TE-CAMO', 'CHAMELEON'][:]),
+            'HRSOD': '+'.join(['DAVIS-S', 'TE-HRSOD', 'TE-UHRSD', 'TE-DUTS', 'DUT-OMRON'][:]),
+            'DIS5K+HRSOD+HRS10K': '+'.join(['DIS-VD'][:]),
+            'P3M-10k': '+'.join(['TE-P3M-500-P', 'TE-P3M-500-NP'][:]),
+        }[config.task])
+    parser.add_argument(
+        '--save_dir', type=str, help='candidate competitors',
+        default='e_results')
+    parser.add_argument(
+        '--check_integrity', type=bool, help='whether to check the file integrity',
+        default=False)
+    parser.add_argument(
+        '--metrics', type=str, help='candidate competitors',
+        default='+'.join(['S', 'MAE', 'E', 'F', 'WF', 'HCE'][:100 if 'DIS5K' in config.task else -1]))
+    args = parser.parse_args()
+
+    os.makedirs(args.save_dir, exist_ok=True)
+    try:
+        args.model_lst = [m for m in sorted(os.listdir(args.pred_root), key=lambda x: int(x.split('epoch_')[-1]), reverse=True) if int(m.split('epoch_')[-1]) % 1 == 0]
+    except:
+        args.model_lst = [m for m in sorted(os.listdir(args.pred_root))]
+
+    # check the integrity of each candidates
+    if args.check_integrity:
+        for _data_name in args.data_lst.split('+'):
+            for _model_name in args.model_lst:
+                gt_pth = os.path.join(args.gt_root, _data_name)
+                pred_pth = os.path.join(args.pred_root, _model_name, _data_name)
+                if not sorted(os.listdir(gt_pth)) == sorted(os.listdir(pred_pth)):
+                    print(len(sorted(os.listdir(gt_pth))), len(sorted(os.listdir(pred_pth))))
+                    print('The {} Dataset of {} Model is not matching to the ground-truth'.format(_data_name, _model_name))
+    else:
+        print('>>> skip check the integrity of each candidates')
+
+    # start engine
+    do_eval(args)

BiRefNet_github/evaluation/evaluate.py

@@ -0,0 +1,60 @@
+import os
+import prettytable as pt
+
+from evaluation.metrics import evaluator
+from config import Config
+
+
+config = Config()
+
+def evaluate(pred_dir, method, testset, only_S_MAE=False, epoch=0):
+    filename = os.path.join('evaluation', 'eval-{}.txt'.format(method))
+    if os.path.exists(filename):
+        id_suffix = 1
+        filename = filename.rstrip('.txt') + '_{}.txt'.format(id_suffix)
+        while os.path.exists(filename):
+            id_suffix += 1
+            filename = filename.replace('_{}.txt'.format(id_suffix-1), '_{}.txt'.format(id_suffix))
+    gt_paths = sorted([
+        os.path.join(config.data_root_dir, config.task, testset, 'gt', p)
+        for p in os.listdir(os.path.join(config.data_root_dir, config.task, testset, 'gt'))
+    ])
+    pred_paths = sorted([os.path.join(pred_dir, method, testset, p) for p in os.listdir(os.path.join(pred_dir, method, testset))])
+    with open(filename, 'a+') as file_to_write:
+        tb = pt.PrettyTable()
+        field_names = [
+            "Dataset", "Method", "maxFm", "wFmeasure", 'MAE', "Smeasure", "meanEm", "maxEm", "meanFm",
+            "adpEm", "adpFm", 'HCE'
+        ]
+        tb.field_names = [name for name in field_names if not only_S_MAE or all(metric not in name for metric in ['Em', 'Fm'])]
+        em, sm, fm, mae, wfm, hce = evaluator(
+            gt_paths=gt_paths[:],
+            pred_paths=pred_paths[:],
+            metrics=['S', 'MAE', 'E', 'F', 'HCE'][:10*(not only_S_MAE) + 2],    # , 'WF'
+            verbose=config.verbose_eval,
+        )
+        e_max, e_mean, e_adp = em['curve'].max(), em['curve'].mean(), em['adp'].mean()
+        f_max, f_mean, f_wfm, f_adp = fm['curve'].max(), fm['curve'].mean(), wfm, fm['adp']
+        tb.add_row(
+            [
+                method+str(epoch), testset, f_max.round(3), f_wfm.round(3), mae.round(3), sm.round(3),
+                e_mean.round(3), e_max.round(3), f_mean.round(3), em['adp'].round(3), f_adp.round(3), hce.round(3)
+            ] if not only_S_MAE else [method, testset, mae.round(3), sm.round(3)]
+        )
+        print(tb)
+        file_to_write.write(str(tb).replace('+', '|')+'\n')
+        file_to_write.close()
+    return {'e_max': e_max, 'e_mean': e_mean, 'e_adp': e_adp, 'sm': sm, 'mae': mae, 'f_max': f_max, 'f_mean': f_mean, 'f_wfm': f_wfm, 'f_adp': f_adp, 'hce': hce}
+
+
+def main():
+    only_S_MAE = False
+    pred_dir = '.'
+    method = 'tmp_val'
+    testsets = 'DIS-VD+DIS-TE1'
+    for testset in testsets.split('+'):
+        res_dct = evaluate(pred_dir, method, testset, only_S_MAE=only_S_MAE)
+
+
+if __name__ == '__main__':
+    main()

BiRefNet_github/evaluation/metrics.py

@@ -0,0 +1,612 @@
+import os
+from tqdm import tqdm
+import cv2
+import numpy as np
+from scipy.ndimage import convolve, distance_transform_edt as bwdist
+from skimage.morphology import skeletonize
+from skimage.morphology import disk
+from skimage.measure import label
+
+
+_EPS = np.spacing(1)
+_TYPE = np.float64
+
+
+def evaluator(gt_paths, pred_paths, metrics=['S', 'MAE', 'E', 'F', 'WF', 'HCE'], verbose=False):
+    # define measures
+    if 'E' in metrics:
+        EM = Emeasure()
+    if 'S' in metrics:
+        SM = Smeasure()
+    if 'F' in metrics:
+        FM = Fmeasure()
+    if 'MAE' in metrics:
+        MAE = MAEmeasure()
+    if 'WF' in metrics:
+        WFM = WeightedFmeasure()
+    if 'HCE' in metrics:
+        HCE = HCEMeasure()
+
+    if isinstance(gt_paths, list) and isinstance(pred_paths, list):
+        # print(len(gt_paths), len(pred_paths))
+        assert len(gt_paths) == len(pred_paths)
+
+    for idx_sample in tqdm(range(len(gt_paths)), total=len(gt_paths)) if verbose else range(len(gt_paths)):
+        gt = gt_paths[idx_sample]
+        pred = pred_paths[idx_sample]
+
+        pred = pred[:-4] + '.png'
+        if os.path.exists(pred):
+            pred_ary = cv2.imread(pred, cv2.IMREAD_GRAYSCALE)
+        else:
+            pred_ary = cv2.imread(pred.replace('.png', '.jpg'), cv2.IMREAD_GRAYSCALE)
+        gt_ary = cv2.imread(gt, cv2.IMREAD_GRAYSCALE)
+        pred_ary = cv2.resize(pred_ary, (gt_ary.shape[1], gt_ary.shape[0]))
+
+        if 'E' in metrics:
+            EM.step(pred=pred_ary, gt=gt_ary)
+        if 'S' in metrics:
+            SM.step(pred=pred_ary, gt=gt_ary)
+        if 'F' in metrics:
+            FM.step(pred=pred_ary, gt=gt_ary)
+        if 'MAE' in metrics:
+            MAE.step(pred=pred_ary, gt=gt_ary)
+        if 'WF' in metrics:
+            WFM.step(pred=pred_ary, gt=gt_ary)
+        if 'HCE' in metrics:
+            ske_path = gt.replace('/gt/', '/ske/')
+            if os.path.exists(ske_path):
+                ske_ary = cv2.imread(ske_path, cv2.IMREAD_GRAYSCALE)
+                ske_ary = ske_ary > 128
+            else:
+                ske_ary = skeletonize(gt_ary > 128)
+                ske_save_dir = os.path.join(*ske_path.split(os.sep)[:-1])
+                if ske_path[0] == os.sep:
+                    ske_save_dir = os.sep + ske_save_dir
+                os.makedirs(ske_save_dir, exist_ok=True)
+                cv2.imwrite(ske_path, ske_ary.astype(np.uint8) * 255)
+            HCE.step(pred=pred_ary, gt=gt_ary, gt_ske=ske_ary)
+
+    if 'E' in metrics:
+        em = EM.get_results()['em']
+    else:
+        em = {'curve': np.array([np.float64(-1)]), 'adp': np.float64(-1)}
+    if 'S' in metrics:
+        sm = SM.get_results()['sm']
+    else:
+        sm = np.float64(-1)
+    if 'F' in metrics:
+        fm = FM.get_results()['fm']
+    else:
+        fm = {'curve': np.array([np.float64(-1)]), 'adp': np.float64(-1)}
+    if 'MAE' in metrics:
+        mae = MAE.get_results()['mae']
+    else:
+        mae = np.float64(-1)
+    if 'WF' in metrics:
+        wfm = WFM.get_results()['wfm']
+    else:
+        wfm = np.float64(-1)
+    if 'HCE' in metrics:
+        hce = HCE.get_results()['hce']
+    else:
+        hce = np.float64(-1)
+
+    return em, sm, fm, mae, wfm, hce
+
+
+def _prepare_data(pred: np.ndarray, gt: np.ndarray) -> tuple:
+    gt = gt > 128
+    pred = pred / 255
+    if pred.max() != pred.min():
+        pred = (pred - pred.min()) / (pred.max() - pred.min())
+    return pred, gt
+
+
+def _get_adaptive_threshold(matrix: np.ndarray, max_value: float = 1) -> float:
+    return min(2 * matrix.mean(), max_value)
+
+
+class Fmeasure(object):
+    def __init__(self, beta: float = 0.3):
+        self.beta = beta
+        self.precisions = []
+        self.recalls = []
+        self.adaptive_fms = []
+        self.changeable_fms = []
+
+    def step(self, pred: np.ndarray, gt: np.ndarray):
+        pred, gt = _prepare_data(pred, gt)
+
+        adaptive_fm = self.cal_adaptive_fm(pred=pred, gt=gt)
+        self.adaptive_fms.append(adaptive_fm)
+
+        precisions, recalls, changeable_fms = self.cal_pr(pred=pred, gt=gt)
+        self.precisions.append(precisions)
+        self.recalls.append(recalls)
+        self.changeable_fms.append(changeable_fms)
+
+    def cal_adaptive_fm(self, pred: np.ndarray, gt: np.ndarray) -> float:
+        adaptive_threshold = _get_adaptive_threshold(pred, max_value=1)
+        binary_predcition = pred >= adaptive_threshold
+        area_intersection = binary_predcition[gt].sum()
+        if area_intersection == 0:
+            adaptive_fm = 0
+        else:
+            pre = area_intersection / np.count_nonzero(binary_predcition)
+            rec = area_intersection / np.count_nonzero(gt)
+            adaptive_fm = (1 + self.beta) * pre * rec / (self.beta * pre + rec)
+        return adaptive_fm
+
+    def cal_pr(self, pred: np.ndarray, gt: np.ndarray) -> tuple:
+        pred = (pred * 255).astype(np.uint8)
+        bins = np.linspace(0, 256, 257)
+        fg_hist, _ = np.histogram(pred[gt], bins=bins)
+        bg_hist, _ = np.histogram(pred[~gt], bins=bins)
+        fg_w_thrs = np.cumsum(np.flip(fg_hist), axis=0)
+        bg_w_thrs = np.cumsum(np.flip(bg_hist), axis=0)
+        TPs = fg_w_thrs
+        Ps = fg_w_thrs + bg_w_thrs
+        Ps[Ps == 0] = 1
+        T = max(np.count_nonzero(gt), 1)
+        precisions = TPs / Ps
+        recalls = TPs / T
+        numerator = (1 + self.beta) * precisions * recalls
+        denominator = np.where(numerator == 0, 1, self.beta * precisions + recalls)
+        changeable_fms = numerator / denominator
+        return precisions, recalls, changeable_fms
+
+    def get_results(self) -> dict:
+        adaptive_fm = np.mean(np.array(self.adaptive_fms, _TYPE))
+        changeable_fm = np.mean(np.array(self.changeable_fms, dtype=_TYPE), axis=0)
+        precision = np.mean(np.array(self.precisions, dtype=_TYPE), axis=0)  # N, 256
+        recall = np.mean(np.array(self.recalls, dtype=_TYPE), axis=0)  # N, 256
+        return dict(fm=dict(adp=adaptive_fm, curve=changeable_fm),
+                    pr=dict(p=precision, r=recall))
+
+
+class MAEmeasure(object):
+    def __init__(self):
+        self.maes = []
+
+    def step(self, pred: np.ndarray, gt: np.ndarray):
+        pred, gt = _prepare_data(pred, gt)
+
+        mae = self.cal_mae(pred, gt)
+        self.maes.append(mae)
+
+    def cal_mae(self, pred: np.ndarray, gt: np.ndarray) -> float:
+        mae = np.mean(np.abs(pred - gt))
+        return mae
+
+    def get_results(self) -> dict:
+        mae = np.mean(np.array(self.maes, _TYPE))
+        return dict(mae=mae)
+
+
+class Smeasure(object):
+    def __init__(self, alpha: float = 0.5):
+        self.sms = []
+        self.alpha = alpha
+
+    def step(self, pred: np.ndarray, gt: np.ndarray):
+        pred, gt = _prepare_data(pred=pred, gt=gt)
+
+        sm = self.cal_sm(pred, gt)
+        self.sms.append(sm)
+
+    def cal_sm(self, pred: np.ndarray, gt: np.ndarray) -> float:
+        y = np.mean(gt)
+        if y == 0:
+            sm = 1 - np.mean(pred)
+        elif y == 1:
+            sm = np.mean(pred)
+        else:
+            sm = self.alpha * self.object(pred, gt) + (1 - self.alpha) * self.region(pred, gt)
+            sm = max(0, sm)
+        return sm
+
+    def object(self, pred: np.ndarray, gt: np.ndarray) -> float:
+        fg = pred * gt
+        bg = (1 - pred) * (1 - gt)
+        u = np.mean(gt)
+        object_score = u * self.s_object(fg, gt) + (1 - u) * self.s_object(bg, 1 - gt)
+        return object_score
+
+    def s_object(self, pred: np.ndarray, gt: np.ndarray) -> float:
+        x = np.mean(pred[gt == 1])
+        sigma_x = np.std(pred[gt == 1], ddof=1)
+        score = 2 * x / (np.power(x, 2) + 1 + sigma_x + _EPS)
+        return score
+
+    def region(self, pred: np.ndarray, gt: np.ndarray) -> float:
+        x, y = self.centroid(gt)
+        part_info = self.divide_with_xy(pred, gt, x, y)
+        w1, w2, w3, w4 = part_info['weight']
+        pred1, pred2, pred3, pred4 = part_info['pred']
+        gt1, gt2, gt3, gt4 = part_info['gt']
+        score1 = self.ssim(pred1, gt1)
+        score2 = self.ssim(pred2, gt2)
+        score3 = self.ssim(pred3, gt3)
+        score4 = self.ssim(pred4, gt4)
+
+        return w1 * score1 + w2 * score2 + w3 * score3 + w4 * score4
+
+    def centroid(self, matrix: np.ndarray) -> tuple:
+        h, w = matrix.shape
+        area_object = np.count_nonzero(matrix)
+        if area_object == 0:
+            x = np.round(w / 2)
+            y = np.round(h / 2)
+        else:
+            # More details can be found at: https://www.yuque.com/lart/blog/gpbigm
+            y, x = np.argwhere(matrix).mean(axis=0).round()
+        return int(x) + 1, int(y) + 1
+
+    def divide_with_xy(self, pred: np.ndarray, gt: np.ndarray, x, y) -> dict:
+        h, w = gt.shape
+        area = h * w
+
+        gt_LT = gt[0:y, 0:x]
+        gt_RT = gt[0:y, x:w]
+        gt_LB = gt[y:h, 0:x]
+        gt_RB = gt[y:h, x:w]
+
+        pred_LT = pred[0:y, 0:x]
+        pred_RT = pred[0:y, x:w]
+        pred_LB = pred[y:h, 0:x]
+        pred_RB = pred[y:h, x:w]
+
+        w1 = x * y / area
+        w2 = y * (w - x) / area
+        w3 = (h - y) * x / area
+        w4 = 1 - w1 - w2 - w3
+
+        return dict(gt=(gt_LT, gt_RT, gt_LB, gt_RB),
+                    pred=(pred_LT, pred_RT, pred_LB, pred_RB),
+                    weight=(w1, w2, w3, w4))
+
+    def ssim(self, pred: np.ndarray, gt: np.ndarray) -> float:
+        h, w = pred.shape
+        N = h * w
+
+        x = np.mean(pred)
+        y = np.mean(gt)
+
+        sigma_x = np.sum((pred - x) ** 2) / (N - 1)
+        sigma_y = np.sum((gt - y) ** 2) / (N - 1)
+        sigma_xy = np.sum((pred - x) * (gt - y)) / (N - 1)
+
+        alpha = 4 * x * y * sigma_xy
+        beta = (x ** 2 + y ** 2) * (sigma_x + sigma_y)
+
+        if alpha != 0:
+            score = alpha / (beta + _EPS)
+        elif alpha == 0 and beta == 0:
+            score = 1
+        else:
+            score = 0
+        return score
+
+    def get_results(self) -> dict:
+        sm = np.mean(np.array(self.sms, dtype=_TYPE))
+        return dict(sm=sm)
+
+
+class Emeasure(object):
+    def __init__(self):
+        self.adaptive_ems = []
+        self.changeable_ems = []
+
+    def step(self, pred: np.ndarray, gt: np.ndarray):
+        pred, gt = _prepare_data(pred=pred, gt=gt)
+        self.gt_fg_numel = np.count_nonzero(gt)
+        self.gt_size = gt.shape[0] * gt.shape[1]
+
+        changeable_ems = self.cal_changeable_em(pred, gt)
+        self.changeable_ems.append(changeable_ems)
+        adaptive_em = self.cal_adaptive_em(pred, gt)
+        self.adaptive_ems.append(adaptive_em)
+
+    def cal_adaptive_em(self, pred: np.ndarray, gt: np.ndarray) -> float:
+        adaptive_threshold = _get_adaptive_threshold(pred, max_value=1)
+        adaptive_em = self.cal_em_with_threshold(pred, gt, threshold=adaptive_threshold)
+        return adaptive_em
+
+    def cal_changeable_em(self, pred: np.ndarray, gt: np.ndarray) -> np.ndarray:
+        changeable_ems = self.cal_em_with_cumsumhistogram(pred, gt)
+        return changeable_ems
+
+    def cal_em_with_threshold(self, pred: np.ndarray, gt: np.ndarray, threshold: float) -> float:
+        binarized_pred = pred >= threshold
+        fg_fg_numel = np.count_nonzero(binarized_pred & gt)
+        fg_bg_numel = np.count_nonzero(binarized_pred & ~gt)
+
+        fg___numel = fg_fg_numel + fg_bg_numel
+        bg___numel = self.gt_size - fg___numel
+
+        if self.gt_fg_numel == 0:
+            enhanced_matrix_sum = bg___numel
+        elif self.gt_fg_numel == self.gt_size:
+            enhanced_matrix_sum = fg___numel
+        else:
+            parts_numel, combinations = self.generate_parts_numel_combinations(
+                fg_fg_numel=fg_fg_numel, fg_bg_numel=fg_bg_numel,
+                pred_fg_numel=fg___numel, pred_bg_numel=bg___numel,
+            )
+
+            results_parts = []
+            for i, (part_numel, combination) in enumerate(zip(parts_numel, combinations)):
+                align_matrix_value = 2 * (combination[0] * combination[1]) / \
+                                     (combination[0] ** 2 + combination[1] ** 2 + _EPS)
+                enhanced_matrix_value = (align_matrix_value + 1) ** 2 / 4
+                results_parts.append(enhanced_matrix_value * part_numel)
+            enhanced_matrix_sum = sum(results_parts)
+
+        em = enhanced_matrix_sum / (self.gt_size - 1 + _EPS)
+        return em
+
+    def cal_em_with_cumsumhistogram(self, pred: np.ndarray, gt: np.ndarray) -> np.ndarray:
+        pred = (pred * 255).astype(np.uint8)
+        bins = np.linspace(0, 256, 257)
+        fg_fg_hist, _ = np.histogram(pred[gt], bins=bins)
+        fg_bg_hist, _ = np.histogram(pred[~gt], bins=bins)
+        fg_fg_numel_w_thrs = np.cumsum(np.flip(fg_fg_hist), axis=0)
+        fg_bg_numel_w_thrs = np.cumsum(np.flip(fg_bg_hist), axis=0)
+
+        fg___numel_w_thrs = fg_fg_numel_w_thrs + fg_bg_numel_w_thrs
+        bg___numel_w_thrs = self.gt_size - fg___numel_w_thrs
+
+        if self.gt_fg_numel == 0:
+            enhanced_matrix_sum = bg___numel_w_thrs
+        elif self.gt_fg_numel == self.gt_size:
+            enhanced_matrix_sum = fg___numel_w_thrs
+        else:
+            parts_numel_w_thrs, combinations = self.generate_parts_numel_combinations(
+                fg_fg_numel=fg_fg_numel_w_thrs, fg_bg_numel=fg_bg_numel_w_thrs,
+                pred_fg_numel=fg___numel_w_thrs, pred_bg_numel=bg___numel_w_thrs,
+            )
+
+            results_parts = np.empty(shape=(4, 256), dtype=np.float64)
+            for i, (part_numel, combination) in enumerate(zip(parts_numel_w_thrs, combinations)):
+                align_matrix_value = 2 * (combination[0] * combination[1]) / \
+                                     (combination[0] ** 2 + combination[1] ** 2 + _EPS)
+                enhanced_matrix_value = (align_matrix_value + 1) ** 2 / 4
+                results_parts[i] = enhanced_matrix_value * part_numel
+            enhanced_matrix_sum = results_parts.sum(axis=0)
+
+        em = enhanced_matrix_sum / (self.gt_size - 1 + _EPS)
+        return em
+
+    def generate_parts_numel_combinations(self, fg_fg_numel, fg_bg_numel, pred_fg_numel, pred_bg_numel):
+        bg_fg_numel = self.gt_fg_numel - fg_fg_numel
+        bg_bg_numel = pred_bg_numel - bg_fg_numel
+
+        parts_numel = [fg_fg_numel, fg_bg_numel, bg_fg_numel, bg_bg_numel]
+
+        mean_pred_value = pred_fg_numel / self.gt_size
+        mean_gt_value = self.gt_fg_numel / self.gt_size
+
+        demeaned_pred_fg_value = 1 - mean_pred_value
+        demeaned_pred_bg_value = 0 - mean_pred_value
+        demeaned_gt_fg_value = 1 - mean_gt_value
+        demeaned_gt_bg_value = 0 - mean_gt_value
+
+        combinations = [
+            (demeaned_pred_fg_value, demeaned_gt_fg_value),
+            (demeaned_pred_fg_value, demeaned_gt_bg_value),
+            (demeaned_pred_bg_value, demeaned_gt_fg_value),
+            (demeaned_pred_bg_value, demeaned_gt_bg_value)
+        ]
+        return parts_numel, combinations
+
+    def get_results(self) -> dict:
+        adaptive_em = np.mean(np.array(self.adaptive_ems, dtype=_TYPE))
+        changeable_em = np.mean(np.array(self.changeable_ems, dtype=_TYPE), axis=0)
+        return dict(em=dict(adp=adaptive_em, curve=changeable_em))
+
+
+class WeightedFmeasure(object):
+    def __init__(self, beta: float = 1):
+        self.beta = beta
+        self.weighted_fms = []
+
+    def step(self, pred: np.ndarray, gt: np.ndarray):
+        pred, gt = _prepare_data(pred=pred, gt=gt)
+
+        if np.all(~gt):
+            wfm = 0
+        else:
+            wfm = self.cal_wfm(pred, gt)
+        self.weighted_fms.append(wfm)
+
+    def cal_wfm(self, pred: np.ndarray, gt: np.ndarray) -> float:
+        # [Dst,IDXT] = bwdist(dGT);
+        Dst, Idxt = bwdist(gt == 0, return_indices=True)
+
+        # %Pixel dependency
+        # E = abs(FG-dGT);
+        E = np.abs(pred - gt)
+        Et = np.copy(E)
+        Et[gt == 0] = Et[Idxt[0][gt == 0], Idxt[1][gt == 0]]
+
+        # K = fspecial('gaussian',7,5);
+        # EA = imfilter(Et,K);
+        K = self.matlab_style_gauss2D((7, 7), sigma=5)
+        EA = convolve(Et, weights=K, mode="constant", cval=0)
+        # MIN_E_EA = E;
+        # MIN_E_EA(GT & EA<E) = EA(GT & EA<E);
+        MIN_E_EA = np.where(gt & (EA < E), EA, E)
+
+        # %Pixel importance
+        B = np.where(gt == 0, 2 - np.exp(np.log(0.5) / 5 * Dst), np.ones_like(gt))
+        Ew = MIN_E_EA * B
+
+        TPw = np.sum(gt) - np.sum(Ew[gt == 1])
+        FPw = np.sum(Ew[gt == 0])
+
+
+        R = 1 - np.mean(Ew[gt == 1])
+        P = TPw / (TPw + FPw + _EPS)
+
+        # % Q = (1+Beta^2)*(R*P)./(eps+R+(Beta.*P));
+        Q = (1 + self.beta) * R * P / (R + self.beta * P + _EPS)
+
+        return Q
+
+    def matlab_style_gauss2D(self, shape: tuple = (7, 7), sigma: int = 5) -> np.ndarray:
+        """
+        2D gaussian mask - should give the same result as MATLAB's
+        fspecial('gaussian',[shape],[sigma])
+        """
+        m, n = [(ss - 1) / 2 for ss in shape]
+        y, x = np.ogrid[-m: m + 1, -n: n + 1]
+        h = np.exp(-(x * x + y * y) / (2 * sigma * sigma))
+        h[h < np.finfo(h.dtype).eps * h.max()] = 0
+        sumh = h.sum()
+        if sumh != 0:
+            h /= sumh
+        return h
+
+    def get_results(self) -> dict:
+        weighted_fm = np.mean(np.array(self.weighted_fms, dtype=_TYPE))
+        return dict(wfm=weighted_fm)
+
+
+class HCEMeasure(object):
+    def __init__(self):
+        self.hces = []
+
+    def step(self, pred: np.ndarray, gt: np.ndarray, gt_ske):
+        # pred, gt = _prepare_data(pred, gt)
+
+        hce = self.cal_hce(pred, gt, gt_ske)
+        self.hces.append(hce)
+
+    def get_results(self) -> dict:
+        hce = np.mean(np.array(self.hces, _TYPE))
+        return dict(hce=hce)
+
+
+    def cal_hce(self, pred: np.ndarray, gt: np.ndarray, gt_ske: np.ndarray, relax=5, epsilon=2.0) -> float:
+        # Binarize gt
+        if(len(gt.shape)>2):
+            gt = gt[:, :, 0]
+
+        epsilon_gt = 128#(np.amin(gt)+np.amax(gt))/2.0
+        gt = (gt>epsilon_gt).astype(np.uint8)
+
+        # Binarize pred
+        if(len(pred.shape)>2):
+            pred = pred[:, :, 0]
+        epsilon_pred = 128#(np.amin(pred)+np.amax(pred))/2.0
+        pred = (pred>epsilon_pred).astype(np.uint8)
+
+        Union = np.logical_or(gt, pred)
+        TP = np.logical_and(gt, pred)
+        FP = pred - TP
+        FN = gt - TP
+
+        # relax the Union of gt and pred
+        Union_erode = Union.copy()
+        Union_erode = cv2.erode(Union_erode.astype(np.uint8), disk(1), iterations=relax)
+
+        # --- get the relaxed False Positive regions for computing the human efforts in correcting them ---
+        FP_ = np.logical_and(FP, Union_erode) # get the relaxed FP
+        for i in range(0, relax):
+            FP_ = cv2.dilate(FP_.astype(np.uint8), disk(1))
+            FP_ = np.logical_and(FP_, 1-np.logical_or(TP, FN))
+        FP_ = np.logical_and(FP, FP_)
+
+        # --- get the relaxed False Negative regions for computing the human efforts in correcting them ---
+        FN_ = np.logical_and(FN, Union_erode) # preserve the structural components of FN
+        ## recover the FN, where pixels are not close to the TP borders
+        for i in range(0, relax):
+            FN_ = cv2.dilate(FN_.astype(np.uint8), disk(1))
+            FN_ = np.logical_and(FN_, 1-np.logical_or(TP, FP))
+        FN_ = np.logical_and(FN, FN_)
+        FN_ = np.logical_or(FN_, np.logical_xor(gt_ske, np.logical_and(TP, gt_ske))) # preserve the structural components of FN
+
+        ## 2. =============Find exact polygon control points and independent regions==============
+        ## find contours from FP_
+        ctrs_FP, hier_FP = cv2.findContours(FP_.astype(np.uint8), cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE)
+        ## find control points and independent regions for human correction
+        bdies_FP, indep_cnt_FP = self.filter_bdy_cond(ctrs_FP, FP_, np.logical_or(TP,FN_))
+        ## find contours from FN_
+        ctrs_FN, hier_FN = cv2.findContours(FN_.astype(np.uint8), cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE)
+        ## find control points and independent regions for human correction
+        bdies_FN, indep_cnt_FN = self.filter_bdy_cond(ctrs_FN, FN_, 1-np.logical_or(np.logical_or(TP, FP_), FN_))
+
+        poly_FP, poly_FP_len, poly_FP_point_cnt = self.approximate_RDP(bdies_FP, epsilon=epsilon)
+        poly_FN, poly_FN_len, poly_FN_point_cnt = self.approximate_RDP(bdies_FN, epsilon=epsilon)
+
+        # FP_points+FP_indep+FN_points+FN_indep
+        return poly_FP_point_cnt+indep_cnt_FP+poly_FN_point_cnt+indep_cnt_FN
+
+    def filter_bdy_cond(self, bdy_, mask, cond):
+
+        cond = cv2.dilate(cond.astype(np.uint8), disk(1))
+        labels = label(mask) # find the connected regions
+        lbls = np.unique(labels) # the indices of the connected regions
+        indep = np.ones(lbls.shape[0]) # the label of each connected regions
+        indep[0] = 0 # 0 indicate the background region
+
+        boundaries = []
+        h,w = cond.shape[0:2]
+        ind_map = np.zeros((h, w))
+        indep_cnt = 0
+
+        for i in range(0, len(bdy_)):
+            tmp_bdies = []
+            tmp_bdy = []
+            for j in range(0, bdy_[i].shape[0]):
+                r, c = bdy_[i][j,0,1],bdy_[i][j,0,0]
+
+                if(np.sum(cond[r, c])==0 or ind_map[r, c]!=0):
+                    if(len(tmp_bdy)>0):
+                        tmp_bdies.append(tmp_bdy)
+                        tmp_bdy = []
+                    continue
+                tmp_bdy.append([c, r])
+                ind_map[r, c] =  ind_map[r, c] + 1
+                indep[labels[r, c]] = 0 # indicates part of the boundary of this region needs human correction
+            if(len(tmp_bdy)>0):
+                tmp_bdies.append(tmp_bdy)
+
+            # check if the first and the last boundaries are connected
+            # if yes, invert the first boundary and attach it after the last boundary
+            if(len(tmp_bdies)>1):
+                first_x, first_y = tmp_bdies[0][0]
+                last_x, last_y = tmp_bdies[-1][-1]
+                if((abs(first_x-last_x)==1 and first_y==last_y) or
+                (first_x==last_x and abs(first_y-last_y)==1) or
+                (abs(first_x-last_x)==1 and abs(first_y-last_y)==1)
+                ):
+                    tmp_bdies[-1].extend(tmp_bdies[0][::-1])
+                    del tmp_bdies[0]
+
+            for k in range(0, len(tmp_bdies)):
+                tmp_bdies[k] =  np.array(tmp_bdies[k])[:, np.newaxis, :]
+            if(len(tmp_bdies)>0):
+                boundaries.extend(tmp_bdies)
+
+        return boundaries, np.sum(indep)
+
+    # this function approximate each boundary by DP algorithm
+    # https://en.wikipedia.org/wiki/Ramer%E2%80%93Douglas%E2%80%93Peucker_algorithm
+    def approximate_RDP(self, boundaries, epsilon=1.0):
+
+        boundaries_ = []
+        boundaries_len_ = []
+        pixel_cnt_ = 0
+
+        # polygon approximate of each boundary
+        for i in range(0, len(boundaries)):
+            boundaries_.append(cv2.approxPolyDP(boundaries[i], epsilon, False))
+
+        # count the control points number of each boundary and the total control points number of all the boundaries
+        for i in range(0, len(boundaries_)):
+            boundaries_len_.append(len(boundaries_[i]))
+            pixel_cnt_ = pixel_cnt_ + len(boundaries_[i])
+
+        return boundaries_, boundaries_len_, pixel_cnt_

BiRefNet_github/evaluation/valid.py

@@ -0,0 +1,9 @@
+from inference import inference
+from evaluation.evaluate import evaluate
+
+
+def valid(model, data_loader_test, pred_dir, method='tmp_val', testset='DIS-VD', only_S_MAE=True, device=0):
+    model.eval()
+    inference(model, data_loader_test, pred_dir, method, testset, device=device)
+    performance_dict = evaluate(pred_dir, method, testset, only_S_MAE=only_S_MAE, epoch=model.epoch)
+    return performance_dict

BiRefNet_github/gen_best_ep.py

@@ -0,0 +1,85 @@
+import os
+from glob import glob
+import numpy as np
+from config import Config
+
+
+config = Config()
+
+eval_txts = sorted(glob('e_results/*_eval.txt'))
+print('eval_txts:', [_.split(os.sep)[-1] for _ in eval_txts])
+score_panel = {}
+sep = '&'
+metrics = ['sm', 'wfm', 'hce']    # we used HCE for DIS and wFm for others.
+if 'DIS5K' not in config.task:
+    metrics.remove('hce')
+
+for metric in metrics:
+    print('Metric:', metric)
+    current_line_nums = []
+    for idx_et, eval_txt in enumerate(eval_txts):
+        with open(eval_txt, 'r') as f:
+            lines = [l for l in f.readlines()[3:] if '.' in l]
+        current_line_nums.append(len(lines))
+    for idx_et, eval_txt in enumerate(eval_txts):
+        with open(eval_txt, 'r') as f:
+            lines = [l for l in f.readlines()[3:] if '.' in l]
+        for idx_line, line in enumerate(lines[:min(current_line_nums)]):    # Consist line numbers by the minimal result file.
+            properties = line.strip().strip(sep).split(sep)
+            dataset = properties[0].strip()
+            ckpt = properties[1].strip()
+            if int(ckpt.split('--epoch_')[-1].strip()) < 0:
+                continue
+            targe_idx = {
+                'sm': [5, 2, 2, 5, 2],
+                'wfm': [3, 3, 8, 3, 8],
+                'hce': [7, -1, -1, 7, -1]
+            }[metric][['DIS5K', 'COD', 'HRSOD', 'DIS5K+HRSOD+HRS10K', 'P3M-10k'].index(config.task)]
+            if metric != 'hce':
+                score_sm = float(properties[targe_idx].strip())
+            else:
+                score_sm = int(properties[targe_idx].strip().strip('.'))
+            if idx_et == 0:
+                score_panel[ckpt] = []
+            score_panel[ckpt].append(score_sm)
+
+    metrics_min = ['hce', 'mae']
+    max_or_min = min if metric in metrics_min else max
+    score_max = max_or_min(score_panel.values(), key=lambda x: np.sum(x))
+
+    good_models = []
+    for k, v in score_panel.items():
+        if (np.sum(v) <= np.sum(score_max)) if metric in metrics_min else (np.sum(v) >= np.sum(score_max)):
+            print(k, v)
+            good_models.append(k)
+
+    # Write
+    with open(eval_txt, 'r') as f:
+        lines = f.readlines()
+    info4good_models = lines[:3]
+    metric_names = [m.strip() for m in lines[1].strip().strip('&').split('&')[2:]]
+    testset_mean_values = {metric_name: [] for metric_name in metric_names}
+    for good_model in good_models:
+        for idx_et, eval_txt in enumerate(eval_txts):
+            with open(eval_txt, 'r') as f:
+                lines = f.readlines()
+            for line in lines:
+                if set([good_model]) & set([_.strip() for _ in line.split(sep)]):
+                    info4good_models.append(line)
+                    metric_scores = [float(m.strip()) for m in line.strip().strip('&').split('&')[2:]]
+                    for idx_score, metric_score in enumerate(metric_scores):
+                        testset_mean_values[metric_names[idx_score]].append(metric_score)
+
+    if 'DIS5K' in config.task:
+        testset_mean_values_lst = ['{:<4}'.format(int(np.mean(v_lst[:-1]).round())) if name == 'HCE' else '{:.3f}'.format(np.mean(v_lst[:-1])).lstrip('0') for name, v_lst in testset_mean_values.items()]  # [:-1] to remove DIS-VD
+        sample_line_for_placing_mean_values = info4good_models[-2]
+        numbers_placed_well = sample_line_for_placing_mean_values.replace(sample_line_for_placing_mean_values.split('&')[1].strip(), 'DIS-TEs').strip().split('&')[3:]
+        for idx_number, (number_placed_well, testset_mean_value) in enumerate(zip(numbers_placed_well, testset_mean_values_lst)):
+            numbers_placed_well[idx_number] = number_placed_well.replace(number_placed_well.strip(), testset_mean_value)
+        testset_mean_line = '&'.join(sample_line_for_placing_mean_values.replace(sample_line_for_placing_mean_values.split('&')[1].strip(), 'DIS-TEs').split('&')[:3] + numbers_placed_well) + '\n'
+        info4good_models.append(testset_mean_line)
+    info4good_models.append(lines[-1])
+    info = ''.join(info4good_models)
+    print(info)
+    with open(os.path.join('e_results', 'eval-{}_best_on_{}.txt'.format(config.task, metric)), 'w') as f:
+        f.write(info + '\n')

BiRefNet_github/inference.py

@@ -0,0 +1,105 @@
+import os
+import argparse
+from glob import glob
+from tqdm import tqdm
+import cv2
+import torch
+
+from dataset import MyData
+from models.birefnet import BiRefNet
+from utils import save_tensor_img, check_state_dict
+from config import Config
+
+
+config = Config()
+
+
+def inference(model, data_loader_test, pred_root, method, testset, device=0):
+    model_training = model.training
+    if model_training:
+        model.eval()
+    for batch in tqdm(data_loader_test, total=len(data_loader_test)) if 1 or config.verbose_eval else data_loader_test:
+        inputs = batch[0].to(device)
+        # gts = batch[1].to(device)
+        label_paths = batch[-1]
+        with torch.no_grad():
+            scaled_preds = model(inputs)[-1].sigmoid()
+
+        os.makedirs(os.path.join(pred_root, method, testset), exist_ok=True)
+
+        for idx_sample in range(scaled_preds.shape[0]):
+            res = torch.nn.functional.interpolate(
+                scaled_preds[idx_sample].unsqueeze(0),
+                size=cv2.imread(label_paths[idx_sample], cv2.IMREAD_GRAYSCALE).shape[:2],
+                mode='bilinear',
+                align_corners=True
+            )
+            save_tensor_img(res, os.path.join(os.path.join(pred_root, method, testset), label_paths[idx_sample].replace('\\', '/').split('/')[-1]))   # test set dir + file name
+    if model_training:
+        model.train()
+    return None
+
+
+def main(args):
+    # Init model
+
+    device = config.device
+    if args.ckpt_folder:
+        print('Testing with models in {}'.format(args.ckpt_folder))
+    else:
+        print('Testing with model {}'.format(args.ckpt))
+
+    if config.model == 'BiRefNet':
+        model = BiRefNet(bb_pretrained=False)
+    weights_lst = sorted(
+        glob(os.path.join(args.ckpt_folder, '*.pth')) if args.ckpt_folder else [args.ckpt],
+        key=lambda x: int(x.split('epoch_')[-1].split('.pth')[0]),
+        reverse=True
+    )
+    for testset in args.testsets.split('+'):
+        print('>>>> Testset: {}...'.format(testset))
+        data_loader_test = torch.utils.data.DataLoader(
+            dataset=MyData(testset, image_size=config.size, is_train=False),
+            batch_size=config.batch_size_valid, shuffle=False, num_workers=config.num_workers, pin_memory=True
+        )
+        for weights in weights_lst:
+            if int(weights.strip('.pth').split('epoch_')[-1]) % 1 != 0:
+                continue
+            print('\tInferencing {}...'.format(weights))
+            # model.load_state_dict(torch.load(weights, map_location='cpu'))
+            state_dict = torch.load(weights, map_location='cpu')
+            state_dict = check_state_dict(state_dict)
+            model.load_state_dict(state_dict)
+            model = model.to(device)
+            inference(
+                model, data_loader_test=data_loader_test, pred_root=args.pred_root,
+                method='--'.join([w.rstrip('.pth') for w in weights.split(os.sep)[-2:]]),
+                testset=testset, device=config.device
+            )
+
+
+if __name__ == '__main__':
+    # Parameter from command line
+    parser = argparse.ArgumentParser(description='')
+    parser.add_argument('--ckpt', type=str, help='model folder')
+    parser.add_argument('--ckpt_folder', default=sorted(glob(os.path.join('ckpt', '*')))[-1], type=str, help='model folder')
+    parser.add_argument('--pred_root', default='e_preds', type=str, help='Output folder')
+    parser.add_argument('--testsets',
+                        default={
+                            'DIS5K': 'DIS-VD+DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4',
+                            'COD': 'TE-COD10K+NC4K+TE-CAMO+CHAMELEON',
+                            'HRSOD': 'DAVIS-S+TE-HRSOD+TE-UHRSD+TE-DUTS+DUT-OMRON',
+                            'DIS5K+HRSOD+HRS10K': 'DIS-VD',
+                            'P3M-10k': 'TE-P3M-500-P+TE-P3M-500-NP',
+                            'DIS5K-': 'DIS-VD',
+                            'COD-': 'TE-COD10K',
+                            'SOD-': 'DAVIS-S+TE-HRSOD+TE-UHRSD',
+                        }[config.task + ''],
+                        type=str,
+                        help="Test all sets: , 'DIS-VD+DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4'")
+
+    args = parser.parse_args()
+
+    if config.precisionHigh:
+        torch.set_float32_matmul_precision('high')
+    main(args)

BiRefNet_github/loss.py

@@ -0,0 +1,274 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+from torch.autograd import Variable
+from math import exp
+from config import Config
+
+
+class Discriminator(nn.Module):
+    def __init__(self, channels=1, img_size=256):
+        super(Discriminator, self).__init__()
+
+        def discriminator_block(in_filters, out_filters, bn=Config().batch_size > 1):
+            block = [nn.Conv2d(in_filters, out_filters, 3, 2, 1), nn.LeakyReLU(0.2, inplace=True), nn.Dropout2d(0.25)]
+            if bn:
+                block.append(nn.BatchNorm2d(out_filters, 0.8))
+            return block
+
+        self.model = nn.Sequential(
+            *discriminator_block(channels, 16, bn=False),
+            *discriminator_block(16, 32),
+            *discriminator_block(32, 64),
+            *discriminator_block(64, 128),
+        )
+
+        # The height and width of downsampled image
+        ds_size = img_size // 2 ** 4
+        self.adv_layer = nn.Sequential(nn.Linear(128 * ds_size ** 2, 1), nn.Sigmoid())
+
+    def forward(self, img):
+        out = self.model(img)
+        out = out.view(out.shape[0], -1)
+        validity = self.adv_layer(out)
+
+        return validity
+
+
+class ContourLoss(torch.nn.Module):
+    def __init__(self):
+        super(ContourLoss, self).__init__()
+
+    def forward(self, pred, target, weight=10):
+        '''
+        target, pred: tensor of shape (B, C, H, W), where target[:,:,region_in_contour] == 1,
+                        target[:,:,region_out_contour] == 0.
+        weight: scalar, length term weight.
+        '''
+        # length term
+        delta_r = pred[:,:,1:,:] - pred[:,:,:-1,:] # horizontal gradient (B, C, H-1, W) 
+        delta_c = pred[:,:,:,1:] - pred[:,:,:,:-1] # vertical gradient   (B, C, H,   W-1)
+
+        delta_r    = delta_r[:,:,1:,:-2]**2  # (B, C, H-2, W-2)
+        delta_c    = delta_c[:,:,:-2,1:]**2  # (B, C, H-2, W-2)
+        delta_pred = torch.abs(delta_r + delta_c) 
+
+        epsilon = 1e-8 # where is a parameter to avoid square root is zero in practice.
+        length = torch.mean(torch.sqrt(delta_pred + epsilon)) # eq.(11) in the paper, mean is used instead of sum.
+
+        c_in  = torch.ones_like(pred)
+        c_out = torch.zeros_like(pred)
+
+        region_in  = torch.mean( pred     * (target - c_in )**2 ) # equ.(12) in the paper, mean is used instead of sum.
+        region_out = torch.mean( (1-pred) * (target - c_out)**2 ) 
+        region = region_in + region_out
+
+        loss =  weight * length + region
+
+        return loss
+
+
+class IoULoss(torch.nn.Module):
+    def __init__(self):
+        super(IoULoss, self).__init__()
+
+    def forward(self, pred, target):
+        b = pred.shape[0]
+        IoU = 0.0
+        for i in range(0, b):
+            # compute the IoU of the foreground
+            Iand1 = torch.sum(target[i, :, :, :] * pred[i, :, :, :])
+            Ior1 = torch.sum(target[i, :, :, :]) + torch.sum(pred[i, :, :, :]) - Iand1
+            IoU1 = Iand1 / Ior1
+            # IoU loss is (1-IoU1)
+            IoU = IoU + (1-IoU1)
+        # return IoU/b
+        return IoU
+
+
+class StructureLoss(torch.nn.Module):
+    def __init__(self):
+        super(StructureLoss, self).__init__()
+
+    def forward(self, pred, target):
+        weit  = 1+5*torch.abs(F.avg_pool2d(target, kernel_size=31, stride=1, padding=15)-target)
+        wbce  = F.binary_cross_entropy_with_logits(pred, target, reduction='none')
+        wbce  = (weit*wbce).sum(dim=(2,3))/weit.sum(dim=(2,3))
+
+        pred  = torch.sigmoid(pred)
+        inter = ((pred * target) * weit).sum(dim=(2, 3))
+        union = ((pred + target) * weit).sum(dim=(2, 3))
+        wiou  = 1-(inter+1)/(union-inter+1)
+
+        return (wbce+wiou).mean()
+
+
+class PatchIoULoss(torch.nn.Module):
+    def __init__(self):
+        super(PatchIoULoss, self).__init__()
+        self.iou_loss = IoULoss()
+
+    def forward(self, pred, target):
+        win_y, win_x = 64, 64
+        iou_loss = 0.
+        for anchor_y in range(0, target.shape[0], win_y):
+            for anchor_x in range(0, target.shape[1], win_y):
+                patch_pred = pred[:, :, anchor_y:anchor_y+win_y, anchor_x:anchor_x+win_x]
+                patch_target = target[:, :, anchor_y:anchor_y+win_y, anchor_x:anchor_x+win_x]
+                patch_iou_loss = self.iou_loss(patch_pred, patch_target)
+                iou_loss += patch_iou_loss
+        return iou_loss
+
+
+class ThrReg_loss(torch.nn.Module):
+    def __init__(self):
+        super(ThrReg_loss, self).__init__()
+
+    def forward(self, pred, gt=None):
+        return torch.mean(1 - ((pred - 0) ** 2 + (pred - 1) ** 2))
+
+
+class ClsLoss(nn.Module):
+    """
+    Auxiliary classification loss for each refined class output.
+    """
+    def __init__(self):
+        super(ClsLoss, self).__init__()
+        self.config = Config()
+        self.lambdas_cls = self.config.lambdas_cls
+
+        self.criterions_last = {
+            'ce': nn.CrossEntropyLoss()
+        }
+
+    def forward(self, preds, gt):
+        loss = 0.
+        for _, pred_lvl in enumerate(preds):
+            if pred_lvl is None:
+                continue
+            for criterion_name, criterion in self.criterions_last.items():
+                loss += criterion(pred_lvl, gt) * self.lambdas_cls[criterion_name]
+        return loss
+
+
+class PixLoss(nn.Module):
+    """
+    Pixel loss for each refined map output.
+    """
+    def __init__(self):
+        super(PixLoss, self).__init__()
+        self.config = Config()
+        self.lambdas_pix_last = self.config.lambdas_pix_last
+
+        self.criterions_last = {}
+        if 'bce' in self.lambdas_pix_last and self.lambdas_pix_last['bce']:
+            self.criterions_last['bce'] = nn.BCELoss() if not self.config.use_fp16 else nn.BCEWithLogitsLoss()
+        if 'iou' in self.lambdas_pix_last and self.lambdas_pix_last['iou']:
+            self.criterions_last['iou'] = IoULoss()
+        if 'iou_patch' in self.lambdas_pix_last and self.lambdas_pix_last['iou_patch']:
+            self.criterions_last['iou_patch'] = PatchIoULoss()
+        if 'ssim' in self.lambdas_pix_last and self.lambdas_pix_last['ssim']:
+            self.criterions_last['ssim'] = SSIMLoss()
+        if 'mse' in self.lambdas_pix_last and self.lambdas_pix_last['mse']:
+            self.criterions_last['mse'] = nn.MSELoss()
+        if 'reg' in self.lambdas_pix_last and self.lambdas_pix_last['reg']:
+            self.criterions_last['reg'] = ThrReg_loss()
+        if 'cnt' in self.lambdas_pix_last and self.lambdas_pix_last['cnt']:
+            self.criterions_last['cnt'] = ContourLoss()
+        if 'structure' in self.lambdas_pix_last and self.lambdas_pix_last['structure']:
+            self.criterions_last['structure'] = StructureLoss()
+
+    def forward(self, scaled_preds, gt):
+        loss = 0.
+        criterions_embedded_with_sigmoid = ['structure', ] + ['bce'] if self.config.use_fp16 else []
+        for _, pred_lvl in enumerate(scaled_preds):
+            if pred_lvl.shape != gt.shape:
+                pred_lvl = nn.functional.interpolate(pred_lvl, size=gt.shape[2:], mode='bilinear', align_corners=True)
+            for criterion_name, criterion in self.criterions_last.items():
+                _loss = criterion(pred_lvl.sigmoid() if criterion_name not in criterions_embedded_with_sigmoid else pred_lvl, gt) * self.lambdas_pix_last[criterion_name]
+                loss += _loss
+                # print(criterion_name, _loss.item())
+        return loss
+
+
+class SSIMLoss(torch.nn.Module):
+    def __init__(self, window_size=11, size_average=True):
+        super(SSIMLoss, self).__init__()
+        self.window_size = window_size
+        self.size_average = size_average
+        self.channel = 1
+        self.window = create_window(window_size, self.channel)
+
+    def forward(self, img1, img2):
+        (_, channel, _, _) = img1.size()
+        if channel == self.channel and self.window.data.type() == img1.data.type():
+            window = self.window
+        else:
+            window = create_window(self.window_size, channel)
+            if img1.is_cuda:
+                window = window.cuda(img1.get_device())
+            window = window.type_as(img1)
+            self.window = window
+            self.channel = channel
+        return 1 - _ssim(img1, img2, window, self.window_size, channel, self.size_average)
+
+
+def gaussian(window_size, sigma):
+    gauss = torch.Tensor([exp(-(x - window_size//2)**2/float(2*sigma**2)) for x in range(window_size)])
+    return gauss/gauss.sum()
+
+
+def create_window(window_size, channel):
+    _1D_window = gaussian(window_size, 1.5).unsqueeze(1)
+    _2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0)
+    window = Variable(_2D_window.expand(channel, 1, window_size, window_size).contiguous())
+    return window
+
+
+def _ssim(img1, img2, window, window_size, channel, size_average=True):
+    mu1 = F.conv2d(img1, window, padding = window_size//2, groups=channel)
+    mu2 = F.conv2d(img2, window, padding = window_size//2, groups=channel)
+
+    mu1_sq = mu1.pow(2)
+    mu2_sq = mu2.pow(2)
+    mu1_mu2 = mu1*mu2
+
+    sigma1_sq = F.conv2d(img1*img1, window, padding=window_size//2, groups=channel) - mu1_sq
+    sigma2_sq = F.conv2d(img2*img2, window, padding=window_size//2, groups=channel) - mu2_sq
+    sigma12 = F.conv2d(img1*img2, window, padding=window_size//2, groups=channel) - mu1_mu2
+
+    C1 = 0.01**2
+    C2 = 0.03**2
+
+    ssim_map = ((2*mu1_mu2 + C1)*(2*sigma12 + C2))/((mu1_sq + mu2_sq + C1)*(sigma1_sq + sigma2_sq + C2))
+
+    if size_average:
+        return ssim_map.mean()
+    else:
+        return ssim_map.mean(1).mean(1).mean(1)
+
+
+def SSIM(x, y):
+    C1 = 0.01 ** 2
+    C2 = 0.03 ** 2
+
+    mu_x = nn.AvgPool2d(3, 1, 1)(x)
+    mu_y = nn.AvgPool2d(3, 1, 1)(y)
+    mu_x_mu_y = mu_x * mu_y
+    mu_x_sq = mu_x.pow(2)
+    mu_y_sq = mu_y.pow(2)
+
+    sigma_x = nn.AvgPool2d(3, 1, 1)(x * x) - mu_x_sq
+    sigma_y = nn.AvgPool2d(3, 1, 1)(y * y) - mu_y_sq
+    sigma_xy = nn.AvgPool2d(3, 1, 1)(x * y) - mu_x_mu_y
+
+    SSIM_n = (2 * mu_x_mu_y + C1) * (2 * sigma_xy + C2)
+    SSIM_d = (mu_x_sq + mu_y_sq + C1) * (sigma_x + sigma_y + C2)
+    SSIM = SSIM_n / SSIM_d
+
+    return torch.clamp((1 - SSIM) / 2, 0, 1)
+
+
+def saliency_structure_consistency(x, y):
+    ssim = torch.mean(SSIM(x,y))
+    return ssim

BiRefNet_github/make_a_copy.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+# Set dst repo here.
+repo=$1
+mkdir ../${repo}
+mkdir ../${repo}/evaluation
+mkdir ../${repo}/models
+mkdir ../${repo}/models/backbones
+mkdir ../${repo}/models/modules
+mkdir ../${repo}/models/refinement
+
+cp ./*.sh ../${repo}
+cp ./*.py ../${repo}
+cp ./evaluation/*.py ../${repo}/evaluation
+cp ./models/*.py ../${repo}/models
+cp ./models/backbones/*.py ../${repo}/models/backbones
+cp ./models/modules/*.py ../${repo}/models/modules
+cp ./models/refinement/*.py ../${repo}/models/refinement
+cp -r ./.git* ../${repo}

BiRefNet_github/models/backbones/build_backbone.py

@@ -0,0 +1,44 @@
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+from torchvision.models import vgg16, vgg16_bn, VGG16_Weights, VGG16_BN_Weights, resnet50, ResNet50_Weights
+from models.backbones.pvt_v2 import pvt_v2_b0, pvt_v2_b1, pvt_v2_b2, pvt_v2_b5
+from models.backbones.swin_v1 import swin_v1_t, swin_v1_s, swin_v1_b, swin_v1_l
+from config import Config
+
+
+config = Config()
+
+def build_backbone(bb_name, pretrained=True, params_settings=''):
+    if bb_name == 'vgg16':
+        bb_net = list(vgg16(pretrained=VGG16_Weights.DEFAULT if pretrained else None).children())[0]
+        bb = nn.Sequential(OrderedDict({'conv1': bb_net[:4], 'conv2': bb_net[4:9], 'conv3': bb_net[9:16], 'conv4': bb_net[16:23]}))
+    elif bb_name == 'vgg16bn':
+        bb_net = list(vgg16_bn(pretrained=VGG16_BN_Weights.DEFAULT if pretrained else None).children())[0]
+        bb = nn.Sequential(OrderedDict({'conv1': bb_net[:6], 'conv2': bb_net[6:13], 'conv3': bb_net[13:23], 'conv4': bb_net[23:33]}))
+    elif bb_name == 'resnet50':
+        bb_net = list(resnet50(pretrained=ResNet50_Weights.DEFAULT if pretrained else None).children())
+        bb = nn.Sequential(OrderedDict({'conv1': nn.Sequential(*bb_net[0:3]), 'conv2': bb_net[4], 'conv3': bb_net[5], 'conv4': bb_net[6]}))
+    else:
+        bb = eval('{}({})'.format(bb_name, params_settings))
+        if pretrained:
+            bb = load_weights(bb, bb_name)
+    return bb
+
+def load_weights(model, model_name):
+    save_model = torch.load(config.weights[model_name], map_location='cpu')
+    model_dict = model.state_dict()
+    state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model.items() if k in model_dict.keys()}
+    # to ignore the weights with mismatched size when I modify the backbone itself.
+    if not state_dict:
+        save_model_keys = list(save_model.keys())
+        sub_item = save_model_keys[0] if len(save_model_keys) == 1 else None
+        state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model[sub_item].items() if k in model_dict.keys()}
+        if not state_dict or not sub_item:
+            print('Weights are not successully loaded. Check the state dict of weights file.')
+            return None
+        else:
+            print('Found correct weights in the "{}" item of loaded state_dict.'.format(sub_item))
+    model_dict.update(state_dict)
+    model.load_state_dict(model_dict)
+    return model

BiRefNet_github/models/backbones/pvt_v2.py

@@ -0,0 +1,435 @@
+import torch
+import torch.nn as nn
+from functools import partial
+
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+from timm.models.registry import register_model
+
+import math
+
+from config import Config
+
+config = Config()
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.dwconv = DWConv(hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = self.fc1(x)
+        x = self.dwconv(x, H, W)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
+        super().__init__()
+        assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
+
+        self.dim = dim
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.q = nn.Linear(dim, dim, bias=qkv_bias)
+        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
+        self.attn_drop_prob = attn_drop
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        self.sr_ratio = sr_ratio
+        if sr_ratio > 1:
+            self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
+            self.norm = nn.LayerNorm(dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+
+        if self.sr_ratio > 1:
+            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
+            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
+            x_ = self.norm(x_)
+            kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        else:
+            kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        k, v = kv[0], kv[1]
+
+        if config.SDPA_enabled:
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
+            ).transpose(1, 2).reshape(B, N, C)
+        else:
+            attn = (q @ k.transpose(-2, -1)) * self.scale
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+
+            x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+
+        return x
+
+
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = x + self.drop_path(self.attn(self.norm1(x), H, W))
+        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
+
+        return x
+
+
+class OverlapPatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+
+    def __init__(self, img_size=224, patch_size=7, stride=4, in_channels=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
+        self.num_patches = self.H * self.W
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=stride,
+                              padding=(patch_size[0] // 2, patch_size[1] // 2))
+        self.norm = nn.LayerNorm(embed_dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x):
+        x = self.proj(x)
+        _, _, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = self.norm(x)
+
+        return x, H, W
+
+
+class PyramidVisionTransformerImpr(nn.Module):
+    def __init__(self, img_size=224, patch_size=16, in_channels=3, num_classes=1000, embed_dims=[64, 128, 256, 512],
+                 num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,
+                 attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
+                 depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]):
+        super().__init__()
+        self.num_classes = num_classes
+        self.depths = depths
+
+        # patch_embed
+        self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_channels=in_channels,
+                                              embed_dim=embed_dims[0])
+        self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_channels=embed_dims[0],
+                                              embed_dim=embed_dims[1])
+        self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_channels=embed_dims[1],
+                                              embed_dim=embed_dims[2])
+        self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_channels=embed_dims[2],
+                                              embed_dim=embed_dims[3])
+
+        # transformer encoder
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+        cur = 0
+        self.block1 = nn.ModuleList([Block(
+            dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[0])
+            for i in range(depths[0])])
+        self.norm1 = norm_layer(embed_dims[0])
+
+        cur += depths[0]
+        self.block2 = nn.ModuleList([Block(
+            dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[1])
+            for i in range(depths[1])])
+        self.norm2 = norm_layer(embed_dims[1])
+
+        cur += depths[1]
+        self.block3 = nn.ModuleList([Block(
+            dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[2])
+            for i in range(depths[2])])
+        self.norm3 = norm_layer(embed_dims[2])
+
+        cur += depths[2]
+        self.block4 = nn.ModuleList([Block(
+            dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[3])
+            for i in range(depths[3])])
+        self.norm4 = norm_layer(embed_dims[3])
+
+        # classification head
+        # self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity()
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def init_weights(self, pretrained=None):
+        if isinstance(pretrained, str):
+            logger = 1
+            #load_checkpoint(self, pretrained, map_location='cpu', strict=False, logger=logger)
+
+    def reset_drop_path(self, drop_path_rate):
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))]
+        cur = 0
+        for i in range(self.depths[0]):
+            self.block1[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[0]
+        for i in range(self.depths[1]):
+            self.block2[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[1]
+        for i in range(self.depths[2]):
+            self.block3[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[2]
+        for i in range(self.depths[3]):
+            self.block4[i].drop_path.drop_prob = dpr[cur + i]
+
+    def freeze_patch_emb(self):
+        self.patch_embed1.requires_grad = False
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'}  # has pos_embed may be better
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+    def forward_features(self, x):
+        B = x.shape[0]
+        outs = []
+
+        # stage 1
+        x, H, W = self.patch_embed1(x)
+        for i, blk in enumerate(self.block1):
+            x = blk(x, H, W)
+        x = self.norm1(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 2
+        x, H, W = self.patch_embed2(x)
+        for i, blk in enumerate(self.block2):
+            x = blk(x, H, W)
+        x = self.norm2(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 3
+        x, H, W = self.patch_embed3(x)
+        for i, blk in enumerate(self.block3):
+            x = blk(x, H, W)
+        x = self.norm3(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 4
+        x, H, W = self.patch_embed4(x)
+        for i, blk in enumerate(self.block4):
+            x = blk(x, H, W)
+        x = self.norm4(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        return outs
+
+        # return x.mean(dim=1)
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        # x = self.head(x)
+
+        return x
+
+
+class DWConv(nn.Module):
+    def __init__(self, dim=768):
+        super(DWConv, self).__init__()
+        self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        x = x.transpose(1, 2).view(B, C, H, W).contiguous()
+        x = self.dwconv(x)
+        x = x.flatten(2).transpose(1, 2)
+
+        return x
+
+
+def _conv_filter(state_dict, patch_size=16):
+    """ convert patch embedding weight from manual patchify + linear proj to conv"""
+    out_dict = {}
+    for k, v in state_dict.items():
+        if 'patch_embed.proj.weight' in k:
+            v = v.reshape((v.shape[0], 3, patch_size, patch_size))
+        out_dict[k] = v
+
+    return out_dict
+
+
+## @register_model
+class pvt_v2_b0(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b0, self).__init__(
+            patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+
+## @register_model
+class pvt_v2_b1(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b1, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+## @register_model
+class pvt_v2_b2(PyramidVisionTransformerImpr):
+    def __init__(self, in_channels=3, **kwargs):
+        super(pvt_v2_b2, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1, in_channels=in_channels)
+
+## @register_model
+class pvt_v2_b3(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b3, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+## @register_model
+class pvt_v2_b4(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b4, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+## @register_model
+class pvt_v2_b5(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b5, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)

BiRefNet_github/models/backbones/swin_v1.py

@@ -0,0 +1,627 @@
+# --------------------------------------------------------
+# Swin Transformer
+# Copyright (c) 2021 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ze Liu, Yutong Lin, Yixuan Wei
+# --------------------------------------------------------
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+import numpy as np
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+from config import Config
+
+
+config = Config()
+
+class Mlp(nn.Module):
+    """ Multilayer perceptron."""
+
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    """ Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w], indexing='ij'))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop_prob = attn_drop
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """ Forward function.
+
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+
+        if config.SDPA_enabled:
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
+            ).transpose(1, 2).reshape(B_, N, C)
+        else:
+            attn = (q @ k.transpose(-2, -1))
+
+            relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+                self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+            relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+            attn = attn + relative_position_bias.unsqueeze(0)
+
+            if mask is not None:
+                nW = mask.shape[0]
+                attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+                attn = attn.view(-1, self.num_heads, N, N)
+                attn = self.softmax(attn)
+            else:
+                attn = self.softmax(attn)
+
+            attn = self.attn_drop(attn)
+
+            x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class SwinTransformerBlock(nn.Module):
+    """ Swin Transformer Block.
+
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.H = None
+        self.W = None
+
+    def forward(self, x, mask_matrix):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+            mask_matrix: Attention mask for cyclic shift.
+        """
+        B, L, C = x.shape
+        H, W = self.H, self.W
+        assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # pad feature maps to multiples of window size
+        pad_l = pad_t = 0
+        pad_r = (self.window_size - W % self.window_size) % self.window_size
+        pad_b = (self.window_size - H % self.window_size) % self.window_size
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+        _, Hp, Wp, _ = x.shape
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+
+        if pad_r > 0 or pad_b > 0:
+            x = x[:, :H, :W, :].contiguous()
+
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+
+class PatchMerging(nn.Module):
+    """ Patch Merging Layer
+
+    Args:
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+    def __init__(self, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x, H, W):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+
+        x = x.view(B, H, W, C)
+
+        # padding
+        pad_input = (H % 2 == 1) or (W % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+
+    Args:
+        dim (int): Number of feature channels
+        depth (int): Depths of this stage.
+        num_heads (int): Number of attention head.
+        window_size (int): Local window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 dim,
+                 depth,
+                 num_heads,
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 norm_layer=nn.LayerNorm,
+                 downsample=None,
+                 use_checkpoint=False):
+        super().__init__()
+        self.window_size = window_size
+        self.shift_size = window_size // 2
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(
+                dim=dim,
+                num_heads=num_heads,
+                window_size=window_size,
+                shift_size=0 if (i % 2 == 0) else window_size // 2,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop,
+                attn_drop=attn_drop,
+                drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                norm_layer=norm_layer)
+            for i in range(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, H, W):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+
+        # calculate attention mask for SW-MSA
+        Hp = int(np.ceil(H / self.window_size)) * self.window_size
+        Wp = int(np.ceil(W / self.window_size)) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device)  # 1 Hp Wp 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        for blk in self.blocks:
+            blk.H, blk.W = H, W
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, attn_mask)
+            else:
+                x = blk(x, attn_mask)
+        if self.downsample is not None:
+            x_down = self.downsample(x, H, W)
+            Wh, Ww = (H + 1) // 2, (W + 1) // 2
+            return x, H, W, x_down, Wh, Ww
+        else:
+            return x, H, W, x, H, W
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+
+    Args:
+        patch_size (int): Patch token size. Default: 4.
+        in_channels (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, patch_size=4, in_channels=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+
+        self.in_channels = in_channels
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        """Forward function."""
+        # padding
+        _, _, H, W = x.size()
+        if W % self.patch_size[1] != 0:
+            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
+        if H % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
+
+        x = self.proj(x)  # B C Wh Ww
+        if self.norm is not None:
+            Wh, Ww = x.size(2), x.size(3)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
+
+        return x
+
+
+class SwinTransformer(nn.Module):
+    """ Swin Transformer backbone.
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+
+    Args:
+        pretrain_img_size (int): Input image size for training the pretrained model,
+            used in absolute postion embedding. Default 224.
+        patch_size (int | tuple(int)): Patch size. Default: 4.
+        in_channels (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        depths (tuple[int]): Depths of each Swin Transformer stage.
+        num_heads (tuple[int]): Number of attention head of each stage.
+        window_size (int): Window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
+        drop_rate (float): Dropout rate.
+        attn_drop_rate (float): Attention dropout rate. Default: 0.
+        drop_path_rate (float): Stochastic depth rate. Default: 0.2.
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True.
+        out_indices (Sequence[int]): Output from which stages.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 pretrain_img_size=224,
+                 patch_size=4,
+                 in_channels=3,
+                 embed_dim=96,
+                 depths=[2, 2, 6, 2],
+                 num_heads=[3, 6, 12, 24],
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop_rate=0.,
+                 attn_drop_rate=0.,
+                 drop_path_rate=0.2,
+                 norm_layer=nn.LayerNorm,
+                 ape=False,
+                 patch_norm=True,
+                 out_indices=(0, 1, 2, 3),
+                 frozen_stages=-1,
+                 use_checkpoint=False):
+        super().__init__()
+
+        self.pretrain_img_size = pretrain_img_size
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.out_indices = out_indices
+        self.frozen_stages = frozen_stages
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            patch_size=patch_size, in_channels=in_channels, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        # absolute position embedding
+        if self.ape:
+            pretrain_img_size = to_2tuple(pretrain_img_size)
+            patch_size = to_2tuple(patch_size)
+            patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]
+
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+
+        # build layers
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(embed_dim * 2 ** i_layer),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
+                norm_layer=norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint)
+            self.layers.append(layer)
+
+        num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
+        self.num_features = num_features
+
+        # add a norm layer for each output
+        for i_layer in out_indices:
+            layer = norm_layer(num_features[i_layer])
+            layer_name = f'norm{i_layer}'
+            self.add_module(layer_name, layer)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        if self.frozen_stages >= 1 and self.ape:
+            self.absolute_pos_embed.requires_grad = False
+
+        if self.frozen_stages >= 2:
+            self.pos_drop.eval()
+            for i in range(0, self.frozen_stages - 1):
+                m = self.layers[i]
+                m.eval()
+                for param in m.parameters():
+                    param.requires_grad = False
+
+
+    def forward(self, x):
+        """Forward function."""
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+        if self.ape:
+            # interpolate the position embedding to the corresponding size
+            absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
+            x = (x + absolute_pos_embed) # B Wh*Ww C
+            
+        outs = []#x.contiguous()]
+        x = x.flatten(2).transpose(1, 2)
+        x = self.pos_drop(x)
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f'norm{i}')
+                x_out = norm_layer(x_out)
+
+                out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
+                outs.append(out)
+
+        return tuple(outs)
+
+    def train(self, mode=True):
+        """Convert the model into training mode while keep layers freezed."""
+        super(SwinTransformer, self).train(mode)
+        self._freeze_stages()
+
+def swin_v1_t():
+    model = SwinTransformer(embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7)
+    return model
+
+def swin_v1_s():
+    model = SwinTransformer(embed_dim=96, depths=[2, 2, 18, 2], num_heads=[3, 6, 12, 24], window_size=7)
+    return model
+
+def swin_v1_b():
+    model = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12)
+    return model
+
+def swin_v1_l():
+    model = SwinTransformer(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=12)
+    return model

BiRefNet_github/models/birefnet.py

@@ -0,0 +1,287 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from kornia.filters import laplacian
+from huggingface_hub import PyTorchModelHubMixin
+
+from config import Config
+from dataset import class_labels_TR_sorted
+from models.backbones.build_backbone import build_backbone
+from models.modules.decoder_blocks import BasicDecBlk, ResBlk, HierarAttDecBlk
+from models.modules.lateral_blocks import BasicLatBlk
+from models.modules.aspp import ASPP, ASPPDeformable
+from models.modules.ing import *
+from models.refinement.refiner import Refiner, RefinerPVTInChannels4, RefUNet
+from models.refinement.stem_layer import StemLayer
+
+
+class BiRefNet(
+    nn.Module,
+    PyTorchModelHubMixin,
+    library_name="birefnet",
+    repo_url="https://github.com/ZhengPeng7/BiRefNet",
+    tags=['Image Segmentation', 'Background Removal', 'Mask Generation', 'Dichotomous Image Segmentation', 'Camouflaged Object Detection', 'Salient Object Detection']
+):
+    def __init__(self, bb_pretrained=True):
+        super(BiRefNet, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        self.bb = build_backbone(self.config.bb, pretrained=bb_pretrained)
+
+        channels = self.config.lateral_channels_in_collection
+
+        if self.config.auxiliary_classification:
+            self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+            self.cls_head = nn.Sequential(
+                nn.Linear(channels[0], len(class_labels_TR_sorted))
+            )
+
+        if self.config.squeeze_block:
+            self.squeeze_module = nn.Sequential(*[
+                eval(self.config.squeeze_block.split('_x')[0])(channels[0]+sum(self.config.cxt), channels[0])
+                for _ in range(eval(self.config.squeeze_block.split('_x')[1]))
+            ])
+
+        self.decoder = Decoder(channels)
+
+        if self.config.ender:
+            self.dec_end = nn.Sequential(
+                nn.Conv2d(1, 16, 3, 1, 1),
+                nn.Conv2d(16, 1, 3, 1, 1),
+                nn.ReLU(inplace=True),
+            )
+
+        # refine patch-level segmentation
+        if self.config.refine:
+            if self.config.refine == 'itself':
+                self.stem_layer = StemLayer(in_channels=3+1, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
+            else:
+                self.refiner = eval('{}({})'.format(self.config.refine, 'in_channels=3+1'))
+
+        if self.config.freeze_bb:
+            # Freeze the backbone...
+            print(self.named_parameters())
+            for key, value in self.named_parameters():
+                if 'bb.' in key and 'refiner.' not in key:
+                    value.requires_grad = False
+
+    def forward_enc(self, x):
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x); x2 = self.bb.conv2(x1); x3 = self.bb.conv3(x2); x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+            if self.config.mul_scl_ipt == 'cat':
+                B, C, H, W = x.shape
+                x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+                x1 = torch.cat([x1, F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x2 = torch.cat([x2, F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x3 = torch.cat([x3, F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x4 = torch.cat([x4, F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+            elif self.config.mul_scl_ipt == 'add':
+                B, C, H, W = x.shape
+                x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+                x1 = x1 + F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)
+                x2 = x2 + F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)
+                x3 = x3 + F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)
+                x4 = x4 + F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)
+        class_preds = self.cls_head(self.avgpool(x4).view(x4.shape[0], -1)) if self.training and self.config.auxiliary_classification else None
+        if self.config.cxt:
+            x4 = torch.cat(
+                (
+                    *[
+                        F.interpolate(x1, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                        F.interpolate(x2, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                        F.interpolate(x3, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                    ][-len(self.config.cxt):],
+                    x4
+                ),
+                dim=1
+            )
+        return (x1, x2, x3, x4), class_preds
+
+    def forward_ori(self, x):
+        ########## Encoder ##########
+        (x1, x2, x3, x4), class_preds = self.forward_enc(x)
+        if self.config.squeeze_block:
+            x4 = self.squeeze_module(x4)
+        ########## Decoder ##########
+        features = [x, x1, x2, x3, x4]
+        if self.training and self.config.out_ref:
+            features.append(laplacian(torch.mean(x, dim=1).unsqueeze(1), kernel_size=5))
+        scaled_preds = self.decoder(features)
+        return scaled_preds, class_preds
+
+    def forward(self, x):
+        scaled_preds, class_preds = self.forward_ori(x)
+        class_preds_lst = [class_preds]
+        return [scaled_preds, class_preds_lst] if self.training else scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels):
+        super(Decoder, self).__init__()
+        self.config = Config()
+        DecoderBlock = eval(self.config.dec_blk)
+        LateralBlock = eval(self.config.lat_blk)
+
+        if self.config.dec_ipt:
+            self.split = self.config.dec_ipt_split
+            N_dec_ipt = 64
+            DBlock = SimpleConvs
+            ic = 64
+            ipt_cha_opt = 1
+            self.ipt_blk5 = DBlock(2**10*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk4 = DBlock(2**8*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk3 = DBlock(2**6*3 if self.split else 3, [N_dec_ipt, channels[1]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk2 = DBlock(2**4*3 if self.split else 3, [N_dec_ipt, channels[2]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk1 = DBlock(2**0*3 if self.split else 3, [N_dec_ipt, channels[3]//8][ipt_cha_opt], inter_channels=ic)
+        else:
+            self.split = None
+
+        self.decoder_block4 = DecoderBlock(channels[0]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2]+([N_dec_ipt, channels[1]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3])
+        self.decoder_block1 = DecoderBlock(channels[3]+([N_dec_ipt, channels[2]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3]//2)
+        self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2+([N_dec_ipt, channels[3]//8][ipt_cha_opt] if self.config.dec_ipt else 0), 1, 1, 1, 0))
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        if self.config.ms_supervision:
+            self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
+            self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
+            self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
+
+            if self.config.out_ref:
+                _N = 16
+                self.gdt_convs_4 = nn.Sequential(nn.Conv2d(channels[1], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+                self.gdt_convs_3 = nn.Sequential(nn.Conv2d(channels[2], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+                self.gdt_convs_2 = nn.Sequential(nn.Conv2d(channels[3], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
+
+                self.gdt_convs_pred_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_pred_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_pred_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                
+                self.gdt_convs_attn_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_attn_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_attn_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+
+    def get_patches_batch(self, x, p):
+        _size_h, _size_w = p.shape[2:]
+        patches_batch = []
+        for idx in range(x.shape[0]):
+            columns_x = torch.split(x[idx], split_size_or_sections=_size_w, dim=-1)
+            patches_x = []
+            for column_x in columns_x:
+                patches_x += [p.unsqueeze(0) for p in torch.split(column_x, split_size_or_sections=_size_h, dim=-2)]
+            patch_sample = torch.cat(patches_x, dim=1)
+            patches_batch.append(patch_sample)
+        return torch.cat(patches_batch, dim=0)
+
+    def forward(self, features):
+        if self.training and self.config.out_ref:
+            outs_gdt_pred = []
+            outs_gdt_label = []
+            x, x1, x2, x3, x4, gdt_gt = features
+        else:
+            x, x1, x2, x3, x4 = features
+        outs = []
+
+        if self.config.dec_ipt:
+            patches_batch = self.get_patches_batch(x, x4) if self.split else x
+            x4 = torch.cat((x4, self.ipt_blk5(F.interpolate(patches_batch, size=x4.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p4 = self.decoder_block4(x4)
+        m4 = self.conv_ms_spvn_4(p4) if self.config.ms_supervision else None
+        if self.config.out_ref:
+            p4_gdt = self.gdt_convs_4(p4)
+            if self.training:
+                # >> GT:
+                m4_dia = m4
+                gdt_label_main_4 = gdt_gt * F.interpolate(m4_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_4)
+                # >> Pred:
+                gdt_pred_4 = self.gdt_convs_pred_4(p4_gdt)
+                outs_gdt_pred.append(gdt_pred_4)
+            gdt_attn_4 = self.gdt_convs_attn_4(p4_gdt).sigmoid()
+            # >> Finally:
+            p4 = p4 * gdt_attn_4
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        if self.config.dec_ipt:
+            patches_batch = self.get_patches_batch(x, _p3) if self.split else x
+            _p3 = torch.cat((_p3, self.ipt_blk4(F.interpolate(patches_batch, size=x3.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p3 = self.decoder_block3(_p3)
+        m3 = self.conv_ms_spvn_3(p3) if self.config.ms_supervision else None
+        if self.config.out_ref:
+            p3_gdt = self.gdt_convs_3(p3)
+            if self.training:
+                # >> GT:
+                # m3 --dilation--> m3_dia
+                # G_3^gt * m3_dia --> G_3^m, which is the label of gradient
+                m3_dia = m3
+                gdt_label_main_3 = gdt_gt * F.interpolate(m3_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_3)
+                # >> Pred:
+                # p3 --conv--BN--> F_3^G, where F_3^G predicts the \hat{G_3} with xx
+                # F_3^G --sigmoid--> A_3^G
+                gdt_pred_3 = self.gdt_convs_pred_3(p3_gdt)
+                outs_gdt_pred.append(gdt_pred_3)
+            gdt_attn_3 = self.gdt_convs_attn_3(p3_gdt).sigmoid()
+            # >> Finally:
+            # p3 = p3 * A_3^G
+            p3 = p3 * gdt_attn_3
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        if self.config.dec_ipt:
+            patches_batch = self.get_patches_batch(x, _p2) if self.split else x
+            _p2 = torch.cat((_p2, self.ipt_blk3(F.interpolate(patches_batch, size=x2.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p2 = self.decoder_block2(_p2)
+        m2 = self.conv_ms_spvn_2(p2) if self.config.ms_supervision else None
+        if self.config.out_ref:
+            p2_gdt = self.gdt_convs_2(p2)
+            if self.training:
+                # >> GT:
+                m2_dia = m2
+                gdt_label_main_2 = gdt_gt * F.interpolate(m2_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_2)
+                # >> Pred:
+                gdt_pred_2 = self.gdt_convs_pred_2(p2_gdt)
+                outs_gdt_pred.append(gdt_pred_2)
+            gdt_attn_2 = self.gdt_convs_attn_2(p2_gdt).sigmoid()
+            # >> Finally:
+            p2 = p2 * gdt_attn_2
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        if self.config.dec_ipt:
+            patches_batch = self.get_patches_batch(x, _p1) if self.split else x
+            _p1 = torch.cat((_p1, self.ipt_blk2(F.interpolate(patches_batch, size=x1.shape[2:], mode='bilinear', align_corners=True))), 1)
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+
+        if self.config.dec_ipt:
+            patches_batch = self.get_patches_batch(x, _p1) if self.split else x
+            _p1 = torch.cat((_p1, self.ipt_blk1(F.interpolate(patches_batch, size=x.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p1_out = self.conv_out1(_p1)
+
+        if self.config.ms_supervision:
+            outs.append(m4)
+            outs.append(m3)
+            outs.append(m2)
+        outs.append(p1_out)
+        return outs if not (self.config.out_ref and self.training) else ([outs_gdt_pred, outs_gdt_label], outs)
+
+
+class SimpleConvs(nn.Module):
+    def __init__(
+        self, in_channels: int, out_channels: int, inter_channels=64
+    ) -> None:
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, 1)
+
+    def forward(self, x):
+        return self.conv_out(self.conv1(x))

BiRefNet_github/models/modules/aspp.py

@@ -0,0 +1,119 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from models.modules.deform_conv import DeformableConv2d
+from config import Config
+
+
+config = Config()
+
+
+class _ASPPModule(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding, dilation):
+        super(_ASPPModule, self).__init__()
+        self.atrous_conv = nn.Conv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, dilation=dilation, bias=False)
+        self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPP(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, output_stride=16):
+        super(ASPP, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+        if output_stride == 16:
+            dilations = [1, 6, 12, 18]
+        elif output_stride == 8:
+            dilations = [1, 12, 24, 36]
+        else:
+            raise NotImplementedError
+
+        self.aspp1 = _ASPPModule(in_channels, self.in_channelster, 1, padding=0, dilation=dilations[0])
+        self.aspp2 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[1], dilation=dilations[1])
+        self.aspp3 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[2], dilation=dilations[2])
+        self.aspp4 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[3], dilation=dilations[3])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
+                                             nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = nn.Conv2d(self.in_channelster * 5, out_channels, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(0.5)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x2 = self.aspp2(x)
+        x3 = self.aspp3(x)
+        x4 = self.aspp4(x)
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, x2, x3, x4, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return self.dropout(x)
+
+
+##################### Deformable
+class _ASPPModuleDeformable(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding):
+        super(_ASPPModuleDeformable, self).__init__()
+        self.atrous_conv = DeformableConv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, bias=False)
+        self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPPDeformable(nn.Module):
+    def __init__(self, in_channels, out_channels=None, parallel_block_sizes=[1, 3, 7]):
+        super(ASPPDeformable, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+
+        self.aspp1 = _ASPPModuleDeformable(in_channels, self.in_channelster, 1, padding=0)
+        self.aspp_deforms = nn.ModuleList([
+            _ASPPModuleDeformable(in_channels, self.in_channelster, conv_size, padding=int(conv_size//2)) for conv_size in parallel_block_sizes
+        ])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
+                                             nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = nn.Conv2d(self.in_channelster * (2 + len(self.aspp_deforms)), out_channels, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(0.5)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x_aspp_deforms = [aspp_deform(x) for aspp_deform in self.aspp_deforms]
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, *x_aspp_deforms, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return self.dropout(x)

BiRefNet_github/models/modules/attentions.py

@@ -0,0 +1,93 @@
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import init
+
+
+class SEWeightModule(nn.Module):
+    def __init__(self, channels, reduction=16):
+        super(SEWeightModule, self).__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d(1)
+        self.fc1 = nn.Conv2d(channels, channels//reduction, kernel_size=1, padding=0)
+        self.relu = nn.ReLU(inplace=True)
+        self.fc2 = nn.Conv2d(channels//reduction, channels, kernel_size=1, padding=0)
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x):
+        out = self.avg_pool(x)
+        out = self.fc1(out)
+        out = self.relu(out)
+        out = self.fc2(out)
+        weight = self.sigmoid(out)
+        return weight
+
+
+class PSA(nn.Module):
+
+    def __init__(self, in_channels, S=4, reduction=4):
+        super().__init__()
+        self.S = S
+
+        _convs = []
+        for i in range(S):
+            _convs.append(nn.Conv2d(in_channels//S, in_channels//S, kernel_size=2*(i+1)+1, padding=i+1))
+        self.convs = nn.ModuleList(_convs)
+
+        self.se_block = SEWeightModule(in_channels//S, reduction=S*reduction)
+
+        self.softmax = nn.Softmax(dim=1)
+
+    def forward(self, x):
+        b, c, h, w = x.size()
+
+        # Step1: SPC module
+        SPC_out = x.view(b, self.S, c//self.S, h, w) #bs,s,ci,h,w
+        for idx, conv in enumerate(self.convs):
+            SPC_out[:,idx,:,:,:] = conv(SPC_out[:,idx,:,:,:].clone())
+
+        # Step2: SE weight
+        se_out=[]
+        for idx in range(self.S):
+            se_out.append(self.se_block(SPC_out[:, idx, :, :, :]))
+        SE_out = torch.stack(se_out, dim=1)
+        SE_out = SE_out.expand_as(SPC_out)
+
+        # Step3: Softmax
+        softmax_out = self.softmax(SE_out)
+
+        # Step4: SPA
+        PSA_out = SPC_out * softmax_out
+        PSA_out = PSA_out.view(b, -1, h, w)
+
+        return PSA_out
+
+
+class SGE(nn.Module):
+
+    def __init__(self, groups):
+        super().__init__()
+        self.groups=groups
+        self.avg_pool = nn.AdaptiveAvgPool2d(1)
+        self.weight=nn.Parameter(torch.zeros(1,groups,1,1))
+        self.bias=nn.Parameter(torch.zeros(1,groups,1,1))
+        self.sig=nn.Sigmoid()
+
+    def forward(self, x):
+        b, c, h,w=x.shape
+        x=x.view(b*self.groups,-1,h,w) #bs*g,dim//g,h,w
+        xn=x*self.avg_pool(x) #bs*g,dim//g,h,w
+        xn=xn.sum(dim=1,keepdim=True) #bs*g,1,h,w
+        t=xn.view(b*self.groups,-1) #bs*g,h*w
+
+        t=t-t.mean(dim=1,keepdim=True) #bs*g,h*w
+        std=t.std(dim=1,keepdim=True)+1e-5
+        t=t/std #bs*g,h*w
+        t=t.view(b,self.groups,h,w) #bs,g,h*w
+        
+        t=t*self.weight+self.bias #bs,g,h*w
+        t=t.view(b*self.groups,1,h,w) #bs*g,1,h*w
+        x=x*self.sig(t)
+        x=x.view(b,c,h,w)
+
+        return x 
+

BiRefNet_github/models/modules/decoder_blocks.py

@@ -0,0 +1,101 @@
+import torch
+import torch.nn as nn
+from models.modules.aspp import ASPP, ASPPDeformable
+from models.modules.attentions import PSA, SGE
+from config import Config
+
+
+config = Config()
+
+
+class BasicDecBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
+        super(BasicDecBlk, self).__init__()
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
+        self.relu_in = nn.ReLU(inplace=True)
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
+        self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
+        self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x
+
+
+class ResBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, inter_channels=64):
+        super(ResBlk, self).__init__()
+        if out_channels is None:
+            out_channels = in_channels
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
+        self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu_in = nn.ReLU(inplace=True)
+
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
+        self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        
+        self.conv_resi = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
+
+    def forward(self, x):
+        _x = self.conv_resi(x)
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x + _x
+
+
+class HierarAttDecBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, inter_channels=64):
+        super(HierarAttDecBlk, self).__init__()
+        if out_channels is None:
+            out_channels = in_channels
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.split_y = 8     # must be divided by channels of all intermediate features
+        self.split_x = 8
+
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
+
+        self.psa = PSA(inter_channels*self.split_y*self.split_x, S=config.batch_size)
+        self.sge = SGE(groups=config.batch_size)
+
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, 1)
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        N, C, H, W = x.shape
+        x_patchs = x.reshape(N, -1, H//self.split_y, W//self.split_x)
+
+        # Hierarchical attention: group attention X patch spatial attention
+        x_patchs = self.psa(x_patchs)   # Group Channel Attention -- each group is a single image
+        x_patchs = self.sge(x_patchs)   # Patch Spatial Attention
+        x = x.reshape(N, C, H, W)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        return x

BiRefNet_github/models/modules/deform_conv.py

@@ -0,0 +1,66 @@
+import torch
+import torch.nn as nn
+from torchvision.ops import deform_conv2d
+
+
+class DeformableConv2d(nn.Module):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 stride=1,
+                 padding=1,
+                 bias=False):
+
+        super(DeformableConv2d, self).__init__()
+        
+        assert type(kernel_size) == tuple or type(kernel_size) == int
+
+        kernel_size = kernel_size if type(kernel_size) == tuple else (kernel_size, kernel_size)
+        self.stride = stride if type(stride) == tuple else (stride, stride)
+        self.padding = padding
+        
+        self.offset_conv = nn.Conv2d(in_channels,
+                                     2 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True)
+
+        nn.init.constant_(self.offset_conv.weight, 0.)
+        nn.init.constant_(self.offset_conv.bias, 0.)
+        
+        self.modulator_conv = nn.Conv2d(in_channels,
+                                     1 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True)
+
+        nn.init.constant_(self.modulator_conv.weight, 0.)
+        nn.init.constant_(self.modulator_conv.bias, 0.)
+
+        self.regular_conv = nn.Conv2d(in_channels,
+                                      out_channels=out_channels,
+                                      kernel_size=kernel_size,
+                                      stride=stride,
+                                      padding=self.padding,
+                                      bias=bias)
+
+    def forward(self, x):
+        #h, w = x.shape[2:]
+        #max_offset = max(h, w)/4.
+
+        offset = self.offset_conv(x)#.clamp(-max_offset, max_offset)
+        modulator = 2. * torch.sigmoid(self.modulator_conv(x))
+        
+        x = deform_conv2d(
+            input=x,
+            offset=offset,
+            weight=self.regular_conv.weight,
+            bias=self.regular_conv.bias,
+            padding=self.padding,
+            mask=modulator,
+            stride=self.stride,
+        )
+        return x

BiRefNet_github/models/modules/ing.py

@@ -0,0 +1,29 @@
+import torch.nn as nn
+from models.modules.mlp import MLPLayer
+
+
+class BlockA(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64, mlp_ratio=4.):
+        super(BlockA, self).__init__()
+        inter_channels = in_channels
+        self.conv = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
+        self.norm1 = nn.LayerNorm(inter_channels)
+        self.ffn = MLPLayer(in_features=inter_channels,
+                            hidden_features=int(inter_channels * mlp_ratio),
+                            act_layer=nn.GELU,
+                            drop=0.)
+        self.norm2 = nn.LayerNorm(inter_channels)
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        _x = self.conv(x)
+        _x = _x.flatten(2).transpose(1, 2)
+        _x = self.norm1(_x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+
+        x = x + _x
+        _x1 = self.ffn(x)
+        _x1 = self.norm2(_x1)
+        _x1 = _x1.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        x = x + _x1
+        return x

BiRefNet_github/models/modules/lateral_blocks.py

@@ -0,0 +1,21 @@
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+
+from config import Config
+
+
+config = Config()
+
+
+class BasicLatBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
+        super(BasicLatBlk, self).__init__()
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.conv = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return x

BiRefNet_github/models/modules/mlp.py

@@ -0,0 +1,118 @@
+import torch
+import torch.nn as nn
+from functools import partial
+
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+from timm.models.registry import register_model
+
+import math
+
+
+class MLPLayer(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
+        super().__init__()
+        assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
+
+        self.dim = dim
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.q = nn.Linear(dim, dim, bias=qkv_bias)
+        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        self.sr_ratio = sr_ratio
+        if sr_ratio > 1:
+            self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
+            self.norm = nn.LayerNorm(dim)
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+
+        if self.sr_ratio > 1:
+            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
+            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
+            x_ = self.norm(x_)
+            kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        else:
+            kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        k, v = kv[0], kv[1]
+
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class Block(nn.Module):
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = MLPLayer(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x, H, W):
+        x = x + self.drop_path(self.attn(self.norm1(x), H, W))
+        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
+        return x
+
+
+class OverlapPatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+
+    def __init__(self, img_size=224, patch_size=7, stride=4, in_channels=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
+        self.num_patches = self.H * self.W
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=stride,
+                              padding=(patch_size[0] // 2, patch_size[1] // 2))
+        self.norm = nn.LayerNorm(embed_dim)
+
+    def forward(self, x):
+        x = self.proj(x)
+        _, _, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = self.norm(x)
+        return x, H, W
+

BiRefNet_github/models/modules/prompt_encoder.py

@@ -0,0 +1,222 @@
+import numpy as np
+import torch
+import torch.nn as nn
+from typing import Any, Optional, Tuple, Type
+
+
+class PromptEncoder(nn.Module):
+    def __init__(
+            self,
+            embed_dim=256,
+            image_embedding_size=1024,
+            input_image_size=(1024, 1024),
+            mask_in_chans=16,
+            activation=nn.GELU
+        ) -> None:
+        super().__init__()
+        """
+        Codes are partially from SAM: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/prompt_encoder.py.
+
+        Arguments:
+          embed_dim (int): The prompts' embedding dimension
+          image_embedding_size (tuple(int, int)): The spatial size of the
+            image embedding, as (H, W).
+          input_image_size (int): The padded size of the image as input
+            to the image encoder, as (H, W).
+          mask_in_chans (int): The number of hidden channels used for
+            encoding input masks.
+          activation (nn.Module): The activation to use when encoding
+            input masks.
+        """
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.input_image_size = input_image_size
+        self.image_embedding_size = image_embedding_size
+        self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)
+
+        self.num_point_embeddings: int = 4  # pos/neg point + 2 box corners
+        point_embeddings = [nn.Embedding(1, embed_dim) for i in range(self.num_point_embeddings)]
+        self.point_embeddings = nn.ModuleList(point_embeddings)
+        self.not_a_point_embed = nn.Embedding(1, embed_dim)
+
+        self.mask_input_size = (4 * image_embedding_size[0], 4 * image_embedding_size[1])
+        self.mask_downscaling = nn.Sequential(
+            nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans // 4),
+            activation(),
+            nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans),
+            activation(),
+            nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),
+        )
+        self.no_mask_embed = nn.Embedding(1, embed_dim)
+
+    def get_dense_pe(self) -> torch.Tensor:
+        """
+        Returns the positional encoding used to encode point prompts,
+        applied to a dense set of points the shape of the image encoding.
+
+        Returns:
+          torch.Tensor: Positional encoding with shape
+            1x(embed_dim)x(embedding_h)x(embedding_w)
+        """
+        return self.pe_layer(self.image_embedding_size).unsqueeze(0)
+
+    def _embed_points(
+        self,
+        points: torch.Tensor,
+        labels: torch.Tensor,
+        pad: bool,
+    ) -> torch.Tensor:
+        """Embeds point prompts."""
+        points = points + 0.5  # Shift to center of pixel
+        if pad:
+            padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)
+            padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)
+            points = torch.cat([points, padding_point], dim=1)
+            labels = torch.cat([labels, padding_label], dim=1)
+        point_embedding = self.pe_layer.forward_with_coords(points, self.input_image_size)
+        point_embedding[labels == -1] = 0.0
+        point_embedding[labels == -1] += self.not_a_point_embed.weight
+        point_embedding[labels == 0] += self.point_embeddings[0].weight
+        point_embedding[labels == 1] += self.point_embeddings[1].weight
+        return point_embedding
+
+    def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
+        """Embeds box prompts."""
+        boxes = boxes + 0.5  # Shift to center of pixel
+        coords = boxes.reshape(-1, 2, 2)
+        corner_embedding = self.pe_layer.forward_with_coords(coords, self.input_image_size)
+        corner_embedding[:, 0, :] += self.point_embeddings[2].weight
+        corner_embedding[:, 1, :] += self.point_embeddings[3].weight
+        return corner_embedding
+
+    def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:
+        """Embeds mask inputs."""
+        mask_embedding = self.mask_downscaling(masks)
+        return mask_embedding
+
+    def _get_batch_size(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> int:
+        """
+        Gets the batch size of the output given the batch size of the input prompts.
+        """
+        if points is not None:
+            return points[0].shape[0]
+        elif boxes is not None:
+            return boxes.shape[0]
+        elif masks is not None:
+            return masks.shape[0]
+        else:
+            return 1
+
+    def _get_device(self) -> torch.device:
+        return self.point_embeddings[0].weight.device
+
+    def forward(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Embeds different types of prompts, returning both sparse and dense
+        embeddings.
+
+        Arguments:
+          points (tuple(torch.Tensor, torch.Tensor) or none): point coordinates
+            and labels to embed.
+          boxes (torch.Tensor or none): boxes to embed
+          masks (torch.Tensor or none): masks to embed
+
+        Returns:
+          torch.Tensor: sparse embeddings for the points and boxes, with shape
+            BxNx(embed_dim), where N is determined by the number of input points
+            and boxes.
+          torch.Tensor: dense embeddings for the masks, in the shape
+            Bx(embed_dim)x(embed_H)x(embed_W)
+        """
+        bs = self._get_batch_size(points, boxes, masks)
+        sparse_embeddings = torch.empty((bs, 0, self.embed_dim), device=self._get_device())
+        if points is not None:
+            coords, labels = points
+            point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))
+            sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)
+        if boxes is not None:
+            box_embeddings = self._embed_boxes(boxes)
+            sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)
+
+        if masks is not None:
+            dense_embeddings = self._embed_masks(masks)
+        else:
+            dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(
+                bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]
+            )
+
+        return sparse_embeddings, dense_embeddings
+
+
+class PositionEmbeddingRandom(nn.Module):
+    """
+    Positional encoding using random spatial frequencies.
+    """
+
+    def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
+        super().__init__()
+        if scale is None or scale <= 0.0:
+            scale = 1.0
+        self.register_buffer(
+            "positional_encoding_gaussian_matrix",
+            scale * torch.randn((2, num_pos_feats)),
+        )
+
+    def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
+        """Positionally encode points that are normalized to [0,1]."""
+        # assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape
+        coords = 2 * coords - 1
+        coords = coords @ self.positional_encoding_gaussian_matrix
+        coords = 2 * np.pi * coords
+        # outputs d_1 x ... x d_n x C shape
+        return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
+
+    def forward(self, size: Tuple[int, int]) -> torch.Tensor:
+        """Generate positional encoding for a grid of the specified size."""
+        h, w = size
+        device: Any = self.positional_encoding_gaussian_matrix.device
+        grid = torch.ones((h, w), device=device, dtype=torch.float32)
+        y_embed = grid.cumsum(dim=0) - 0.5
+        x_embed = grid.cumsum(dim=1) - 0.5
+        y_embed = y_embed / h
+        x_embed = x_embed / w
+
+        pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
+        return pe.permute(2, 0, 1)  # C x H x W
+
+    def forward_with_coords(
+        self, coords_input: torch.Tensor, image_size: Tuple[int, int]
+    ) -> torch.Tensor:
+        """Positionally encode points that are not normalized to [0,1]."""
+        coords = coords_input.clone()
+        coords[:, :, 0] = coords[:, :, 0] / image_size[1]
+        coords[:, :, 1] = coords[:, :, 1] / image_size[0]
+        return self._pe_encoding(coords.to(torch.float))  # B x N x C
+
+
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x
+    

BiRefNet_github/models/modules/utils.py

@@ -0,0 +1,54 @@
+import torch.nn as nn
+
+
+def build_act_layer(act_layer):
+    if act_layer == 'ReLU':
+        return nn.ReLU(inplace=True)
+    elif act_layer == 'SiLU':
+        return nn.SiLU(inplace=True)
+    elif act_layer == 'GELU':
+        return nn.GELU()
+
+    raise NotImplementedError(f'build_act_layer does not support {act_layer}')
+
+
+def build_norm_layer(dim,
+                     norm_layer,
+                     in_format='channels_last',
+                     out_format='channels_last',
+                     eps=1e-6):
+    layers = []
+    if norm_layer == 'BN':
+        if in_format == 'channels_last':
+            layers.append(to_channels_first())
+        layers.append(nn.BatchNorm2d(dim))
+        if out_format == 'channels_last':
+            layers.append(to_channels_last())
+    elif norm_layer == 'LN':
+        if in_format == 'channels_first':
+            layers.append(to_channels_last())
+        layers.append(nn.LayerNorm(dim, eps=eps))
+        if out_format == 'channels_first':
+            layers.append(to_channels_first())
+    else:
+        raise NotImplementedError(
+            f'build_norm_layer does not support {norm_layer}')
+    return nn.Sequential(*layers)
+
+
+class to_channels_first(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.permute(0, 3, 1, 2)
+
+
+class to_channels_last(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.permute(0, 2, 3, 1)

BiRefNet_github/models/refinement/refiner.py

@@ -0,0 +1,253 @@
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.models import vgg16, vgg16_bn
+from torchvision.models import resnet50
+
+from config import Config
+from dataset import class_labels_TR_sorted
+from models.backbones.build_backbone import build_backbone
+from models.modules.decoder_blocks import BasicDecBlk
+from models.modules.lateral_blocks import BasicLatBlk
+from models.modules.ing import *
+from models.refinement.stem_layer import StemLayer
+
+
+class RefinerPVTInChannels4(nn.Module):
+    def __init__(self, in_channels=3+1):
+        super(RefinerPVTInChannels4, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        self.bb = build_backbone(self.config.bb, params_settings='in_channels=4')
+
+        lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+        }
+        channels = lateral_channels_in_collection[self.config.bb]
+        self.squeeze_module = BasicDecBlk(channels[0], channels[0])
+
+        self.decoder = Decoder(channels)
+
+        if 0:
+            for key, value in self.named_parameters():
+                if 'bb.' in key:
+                    value.requires_grad = False
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        ########## Encoder ##########
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x)
+            x2 = self.bb.conv2(x1)
+            x3 = self.bb.conv3(x2)
+            x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+
+        x4 = self.squeeze_module(x4)
+
+        ########## Decoder ##########
+
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+
+        return scaled_preds
+
+
+class Refiner(nn.Module):
+    def __init__(self, in_channels=3+1):
+        super(Refiner, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        self.stem_layer = StemLayer(in_channels=in_channels, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
+        self.bb = build_backbone(self.config.bb)
+
+        lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+        }
+        channels = lateral_channels_in_collection[self.config.bb]
+        self.squeeze_module = BasicDecBlk(channels[0], channels[0])
+
+        self.decoder = Decoder(channels)
+
+        if 0:
+            for key, value in self.named_parameters():
+                if 'bb.' in key:
+                    value.requires_grad = False
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        x = self.stem_layer(x)
+        ########## Encoder ##########
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x)
+            x2 = self.bb.conv2(x1)
+            x3 = self.bb.conv3(x2)
+            x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+
+        x4 = self.squeeze_module(x4)
+
+        ########## Decoder ##########
+
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+
+        return scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels):
+        super(Decoder, self).__init__()
+        self.config = Config()
+        DecoderBlock = eval('BasicDecBlk')
+        LateralBlock = eval('BasicLatBlk')
+
+        self.decoder_block4 = DecoderBlock(channels[0], channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1], channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2], channels[3])
+        self.decoder_block1 = DecoderBlock(channels[3], channels[3]//2)
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        if self.config.ms_supervision:
+            self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
+            self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
+            self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
+        self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2, 1, 1, 1, 0))
+
+    def forward(self, features):
+        x, x1, x2, x3, x4 = features
+        outs = []
+        p4 = self.decoder_block4(x4)
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        p3 = self.decoder_block3(_p3)
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        p2 = self.decoder_block2(_p2)
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+        p1_out = self.conv_out1(_p1)
+
+        if self.config.ms_supervision:
+            outs.append(self.conv_ms_spvn_4(p4))
+            outs.append(self.conv_ms_spvn_3(p3))
+            outs.append(self.conv_ms_spvn_2(p2))
+        outs.append(p1_out)
+        return outs
+
+
+class RefUNet(nn.Module):
+    # Refinement
+    def __init__(self, in_channels=3+1):
+        super(RefUNet, self).__init__()
+        self.encoder_1 = nn.Sequential(
+            nn.Conv2d(in_channels, 64, 3, 1, 1),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_2 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_3 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_4 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.pool4 = nn.MaxPool2d(2, 2, ceil_mode=True)
+        #####
+        self.decoder_5 = nn.Sequential(
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+        #####
+        self.decoder_4 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_3 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_2 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_1 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.conv_d0 = nn.Conv2d(64, 1, 3, 1, 1)
+
+        self.upscore2 = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
+
+    def forward(self, x):
+        outs = []
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        hx = x
+
+        hx1 = self.encoder_1(hx)
+        hx2 = self.encoder_2(hx1)
+        hx3 = self.encoder_3(hx2)
+        hx4 = self.encoder_4(hx3)
+
+        hx = self.decoder_5(self.pool4(hx4))
+        hx = torch.cat((self.upscore2(hx), hx4), 1)
+
+        d4 = self.decoder_4(hx)
+        hx = torch.cat((self.upscore2(d4), hx3), 1)
+
+        d3 = self.decoder_3(hx)
+        hx = torch.cat((self.upscore2(d3), hx2), 1)
+
+        d2 = self.decoder_2(hx)
+        hx = torch.cat((self.upscore2(d2), hx1), 1)
+
+        d1 = self.decoder_1(hx)
+
+        x = self.conv_d0(d1)
+        outs.append(x)
+        return outs

BiRefNet_github/models/refinement/stem_layer.py

@@ -0,0 +1,45 @@
+import torch.nn as nn
+from models.modules.utils import build_act_layer, build_norm_layer
+
+
+class StemLayer(nn.Module):
+    r""" Stem layer of InternImage
+    Args:
+        in_channels (int): number of input channels
+        out_channels (int): number of output channels
+        act_layer (str): activation layer
+        norm_layer (str): normalization layer
+    """
+
+    def __init__(self,
+                 in_channels=3+1,
+                 inter_channels=48,
+                 out_channels=96,
+                 act_layer='GELU',
+                 norm_layer='BN'):
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels,
+                               inter_channels,
+                               kernel_size=3,
+                               stride=1,
+                               padding=1)
+        self.norm1 = build_norm_layer(
+            inter_channels, norm_layer, 'channels_first', 'channels_first'
+        )
+        self.act = build_act_layer(act_layer)
+        self.conv2 = nn.Conv2d(inter_channels,
+                               out_channels,
+                               kernel_size=3,
+                               stride=1,
+                               padding=1)
+        self.norm2 = build_norm_layer(
+            out_channels, norm_layer, 'channels_first', 'channels_first'
+        )
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.act(x)
+        x = self.conv2(x)
+        x = self.norm2(x)
+        return x

BiRefNet_github/preproc.py

@@ -0,0 +1,85 @@
+from PIL import Image, ImageEnhance
+import random
+import numpy as np
+import random
+
+
+def preproc(image, label, preproc_methods=['flip']):
+    if 'flip' in preproc_methods:
+        image, label = cv_random_flip(image, label)
+    if 'crop' in preproc_methods:
+        image, label = random_crop(image, label)
+    if 'rotate' in preproc_methods:
+        image, label = random_rotate(image, label)
+    if 'enhance' in preproc_methods:
+        image = color_enhance(image)
+    if 'pepper' in preproc_methods:
+        label = random_pepper(label)
+    return image, label
+
+
+def cv_random_flip(img, label):
+    if random.random() > 0.5:
+        img = img.transpose(Image.FLIP_LEFT_RIGHT)
+        label = label.transpose(Image.FLIP_LEFT_RIGHT)
+    return img, label
+
+
+def random_crop(image, label):
+    border = 30
+    image_width = image.size[0]
+    image_height = image.size[1]
+    border = int(min(image_width, image_height) * 0.1)
+    crop_win_width = np.random.randint(image_width - border, image_width)
+    crop_win_height = np.random.randint(image_height - border, image_height)
+    random_region = (
+        (image_width - crop_win_width) >> 1, (image_height - crop_win_height) >> 1, (image_width + crop_win_width) >> 1,
+        (image_height + crop_win_height) >> 1)
+    return image.crop(random_region), label.crop(random_region)
+
+
+def random_rotate(image, label, angle=15):
+    mode = Image.BICUBIC
+    if random.random() > 0.8:
+        random_angle = np.random.randint(-angle, angle)
+        image = image.rotate(random_angle, mode)
+        label = label.rotate(random_angle, mode)
+    return image, label
+
+
+def color_enhance(image):
+    bright_intensity = random.randint(5, 15) / 10.0
+    image = ImageEnhance.Brightness(image).enhance(bright_intensity)
+    contrast_intensity = random.randint(5, 15) / 10.0
+    image = ImageEnhance.Contrast(image).enhance(contrast_intensity)
+    color_intensity = random.randint(0, 20) / 10.0
+    image = ImageEnhance.Color(image).enhance(color_intensity)
+    sharp_intensity = random.randint(0, 30) / 10.0
+    image = ImageEnhance.Sharpness(image).enhance(sharp_intensity)
+    return image
+
+
+def random_gaussian(image, mean=0.1, sigma=0.35):
+    def gaussianNoisy(im, mean=mean, sigma=sigma):
+        for _i in range(len(im)):
+            im[_i] += random.gauss(mean, sigma)
+        return im
+
+    img = np.asarray(image)
+    width, height = img.shape
+    img = gaussianNoisy(img[:].flatten(), mean, sigma)
+    img = img.reshape([width, height])
+    return Image.fromarray(np.uint8(img))
+
+
+def random_pepper(img, N=0.0015):
+    img = np.array(img)
+    noiseNum = int(N * img.shape[0] * img.shape[1])
+    for i in range(noiseNum):
+        randX = random.randint(0, img.shape[0] - 1)
+        randY = random.randint(0, img.shape[1] - 1)
+        if random.randint(0, 1) == 0:
+            img[randX, randY] = 0
+        else:
+            img[randX, randY] = 255
+    return Image.fromarray(img)

BiRefNet_github/requirements.txt

@@ -0,0 +1,15 @@
+--extra-index-url https://download.pytorch.org/whl/cu118
+torch==2.0.1
+--extra-index-url https://download.pytorch.org/whl/cu118
+torchvision==0.15.2
+numpy<2
+opencv-python
+timm
+scipy
+scikit-image
+kornia
+
+tqdm
+prettytable
+
+huggingface_hub

BiRefNet_github/rm_cache.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+rm -rf __pycache__ */__pycache__
+
+# Val
+rm -r tmp*
+
+# Train
+rm slurm*
+rm -r ckpt
+rm nohup.out*
+
+# Eval
+rm -r evaluation/eval-*
+rm -r tmp*
+rm -r e_logs/
+
+# System
+rm core-*-python-*
+
+clear

BiRefNet_github/sub.sh

@@ -0,0 +1,19 @@
+#!/bin/sh
+# Example: ./sub.sh tmp_proj 0,1,2,3 3 --> Use 0,1,2,3 for training, release GPUs, use GPU:3 for inference.
+
+module load compilers/cuda/11.8
+
+export PYTHONUNBUFFERED=1
+export LD_PRELOAD=/home/bingxing2/apps/compilers/gcc/12.2.0/lib64/libstdc++.so.6
+export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${HOME}/miniconda3/lib:/home/bingxing2/apps/cudnn/8.4.0.27_cuda11.x/lib
+
+method=${1:-"BSL"}
+devices=${2:-0}
+
+sbatch --nodes=1 -p vip_gpu_ailab -A ai4bio \
+--ntasks-per-node=1 \
+--gres=gpu:$(($(echo ${devices%%,} | grep -o "," | wc -l)+1)) \
+--cpus-per-task=32 \
+./train_test.sh ${method} ${devices}
+
+hostname

BiRefNet_github/test.sh

@@ -0,0 +1,28 @@
+devices=${1:-0}
+pred_root=${2:-e_preds}
+
+# Inference
+
+CUDA_VISIBLE_DEVICES=${devices} python inference.py --pred_root ${pred_root}
+
+echo Inference finished at $(date)
+
+# Evaluation
+log_dir=e_logs && mkdir ${log_dir}
+
+task=$(python3 config.py)
+case "${task}" in
+    "DIS5K") testsets='DIS-VD,DIS-TE1,DIS-TE2,DIS-TE3,DIS-TE4' ;;
+    "COD") testsets='CHAMELEON,NC4K,TE-CAMO,TE-COD10K' ;;
+    "HRSOD") testsets='DAVIS-S,TE-HRSOD,TE-UHRSD,DUT-OMRON,TE-DUTS' ;;
+    "DIS5K+HRSOD+HRS10K") testsets='DIS-VD' ;;
+    "P3M-10k") testsets='TE-P3M-500-P,TE-P3M-500-NP' ;;
+esac
+testsets=(`echo ${testsets} | tr ',' ' '`) && testsets=${testsets[@]}
+
+for testset in ${testsets}; do
+    nohup python eval_existingOnes.py --pred_root ${pred_root} --data_lst ${testset} > ${log_dir}/eval_${testset}.out 2>&1 &
+done
+
+
+echo Evaluation started at $(date)

BiRefNet_github/train.py

@@ -0,0 +1,377 @@
+import os
+import datetime
+import argparse
+import torch
+import torch.nn as nn
+import torch.optim as optim
+from torch.autograd import Variable
+
+from config import Config
+from loss import PixLoss, ClsLoss
+from dataset import MyData
+from models.birefnet import BiRefNet
+from utils import Logger, AverageMeter, set_seed, check_state_dict
+from evaluation.valid import valid
+
+from torch.utils.data.distributed import DistributedSampler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.distributed import init_process_group, destroy_process_group, get_rank
+from torch.cuda import amp
+
+
+parser = argparse.ArgumentParser(description='')
+parser.add_argument('--resume', default=None, type=str, help='path to latest checkpoint')
+parser.add_argument('--epochs', default=120, type=int)
+parser.add_argument('--trainset', default='DIS5K', type=str, help="Options: 'DIS5K'")
+parser.add_argument('--ckpt_dir', default=None, help='Temporary folder')
+parser.add_argument('--testsets', default='DIS-VD+DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4', type=str)
+parser.add_argument('--dist', default=False, type=lambda x: x == 'True')
+args = parser.parse_args()
+
+
+config = Config()
+if config.rand_seed:
+    set_seed(config.rand_seed)
+
+if config.use_fp16:
+    # Half Precision
+    scaler = amp.GradScaler(enabled=config.use_fp16)
+
+# DDP
+to_be_distributed = args.dist
+if to_be_distributed:
+    init_process_group(backend="nccl", timeout=datetime.timedelta(seconds=3600*10))
+    device = int(os.environ["LOCAL_RANK"])
+else:
+    device = config.device
+
+epoch_st = 1
+# make dir for ckpt
+os.makedirs(args.ckpt_dir, exist_ok=True)
+
+# Init log file
+logger = Logger(os.path.join(args.ckpt_dir, "log.txt"))
+logger_loss_idx = 1
+
+# log model and optimizer params
+# logger.info("Model details:"); logger.info(model)
+logger.info("datasets: load_all={}, compile={}.".format(config.load_all, config.compile))
+logger.info("Other hyperparameters:"); logger.info(args)
+print('batch size:', config.batch_size)
+
+
+if os.path.exists(os.path.join(config.data_root_dir, config.task, args.testsets.strip('+').split('+')[0])):
+    args.testsets = args.testsets.strip('+').split('+')
+else:
+    args.testsets = []
+
+# Init model
+def prepare_dataloader(dataset: torch.utils.data.Dataset, batch_size: int, to_be_distributed=False, is_train=True):
+    if to_be_distributed:
+        return torch.utils.data.DataLoader(
+            dataset=dataset, batch_size=batch_size, num_workers=min(config.num_workers, batch_size), pin_memory=True,
+            shuffle=False, sampler=DistributedSampler(dataset), drop_last=True
+        )
+    else:
+        return torch.utils.data.DataLoader(
+            dataset=dataset, batch_size=batch_size, num_workers=min(config.num_workers, batch_size, 0), pin_memory=True,
+            shuffle=is_train, drop_last=True
+        )
+
+
+def init_data_loaders(to_be_distributed):
+    # Prepare dataset
+    train_loader = prepare_dataloader(
+        MyData(datasets=config.training_set, image_size=config.size, is_train=True),
+        config.batch_size, to_be_distributed=to_be_distributed, is_train=True
+    )
+    print(len(train_loader), "batches of train dataloader {} have been created.".format(config.training_set))
+    test_loaders = {}
+    for testset in args.testsets:
+        _data_loader_test = prepare_dataloader(
+            MyData(datasets=testset, image_size=config.size, is_train=False),
+            config.batch_size_valid, is_train=False
+        )
+        print(len(_data_loader_test), "batches of valid dataloader {} have been created.".format(testset))
+        test_loaders[testset] = _data_loader_test
+    return train_loader, test_loaders
+
+
+def init_models_optimizers(epochs, to_be_distributed):
+    model = BiRefNet(bb_pretrained=True)
+    if args.resume:
+        if os.path.isfile(args.resume):
+            logger.info("=> loading checkpoint '{}'".format(args.resume))
+            state_dict = torch.load(args.resume, map_location='cpu')
+            state_dict = check_state_dict(state_dict)
+            model.load_state_dict(state_dict)
+            global epoch_st
+            epoch_st = int(args.resume.rstrip('.pth').split('epoch_')[-1]) + 1
+        else:
+            logger.info("=> no checkpoint found at '{}'".format(args.resume))
+    if to_be_distributed:
+        model = model.to(device)
+        model = DDP(model, device_ids=[device])
+    else:
+        model = model.to(device)
+    if config.compile:
+        model = torch.compile(model, mode=['default', 'reduce-overhead', 'max-autotune'][0])
+    if config.precisionHigh:
+        torch.set_float32_matmul_precision('high')
+
+
+    # Setting optimizer
+    if config.optimizer == 'AdamW':
+        optimizer = optim.AdamW(params=model.parameters(), lr=config.lr, weight_decay=1e-2)
+    elif config.optimizer == 'Adam':
+        optimizer = optim.Adam(params=model.parameters(), lr=config.lr, weight_decay=0)
+    lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
+        optimizer,
+        milestones=[lde if lde > 0 else epochs + lde + 1 for lde in config.lr_decay_epochs],
+        gamma=config.lr_decay_rate
+    )
+    logger.info("Optimizer details:"); logger.info(optimizer)
+    logger.info("Scheduler details:"); logger.info(lr_scheduler)
+
+    return model, optimizer, lr_scheduler
+
+
+class Trainer:
+    def __init__(
+        self, data_loaders, model_opt_lrsch,
+    ):
+        self.model, self.optimizer, self.lr_scheduler = model_opt_lrsch
+        self.train_loader, self.test_loaders = data_loaders
+        if config.out_ref:
+            self.criterion_gdt = nn.BCELoss() if not config.use_fp16 else nn.BCEWithLogitsLoss()
+
+        # Setting Losses
+        self.pix_loss = PixLoss()
+        self.cls_loss = ClsLoss()
+        
+        # Others
+        self.loss_log = AverageMeter()
+        if config.lambda_adv_g:
+            self.optimizer_d, self.lr_scheduler_d, self.disc, self.adv_criterion = self._load_adv_components()
+            self.disc_update_for_odd = 0
+
+    def _load_adv_components(self):
+        # AIL
+        from loss import Discriminator
+        disc = Discriminator(channels=3, img_size=config.size)
+        if to_be_distributed:
+            disc = disc.to(device)
+            disc = DDP(disc, device_ids=[device], broadcast_buffers=False)
+        else:
+            disc = disc.to(device)
+        if config.compile:
+            disc = torch.compile(disc, mode=['default', 'reduce-overhead', 'max-autotune'][0])
+        adv_criterion = nn.BCELoss() if not config.use_fp16 else nn.BCEWithLogitsLoss()
+        if config.optimizer == 'AdamW':
+            optimizer_d = optim.AdamW(params=disc.parameters(), lr=config.lr, weight_decay=1e-2)
+        elif config.optimizer == 'Adam':
+            optimizer_d = optim.Adam(params=disc.parameters(), lr=config.lr, weight_decay=0)
+        lr_scheduler_d = torch.optim.lr_scheduler.MultiStepLR(
+            optimizer_d,
+            milestones=[lde if lde > 0 else args.epochs + lde + 1 for lde in config.lr_decay_epochs],
+            gamma=config.lr_decay_rate
+        )
+        return optimizer_d, lr_scheduler_d, disc, adv_criterion
+
+    def _train_batch(self, batch):
+        inputs = batch[0].to(device)
+        gts = batch[1].to(device)
+        class_labels = batch[2].to(device)
+        if config.use_fp16:
+            with amp.autocast(enabled=config.use_fp16):
+                scaled_preds, class_preds_lst = self.model(inputs)
+                if config.out_ref:
+                    (outs_gdt_pred, outs_gdt_label), scaled_preds = scaled_preds
+                    for _idx, (_gdt_pred, _gdt_label) in enumerate(zip(outs_gdt_pred, outs_gdt_label)):
+                        _gdt_pred = nn.functional.interpolate(_gdt_pred, size=_gdt_label.shape[2:], mode='bilinear', align_corners=True)#.sigmoid()
+                        # _gdt_label = _gdt_label.sigmoid()
+                        loss_gdt = self.criterion_gdt(_gdt_pred, _gdt_label) if _idx == 0 else self.criterion_gdt(_gdt_pred, _gdt_label) + loss_gdt
+                    # self.loss_dict['loss_gdt'] = loss_gdt.item()
+                if None in class_preds_lst:
+                    loss_cls = 0.
+                else:
+                    loss_cls = self.cls_loss(class_preds_lst, class_labels) * 1.0
+                    self.loss_dict['loss_cls'] = loss_cls.item()
+
+                # Loss
+                loss_pix = self.pix_loss(scaled_preds, torch.clamp(gts, 0, 1)) * 1.0
+                self.loss_dict['loss_pix'] = loss_pix.item()
+                # since there may be several losses for sal, the lambdas for them (lambdas_pix) are inside the loss.py
+                loss = loss_pix + loss_cls
+                if config.out_ref:
+                    loss = loss + loss_gdt * 1.0
+
+                if config.lambda_adv_g:
+                    # gen
+                    valid = Variable(torch.cuda.FloatTensor(scaled_preds[-1].shape[0], 1).fill_(1.0), requires_grad=False).to(device)
+                    adv_loss_g = self.adv_criterion(self.disc(scaled_preds[-1] * inputs), valid) * config.lambda_adv_g
+                    loss += adv_loss_g
+                    self.loss_dict['loss_adv'] = adv_loss_g.item()
+                    self.disc_update_for_odd += 1
+            # self.loss_log.update(loss.item(), inputs.size(0))
+            # self.optimizer.zero_grad()
+            # loss.backward()
+            # self.optimizer.step()
+            self.optimizer.zero_grad()
+            scaler.scale(loss).backward()
+            scaler.step(self.optimizer)
+            scaler.update()
+
+            if config.lambda_adv_g and self.disc_update_for_odd % 2 == 0:
+                # disc
+                fake = Variable(torch.cuda.FloatTensor(scaled_preds[-1].shape[0], 1).fill_(0.0), requires_grad=False).to(device)
+                adv_loss_real = self.adv_criterion(self.disc(gts * inputs), valid)
+                adv_loss_fake = self.adv_criterion(self.disc(scaled_preds[-1].detach() * inputs.detach()), fake)
+                adv_loss_d = (adv_loss_real + adv_loss_fake) / 2 * config.lambda_adv_d
+                self.loss_dict['loss_adv_d'] = adv_loss_d.item()
+                # self.optimizer_d.zero_grad()
+                # adv_loss_d.backward()
+                # self.optimizer_d.step()
+                self.optimizer_d.zero_grad()
+                scaler.scale(adv_loss_d).backward()
+                scaler.step(self.optimizer_d)
+                scaler.update()
+        else:
+            scaled_preds, class_preds_lst = self.model(inputs)
+            if config.out_ref:
+                (outs_gdt_pred, outs_gdt_label), scaled_preds = scaled_preds
+                for _idx, (_gdt_pred, _gdt_label) in enumerate(zip(outs_gdt_pred, outs_gdt_label)):
+                    _gdt_pred = nn.functional.interpolate(_gdt_pred, size=_gdt_label.shape[2:], mode='bilinear', align_corners=True).sigmoid()
+                    _gdt_label = _gdt_label.sigmoid()
+                    loss_gdt = self.criterion_gdt(_gdt_pred, _gdt_label) if _idx == 0 else self.criterion_gdt(_gdt_pred, _gdt_label) + loss_gdt
+                # self.loss_dict['loss_gdt'] = loss_gdt.item()
+            if None in class_preds_lst:
+                loss_cls = 0.
+            else:
+                loss_cls = self.cls_loss(class_preds_lst, class_labels) * 1.0
+                self.loss_dict['loss_cls'] = loss_cls.item()
+
+            # Loss
+            loss_pix = self.pix_loss(scaled_preds, torch.clamp(gts, 0, 1)) * 1.0
+            self.loss_dict['loss_pix'] = loss_pix.item()
+            # since there may be several losses for sal, the lambdas for them (lambdas_pix) are inside the loss.py
+            loss = loss_pix + loss_cls
+            if config.out_ref:
+                loss = loss + loss_gdt * 1.0
+
+            if config.lambda_adv_g:
+                # gen
+                valid = Variable(torch.cuda.FloatTensor(scaled_preds[-1].shape[0], 1).fill_(1.0), requires_grad=False).to(device)
+                adv_loss_g = self.adv_criterion(self.disc(scaled_preds[-1] * inputs), valid) * config.lambda_adv_g
+                loss += adv_loss_g
+                self.loss_dict['loss_adv'] = adv_loss_g.item()
+                self.disc_update_for_odd += 1
+            self.loss_log.update(loss.item(), inputs.size(0))
+            self.optimizer.zero_grad()
+            loss.backward()
+            self.optimizer.step()
+
+            if config.lambda_adv_g and self.disc_update_for_odd % 2 == 0:
+                # disc
+                fake = Variable(torch.cuda.FloatTensor(scaled_preds[-1].shape[0], 1).fill_(0.0), requires_grad=False).to(device)
+                adv_loss_real = self.adv_criterion(self.disc(gts * inputs), valid)
+                adv_loss_fake = self.adv_criterion(self.disc(scaled_preds[-1].detach() * inputs.detach()), fake)
+                adv_loss_d = (adv_loss_real + adv_loss_fake) / 2 * config.lambda_adv_d
+                self.loss_dict['loss_adv_d'] = adv_loss_d.item()
+                self.optimizer_d.zero_grad()
+                adv_loss_d.backward()
+                self.optimizer_d.step()
+
+    def train_epoch(self, epoch):
+        global logger_loss_idx
+        self.model.train()
+        self.loss_dict = {}
+        if epoch > args.epochs + config.IoU_finetune_last_epochs:
+            self.pix_loss.lambdas_pix_last['bce'] *= 0
+            self.pix_loss.lambdas_pix_last['ssim'] *= 1
+            self.pix_loss.lambdas_pix_last['iou'] *= 0.5
+
+        for batch_idx, batch in enumerate(self.train_loader):
+            self._train_batch(batch)
+            # Logger
+            if batch_idx % 20 == 0:
+                info_progress = 'Epoch[{0}/{1}] Iter[{2}/{3}].'.format(epoch, args.epochs, batch_idx, len(self.train_loader))
+                info_loss = 'Training Losses'
+                for loss_name, loss_value in self.loss_dict.items():
+                    info_loss += ', {}: {:.3f}'.format(loss_name, loss_value)
+                logger.info(' '.join((info_progress, info_loss)))
+        info_loss = '@==Final== Epoch[{0}/{1}]  Training Loss: {loss.avg:.3f}  '.format(epoch, args.epochs, loss=self.loss_log)
+        logger.info(info_loss)
+
+        self.lr_scheduler.step()
+        if config.lambda_adv_g:
+            self.lr_scheduler_d.step()
+        return self.loss_log.avg
+
+    def validate_model(self, epoch):
+        num_image_testset_all = {'DIS-VD': 470, 'DIS-TE1': 500, 'DIS-TE2': 500, 'DIS-TE3': 500, 'DIS-TE4': 500}
+        num_image_testset = {}
+        for testset in args.testsets:
+            if 'DIS-TE' in testset:
+                num_image_testset[testset] = num_image_testset_all[testset]
+        weighted_scores = {'f_max': 0, 'f_mean': 0, 'f_wfm': 0, 'sm': 0, 'e_max': 0, 'e_mean': 0, 'mae': 0}
+        len_all_data_loaders = 0
+        self.model.epoch = epoch
+        for testset, data_loader_test in self.test_loaders.items():
+            print('Validating {}...'.format(testset))
+            performance_dict = valid(
+                self.model,
+                data_loader_test,
+                pred_dir='.',
+                method=args.ckpt_dir.split('/')[-1] if args.ckpt_dir.split('/')[-1].strip('.').strip('/') else 'tmp_val',
+                testset=testset,
+                only_S_MAE=config.only_S_MAE,
+                device=device
+            )
+            print('Test set: {}:'.format(testset))
+            if config.only_S_MAE:
+                print('Smeasure: {:.4f}, MAE: {:.4f}'.format(
+                    performance_dict['sm'], performance_dict['mae']
+                ))
+            else:
+                print('Fmax: {:.4f}, Fwfm: {:.4f}, Smeasure: {:.4f}, Emean: {:.4f}, MAE: {:.4f}'.format(
+                    performance_dict['f_max'], performance_dict['f_wfm'], performance_dict['sm'], performance_dict['e_mean'], performance_dict['mae']
+                ))
+            if '-TE' in testset:
+                for metric in ['sm', 'mae'] if config.only_S_MAE else ['f_max', 'f_mean', 'f_wfm', 'sm', 'e_max', 'e_mean', 'mae']:
+                    weighted_scores[metric] += performance_dict[metric] * len(data_loader_test)
+                len_all_data_loaders += len(data_loader_test)
+        print('Weighted Scores:')
+        for metric, score in weighted_scores.items():
+            if score:
+                print('\t{}: {:.4f}.'.format(metric, score / len_all_data_loaders))
+
+
+def main():
+
+    trainer = Trainer(
+        data_loaders=init_data_loaders(to_be_distributed),
+        model_opt_lrsch=init_models_optimizers(args.epochs, to_be_distributed)
+    )
+
+    for epoch in range(epoch_st, args.epochs+1):
+        train_loss = trainer.train_epoch(epoch)
+        # Save checkpoint
+        # DDP
+        if epoch >= args.epochs - config.save_last and epoch % config.save_step == 0:
+            torch.save(
+                trainer.model.module.state_dict() if to_be_distributed else trainer.model.state_dict(),
+                os.path.join(args.ckpt_dir, 'epoch_{}.pth'.format(epoch))
+            )
+        if config.val_step and epoch >= args.epochs - config.save_last and (args.epochs - epoch) % config.val_step == 0:
+            if to_be_distributed:
+                if get_rank() == 0:
+                    print('Validating at rank-{}...'.format(get_rank()))
+                    trainer.validate_model(epoch)
+            else:
+                trainer.validate_model(epoch)
+    if to_be_distributed:
+        destroy_process_group()
+
+if __name__ == '__main__':
+    main()

BiRefNet_github/train.sh

@@ -0,0 +1,41 @@
+#!/bin/bash
+# Run script
+# Settings of training & test for different tasks.
+method="$1"
+task=$(python3 config.py)
+case "${task}" in
+    "DIS5K") epochs=600 && val_last=100 && step=5 ;;
+    "COD") epochs=150 && val_last=50 && step=5 ;;
+    "HRSOD") epochs=150 && val_last=50 && step=5 ;;
+    "DIS5K+HRSOD+HRS10K") epochs=250 && val_last=50 && step=5 ;;
+    "P3M-10k") epochs=150 && val_last=50 && step=5 ;;
+esac
+testsets=NO     # Non-existing folder to skip.
+# testsets=TE-COD10K   # for COD
+
+# Train
+devices=$2
+nproc_per_node=$(echo ${devices%%,} | grep -o "," | wc -l)
+
+to_be_distributed=`echo ${nproc_per_node} | awk '{if($e > 0) print "True"; else print "False";}'`
+
+echo Training started at $(date)
+if [ ${to_be_distributed} == "True" ]
+then
+    # Adapt the nproc_per_node by the number of GPUs. Give 8989 as the default value of master_port.
+    echo "Multi-GPU mode received..."
+    CUDA_VISIBLE_DEVICES=${devices} \
+    torchrun --nproc_per_node $((nproc_per_node+1)) --master_port=${3:-8989} \
+    train.py --ckpt_dir ckpt/${method} --epochs ${epochs} \
+        --testsets ${testsets} \
+        --dist ${to_be_distributed}
+else
+    echo "Single-GPU mode received..."
+    CUDA_VISIBLE_DEVICES=${devices} \
+    python train.py --ckpt_dir ckpt/${method} --epochs ${epochs} \
+        --testsets ${testsets} \
+        --dist ${to_be_distributed} \
+        --resume ckpt/xx/ep100.pth
+fi
+
+echo Training finished at $(date)

BiRefNet_github/train_test.sh

@@ -0,0 +1,11 @@
+#!/bin/sh
+
+method=${1:-"BSL"}
+devices=${2:-"0,1,2,3,4,5,6,7"}
+
+bash train.sh ${method} ${devices}
+
+devices_test=${3:-0}
+bash test.sh ${devices_test}
+
+hostname

BiRefNet_github/utils.py

@@ -0,0 +1,97 @@
+import logging
+import os
+import torch
+from torchvision import transforms
+import numpy as np
+import random
+import cv2
+from PIL import Image
+
+
+def path_to_image(path, size=(1024, 1024), color_type=['rgb', 'gray'][0]):
+    if color_type.lower() == 'rgb':
+        image = cv2.imread(path)
+    elif color_type.lower() == 'gray':
+        image = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
+    else:
+        print('Select the color_type to return, either to RGB or gray image.')
+        return
+    if size:
+        image = cv2.resize(image, size, interpolation=cv2.INTER_LINEAR)
+    if color_type.lower() == 'rgb':
+        image = Image.fromarray(cv2.cvtColor(image, cv2.COLOR_BGR2RGB)).convert('RGB')
+    else:
+        image = Image.fromarray(image).convert('L')
+    return image
+
+
+
+def check_state_dict(state_dict, unwanted_prefix='_orig_mod.'):
+    for k, v in list(state_dict.items()):
+        if k.startswith(unwanted_prefix):
+            state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)
+    return state_dict
+
+
+def generate_smoothed_gt(gts):
+    epsilon = 0.001
+    new_gts = (1-epsilon)*gts+epsilon/2
+    return new_gts
+
+
+class Logger():
+    def __init__(self, path="log.txt"):
+        self.logger = logging.getLogger('BiRefNet')
+        self.file_handler = logging.FileHandler(path, "w")
+        self.stdout_handler = logging.StreamHandler()
+        self.stdout_handler.setFormatter(logging.Formatter('%(asctime)s %(levelname)s %(message)s'))
+        self.file_handler.setFormatter(logging.Formatter('%(asctime)s %(levelname)s %(message)s'))
+        self.logger.addHandler(self.file_handler)
+        self.logger.addHandler(self.stdout_handler)
+        self.logger.setLevel(logging.INFO)
+        self.logger.propagate = False
+    
+    def info(self, txt):
+        self.logger.info(txt)
+    
+    def close(self):
+        self.file_handler.close()
+        self.stdout_handler.close()
+
+
+class AverageMeter(object):
+    """Computes and stores the average and current value"""
+    def __init__(self):
+        self.reset()
+
+    def reset(self):
+        self.val = 0.0
+        self.avg = 0.0
+        self.sum = 0.0
+        self.count = 0.0
+
+    def update(self, val, n=1):
+        self.val = val
+        self.sum += val * n
+        self.count += n
+        self.avg = self.sum / self.count
+
+
+def save_checkpoint(state, path, filename="latest.pth"):
+    torch.save(state, os.path.join(path, filename))
+
+
+def save_tensor_img(tenor_im, path):
+    im = tenor_im.cpu().clone()
+    im = im.squeeze(0)
+    tensor2pil = transforms.ToPILImage()
+    im = tensor2pil(im)
+    im.save(path)
+
+
+def set_seed(seed):
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+    np.random.seed(seed)
+    random.seed(seed)
+    torch.backends.cudnn.deterministic = True

BiRefNet_github/waiting4eval.py

@@ -0,0 +1,141 @@
+# --------------------------------------------------------
+# Make evaluation along with training. Swith time with space/computation.
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Peng Zheng
+# --------------------------------------------------------
+import os
+from glob import glob
+from time import sleep
+import argparse
+import torch
+
+from config import Config
+from models.birefnet import BiRefNet
+from dataset import MyData
+from evaluation.valid import valid
+
+
+parser = argparse.ArgumentParser(description='')
+parser.add_argument('--cuda_idx', default=-1, type=int)
+parser.add_argument('--val_step', default=5*1, type=int)
+parser.add_argument('--program_id', default=0, type=int)
+# id-th one of this program will evaluate  val_step * N + program_id -th epoch model.
+# Test more models, number of programs == number of GPUs: [models[num_all - program_id_1], models[num_all - program_id_max(n, val_step-1)], ...] programs with id>val_step will speed up the evaluation on (val_step - id)%val_step -th epoch models.
+# Test fastest, only sequentially searched val_step*N -th models -- set all program_id as the same.
+parser.add_argument('--testsets', default='DIS-VD+DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4', type=str)
+args_eval = parser.parse_args()
+
+args_eval.program_id = (args_eval.val_step - args_eval.program_id) % args_eval.val_step
+
+config = Config()
+config.only_S_MAE = True
+device = 'cpu' if args_eval.cuda_idx < 0 else 'cuda:{}'.format(args_eval.cuda_idx)
+ckpt_dir, testsets = glob(os.path.join('ckpt', '*'))[0], args_eval.testsets
+
+
+def validate_model(model, test_loaders, epoch):
+    num_image_testset_all = {'DIS-VD': 470, 'DIS-TE1': 500, 'DIS-TE2': 500, 'DIS-TE3': 500, 'DIS-TE4': 500}
+    num_image_testset = {}
+    for testset in testsets.split('+'):
+        if 'DIS-TE' in testset:
+            num_image_testset[testset] = num_image_testset_all[testset]
+    weighted_scores = {'f_max': 0, 'sm': 0, 'e_max': 0, 'mae': 0}
+    len_all_data_loaders = 0
+    model.epoch = epoch
+    for testset, data_loader_test in test_loaders.items():
+        print('Validating {}...'.format(testset))
+        performance_dict = valid(
+            model,
+            data_loader_test,
+            pred_dir='.',
+            method=ckpt_dir.split('/')[-1] if ckpt_dir.split('/')[-1].strip('.').strip('/') else 'tmp_val',
+            testset=testset,
+            only_S_MAE=config.only_S_MAE,
+            device=device
+        )
+        print('Test set: {}:'.format(testset))
+        if config.only_S_MAE:
+            print('Smeasure: {:.4f}, MAE: {:.4f}'.format(
+                performance_dict['sm'], performance_dict['mae']
+            ))
+        else:
+            print('Fmax: {:.4f}, Fwfm: {:.4f}, Smeasure: {:.4f}, Emean: {:.4f}, MAE: {:.4f}'.format(
+                performance_dict['f_max'], performance_dict['f_wfm'], performance_dict['sm'], performance_dict['e_mean'], performance_dict['mae']
+            ))
+        if '-TE' in testset:
+            for metric in ['sm', 'mae'] if config.only_S_MAE else ['f_max', 'f_wfm', 'sm', 'e_mean', 'mae']:
+                weighted_scores[metric] += performance_dict[metric] * len(data_loader_test)
+            len_all_data_loaders += len(data_loader_test)
+    print('Weighted Scores:')
+    for metric, score in weighted_scores.items():
+        if score:
+            print('\t{}: {:.4f}.'.format(metric, score / len_all_data_loaders))
+
+@torch.no_grad()
+def main():
+    config = Config()
+    # Dataloader
+    test_loaders = {}
+    for testset in testsets.split('+'):
+        dataset = MyData(
+            datasets=testset,
+            image_size=config.size, is_train=False
+        )
+        _data_loader_test = torch.utils.data.DataLoader(
+            dataset=dataset, batch_size=config.batch_size_valid, num_workers=min(config.num_workers, config.batch_size_valid),
+            pin_memory=device != 'cpu', shuffle=False
+        )
+        print(len(_data_loader_test), "batches of valid dataloader {} have been created.".format(testset))
+        test_loaders[testset] = _data_loader_test
+
+    # Model, 3070MiB GPU memory for inference
+    model = BiRefNet(bb_pretrained=False).to(device)
+    models_evaluated = []
+    continous_sleep_time = 0
+    while True:
+        if (
+            (models_evaluated and continous_sleep_time > 60*60*2) or
+            (not models_evaluated and continous_sleep_time > 60*60*24)
+        ):
+            # If no ckpt has been saved, we wait for 24h;
+            # elif some ckpts have been saved, we wait for 2h for new ones;
+            # else: exit this waiting.
+            print('Exiting the waiting for evaluation.')
+            break
+        models_evaluated_record = 'tmp_models_evaluated.txt'
+        if os.path.exists(models_evaluated_record):
+            with open(models_evaluated_record, 'r') as f:
+                models_evaluated_global = f.read().splitlines()
+        else:
+            models_evaluated_global = []
+        models_detected = [
+            m for idx_m, m in enumerate(sorted(
+                glob(os.path.join(ckpt_dir, '*.pth')),
+                key=lambda x: int(x.rstrip('.pth').split('epoch_')[-1]), reverse=True
+            )) if idx_m % args_eval.val_step == args_eval.program_id and m not in models_evaluated + models_evaluated_global
+        ]
+        if models_detected:
+            from time import time
+            time_st = time()
+            # register the evaluated models
+            model_not_evaluated_latest = models_detected[0]
+            with open('tmp_models_evaluated.txt', 'a') as f:
+                f.write(model_not_evaluated_latest + '\n')
+            models_evaluated.append(model_not_evaluated_latest)
+            print('Loading {} for validation...'.format(model_not_evaluated_latest))
+
+            # evaluate the current model
+            state_dict = torch.load(model_not_evaluated_latest, map_location=device)
+            model.load_state_dict(state_dict, strict=False)
+            validate_model(model, test_loaders, int(model_not_evaluated_latest.rstrip('.pth').split('epoch_')[-1]))
+            continous_sleep_time = 0
+            print('Duration of this evaluation:', time() - time_st)
+        else:
+            sleep_interval = 60 * 2
+            sleep(sleep_interval)
+            continous_sleep_time += sleep_interval
+            continue
+
+
+if __name__ == '__main__':
+    main()