README.md

---
license: mit
---

<p align="center" width="100%">
<img src="https://i.postimg.cc/MKmyP9wH/new-banner.png"  width="80%" height="80%">
</p>


<div>
<div align="center">
    <a href='https://brianboli.com/' target='_blank'>Bo Li*<sup>1</sup></a>&emsp;
    <a href='https://zhangyuanhan-ai.github.io/' target='_blank'>Yuanhan Zhang*<sup>,1</sup></a>&emsp;
    <a href='https://cliangyu.com/' target='_blank'>Liangyu Chen*<sup>,1</sup></a>&emsp;
    <a href='https://king159.github.io/' target='_blank'>Jinghao Wang*<sup>,1</sup></a>&emsp;
    <a href='https://pufanyi.github.io/' target='_blank'>Fanyi Pu*<sup>,1</sup></a>&emsp;
    </br>
    <a href='https://jingkang50.github.io/' target='_blank'>Jingkang Yang<sup>1</sup></a>&emsp;
    <a href='https://chunyuan.li/' target='_blank'>Chunyuan Li<sup>2</sup></a>&emsp;
    <a href='https://liuziwei7.github.io/' target='_blank'>Ziwei Liu<sup>1</sup></a>
</div>
<div>
<div align="center">
    <sup>1</sup>S-Lab, Nanyang Technological University&emsp;
    <sup>2</sup>Microsoft Research, Redmond
</div>

This weight is for **initilizing training for Otter**. It's directly converted from Openflamingo.

You can load and try this model using
```python
model = OtterForConditionalGeneration.from_pretrained("luodian/OTTER-MPT7B-Init", device_map="sequential")
model.text_tokenizer.padding_side = "left"
tokenizer = model.text_tokenizer
image_processor = transformers.CLIPImageProcessor()
model.eval()
```

You can also start training Otter via the commands
```python
python -m accelerate.commands.launch --config_file=./pipeline/accelerate_configs/accelerate_config_fsdp.yaml \
pipeline/train/instruction_following.py \
--pretrained_model_name_or_path=luodian/OTTER-MPT7B-Init \
--mimicit_path=/data/azure_storage/otter/mimicit/xx/xx_instructions.json \
--images_path=/data/azure_storage/otter/mimicit/xx/xx.json \
--batch_size=4 --num_epochs=1 --report_to_wandb \
--wandb_entity=ntu-slab \
--external_save_dir=/data/bli/checkpoints \
--save_hf_model \
--run_name=OTTER-MPT1B \
--wandb_project=OTTER-MPT1B \
--workers=4 \
--lr_scheduler=cosine \
--learning_rate=1e-5 \
--warmup_steps_ratio=0.01
```

If you wish to init a video instruction tuning, you should add 
```json
"max_num_frames": 128
```
to `config.json` inside the folder.