app.py

import os
import cv2
import numpy as np
import torch
import gradio as gr
import spaces

from glob import glob
from typing import Optional, Tuple

from PIL import Image
from gradio_imageslider import ImageSlider
from transformers import AutoModelForImageSegmentation
from torchvision import transforms

import requests
from io import BytesIO
import zipfile


torch.set_float32_matmul_precision('high')
torch.jit.script = lambda f: f

device = "cuda" if torch.cuda.is_available() else "CPU"


def array_to_pil_image(image: np.ndarray, size: Tuple[int, int] = (1024, 1024)) -> Image.Image:
    image = cv2.resize(image, size, interpolation=cv2.INTER_LINEAR)
    image = Image.fromarray(image).convert('RGB')
    return image


class ImagePreprocessor():
    def __init__(self, resolution: Tuple[int, int] = (1024, 1024)) -> None:
        self.transform_image = transforms.Compose([
            # transforms.Resize(resolution),    # 1. keep consistent with the cv2.resize used in training 2. redundant with that in path_to_image()
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
        ])

    def proc(self, image: Image.Image) -> torch.Tensor:
        image = self.transform_image(image)
        return image


usage_to_weights_file = {
    'General': 'BiRefNet',
    'General-Lite': 'BiRefNet_lite',
    'Portrait': 'BiRefNet-portrait',
    'DIS': 'BiRefNet-DIS5K',
    'HRSOD': 'BiRefNet-HRSOD',
    'COD': 'BiRefNet-COD',
    'DIS-TR_TEs': 'BiRefNet-DIS5K-TR_TEs',
    'General-legacy': 'BiRefNet-legacy'
}

birefnet = AutoModelForImageSegmentation.from_pretrained('/'.join(('zhengpeng7', usage_to_weights_file['General'])), trust_remote_code=True)
birefnet.to(device)
birefnet.eval()


@spaces.GPU
def predict(images, resolution, weights_file):
    assert (images is not None), 'AssertionError: images cannot be None.'

    global birefnet
    # Load BiRefNet with chosen weights
    _weights_file = '/'.join(('zhengpeng7', usage_to_weights_file[weights_file] if weights_file is not None else usage_to_weights_file['General']))
    print('Using weights: {}.'.format(_weights_file))
    birefnet = AutoModelForImageSegmentation.from_pretrained(_weights_file, trust_remote_code=True)
    birefnet.to(device)
    birefnet.eval()

    try:
        resolution = [int(int(reso)//32*32) for reso in resolution.strip().split('x')]
    except:
        resolution = [1024, 1024]
        print('Invalid resolution input. Automatically changed to 1024x1024.')

    if isinstance(images, list):
        # For tab_batch
        save_paths = []
        save_dir = 'preds-BiRefNet'
        if not os.path.exists(save_dir):
            os.makedirs(save_dir)
        tab_is_batch = True
    else:
        images = [images]
        tab_is_batch = False

    for idx_image, image_src in enumerate(images):
        if isinstance(image_src, str):
            if os.path.isfile(image_src):
                image = np.array(Image.open(image_src))
            else:
                response = requests.get(image_src)
                image_data = BytesIO(response.content)
                image = np.array(Image.open(image_data))
        else:
            image = image_src

        image_shape = image.shape[:2]
        image_pil = array_to_pil_image(image, tuple(resolution))

        # Preprocess the image
        image_preprocessor = ImagePreprocessor(resolution=tuple(resolution))
        image_proc = image_preprocessor.proc(image_pil)
        image_proc = image_proc.unsqueeze(0)

        # Perform the prediction
        with torch.no_grad():
            scaled_pred_tensor = birefnet(image_proc.to(device))[-1].sigmoid()

        if device == 'cuda':
            scaled_pred_tensor = scaled_pred_tensor.cpu()

        # Resize the prediction to match the original image shape
        pred = torch.nn.functional.interpolate(scaled_pred_tensor, size=image_shape, mode='bilinear', align_corners=True).squeeze().numpy()

        # Apply the prediction mask to the original image
        image_pil = image_pil.resize(pred.shape[::-1])
        pred = np.repeat(np.expand_dims(pred, axis=-1), 3, axis=-1)
        image_masked = (pred * np.array(image_pil)).astype(np.uint8)

        torch.cuda.empty_cache()

        if tab_is_batch:
            save_file_path = os.path.join(save_dir, "{}.png".format(os.path.splitext(os.path.basename(image_src))[0]))
            cv2.imwrite(save_file_path, image_masked)
            save_paths.append(save_file_path)

    if tab_is_batch:
        zip_file_path = os.path.join(save_dir, "{}.zip".format(save_dir))
        with zipfile.ZipFile(zip_file_path, 'w') as zipf:
            for file in save_paths:
                zipf.write(file, os.path.basename(file))

    return image, image_masked


examples = [[_] for _ in glob('examples/*')][:]
# Add the option of resolution in a text box.
for idx_example, example in enumerate(examples):
    examples[idx_example].append('1024x1024')
examples.append(examples[-1].copy())
examples[-1][1] = '512x512'

examples_url = [
    ['https://hips.hearstapps.com/hmg-prod/images/gettyimages-1229892983-square.jpg'],
]
for idx_example_url, example_url in enumerate(examples_url):
    examples_url[idx_example_url].append('1024x1024')

descriptions = ('Upload a picture, our model will extract a highly accurate segmentation of the subject in it.\n)'
                 ' The resolution used in our training was `1024x1024`, thus the suggested resolution to obtain good results!\n'
                 ' Our codes can be found at https://github.com/ZhengPeng7/BiRefNet.\n'
                 ' We also maintain the HF model of BiRefNet at https://huggingface.co/ZhengPeng7/BiRefNet for easier access.')

tab_image = gr.Interface(
    fn=predict,
    inputs=[
        gr.Image(label='Upload an image'),
        gr.Textbox(lines=1, placeholder="Type the resolution (`WxH`) you want, e.g., `1024x1024`. Higher resolutions can be much slower for inference.", label="Resolution"),
        gr.Radio(list(usage_to_weights_file.keys()), value='General', label="Weights", info="Choose the weights you want.")
    ],
    outputs=ImageSlider(label="BiRefNet's prediction", type="pil"),
    examples=examples,
    api_name="image",
    description=descriptions,
)

tab_text = gr.Interface(
    fn=predict,
    inputs=[
        gr.Textbox(label="Paste an image URL"),
        gr.Textbox(lines=1, placeholder="Type the resolution (`WxH`) you want, e.g., `1024x1024`. Higher resolutions can be much slower for inference.", label="Resolution"),
        gr.Radio(list(usage_to_weights_file.keys()), value='General', label="Weights", info="Choose the weights you want.")
    ],
    outputs=ImageSlider(label="BiRefNet's prediction", type="pil"),
    examples=examples_url,
    api_name="text",
    description=descriptions+'\nTab-URL is partially modified from https://huggingface.co/spaces/not-lain/background-removal, thanks to this great work!',
)

tab_batch = gr.Interface(
    fn=predict,
    inputs=gr.File(label="Upload multiple images", type="filepath", file_count="multiple"),
    outputs=[gr.Gallery(label="BiRefNet's predictions"), gr.File(label="Download masked images.")],
    api_name="batch",
    description=descriptions+'\nTab-batch is partially modified from https://huggingface.co/spaces/NegiTurkey/Multi_Birefnetfor_Background_Removal, thanks to this great work!',
)

demo = gr.TabbedInterface(
    [tab_image, tab_text, tab_batch],
    ['image', 'text', 'batch'],
    title="BiRefNet demo for subject extraction (general / salient / camouflaged / portrait).",
)

if __name__ == "__main__":
    demo.launch(debug=True)