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

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_T',
    '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(
    image: np.ndarray,
    resolution: str,
    weights_file: Optional[str]
) -> Tuple[np.ndarray, np.ndarray]:
    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:', _weights_file)
    birefnet = AutoModelForImageSegmentation.from_pretrained(_weights_file, trust_remote_code=True)
    birefnet.to(device)
    birefnet.eval()

    resolution = f"{image.shape[1]}x{image.shape[0]}" if resolution == '' else resolution
    resolution = [int(int(reso)//32*32) for reso in resolution.strip().split('x')]
    
    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_pred = (pred * np.array(image_pil)).astype(np.uint8)

    torch.cuda.empty_cache()

    return image, image_pred


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'

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"
)

tab_text = gr.Interface(
    fn=predict,
    inputs=gr.Textbox(label="Paste an image URL"),
    outputs=ImageSlider(label="BiRefNet's prediction", type="pil"),
    examples=["https://hips.hearstapps.com/hmg-prod/images/gettyimages-1229892983-square.jpg"],
    api_name="text"
)

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

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