app.py

import os
from glob import glob
import cv2
import numpy as np
from PIL import Image
import torch
from torchvision import transforms
import gradio as gr

from models.baseline import BiRefNet
from config import Config


config = Config()
device = config.device


class ImagePreprocessor():
    def __init__(self, resolution=(1024, 1024)) -> None:
        self.transform_image = transforms.Compose([
            transforms.Resize(resolution),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
        ])

    def proc(self, image):
        image = self.transform_image(image)
        return image


model = BiRefNet().to(device)
state_dict = './birefnet_dis.pth'
if os.path.exists(state_dict):
    birefnet_dict = torch.load(state_dict, map_location=device)
    unwanted_prefix = '_orig_mod.'
    for k, v in list(birefnet_dict.items()):
        if k.startswith(unwanted_prefix):
            birefnet_dict[k[len(unwanted_prefix):]] = birefnet_dict.pop(k)
    model.load_state_dict(birefnet_dict)
model.eval()


# def predict(image_1, image_2):
#     images = [image_1, image_2]
def predict(image, resolution='1024x1024'):
    images = [image]
    image_shapes = [image.shape[:2] for image in images]
    images = [Image.fromarray(image) for image in images]

    resolution = [int(int(reso)//32*32) for reso in resolution.strip().split('x')]
    image_preprocessor = ImagePreprocessor(resolution=resolution)
    images_proc = []
    for image in images:
        images_proc.append(image_preprocessor.proc(image))
    images_proc = torch.cat([image_proc.unsqueeze(0) for image_proc in images_proc])

    with torch.no_grad():
        scaled_preds_tensor = model(images_proc.to(device))[-1].sigmoid()   # BiRefNet needs an sigmoid activation outside the forward.
    preds = []
    for image_shape, pred_tensor in zip(image_shapes, scaled_preds_tensor):
        if device == 'cuda':
            pred_tensor = pred_tensor.cpu()
        preds.append(torch.nn.functional.interpolate(pred_tensor.unsqueeze(0), size=image_shape, mode='bilinear', align_corners=True).squeeze().numpy())
    image_preds = []
    for image, pred in zip(images, preds):
        image_preds.append(
            cv2.cvtColor((pred*255).astype(np.uint8), cv2.COLOR_GRAY2RGB)
        )
    return image_preds[:] if len(images) > 1 else image_preds[0]


examples = [[_] for _ in glob('materials/examples/*')][:]

N = 1
ipt = [gr.Image() for _ in range(N)]
opt = [gr.Image() for _ in range(N)]

# Add the option of resolution in a text box.
ipt += [gr.Textbox(lines=1, placeholder="Type the resolution (`WxH`) you want, e.g., `512x512`. Higher resolutions can be much slower for inference.", label="Resolution")]
for idx_example, example in enumerate(examples):
    examples[idx_example].append('1024x1024')
examples.append(examples[-1].copy())
examples[-1][1] = '512x512'

demo = gr.Interface(
    fn=predict,
    inputs=ipt,
    outputs=opt,
    examples=examples,
    title='Online demo for `Bilateral Reference for High-Resolution Dichotomous Image Segmentation`',
    description=('Upload a picture, our model will give you the binary maps of the highly accurate segmentation of the salient objects in it. :)'
                 '\nThe resolution used in our training was `1024x1024`, which is too much burden for the huggingface free spaces like this one (cost ~500s). Please set resolution as more than `768x768` for images with many texture details to obtain good results!')
)
demo.launch(debug=True)