xdecoder/Instruct-X-Decoder

1# --------------------------------------------------------2# X-Decoder -- Generalized Decoding for Pixel, Image, and Language3# Copyright (c) 2022 Microsoft4# Licensed under The MIT License [see LICENSE for details]5# Written by Jianwei Yang (jianwyan@microsoft.com), Xueyan Zou (xueyan@cs.wisc.edu)6# --------------------------------------------------------7 8import torch9import numpy as np10from PIL import Image11from utils.inpainting import pad_image12from torchvision import transforms13from utils.visualizer import Visualizer14from diffusers import StableDiffusionInpaintPipeline15from detectron2.utils.colormap import random_color16from detectron2.data import MetadataCatalog17from scipy import ndimage18 19 20t = []21t.append(transforms.Resize(512, interpolation=Image.BICUBIC))22transform = transforms.Compose(t)23metadata = MetadataCatalog.get('ade20k_panoptic_train')24 25pipe = StableDiffusionInpaintPipeline.from_pretrained(26    # "stabilityai/stable-diffusion-2-inpainting",27    "runwayml/stable-diffusion-inpainting",28    revision="fp16", 29    torch_dtype=torch.float16,30).to("cuda")31 32def crop_image(input_image):33    crop_w, crop_h = np.floor(np.array(input_image.size) / 64).astype(int) * 6434    im_cropped = Image.fromarray(np.array(input_image)[:crop_h, :crop_w])35    return im_cropped36 37def referring_inpainting(model, image, texts, inpainting_text, *args, **kwargs):38    model.model.metadata = metadata39    texts = [[texts if texts.strip().endswith('.') else (texts.strip() + '.')]]40    image_ori = crop_image(transform(image))41 42    with torch.no_grad():43        width = image_ori.size[0]44        height = image_ori.size[1]45        image = np.asarray(image_ori)46        image_ori_np = np.asarray(image_ori)47        images = torch.from_numpy(image.copy()).permute(2,0,1).cuda()48 49        batch_inputs = [{'image': images, 'height': height, 'width': width, 'groundings': {'texts': texts}}]        50        outputs = model.model.evaluate_grounding(batch_inputs, None)51        visual = Visualizer(image_ori_np, metadata=metadata)52 53        grd_mask = (outputs[0]['grounding_mask'] > 0).float().cpu().numpy()54        for idx, mask in enumerate(grd_mask):55            color = random_color(rgb=True, maximum=1).astype(np.int32).tolist()56            demo = visual.draw_binary_mask(mask, color=color, text=texts[idx])57        res = demo.get_image()58    59    if inpainting_text not in ['no', '']:60        # if we want to do inpainting61        image_crop = image_ori62        struct2 = ndimage.generate_binary_structure(2, 2)63        mask_dilated = ndimage.binary_dilation(grd_mask[0], structure=struct2, iterations=3).astype(grd_mask[0].dtype)64        mask = Image.fromarray(mask_dilated * 255).convert('RGB')65        image_and_mask = {66            "image": image_crop,67            "mask": mask,68        }69        width = image_crop.size[0]; height = image_crop.size[1]70        images_inpainting = pipe(prompt = inpainting_text.strip(), image=image_and_mask['image'], mask_image=image_and_mask['mask'], height=height, width=width).images[0]71        # put images_inpainting back to original image72        # image_ori.paste(images_inpainting)        73        torch.cuda.empty_cache()74        return Image.fromarray(res) ,'' , images_inpainting75    else:76        torch.cuda.empty_cache()77        return image_ori, 'text', Image.fromarray(res)