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)6# --------------------------------------------------------7import os8import openai9import torch10import numpy as np11from scipy import ndimage12from PIL import Image13from utils.inpainting import pad_image, crop_image14from torchvision import transforms15from utils.visualizer import Visualizer16from diffusers import StableDiffusionInpaintPipeline17from detectron2.utils.colormap import random_color18from detectron2.data import MetadataCatalog19 20 21t = []22t.append(transforms.Resize(512, interpolation=Image.BICUBIC))23transform = transforms.Compose(t)24metadata = MetadataCatalog.get('ade20k_panoptic_train')25 26pipe = StableDiffusionInpaintPipeline.from_pretrained(27    # "stabilityai/stable-diffusion-2-inpainting",28    "runwayml/stable-diffusion-inpainting",29    revision="fp16", 30    torch_dtype=torch.float16,31).to("cuda")32 33prompts = []34prompts.append("instruction: remove the person, task: (referring editing), source: [person], target:<clean and empty scene>.")35prompts.append("instruction: remove the person in the middle, task: (referring editing), source: [person in the middle], target:<clean and empty scene>.")36prompts.append("instruction: remove the dog on the left side, task: (referring editing), source: [dog on the left side], target:<clean and empty scene>.")37prompts.append("instruction: change the apple to a pear, task: (referring editing), source: [apple], target: <pear>.")38prompts.append("instruction: change the red apple to a green one, task: (referring editing), source: [red apple], target: <green apple>.")39prompts.append("instruction: change the color of bird's feathers from white to blue, task: (referring editing), source: [white bird], target: <blue bird>.")40prompts.append("instruction: replace the dog with a cat, task: (referring editing), source: [dot], target: <cat>.")41prompts.append("instruction: replace the red apple with a green one, task: (referring editing), source: [red apple], target: <green apple>.")42 43openai.api_type = "azure"44openai.api_base = "https://xdecoder.openai.azure.com/"45openai.api_version = "2022-12-01"46openai.api_key = os.environ["OPENAI_API_KEY"]47 48def get_gpt3_response(prompt):49    response = openai.Completion.create(50      engine="text001",51      prompt=prompt,52      temperature=0.7,53      max_tokens=512,54      top_p=1,55      frequency_penalty=0,56      presence_penalty=0,57    )58    59    return response60 61def referring_inpainting_gpt3(model, image, instruction, *args, **kwargs):    62    # convert instruction to source and target63    instruction = instruction.replace('.', '')64    print(instruction)65    resp = get_gpt3_response(' '.join(prompts) + ' instruction: ' + instruction + ',')66    resp_text = resp['choices'][0]['text']67    print(resp_text)68    ref_text = resp_text[resp_text.find('[')+1:resp_text.find(']')]69    inp_text = resp_text[resp_text.find('<')+1:resp_text.find('>')]70 71    model.model.metadata = metadata72    texts = [[ref_text if ref_text.strip().endswith('.') else (ref_text.strip() + '.')]]73    image_ori = crop_image(transform(image))74 75    with torch.no_grad():76        width = image_ori.size[0]77        height = image_ori.size[1]78        image = np.asarray(image_ori)79        image_ori_np = np.asarray(image_ori)80        images = torch.from_numpy(image.copy()).permute(2,0,1).cuda()81 82        batch_inputs = [{'image': images, 'height': height, 'width': width, 'groundings': {'texts': texts}}]        83        outputs = model.model.evaluate_grounding(batch_inputs, None)84        visual = Visualizer(image_ori_np, metadata=metadata)85 86        grd_mask = (outputs[0]['grounding_mask'] > 0).float().cpu().numpy()87        for idx, mask in enumerate(grd_mask):88            color = random_color(rgb=True, maximum=1).astype(np.int32).tolist()89            demo = visual.draw_binary_mask(mask, color=color, text=texts[idx])90        res = demo.get_image()91    92    if inp_text not in ['no', '']:93        image_crop = image_ori94        struct2 = ndimage.generate_binary_structure(2, 2)95        mask_dilated = ndimage.binary_dilation(grd_mask[0], structure=struct2, iterations=3).astype(grd_mask[0].dtype)96        mask = Image.fromarray(mask_dilated * 255).convert('RGB')97        image_and_mask = {98            "image": image_crop,99            "mask": mask,100        }101        # images_inpainting = inpainting(inpainting_model, image_and_mask, inp_text, ddim_steps, num_samples, scale, seed)102        width = image_ori.size[0]; height = image_ori.size[1]103        images_inpainting = pipe(prompt = inp_text.strip(), image=image_and_mask['image'], mask_image=image_and_mask['mask'], height=height, width=width).images104        torch.cuda.empty_cache()105        return images_inpainting[0]106    else:107        torch.cuda.empty_cache()108        return Image.fromarray(res)