Shawn87377/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 Xueyan Zou (xueyan@cs.wisc.edu)6# --------------------------------------------------------7 8import torch9import numpy as np10from PIL import Image11from torchvision import transforms12from utils.visualizer import Visualizer13from detectron2.utils.colormap import random_color14from detectron2.data import MetadataCatalog15 16 17t = []18t.append(transforms.Resize(512, interpolation=Image.BICUBIC))19transform = transforms.Compose(t)20metadata = MetadataCatalog.get('ade20k_panoptic_train')21 22def open_panoseg(model, image, texts, inpainting_text, *args, **kwargs):23    stuff_classes = [x.strip() for x in texts.split(';')[0].replace('stuff:','').split(',')]24    thing_classes = [x.strip() for x in texts.split(';')[1].replace('thing:','').split(',')]25    thing_colors = [random_color(rgb=True, maximum=255).astype(np.int32).tolist() for _ in range(len(thing_classes))]26    stuff_colors = [random_color(rgb=True, maximum=255).astype(np.int32).tolist() for _ in range(len(stuff_classes))]27    thing_dataset_id_to_contiguous_id = {x:x for x in range(len(thing_classes))}28    stuff_dataset_id_to_contiguous_id = {x+len(thing_classes):x for x in range(len(stuff_classes))}29 30    MetadataCatalog.get("demo").set(31        thing_colors=thing_colors,32        thing_classes=thing_classes,33        thing_dataset_id_to_contiguous_id=thing_dataset_id_to_contiguous_id,34        stuff_colors=stuff_colors,35        stuff_classes=stuff_classes,36        stuff_dataset_id_to_contiguous_id=stuff_dataset_id_to_contiguous_id,37    )38    model.model.sem_seg_head.predictor.lang_encoder.get_text_embeddings(thing_classes + stuff_classes + ["background"], is_eval=True)39    metadata = MetadataCatalog.get('demo')40    model.model.metadata = metadata41    model.model.sem_seg_head.num_classes = len(thing_classes + stuff_classes)42 43    with torch.no_grad():44        image_ori = transform(image)45        width = image_ori.size[0]46        height = image_ori.size[1]47        image = transform(image_ori)48        image = np.asarray(image)49        images = torch.from_numpy(image.copy()).permute(2,0,1).cuda()50 51        batch_inputs = [{'image': images, 'height': height, 'width': width}]52        outputs = model.forward(batch_inputs)53        visual = Visualizer(image_ori, metadata=metadata)54 55        pano_seg = outputs[-1]['panoptic_seg'][0]56        pano_seg_info = outputs[-1]['panoptic_seg'][1]57 58        for i in range(len(pano_seg_info)):59            if pano_seg_info[i]['category_id'] in metadata.thing_dataset_id_to_contiguous_id.keys():60                pano_seg_info[i]['category_id'] = metadata.thing_dataset_id_to_contiguous_id[pano_seg_info[i]['category_id']]61            else:62                pano_seg_info[i]['isthing'] = False63                pano_seg_info[i]['category_id'] = metadata.stuff_dataset_id_to_contiguous_id[pano_seg_info[i]['category_id']]64 65        demo = visual.draw_panoptic_seg(pano_seg.cpu(), pano_seg_info) # rgb Image66        res = demo.get_image()67 68    MetadataCatalog.remove('demo')69    torch.cuda.empty_cache()70    return Image.fromarray(res), '', None