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 Xueyan Zou (xueyan@cs.wisc.edu)6# --------------------------------------------------------7 8import torch9import torch.nn.functional as F10import numpy as np11from PIL import Image12from torchvision import transforms13from utils.visualizer import Visualizer14from detectron2.data import MetadataCatalog15 16t = []17t.append(transforms.Resize(224, interpolation=Image.BICUBIC))18transform_ret = transforms.Compose(t)19t = []20t.append(transforms.Resize(512, interpolation=Image.BICUBIC))21transform_grd = transforms.Compose(t)22 23metedata = MetadataCatalog.get('coco_2017_train_panoptic')24 25def referring_captioning(model, image, texts, inpainting_text, *args, **kwargs):26    model_last, model_cap = model27    with torch.no_grad():28        image_ori = image29        image = transform_grd(image)30        width = image.size[0]31        height = image.size[1]32        image = np.asarray(image)33        image_ori_ = image34        images = torch.from_numpy(image.copy()).permute(2,0,1).cuda()35        texts_input = [[texts.strip() if texts.endswith('.') else (texts + '.')]]36 37        batch_inputs = [{'image': images, 'groundings': {'texts':texts_input}, 'height': height, 'width': width}]38        outputs = model_last.model.evaluate_grounding(batch_inputs, None)39 40        grd_mask = (outputs[-1]['grounding_mask'] > 0).float()41        grd_mask_ = (1 - F.interpolate(grd_mask[None,], (224, 224), mode='nearest')[0]).bool()42 43        color = [252/255, 91/255, 129/255]44        visual = Visualizer(image_ori_, metadata=metedata)45        demo = visual.draw_binary_mask(grd_mask.cpu().numpy()[0], color=color, text=texts)46        res = demo.get_image()47 48        if (1 - grd_mask_.float()).sum() < 5:49            torch.cuda.empty_cache()50            return Image.fromarray(res), 'n/a', None51 52        grd_mask_ = grd_mask_ * 053        image = transform_ret(image_ori)54        image_ori = np.asarray(image_ori)55        image = np.asarray(image)56        images = torch.from_numpy(image.copy()).permute(2,0,1).cuda()57        batch_inputs = [{'image': images, 'image_id': 0, 'captioning_mask': grd_mask_}]58 59        token_text = texts.replace('.','') if texts.endswith('.') else texts60        token = model_cap.model.sem_seg_head.predictor.lang_encoder.tokenizer.encode(token_text)61        token = torch.tensor(token)[None,:-1]62 63        outputs = model_cap.model.evaluate_captioning(batch_inputs, extra={'token': token})64        # outputs = model_cap.model.evaluate_captioning(batch_inputs, extra={})65        text = outputs[-1]['captioning_text']66 67    torch.cuda.empty_cache()68    return Image.fromarray(res), text, None