tasks/open_sem.py

# --------------------------------------------------------
# X-Decoder -- Generalized Decoding for Pixel, Image, and Language
# Copyright (c) 2022 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Written by Xueyan Zou (xueyan@cs.wisc.edu)
# --------------------------------------------------------

import os
import cv2
import torch
import numpy as np
from PIL import Image
from torchvision import transforms
from utils.visualizer import Visualizer
from detectron2.utils.colormap import random_color
from detectron2.data import MetadataCatalog


t = []
t.append(transforms.Resize(512, interpolation=Image.BICUBIC))
transform = transforms.Compose(t)
metadata = MetadataCatalog.get('ade20k_panoptic_train')

def open_semseg(model, image, texts, inpainting_text, *args, **kwargs):    
    stuff_classes = [x.strip() for x in texts.split(',')]
    stuff_colors = [random_color(rgb=True, maximum=255).astype(np.int32).tolist() for _ in range(len(stuff_classes))]
    stuff_dataset_id_to_contiguous_id = {x:x for x in range(len(stuff_classes))}

    MetadataCatalog.get("demo").set(
        stuff_colors=stuff_colors,
        stuff_classes=stuff_classes,
        stuff_dataset_id_to_contiguous_id=stuff_dataset_id_to_contiguous_id,
    )
    model.model.sem_seg_head.predictor.lang_encoder.get_text_embeddings(stuff_classes + ["background"], is_eval=True)
    metadata = MetadataCatalog.get('demo')
    model.model.metadata = metadata
    model.model.sem_seg_head.num_classes = len(stuff_classes)

    with torch.no_grad():
        image_ori = transform(image)
        width = image_ori.size[0]
        height = image_ori.size[1]
        image = transform(image_ori)
        image = np.asarray(image)
        images = torch.from_numpy(image.copy()).permute(2,0,1).cuda()

        batch_inputs = [{'image': images, 'height': height, 'width': width}]
        outputs = model.forward(batch_inputs)
        visual = Visualizer(image_ori, metadata=metadata)

        sem_seg = outputs[-1]['sem_seg'].max(0)[1]
        demo = visual.draw_sem_seg(sem_seg.cpu(), alpha=0.5) # rgb Image
        res = demo.get_image()
        
    MetadataCatalog.remove('demo')
    torch.cuda.empty_cache()
    return Image.fromarray(res), '', None