Create open_inst.py

tasks/open_inst.py

@@ -0,0 +1,60 @@
+# --------------------------------------------------------
+# 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 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
+from detectron2.structures import BitMasks
+
+
+t = []
+t.append(transforms.Resize(512, interpolation=Image.BICUBIC))
+transform = transforms.Compose(t)
+metadata = MetadataCatalog.get('ade20k_panoptic_train')
+
+def open_instseg(model, image, texts, inpainting_text, *args, **kwargs):
+    thing_classes = [x.strip() for x in texts.split(',')]
+    thing_colors = [random_color(rgb=True, maximum=255).astype(np.int32).tolist() for _ in range(len(thing_classes))]
+    thing_dataset_id_to_contiguous_id = {x:x for x in range(len(thing_classes))}
+
+    MetadataCatalog.get("demo").set(
+        thing_colors=thing_colors,
+        thing_classes=thing_classes,
+        thing_dataset_id_to_contiguous_id=thing_dataset_id_to_contiguous_id,
+    )
+
+    with torch.no_grad():
+        model.model.sem_seg_head.predictor.lang_encoder.get_text_embeddings(thing_classes + ["background"], is_eval=True)
+
+        metadata = MetadataCatalog.get('demo')
+        model.model.metadata = metadata
+        model.model.sem_seg_head.num_classes = len(thing_classes)
+
+        image_ori = transform(image)
+        width = image_ori.size[0]
+        height = image_ori.size[1]
+        image = np.asarray(image_ori)
+        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)
+
+        inst_seg = outputs[-1]['instances']
+        inst_seg.pred_masks = inst_seg.pred_masks.cpu()
+        inst_seg.pred_boxes = BitMasks(inst_seg.pred_masks > 0).get_bounding_boxes()
+        demo = visual.draw_instance_predictions(inst_seg) # rgb Image
+        res = demo.get_image()
+
+
+    MetadataCatalog.remove('demo')
+    torch.cuda.empty_cache()
+    return Image.fromarray(res), '', None