File size: 30,373 Bytes
760bde3
6e4c9f7
 
 
72fe59d
9546498
760bde3
434a891
 
72fe59d
 
 
40c7708
 
ed7763f
 
 
760bde3
5c0bd4c
72fe59d
fd01170
6061d17
fd01170
4ba09fa
 
 
 
 
 
 
72fe59d
4ba09fa
 
6061d17
 
 
 
 
4ba09fa
 
 
 
72fe59d
 
4ba09fa
72fe59d
7a7f9d8
4ba09fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7957dbb
4ba09fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5d0da89
 
 
 
 
4ba09fa
 
 
 
 
 
 
72fe59d
 
 
 
4ba09fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c419c35
4ba09fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
72fe59d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ca589ef
4ba09fa
 
 
 
5d0da89
4ba09fa
57ec633
 
91f6a3d
 
 
c419c35
72fe59d
c419c35
 
 
72fe59d
0d1ddf1
 
bf71fd5
c419c35
 
72fe59d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c419c35
 
72fe59d
 
f2bd037
72fe59d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7a7f9d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9988241
7a7f9d8
 
 
 
 
 
 
 
 
 
9988241
7a7f9d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
72fe59d
 
 
7a7f9d8
 
 
 
 
b902809
 
 
91f6a3d
 
 
 
72fe59d
779c33a
7a7f9d8
4ba09fa
 
72fe59d
 
 
 
779c33a
4ba09fa
ed7763f
9546498
72fe59d
4ba09fa
72fe59d
4ba09fa
72fe59d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b902809
7a7f9d8
91f6a3d
72fe59d
4ba09fa
72fe59d
 
 
 
 
 
 
 
 
 
 
4ba09fa
7a7f9d8
72fe59d
 
c419c35
4ba09fa
 
 
 
 
 
 
 
c419c35
4ba09fa
c419c35
4ba09fa
 
 
 
 
72fe59d
4ba09fa
 
 
 
 
 
 
 
 
ed7763f
4ba09fa
72fe59d
 
 
 
7a7f9d8
72fe59d
7a7f9d8
b902809
72fe59d
 
 
 
7a7f9d8
72fe59d
 
 
 
6e4c9f7
72fe59d
 
 
 
 
 
 
7a7f9d8
 
72fe59d
4ba09fa
ed7763f
72fe59d
 
 
 
 
 
 
 
 
 
 
6e4c9f7
72fe59d
 
 
 
 
 
 
 
 
6e4c9f7
72fe59d
 
 
 
 
 
 
 
 
 
 
 
 
 
6e4c9f7
72fe59d
 
 
7a7f9d8
 
72fe59d
 
 
 
 
 
1f8f331
4ba09fa
ed7763f
1f8f331
4ba09fa
ed7763f
7a7f9d8
72fe59d
b902809
4ba09fa
72fe59d
7a7f9d8
b902809
72fe59d
7a7f9d8
72fe59d
 
 
7a7f9d8
 
 
ca589ef
72fe59d
 
 
 
 
7a7f9d8
 
 
 
 
53720d1
 
ca589ef
4ba09fa
 
 
 
 
 
 
 
 
 
 
 
72fe59d
7a7f9d8
 
72fe59d
 
7a7f9d8
b902809
72fe59d
7a7f9d8
4ba09fa
7a7f9d8
4ba09fa
5d0da89
4ba09fa
 
5d0da89
4ba09fa
72fe59d
5d0da89
72fe59d
 
 
 
 
 
 
4ba09fa
 
7a7f9d8
 
 
 
 
 
 
 
 
 
4ba09fa
e5f7fa3
 
 
 
72fe59d
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743

import warnings
warnings.filterwarnings('ignore')

import subprocess, io, os, sys, time
from loguru import logger

os.environ["CUDA_VISIBLE_DEVICES"] = "0"

if os.environ.get('IS_MY_DEBUG') is None:
    result = subprocess.run(['pip', 'install', '-e', 'GroundingDINO'], check=True)
    print(f'pip install GroundingDINO = {result}')

result = subprocess.run(['pip', 'list'], check=True)
print(f'pip list = {result}')

sys.path.insert(0, './GroundingDINO')

if not os.path.exists('./sam_vit_h_4b8939.pth'):
    logger.info(f"get sam_vit_h_4b8939.pth...")
    result = subprocess.run(['wget', 'https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth'], check=True)
    print(f'wget sam_vit_h_4b8939.pth result = {result}')    

import gradio as gr

import argparse
import copy

import numpy as np
import torch
from PIL import Image, ImageDraw, ImageFont, ImageOps

# Grounding DINO
import GroundingDINO.groundingdino.datasets.transforms as T
from GroundingDINO.groundingdino.models import build_model
from GroundingDINO.groundingdino.util import box_ops
from GroundingDINO.groundingdino.util.slconfig import SLConfig
from GroundingDINO.groundingdino.util.utils import clean_state_dict, get_phrases_from_posmap

import cv2
import numpy as np
import matplotlib.pyplot as plt
from lama_cleaner.model_manager import ModelManager
from lama_cleaner.schema import Config

# segment anything
from segment_anything import build_sam, SamPredictor, SamAutomaticMaskGenerator

# diffusers
import PIL
import requests
import torch
from io import BytesIO
from diffusers import StableDiffusionInpaintPipeline
from huggingface_hub import hf_hub_download

def load_model_hf(model_config_path, repo_id, filename, device='cpu'):
    args = SLConfig.fromfile(model_config_path) 
    model = build_model(args)
    args.device = device

    cache_file = hf_hub_download(repo_id=repo_id, filename=filename)
    checkpoint = torch.load(cache_file, map_location=device)
    log = model.load_state_dict(clean_state_dict(checkpoint['model']), strict=False)
    print("Model loaded from {} \n => {}".format(cache_file, log))
    _ = model.eval()
    return model    

def plot_boxes_to_image(image_pil, tgt):
    H, W = tgt["size"]
    boxes = tgt["boxes"]
    labels = tgt["labels"]
    assert len(boxes) == len(labels), "boxes and labels must have same length"

    draw = ImageDraw.Draw(image_pil)
    mask = Image.new("L", image_pil.size, 0)
    mask_draw = ImageDraw.Draw(mask)

    # draw boxes and masks
    for box, label in zip(boxes, labels):
        # from 0..1 to 0..W, 0..H
        box = box * torch.Tensor([W, H, W, H])
        # from xywh to xyxy
        box[:2] -= box[2:] / 2
        box[2:] += box[:2]
        # random color
        color = tuple(np.random.randint(0, 255, size=3).tolist())
        # draw
        x0, y0, x1, y1 = box
        x0, y0, x1, y1 = int(x0), int(y0), int(x1), int(y1)

        draw.rectangle([x0, y0, x1, y1], outline=color, width=6)
        # draw.text((x0, y0), str(label), fill=color)

        font = ImageFont.load_default()
        if hasattr(font, "getbbox"):
            bbox = draw.textbbox((x0, y0), str(label), font)
        else:
            w, h = draw.textsize(str(label), font)
            bbox = (x0, y0, w + x0, y0 + h)
        # bbox = draw.textbbox((x0, y0), str(label))
        draw.rectangle(bbox, fill=color)
        font = os.path.join(cv2.__path__[0],'qt','fonts','DejaVuSans.ttf')
        font_size = 36
        new_font = ImageFont.truetype(font, font_size)

        draw.text((x0+2, y0+2), str(label), font=new_font, fill="white")

        mask_draw.rectangle([x0, y0, x1, y1], fill=255, width=6)

    return image_pil, mask

def load_image(image_path):
    # # load image
    if isinstance(image_path, PIL.Image.Image):
        image_pil = image_path
    else:
        image_pil = Image.open(image_path).convert("RGB")  # load image

    transform = T.Compose(
        [
            T.RandomResize([800], max_size=1333),
            T.ToTensor(),
            T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
        ]
    )
    image, _ = transform(image_pil, None)  # 3, h, w
    return image_pil, image

def load_model(model_config_path, model_checkpoint_path, device):
    args = SLConfig.fromfile(model_config_path)
    args.device = device
    model = build_model(args)
    checkpoint = torch.load(model_checkpoint_path, map_location=device) #"cpu")
    load_res = model.load_state_dict(clean_state_dict(checkpoint["model"]), strict=False)
    print(load_res)
    _ = model.eval()
    return model

def get_grounding_output(model, image, caption, box_threshold, text_threshold, with_logits=True, device="cpu"):
    caption = caption.lower()
    caption = caption.strip()
    if not caption.endswith("."):
        caption = caption + "."
    model = model.to(device)
    image = image.to(device)
    with torch.no_grad():
        outputs = model(image[None], captions=[caption])
    logits = outputs["pred_logits"].cpu().sigmoid()[0]  # (nq, 256)
    boxes = outputs["pred_boxes"].cpu()[0]  # (nq, 4)
    logits.shape[0]

    # filter output
    logits_filt = logits.clone()
    boxes_filt = boxes.clone()
    filt_mask = logits_filt.max(dim=1)[0] > box_threshold
    logits_filt = logits_filt[filt_mask]  # num_filt, 256
    boxes_filt = boxes_filt[filt_mask]  # num_filt, 4
    logits_filt.shape[0]

    # get phrase
    tokenlizer = model.tokenizer
    tokenized = tokenlizer(caption)
    # build pred
    pred_phrases = []
    for logit, box in zip(logits_filt, boxes_filt):
        pred_phrase = get_phrases_from_posmap(logit > text_threshold, tokenized, tokenlizer)
        if with_logits:
            pred_phrases.append(pred_phrase + f"({str(logit.max().item())[:4]})")
        else:
            pred_phrases.append(pred_phrase)

    return boxes_filt, pred_phrases

def show_mask(mask, ax, random_color=False):
    if random_color:
        color = np.concatenate([np.random.random(3), np.array([0.6])], axis=0)
    else:
        color = np.array([30/255, 144/255, 255/255, 0.6])
    h, w = mask.shape[-2:]
    mask_image = mask.reshape(h, w, 1) * color.reshape(1, 1, -1)
    ax.imshow(mask_image)

def show_box(box, ax, label):
    x0, y0 = box[0], box[1]
    w, h = box[2] - box[0], box[3] - box[1]
    ax.add_patch(plt.Rectangle((x0, y0), w, h, edgecolor='green', facecolor=(0,0,0,0), lw=2)) 
    ax.text(x0, y0, label)

def xywh_to_xyxy(box, sizeW, sizeH):
    if isinstance(box, list):
        box = torch.Tensor(box)
    box = box * torch.Tensor([sizeW, sizeH, sizeW, sizeH])
    box[:2] -= box[2:] / 2
    box[2:] += box[:2]
    box = box.numpy()
    return box

def mask_extend(img, box, extend_pixels=10, useRectangle=True):
    box[0] = int(box[0])
    box[1] = int(box[1])
    box[2] = int(box[2])
    box[3] = int(box[3])
    region = img.crop(tuple(box))
    new_width = box[2] - box[0] + 2*extend_pixels
    new_height = box[3] - box[1] + 2*extend_pixels

    region_BILINEAR = region.resize((int(new_width), int(new_height)))
    if useRectangle:
        region_draw = ImageDraw.Draw(region_BILINEAR)
        region_draw.rectangle((0, 0, new_width, new_height), fill=(255, 255, 255))    
    img.paste(region_BILINEAR, (int(box[0]-extend_pixels), int(box[1]-extend_pixels)))
    return img

def mix_masks(imgs):
    re_img =  1 - np.asarray(imgs[0].convert("1"))
    for i in range(len(imgs)-1):
        re_img = np.multiply(re_img, 1 - np.asarray(imgs[i+1].convert("1")))
    re_img =  1 - re_img
    return  Image.fromarray(np.uint8(255*re_img))

config_file = 'GroundingDINO/groundingdino/config/GroundingDINO_SwinT_OGC.py'
ckpt_repo_id = "ShilongLiu/GroundingDINO"
ckpt_filenmae = "groundingdino_swint_ogc.pth"
sam_checkpoint = './sam_vit_h_4b8939.pth' 
output_dir = "outputs"
device = evice = 'cuda' if torch.cuda.is_available() else 'cpu'

print(f'device={device}')

# make dir
os.makedirs(output_dir, exist_ok=True)

# initialize groundingdino model
logger.info(f"initialize groundingdino model...")
groundingdino_model = load_model_hf(config_file, ckpt_repo_id, ckpt_filenmae)

# initialize SAM
logger.info(f"initialize SAM model...")
sam_model = build_sam(checkpoint=sam_checkpoint).to(device)
sam_predictor = SamPredictor(sam_model)
sam_mask_generator = SamAutomaticMaskGenerator(sam_model)

# initialize stable-diffusion-inpainting
logger.info(f"initialize stable-diffusion-inpainting...")
sd_pipe = None
if os.environ.get('IS_MY_DEBUG') is None:
    sd_pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", 
            torch_dtype=torch.float16
    )
    sd_pipe = sd_pipe.to(device)

# initialize lama_cleaner
logger.info(f"initialize lama_cleaner...")
from lama_cleaner.helper import (
    load_img,
    numpy_to_bytes,
    resize_max_size,
)

lama_cleaner_model = ModelManager(
        name='lama',
        device='cpu', # device,
    )

def lama_cleaner_process(image, mask):
    ori_image = image
    if mask.shape[0] == image.shape[1] and mask.shape[1] == image.shape[0] and mask.shape[0] != mask.shape[1]:
        # rotate image
        ori_image = np.transpose(image[::-1, ...][:, ::-1], axes=(1, 0, 2))[::-1, ...]
        image = ori_image
    
    original_shape = ori_image.shape
    interpolation = cv2.INTER_CUBIC
    
    size_limit = 1080
    if size_limit == "Original":
        size_limit = max(image.shape)
    else:
        size_limit = int(size_limit)

    config = Config(
        ldm_steps=25,
        ldm_sampler='plms',
        zits_wireframe=True,
        hd_strategy='Original',
        hd_strategy_crop_margin=196,
        hd_strategy_crop_trigger_size=1280,
        hd_strategy_resize_limit=2048,
        prompt='',
        use_croper=False,
        croper_x=0,
        croper_y=0,
        croper_height=512,
        croper_width=512,
        sd_mask_blur=5,
        sd_strength=0.75,
        sd_steps=50,
        sd_guidance_scale=7.5,
        sd_sampler='ddim',
        sd_seed=42,
        cv2_flag='INPAINT_NS',
        cv2_radius=5,
    )
    
    if config.sd_seed == -1:
        config.sd_seed = random.randint(1, 999999999)

    # logger.info(f"Origin image shape_0_: {original_shape} / {size_limit}")
    image = resize_max_size(image, size_limit=size_limit, interpolation=interpolation)
    # logger.info(f"Resized image shape_1_: {image.shape}")
    
    # logger.info(f"mask image shape_0_: {mask.shape} / {type(mask)}")
    mask = resize_max_size(mask, size_limit=size_limit, interpolation=interpolation)
    # logger.info(f"mask image shape_1_: {mask.shape} / {type(mask)}")

    res_np_img = lama_cleaner_model(image, mask, config)
    torch.cuda.empty_cache()
  
    image = Image.open(io.BytesIO(numpy_to_bytes(res_np_img, 'png')))
    return  image

# relate anything
from ram_utils import iou, sort_and_deduplicate, relation_classes, MLP, show_anns, show_mask
from ram_train_eval import RamModel,RamPredictor
from mmengine.config import Config
input_size = 512
hidden_size = 256
num_classes = 56

# load ram model
model_path = "./checkpoints/ram_epoch12.pth"
config = dict(
    model=dict(
        pretrained_model_name_or_path='bert-base-uncased',
        load_pretrained_weights=False,
        num_transformer_layer=2,
        input_feature_size=256,
        output_feature_size=768,
        cls_feature_size=512,
        num_relation_classes=56,
        pred_type='attention',
        loss_type='multi_label_ce',
    ),
    load_from=model_path,
)
config = Config(config)

class Ram_Predictor(RamPredictor, device='cpu'):
    def __init__(self,config):
        self.config = config
        self.device = torch.device(device)
        self._build_model()

    def _build_model(self):
        self.model = RamModel(**self.config.model).to(self.device)
        if self.config.load_from is not None:
            self.model.load_state_dict(torch.load(self.config.load_from, map_location=self.device))
        self.model.train()
ram_model = Ram_Predictor(config, device)

# visualization
def draw_selected_mask(mask, draw):
    color = (255, 0, 0, 153)
    nonzero_coords = np.transpose(np.nonzero(mask))
    for coord in nonzero_coords:
        draw.point(coord[::-1], fill=color)

def draw_object_mask(mask, draw):
    color = (0, 0, 255, 153)
    nonzero_coords = np.transpose(np.nonzero(mask))
    for coord in nonzero_coords:
        draw.point(coord[::-1], fill=color)

def create_title_image(word1, word2, word3, width, font_path='./assets/OpenSans-Bold.ttf'):
    # Define the colors to use for each word
    color_red = (255, 0, 0)
    color_black = (0, 0, 0)
    color_blue = (0, 0, 255)

    # Define the initial font size and spacing between words
    font_size = 40

    # Create a new image with the specified width and white background
    image = Image.new('RGB', (width, 60), (255, 255, 255))

    # Load the specified font
    font = ImageFont.truetype(font_path, font_size)

    # Keep increasing the font size until all words fit within the desired width
    while True:
        # Create a draw object for the image
        draw = ImageDraw.Draw(image)
        
        word_spacing = font_size / 2
        # Draw each word in the appropriate color
        x_offset = word_spacing
        draw.text((x_offset, 0), word1, color_red, font=font)
        x_offset += font.getsize(word1)[0] + word_spacing
        draw.text((x_offset, 0), word2, color_black, font=font)
        x_offset += font.getsize(word2)[0] + word_spacing
        draw.text((x_offset, 0), word3, color_blue, font=font)
        
        word_sizes = [font.getsize(word) for word in [word1, word2, word3]]
        total_width = sum([size[0] for size in word_sizes]) + word_spacing * 3

        # Stop increasing font size if the image is within the desired width
        if total_width <= width:
            break
            
        # Increase font size and reset the draw object
        font_size -= 1
        image = Image.new('RGB', (width, 50), (255, 255, 255))
        font = ImageFont.truetype(font_path, font_size)
        draw = None

    return image

def concatenate_images_vertical(image1, image2):
    # Get the dimensions of the two images
    width1, height1 = image1.size
    width2, height2 = image2.size

    # Create a new image with the combined height and the maximum width
    new_image = Image.new('RGBA', (max(width1, width2), height1 + height2))

    # Paste the first image at the top of the new image
    new_image.paste(image1, (0, 0))

    # Paste the second image below the first image
    new_image.paste(image2, (0, height1))

    return new_image

def relate_anything(input_image, k):
    w, h = input_image.size
    max_edge = 1500
    if w > max_edge or h > max_edge:
        ratio = max(w, h) / max_edge
        new_size = (int(w / ratio), int(h / ratio))
        input_image.thumbnail(new_size)
    
    # load image
    pil_image = input_image.convert('RGBA')
    image = np.array(input_image)
    sam_masks = sam_mask_generator.generate(image)
    filtered_masks = sort_and_deduplicate(sam_masks)

    feat_list = []
    for fm in filtered_masks:
        feat = torch.Tensor(fm['feat']).unsqueeze(0).unsqueeze(0).to(device)
        feat_list.append(feat)
    feat = torch.cat(feat_list, dim=1).to(device)
    matrix_output, rel_triplets = ram_model.predict(feat)

    pil_image_list = []
    for i, rel in enumerate(rel_triplets[:k]):
        s,o,r = int(rel[0]),int(rel[1]),int(rel[2])
        relation = relation_classes[r]

        mask_image = Image.new('RGBA', pil_image.size, color=(0, 0, 0, 0))
        mask_draw = ImageDraw.Draw(mask_image)
            
        draw_selected_mask(filtered_masks[s]['segmentation'], mask_draw)
        draw_object_mask(filtered_masks[o]['segmentation'], mask_draw)

        current_pil_image = pil_image.copy()
        current_pil_image.alpha_composite(mask_image)
        
        title_image = create_title_image('Red', relation, 'Blue', current_pil_image.size[0])
        concate_pil_image = concatenate_images_vertical(current_pil_image, title_image)
        pil_image_list.append(concate_pil_image)

    yield pil_image_list


mask_source_draw = "draw a mask on input image"
mask_source_segment = "type what to detect below"

def run_anything_task(input_image, text_prompt, task_type, inpaint_prompt, box_threshold, text_threshold, 
            iou_threshold, inpaint_mode, mask_source_radio, remove_mode, remove_mask_extend, num_relation):
    if task_type == "relate anything":
        return relate_anything(input_image['image'], num_relation)

    text_prompt = text_prompt.strip()
    if not ((task_type == 'inpainting' or task_type == 'remove') and mask_source_radio == mask_source_draw):
        if text_prompt == '':
            return [], gr.Gallery.update(label='Detection prompt is not found!πŸ˜‚πŸ˜‚πŸ˜‚πŸ˜‚')

    if input_image is None:
            return [], gr.Gallery.update(label='Please upload a image!πŸ˜‚πŸ˜‚πŸ˜‚πŸ˜‚')

    file_temp = int(time.time())
    logger.info(f'run_anything_task_[{file_temp}]_{task_type}/{inpaint_mode}/[{mask_source_radio}]/{remove_mode}/{remove_mask_extend}_[{text_prompt}]/[{inpaint_prompt}]___1_')

    # load image
    input_mask_pil = input_image['mask']
    input_mask = np.array(input_mask_pil.convert("L"))  
     
    image_pil, image = load_image(input_image['image'].convert("RGB"))
    
    # visualize raw image
    # image_pil.save(os.path.join(output_dir, f"raw_image_{file_temp}.jpg"))

    size = image_pil.size

    output_images = []
    # run grounding dino model
    if (task_type == 'inpainting' or task_type == 'remove') and mask_source_radio == mask_source_draw:
        pass
    else:
        groundingdino_device = 'cpu'
        if device != 'cpu':
            try:
                from groundingdino import _C
                groundingdino_device = 'cuda:0'
            except:
                warnings.warn("Failed to load custom C++ ops. Running on CPU mode Only in groundingdino!")

        groundingdino_device = 'cpu'
        boxes_filt, pred_phrases = get_grounding_output(
            groundingdino_model, image, text_prompt, box_threshold, text_threshold, device=groundingdino_device
        )
        if boxes_filt.size(0) == 0:
            logger.info(f'run_anything_task_[{file_temp}]_{task_type}_[{text_prompt}]_1_[No objects detected, please try others.]_')
            return [], gr.Gallery.update(label='No objects detected, please try others.πŸ˜‚πŸ˜‚πŸ˜‚πŸ˜‚')
        boxes_filt_ori = copy.deepcopy(boxes_filt)

        pred_dict = {
            "boxes": boxes_filt,
            "size": [size[1], size[0]],  # H,W
            "labels": pred_phrases,
        }
        image_with_box = plot_boxes_to_image(copy.deepcopy(image_pil), pred_dict)[0]
        image_path = os.path.join(output_dir, f"grounding_dino_output_{file_temp}.jpg")
        image_with_box.save(image_path)
        detection_image_result = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB)
        os.remove(image_path)
        output_images.append(detection_image_result)

    logger.info(f'run_anything_task_[{file_temp}]_{task_type}_2_')
    if task_type == 'segment' or ((task_type == 'inpainting' or task_type == 'remove') and mask_source_radio == mask_source_segment):
        image = np.array(input_image['image'])
        sam_predictor.set_image(image)

        H, W = size[1], size[0]
        for i in range(boxes_filt.size(0)):
            boxes_filt[i] = boxes_filt[i] * torch.Tensor([W, H, W, H])
            boxes_filt[i][:2] -= boxes_filt[i][2:] / 2
            boxes_filt[i][2:] += boxes_filt[i][:2]

        boxes_filt = boxes_filt.cpu()
        transformed_boxes = sam_predictor.transform.apply_boxes_torch(boxes_filt, image.shape[:2])

        masks, _, _ = sam_predictor.predict_torch(
            point_coords = None,
            point_labels = None,
            boxes = transformed_boxes,
            multimask_output = False,
        )
        # masks: [9, 1, 512, 512]
        assert sam_checkpoint, 'sam_checkpoint is not found!'
        # draw output image
        plt.figure(figsize=(10, 10))
        plt.imshow(image)
        for mask in masks:
            show_mask(mask.cpu().numpy(), plt.gca(), random_color=True)
        for box, label in zip(boxes_filt, pred_phrases):
            show_box(box.numpy(), plt.gca(), label)
        plt.axis('off')
        image_path = os.path.join(output_dir, f"grounding_seg_output_{file_temp}.jpg")
        plt.savefig(image_path, bbox_inches="tight")
        segment_image_result = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB)
        os.remove(image_path)
        output_images.append(segment_image_result)        

    logger.info(f'run_anything_task_[{file_temp}]_{task_type}_3_')
    if task_type == 'detection' or task_type == 'segment':
        logger.info(f'run_anything_task_[{file_temp}]_{task_type}_9_')
        return output_images, gr.Gallery.update(label='result images')
    elif task_type == 'inpainting' or task_type == 'remove':
        if inpaint_prompt.strip() == '' and mask_source_radio == mask_source_segment:
            task_type = 'remove'

        logger.info(f'run_anything_task_[{file_temp}]_{task_type}_4_')  
        if mask_source_radio == mask_source_draw:
            mask_pil = input_mask_pil
            mask = input_mask          
        else:
            masks_ori = copy.deepcopy(masks)
            if inpaint_mode == 'merge':
                masks = torch.sum(masks, dim=0).unsqueeze(0)
                masks = torch.where(masks > 0, True, False)
            mask = masks[0][0].cpu().numpy()
            mask_pil = Image.fromarray(mask)   

        image_path = os.path.join(output_dir, f"image_mask_{file_temp}.jpg")
        # if reverse_mask:
        #     mask_pil = mask_pil.point(lambda _: 255-_)
        mask_pil.convert("RGB").save(image_path)
        image_result = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB)
        os.remove(image_path)
        output_images.append(image_result)

        if task_type == 'inpainting':
            # inpainting pipeline
            image_source_for_inpaint = image_pil.resize((512, 512))
            image_mask_for_inpaint = mask_pil.resize((512, 512))
            image_inpainting = sd_pipe(prompt=inpaint_prompt, image=image_source_for_inpaint, mask_image=image_mask_for_inpaint).images[0]
        else:
            # remove from mask
            if mask_source_radio == mask_source_segment:
                mask_imgs = []
                masks_shape = masks_ori.shape        
                boxes_filt_ori_array = boxes_filt_ori.numpy()
                if inpaint_mode == 'merge':
                    extend_shape_0 = masks_shape[0]
                    extend_shape_1 = masks_shape[1]
                else:
                    extend_shape_0 = 1
                    extend_shape_1 = 1
                for i in range(extend_shape_0):
                    for j in range(extend_shape_1):                
                        mask = masks_ori[i][j].cpu().numpy()
                        mask_pil = Image.fromarray(mask)
                    
                        if remove_mode == 'segment':
                            useRectangle = False
                        else:
                            useRectangle = True

                        try:
                            remove_mask_extend = int(remove_mask_extend)
                        except:
                            remove_mask_extend = 10
                        mask_pil_exp = mask_extend(copy.deepcopy(mask_pil).convert("RGB"), 
                                        xywh_to_xyxy(torch.tensor(boxes_filt_ori_array[i]), size[0], size[1]),
                                        extend_pixels=remove_mask_extend, useRectangle=useRectangle)
                        mask_imgs.append(mask_pil_exp)
                mask_pil = mix_masks(mask_imgs)

                image_path = os.path.join(output_dir, f"image_mask_{file_temp}.jpg")
                # if reverse_mask:
                #     mask_pil = mask_pil.point(lambda _: 255-_)
                mask_pil.convert("RGB").save(image_path)
                image_result = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB)
                os.remove(image_path)
                output_images.append(image_result)               
            image_inpainting = lama_cleaner_process(np.array(image_pil), np.array(mask_pil.convert("L")))

        image_inpainting = image_inpainting.resize((image_pil.size[0], image_pil.size[1]))

        image_path = os.path.join(output_dir, f"grounded_sam_inpainting_output_{file_temp}.jpg")
        image_inpainting.save(image_path)
        image_result = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB)
        os.remove(image_path)
        logger.info(f'run_anything_task_[{file_temp}]_{task_type}_9_')
        output_images.append(image_result)
        return output_images, gr.Gallery.update(label='result images')        
    else:
        logger.info(f"task_type:{task_type} error!")
    logger.info(f'run_anything_task_[{file_temp}]_9_9_')
    return output_images, gr.Gallery.update(label='result images')

def change_radio_display(task_type, mask_source_radio, num_relation): #, gsa_gallery, ram_gallery):
    text_prompt_visible = True
    inpaint_prompt_visible = False
    mask_source_radio_visible = False
    num_relation_visible = False
    # gsa_gallery_visible = True
    # ram_gallery_visible = False
    if task_type == "inpainting":
        inpaint_prompt_visible = True
    if task_type == "inpainting" or task_type == "remove":
        mask_source_radio_visible = True   
        if mask_source_radio == mask_source_draw:
            text_prompt_visible = False
    if task_type == "relate anything":
        text_prompt_visible = False
        num_relation_visible = True
        # gsa_gallery_visible = False
        # ram_gallery_visible = True
    return  gr.Textbox.update(visible=text_prompt_visible), gr.Textbox.update(visible=inpaint_prompt_visible), gr.Radio.update(visible=mask_source_radio_visible), gr.Slider.update(visible=num_relation_visible)
    # gr.Gallery.update(visible=gas_gallery_visible), gr.Gallery.update(visible=ram_gallery_visible)

if __name__ == "__main__":
    parser = argparse.ArgumentParser("Grounded SAM demo", add_help=True)
    parser.add_argument("--debug", action="store_true", help="using debug mode")
    parser.add_argument("--share", action="store_true", help="share the app")
    args = parser.parse_args()

    print(f'args = {args}')

    block = gr.Blocks().queue()
    with block:
        with gr.Row():
            with gr.Column():
                input_image = gr.Image(source='upload', elem_id="image_upload", tool='sketch', type='pil', label="Upload")    
                task_type = gr.Radio(["detection", "segment", "inpainting", "remove", "relate anything"],  value="detection", 
                                                label='Task type', visible=True) 
                mask_source_radio = gr.Radio([mask_source_draw, mask_source_segment], 
                                    value=mask_source_segment, label="Mask from",
                                    visible=False) 
                text_prompt = gr.Textbox(label="Detection Prompt[To detect multiple objects, seperating each name with '.', like this: cat . dog . chair ]", placeholder="Cannot be empty")                                                
                inpaint_prompt = gr.Textbox(label="Inpaint Prompt (if this is empty, then remove)", visible=False)
                num_relation = gr.Slider(label="How many relations do you want to see", minimum=1, maximum=20, value=5, step=1, visible=False)
                run_button = gr.Button(label="Run")
                with gr.Accordion("Advanced options", open=False) as advanced_options:
                    box_threshold = gr.Slider(
                        label="Box Threshold", minimum=0.0, maximum=1.0, value=0.3, step=0.001
                    )
                    text_threshold = gr.Slider(
                        label="Text Threshold", minimum=0.0, maximum=1.0, value=0.25, step=0.001
                    )
                    iou_threshold = gr.Slider(
                        label="IOU Threshold", minimum=0.0, maximum=1.0, value=0.8, step=0.001
                    )                    
                    inpaint_mode = gr.Radio(["merge", "first"], value="merge", label="inpaint_mode")
                    with gr.Row():
                        with gr.Column(scale=1):
                            remove_mode = gr.Radio(["segment", "rectangle"],  value="segment", label='remove mode') 
                        with gr.Column(scale=1):
                            remove_mask_extend = gr.Textbox(label="remove_mask_extend", value='10')

            with gr.Column():
                # gsa_gallery = gr.Gallery(
                #     label="result images", show_label=True, elem_id="gsa_gallery"
                # ).style(grid=[2], full_width=True, full_height=True)
                gallery = gr.Gallery(label="Your Result", show_label=True, elem_id="gallery").style(preview=True, columns=5, object_fit="scale-down")


        run_button.click(fn=run_anything_task, inputs=[
                        input_image, text_prompt, task_type, inpaint_prompt, box_threshold, text_threshold, iou_threshold, inpaint_mode, mask_source_radio, remove_mode, remove_mask_extend, num_relation], outputs=[gsa_gallery, gsa_gallery])
        task_type.change(fn=change_radio_display, inputs=[task_type, mask_source_radio], outputs=[text_prompt, inpaint_prompt, mask_source_radio, num_relation])
        mask_source_radio.change(fn=change_radio_display, inputs=[task_type, mask_source_radio], outputs=[text_prompt, inpaint_prompt, mask_source_radio, num_relation])

        DESCRIPTION = '### This demo from [Grounded-Segment-Anything](https://github.com/IDEA-Research/Grounded-Segment-Anything). Thanks for their excellent work.'
        DESCRIPTION += f'<p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings. <a href="https://huggingface.co/spaces/yizhangliu/Grounded-Segment-Anything?duplicate=true"><img style="display: inline; margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space" /></a></p>'
        gr.Markdown(DESCRIPTION)

    block.launch(server_name='0.0.0.0', debug=args.debug, share=args.share)