Spaces:
Sleeping
Sleeping
File size: 42,157 Bytes
0b86477 94ac4ef 6b32938 0b86477 0b51359 0b86477 94ac4ef d61084a 0b86477 94ac4ef 0ee959c 94ac4ef 0ee959c 94ac4ef 0ee959c 0b86477 0ee959c 3ba0b94 94ac4ef 3ba0b94 548ad6f 530a504 0ee959c 530a504 d61084a 0ee959c 530a504 0ee959c 530a504 81144e1 505d087 81144e1 530a504 0ee959c 0b86477 0b51359 33ea65f 0b51359 0b86477 6b32938 0b51359 33ea65f 0b51359 33ea65f 0b51359 33ea65f 0b51359 33ea65f 0b51359 33ea65f 0b51359 33ea65f 0b51359 33ea65f 0b51359 33ea65f 0b51359 33ea65f 0b51359 33ea65f 0b51359 33ea65f 0b51359 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b51359 6b32938 33ea65f 6b32938 33ea65f 6b32938 0b86477 6b32938 0b86477 6b32938 33ea65f 6b32938 d61084a a513848 d61084a 8cb9ee2 d61084a 2b2d916 d61084a 2b2d916 d61084a 17a1674 d61084a 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 f61a56b 6b32938 f61a56b 6b32938 f61a56b 6b32938 f61a56b 6b32938 f61a56b 6b32938 f61a56b 6b32938 f61a56b 6b32938 f61a56b 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 f61a56b 6b32938 f61a56b 6b32938 f61a56b 6b32938 f61a56b 6b32938 f61a56b 6b32938 f61a56b 6b32938 f61a56b 6b32938 f61a56b d61084a 0b51359 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 6b32938 0b86477 daadbf0 6b32938 0b86477 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 |
"""
HuggingFace Space for SAM / MedSAM Inference
API-compatible with Dense-Captioning-Toolkit backend
Deploy this to: https://huggingface.co/spaces/YOUR_USERNAME/medsam-inference
"""
import gradio as gr
import torch
import numpy as np
from PIL import Image
import io
import json
import base64
import os
import uuid
from huggingface_hub import hf_hub_download
# Import SAM components
from segment_anything import sam_model_registry, SamPredictor, SamAutomaticMaskGenerator
# Initialize model
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")
# -----------------------------------------------------------------------------
# Model configuration
# -----------------------------------------------------------------------------
# 1) MedSAM (ViT-B) for interactive segmentation (points / boxes / multiple boxes)
# We assume medsam_vit_b.pth is committed in this repo (small enough for Spaces).
MEDSAM_CHECKPOINT = os.path.join(os.path.dirname(__file__), "medsam_vit_b.pth")
print("Loading MedSAM model (vit_b) for interactive segmentation...")
try:
# MedSAM checkpoints are typically state_dicts; load and apply to a vit_b SAM backbone.
state_dict = torch.load(MEDSAM_CHECKPOINT, map_location=device)
medsam = sam_model_registry["vit_b"](checkpoint=None)
medsam.load_state_dict(state_dict)
medsam.to(device=device)
medsam.eval()
print("✓ MedSAM model (vit_b) loaded successfully")
except Exception as e:
print(f"✗ Failed to load MedSAM model from {MEDSAM_CHECKPOINT}: {e}")
raise
# SamPredictor for interactive segmentation (point/box prompts) using MedSAM
predictor = SamPredictor(medsam)
print("✓ SamPredictor (MedSAM) initialized for interactive segmentation")
# 2) SAM ViT-H for automatic mask generation and embedding (encode_image)
# We download this large checkpoint from a separate model repo using hf_hub_download.
MODEL_REPO_ID = "Aniketg6/dense-captioning-models"
MODEL_FILENAME = "sam_vit_h_4b8939.pth" # change if your filename is different
MODEL_TYPE = "vit_h" # using SAM ViT-H (general-purpose SAM)
print(f"Downloading SAM (vit_h) checkpoint `{MODEL_FILENAME}` from repo `{MODEL_REPO_ID}`...")
SAM_CHECKPOINT = hf_hub_download(
repo_id=MODEL_REPO_ID,
filename=MODEL_FILENAME,
)
print(f"✓ SAM (vit_h) checkpoint downloaded to: {SAM_CHECKPOINT}")
print("Loading SAM model (vit_h) for auto masks and embeddings...")
# Monkey-patch torch.load to use CPU mapping when needed
original_torch_load = torch.load
def patched_torch_load(f, *args, **kwargs):
if "map_location" not in kwargs and device == "cpu":
kwargs["map_location"] = "cpu"
return original_torch_load(f, *args, **kwargs)
torch.load = patched_torch_load
try:
# Ensure we always load onto CPU when no GPU is available
torch.load = patched_torch_load
sam = sam_model_registry[MODEL_TYPE](checkpoint=SAM_CHECKPOINT)
finally:
torch.load = original_torch_load
sam.to(device=device)
sam.eval()
print("✓ SAM model (vit_h) loaded successfully")
# SamAutomaticMaskGenerator for automatic mask generation (SAM ViT-H)
mask_generator = SamAutomaticMaskGenerator(
model=sam,
points_per_side=16, # Lighter grid (16x16) for faster CPU + smaller responses
pred_iou_thresh=0.7, # IoU threshold for filtering
stability_score_thresh=0.7, # Stability threshold
crop_n_layers=0, # Disable multi-scale crops to avoid IndexError
crop_n_points_downscale_factor=2,
min_mask_region_area=0 # Allow small masks (backend can filter if needed)
)
print("✓ SamAutomaticMaskGenerator (SAM vit-h) initialized for automatic segmentation")
# =============================================================================
# HELPER FUNCTIONS FOR EMBEDDINGS (STATELESS)
# =============================================================================
def set_predictor_features_from_embedding(embedding_tensor: torch.Tensor, image_shape: tuple):
"""
Set SamPredictor's internal features using precomputed embedding
Args:
embedding_tensor: Precomputed embedding tensor [1, C, H, W]
image_shape: Original image shape (height, width)
"""
# SamPredictor stores features in self.features
# We need to set it directly (this is a bit of a hack but necessary)
predictor.features = embedding_tensor
predictor.original_image_size = image_shape
predictor.input_size = (1024, 1024) # SAM default input size
predictor.is_image_set = True
# =============================================================================
# API FUNCTIONS - MATCHING BACKEND FORMAT (backend/app.py)
# =============================================================================
def encode_image(image, request_json):
"""
Encode image using SAM image encoder and return embedding to the client.
This is now a stateless API: it does NOT talk to Supabase. The caller
(your backend) is responsible for storing the embedding if desired.
Args:
image: PIL Image
request_json: JSON string with optional fields:
{
"image_id": "uuid-string" # Optional: image ID from your DB
}
Returns:
JSON string:
{
"success": true/false,
"image_id": "uuid-string" or null,
"embedding_npy_base64": "...", # base64-encoded .npy of [C,H,W]
"embedding_shape": [1, C, H, W]
}
"""
try:
# Parse input (image_id is optional and just echoed back)
data = json.loads(request_json) if request_json else {}
image_id = data.get("image_id")
# Convert PIL to numpy
image_array = np.array(image)
H, W = image_array.shape[:2]
# Resize image to SAM's expected input size (1024x1024)
from skimage import transform
img_resized = transform.resize(
image_array,
(1024, 1024),
order=3,
preserve_range=True,
anti_aliasing=True,
).astype(np.uint8)
# Normalize image (SAM expects normalized input)
img_norm = (img_resized - img_resized.min()) / np.clip(
img_resized.max() - img_resized.min(), 1e-8, None
)
# Convert to tensor and add batch dimension
tensor = (
torch.tensor(img_norm)
.float()
.permute(2, 0, 1)
.unsqueeze(0)
.to(device)
)
# Encode image using SAM image encoder
print(f"Encoding image (image_id={image_id}) original size: {W}x{H} -> 1024x1024")
with torch.no_grad():
embedding = sam.image_encoder(tensor)
# Convert embedding to numpy [C, Hf, Wf]
arr = embedding.squeeze(0).cpu().numpy().astype(np.float32)
# Serialize as .npy in memory and base64-encode it
buf = io.BytesIO()
np.save(buf, arr)
buf.seek(0)
embedding_b64 = base64.b64encode(buf.read()).decode("utf-8")
return json.dumps(
{
"success": True,
"image_id": image_id,
"embedding_npy_base64": embedding_b64,
"embedding_shape": list(embedding.shape),
}
)
except Exception as e:
import traceback
return json.dumps(
{
"success": False,
"error": str(e),
"traceback": traceback.format_exc(),
}
)
def segment_points(image, request_json):
"""
Segment image with point prompts - MATCHES BACKEND /api/medsam/segment_points
Each point gets its own small segment (converted to small bounding box).
This matches the backend behavior where points are converted to small boxes.
Args:
image: PIL Image
request_json: JSON string with format:
{
"points": [[x1, y1], [x2, y2], ...],
"labels": [1, 0, ...] # 1=foreground, 0=background
}
Returns:
JSON string matching backend response format:
{
"success": true,
"masks": [{"mask": [[...]], "confidence": 0.95}, ...],
"confidences": [0.95, ...],
"method": "medsam_points_individual"
}
"""
try:
# Parse input
data = json.loads(request_json)
points = data.get("points", [])
labels = data.get("labels", [])
image_id = data.get("image_id") # Optional: if provided, use precomputed embedding
if not points:
return json.dumps({'success': False, 'error': 'At least one point is required'})
# Convert PIL to numpy
image_array = np.array(image)
H, W = image_array.shape[:2]
# For now, always compute embedding from image (stateless API)
predictor.set_image(image_array)
# Process each point individually (like backend does)
box_size = 20 # Small box size for point-based segmentation
masks_list = []
confidences_list = []
for i, pt in enumerate(points):
x, y = pt
# Create a small bounding box centered on the point (matching backend behavior)
x1 = max(0, x - box_size // 2)
y1 = max(0, y - box_size // 2)
x2 = min(W - 1, x + box_size // 2)
y2 = min(H - 1, y + box_size // 2)
bbox = np.array([x1, y1, x2, y2])
print(f"Processing point {i+1}/{len(points)}: ({x}, {y}) -> bbox: {bbox.tolist()}")
# Run prediction with box
masks, scores, logits = predictor.predict(
point_coords=None,
point_labels=None,
box=bbox,
multimask_output=False
)
if len(masks) > 0:
# Take the best mask
best_idx = np.argmax(scores)
mask = masks[best_idx]
score = float(scores[best_idx])
masks_list.append({
'mask': mask.astype(np.uint8).tolist(),
'confidence': score
})
confidences_list.append(score)
print(f"Point {i+1} segmentation successful, confidence: {score:.4f}")
else:
print(f"Point {i+1} segmentation failed")
if masks_list:
result = {
'success': True,
'masks': masks_list,
'confidences': confidences_list,
'method': 'medsam_points_individual'
}
else:
result = {'success': False, 'error': 'All point segmentations failed'}
return json.dumps(result)
except Exception as e:
import traceback
return json.dumps({
'success': False,
'error': str(e),
'traceback': traceback.format_exc()
})
def segment_box(image, request_json):
"""
Segment image with a single bounding box - MATCHES BACKEND /api/medsam/segment_box
Args:
image: PIL Image
request_json: JSON string with format:
{
"bbox": [x1, y1, x2, y2] # Can be array or object with x1,y1,x2,y2
}
Returns:
JSON string matching backend response format:
{
"success": true,
"mask": [[...]],
"confidence": 0.95,
"method": "medsam_box"
}
"""
try:
# Parse input
data = json.loads(request_json)
bbox = data.get("bbox", [])
# Handle both array format [x1,y1,x2,y2] and object format {x1,y1,x2,y2}
if isinstance(bbox, dict):
bbox = [bbox.get('x1', 0), bbox.get('y1', 0), bbox.get('x2', 0), bbox.get('y2', 0)]
if not bbox or len(bbox) != 4:
return json.dumps({'success': False, 'error': 'Valid bounding box required [x1, y1, x2, y2]'})
box = np.array(bbox)
# Convert PIL to numpy
image_array = np.array(image)
# Stateless: always compute embedding from image
predictor.set_image(image_array)
# Run prediction with box
masks, scores, logits = predictor.predict(
point_coords=None,
point_labels=None,
box=box,
multimask_output=False
)
if len(masks) > 0:
best_idx = np.argmax(scores)
mask = masks[best_idx]
score = float(scores[best_idx])
result = {
'success': True,
'mask': mask.astype(np.uint8).tolist(),
'confidence': score,
'method': 'medsam_box'
}
else:
result = {'success': False, 'error': 'Segmentation failed'}
return json.dumps(result)
except Exception as e:
import traceback
return json.dumps({
'success': False,
'error': str(e),
'traceback': traceback.format_exc()
})
def segment_multiple_boxes(image, request_json):
"""
Segment image with multiple bounding boxes - MATCHES BACKEND /api/medsam/segment_multiple_boxes
This is the main API endpoint used by the frontend for box-based segmentation.
Args:
image: PIL Image
request_json: JSON string with format:
{
"bboxes": [
[x1, y1, x2, y2], # Array format
{"x1": 10, "y1": 20, "x2": 100, "y2": 200} # Object format (also supported)
]
}
Returns:
JSON string matching backend response format:
{
"success": true,
"masks": [{"mask": [[...]], "confidence": 0.95}, ...],
"confidences": [0.95, ...],
"method": "medsam_multiple_boxes"
}
"""
try:
# Parse input
data = json.loads(request_json)
bboxes = data.get("bboxes", [])
if not bboxes:
return json.dumps({'success': False, 'error': 'At least one bounding box is required'})
# Convert PIL to numpy
image_array = np.array(image)
# Stateless: always compute embedding from image
predictor.set_image(image_array)
print(f"Processing {len(bboxes)} boxes for segmentation")
masks_list = []
confidences_list = []
for i, bbox in enumerate(bboxes):
# Handle both array format [x1,y1,x2,y2] and object format {x1,y1,x2,y2}
if isinstance(bbox, dict):
box = np.array([
bbox.get('x1', 0),
bbox.get('y1', 0),
bbox.get('x2', 0),
bbox.get('y2', 0)
])
else:
box = np.array(bbox)
print(f"Processing box {i+1}/{len(bboxes)}: {box.tolist()}")
# Run prediction with box
masks, scores, logits = predictor.predict(
point_coords=None,
point_labels=None,
box=box,
multimask_output=False
)
if len(masks) > 0:
best_idx = np.argmax(scores)
mask = masks[best_idx]
score = float(scores[best_idx])
masks_list.append({
'mask': mask.astype(np.uint8).tolist(),
'confidence': score
})
confidences_list.append(score)
print(f"Box {i+1} segmentation successful, confidence: {score:.4f}")
else:
print(f"Box {i+1} segmentation failed")
if masks_list:
result = {
'success': True,
'masks': masks_list,
'confidences': confidences_list,
'method': 'medsam_multiple_boxes'
}
else:
result = {'success': False, 'error': 'All segmentations failed'}
return json.dumps(result)
except Exception as e:
import traceback
return json.dumps({
'success': False,
'error': str(e),
'traceback': traceback.format_exc()
})
# =============================================================================
# AUTO MASK GENERATION API (replaces local mask_generator.generate())
# =============================================================================
def generate_auto_masks(image, request_json):
"""
Automatically generate all masks for an image using SAM-H model.
This is equivalent to `mask_generator.generate(img_np)` in enhanced_preprocessing.py
Args:
image: PIL Image
request_json: JSON string with optional parameters:
{
"points_per_side": 32, # Grid density (default: 32)
"pred_iou_thresh": 0.88, # IoU threshold (default: 0.88)
"stability_score_thresh": 0.95, # Stability threshold (default: 0.95)
"min_mask_region_area": 0 # Minimum mask area (default: 0)
}
Returns:
JSON string with format matching SamAutomaticMaskGenerator output:
{
"success": true,
"masks": [
{
"segmentation": [[...2D boolean array...]],
"area": 12345,
"bbox": [x, y, width, height],
"predicted_iou": 0.95,
"point_coords": [[x, y]],
"stability_score": 0.98,
"crop_box": [x, y, width, height]
},
...
],
"num_masks": 42,
"image_size": [height, width]
}
"""
try:
if mask_generator is None:
return json.dumps({
'success': False,
'error': 'MedSAM model not loaded. Please ensure medsam_vit_b.pth is available.',
'available': False
})
# Parse optional parameters
params = {}
if request_json:
try:
params = json.loads(request_json) if request_json.strip() else {}
except:
params = {}
# Convert PIL to numpy
image_array = np.array(image)
H, W = image_array.shape[:2]
# Optional downscaling to keep masks smaller / faster
resize_longest = int(params.get("resize_longest", 0) or 0)
if resize_longest > 0 and max(H, W) > resize_longest:
scale = resize_longest / float(max(H, W))
new_w = max(1, int(W * scale))
new_h = max(1, int(H * scale))
print(f"Resizing image from {W}x{H} to {new_w}x{new_h} for auto masks...")
image_array = np.array(Image.fromarray(image_array).resize((new_w, new_h)))
H, W = image_array.shape[:2]
print(f"Generating automatic masks for image of size {W}x{H}...")
# Generate masks using SAM automatic mask generator
masks = mask_generator.generate(image_array)
print(f"Generated {len(masks)} masks")
if len(masks) > 0:
# Log some stats about the masks
areas = [m['area'] for m in masks]
ious = [m['predicted_iou'] for m in masks]
stabilities = [m['stability_score'] for m in masks]
print(f" Area range: {min(areas)} - {max(areas)} pixels")
print(f" IoU range: {min(ious):.3f} - {max(ious):.3f}")
print(f" Stability range: {min(stabilities):.3f} - {max(stabilities):.3f}")
else:
print(" WARNING: No masks generated! This could mean:")
print(" - Image is too uniform/simple")
print(" - Thresholds are still too strict")
print(" - Image size is too small or too large")
# Optionally limit number of masks returned to keep JSON payload reasonable
max_masks = int(params.get("max_masks", 10))
if max_masks > 0 and len(masks) > max_masks:
# Sort by predicted IoU (descending) and keep top-K
print(f"Limiting masks from {len(masks)} to top {max_masks} by predicted_iou")
masks = sorted(
masks,
key=lambda m: float(m.get("predicted_iou", 0.0)),
reverse=True,
)[:max_masks]
print(f"Preparing {len(masks)} masks to return to client...")
# Convert masks to JSON-serializable format
masks_output = []
for m in masks:
mask_data = {
"segmentation": m["segmentation"].astype(np.uint8).tolist(),
"area": int(m["area"]),
"bbox": [int(x) for x in m["bbox"]], # [x, y, width, height]
"predicted_iou": float(m["predicted_iou"]),
"point_coords": [
[int(p[0]), int(p[1])] for p in m["point_coords"]
]
if m["point_coords"] is not None
else [],
"stability_score": float(m["stability_score"]),
"crop_box": [int(x) for x in m["crop_box"]], # [x, y, width, height]
}
masks_output.append(mask_data)
result = {
'success': True,
'masks': masks_output,
'num_masks': len(masks_output),
'image_size': [H, W]
}
print(f"Auto mask generation complete: {len(masks_output)} masks")
return json.dumps(result)
except Exception as e:
import traceback
return json.dumps({
'success': False,
'error': str(e),
'traceback': traceback.format_exc()
})
def check_auto_mask_status():
"""
Check if automatic mask generation is available
"""
return json.dumps({
'available': mask_generator is not None,
'model': MODEL_FILENAME if mask_generator else None,
'model_type': MODEL_TYPE,
'device': str(device)
})
# =============================================================================
# LEGACY API FUNCTIONS (kept for backwards compatibility with test scripts)
# =============================================================================
def segment_with_points_legacy(image, points_json):
"""
Legacy API - Segment with point prompts using true point-based segmentation
Args:
points_json: JSON string with format:
{
"coords": [[x1, y1], [x2, y2], ...],
"labels": [1, 0, ...],
"multimask_output": true/false
}
"""
try:
points_data = json.loads(points_json)
coords = np.array(points_data["coords"])
labels = np.array(points_data["labels"])
multimask_output = points_data.get("multimask_output", True)
image_array = np.array(image)
predictor.set_image(image_array)
masks, scores, logits = predictor.predict(
point_coords=coords,
point_labels=labels,
multimask_output=multimask_output
)
masks_list = []
scores_list = []
for i, (mask, score) in enumerate(zip(masks, scores)):
mask_uint8 = (mask * 255).astype(np.uint8)
mask_image = Image.fromarray(mask_uint8)
buffer = io.BytesIO()
mask_image.save(buffer, format='PNG')
mask_base64 = base64.b64encode(buffer.getvalue()).decode('utf-8')
masks_list.append({
'mask_base64': mask_base64,
'mask_shape': mask.shape,
'mask_data': mask.tolist()
})
scores_list.append(float(score))
return json.dumps({
'success': True,
'masks': masks_list,
'scores': scores_list,
'num_masks': len(masks_list)
})
except Exception as e:
return json.dumps({'success': False, 'error': str(e)})
def segment_with_box_legacy(image, box_json):
"""
Legacy API - Segment with box prompt
Args:
box_json: JSON string with format:
{"box": [x1, y1, x2, y2], "multimask_output": false}
"""
try:
box_data = json.loads(box_json)
box = np.array(box_data["box"])
multimask_output = box_data.get("multimask_output", False)
image_array = np.array(image)
predictor.set_image(image_array)
masks, scores, logits = predictor.predict(
point_coords=None,
point_labels=None,
box=box,
multimask_output=multimask_output
)
masks_list = []
scores_list = []
for i, (mask, score) in enumerate(zip(masks, scores)):
mask_uint8 = (mask * 255).astype(np.uint8)
mask_image = Image.fromarray(mask_uint8)
buffer = io.BytesIO()
mask_image.save(buffer, format='PNG')
mask_base64 = base64.b64encode(buffer.getvalue()).decode('utf-8')
masks_list.append({
'mask_base64': mask_base64,
'mask_shape': mask.shape,
'mask_data': mask.tolist()
})
scores_list.append(float(score))
return json.dumps({
'success': True,
'masks': masks_list,
'scores': scores_list,
'num_masks': len(masks_list),
'box': box.tolist()
})
except Exception as e:
import traceback
return json.dumps({
'success': False,
'error': str(e),
'traceback': traceback.format_exc()
})
def segment_simple(image, x, y, label=1, multimask=True):
"""Simple single-point segmentation for Gradio UI"""
try:
points_json = json.dumps({
"coords": [[int(x), int(y)]],
"labels": [int(label)],
"multimask_output": multimask
})
result_json = segment_with_points_legacy(image, points_json)
result = json.loads(result_json)
if not result['success']:
return None, f"Error: {result['error']}"
best_idx = np.argmax(result['scores'])
best_mask_base64 = result['masks'][best_idx]['mask_base64']
best_score = result['scores'][best_idx]
mask_bytes = base64.b64decode(best_mask_base64)
mask_image = Image.open(io.BytesIO(mask_bytes))
return mask_image, f"Score: {best_score:.4f}"
except Exception as e:
return None, f"Error: {str(e)}"
# =============================================================================
# GRADIO INTERFACE
# =============================================================================
with gr.Blocks(title="MedSAM Inference API") as demo:
gr.Markdown("# 🏥 MedSAM Inference API")
gr.Markdown("Point and box-based segmentation using Fine-Tuned MedSAM")
gr.Markdown("**API-compatible with Dense-Captioning-Toolkit backend**")
with gr.Tabs():
# Tab 1: Backend-Compatible API (Points)
with gr.Tab("Segment Points (Backend API)"):
gr.Markdown("""
## Point-based Segmentation - Backend Compatible
**Matches `/api/medsam/segment_points`**
Each point is converted to a small bounding box for segmentation.
**Input Format:**
```json
{
"points": [[x1, y1], [x2, y2], ...],
"labels": [1, 0, ...]
}
```
**Output Format (matches backend):**
```json
{
"success": true,
"masks": [{"mask": [[...]], "confidence": 0.95}, ...],
"confidences": [0.95, ...],
"method": "medsam_points_individual"
}
```
""")
with gr.Row():
with gr.Column():
points_image = gr.Image(type="pil", label="Input Image")
points_json_input = gr.Textbox(
label="Request JSON",
placeholder='{"points": [[100, 150], [200, 200]], "labels": [1, 1]}',
lines=3
)
points_button = gr.Button("Segment Points", variant="primary")
with gr.Column():
points_output = gr.Textbox(label="Result JSON", lines=15)
points_button.click(
fn=segment_points,
inputs=[points_image, points_json_input],
outputs=points_output,
api_name="segment_points"
)
# Tab 2: Backend-Compatible API (Multiple Boxes)
with gr.Tab("Segment Multiple Boxes (Backend API)"):
gr.Markdown("""
## Multiple Box Segmentation - Backend Compatible
**Matches `/api/medsam/segment_multiple_boxes`** (main frontend API)
**Input Format:**
```json
{
"bboxes": [
[x1, y1, x2, y2],
{"x1": 10, "y1": 20, "x2": 100, "y2": 200}
]
}
```
**Output Format (matches backend):**
```json
{
"success": true,
"masks": [{"mask": [[...]], "confidence": 0.95}, ...],
"confidences": [0.95, ...],
"method": "medsam_multiple_boxes"
}
```
""")
with gr.Row():
with gr.Column():
multi_box_image = gr.Image(type="pil", label="Input Image")
multi_box_json = gr.Textbox(
label="Request JSON",
placeholder='{"bboxes": [[100, 100, 300, 300], [400, 400, 600, 600]]}',
lines=3
)
multi_box_button = gr.Button("Segment Multiple Boxes", variant="primary")
with gr.Column():
multi_box_output = gr.Textbox(label="Result JSON", lines=15)
multi_box_button.click(
fn=segment_multiple_boxes,
inputs=[multi_box_image, multi_box_json],
outputs=multi_box_output,
api_name="segment_multiple_boxes"
)
# Tab 3: Backend-Compatible API (Single Box)
with gr.Tab("Segment Box (Backend API)"):
gr.Markdown("""
## Single Box Segmentation - Backend Compatible
**Matches `/api/medsam/segment_box`**
**Input Format:**
```json
{
"bbox": [x1, y1, x2, y2]
}
```
**Output Format (matches backend):**
```json
{
"success": true,
"mask": [[...]],
"confidence": 0.95,
"method": "medsam_box"
}
```
""")
with gr.Row():
with gr.Column():
box_image = gr.Image(type="pil", label="Input Image")
box_json_input = gr.Textbox(
label="Request JSON",
placeholder='{"bbox": [100, 100, 300, 300]}',
lines=3
)
box_button = gr.Button("Segment Box", variant="primary")
with gr.Column():
box_output = gr.Textbox(label="Result JSON", lines=15)
box_button.click(
fn=segment_box,
inputs=[box_image, box_json_input],
outputs=box_output,
api_name="segment_box"
)
# Tab 4: Legacy API (for test scripts)
with gr.Tab("Legacy API"):
gr.Markdown("""
## Legacy API (for backwards compatibility)
Original API format with `coords`, `mask_data`, `scores`, etc.
Use if you have existing scripts using the old format.
""")
with gr.Row():
with gr.Column():
legacy_image = gr.Image(type="pil", label="Input Image")
legacy_points = gr.Textbox(
label="Points JSON (Legacy Format)",
placeholder='{"coords": [[100, 150]], "labels": [1], "multimask_output": true}',
lines=3
)
legacy_button = gr.Button("Run Segmentation (Legacy)", variant="secondary")
with gr.Column():
legacy_output = gr.Textbox(label="Result JSON", lines=15)
legacy_button.click(
fn=segment_with_points_legacy,
inputs=[legacy_image, legacy_points],
outputs=legacy_output,
api_name="segment_with_points" # Keep old API name for compatibility
)
gr.Markdown("---")
with gr.Row():
with gr.Column():
legacy_box_image = gr.Image(type="pil", label="Input Image")
legacy_box_json = gr.Textbox(
label="Box JSON (Legacy Format)",
placeholder='{"box": [100, 100, 300, 300], "multimask_output": false}',
lines=3
)
legacy_box_button = gr.Button("Run Box Segmentation (Legacy)", variant="secondary")
with gr.Column():
legacy_box_output = gr.Textbox(label="Result JSON", lines=15)
legacy_box_button.click(
fn=segment_with_box_legacy,
inputs=[legacy_box_image, legacy_box_json],
outputs=legacy_box_output,
api_name="segment_with_box" # Keep old API name for compatibility
)
# Tab 5: Auto Mask Generation (for preprocessing)
with gr.Tab("Auto Mask Generation"):
gr.Markdown("""
## Automatic Mask Generation (MedSAM)
**Replaces `mask_generator.generate(img_np)` in preprocessing pipeline**
Uses MedSAM (ViT-B) model with `SamAutomaticMaskGenerator` to automatically
segment all objects in an image. This is used for initial preprocessing
of scientific/medical images.
Uses the same `medsam_vit_b.pth` model as interactive segmentation.
**Output Format:**
```json
{
"success": true,
"masks": [
{
"segmentation": [[...2D array...]],
"area": 12345,
"bbox": [x, y, width, height],
"predicted_iou": 0.95,
"point_coords": [[x, y]],
"stability_score": 0.98,
"crop_box": [x, y, width, height]
}
],
"num_masks": 42
}
```
""")
with gr.Row():
with gr.Column():
auto_image = gr.Image(type="pil", label="Input Image")
auto_params = gr.Textbox(
label="Parameters (optional)",
placeholder='{"points_per_side": 32, "pred_iou_thresh": 0.88}',
lines=2
)
with gr.Row():
auto_button = gr.Button("Generate All Masks", variant="primary")
status_button = gr.Button("Check Status", variant="secondary")
with gr.Column():
auto_output = gr.Textbox(label="Result JSON", lines=20)
status_output = gr.Textbox(label="Status", lines=3)
auto_button.click(
fn=generate_auto_masks,
inputs=[auto_image, auto_params],
outputs=auto_output,
api_name="generate_auto_masks"
)
status_button.click(
fn=check_auto_mask_status,
inputs=[],
outputs=status_output,
api_name="check_auto_mask_status"
)
# Tab 6: Encode Image (for embedding storage)
with gr.Tab("Encode Image"):
gr.Markdown("""
## Image Encoding API
**Encodes image using SAM image encoder and saves embedding to Supabase**
This endpoint is used during preprocessing to compute and store image embeddings
once per image. Later segmentation calls can use these precomputed embeddings
for faster inference (no need to recompute embeddings on each API call).
**Input Format:**
```json
{
"image_id": "uuid-string" # Required: image ID from database
}
```
**Output Format:**
```json
{
"success": true,
"message": "Embedding saved successfully for image_id=...",
"image_id": "uuid-string",
"embedding_shape": [1, 256, 64, 64]
}
```
**Note:** Requires Supabase credentials (SUPABASE_URL and SUPABASE_KEY environment variables)
""")
with gr.Row():
with gr.Column():
encode_image_input = gr.Image(type="pil", label="Input Image")
encode_json_input = gr.Textbox(
label="Request JSON",
placeholder='{"image_id": "123e4567-e89b-12d3-a456-426614174000"}',
lines=2
)
encode_button = gr.Button("Encode Image", variant="primary")
with gr.Column():
encode_output = gr.Textbox(label="Result JSON", lines=10)
encode_button.click(
fn=encode_image,
inputs=[encode_image_input, encode_json_input],
outputs=encode_output,
api_name="encode_image"
)
# Tab 7: Simple UI Interface
with gr.Tab("Simple Interface"):
gr.Markdown("## Click-based Segmentation")
gr.Markdown("Enter X, Y coordinates to segment")
with gr.Row():
with gr.Column():
simple_image = gr.Image(type="pil", label="Input Image")
with gr.Row():
simple_x = gr.Number(label="X Coordinate", value=100)
simple_y = gr.Number(label="Y Coordinate", value=100)
with gr.Row():
simple_label = gr.Radio(
choices=[1, 0],
value=1,
label="Point Label (1=foreground, 0=background)"
)
simple_multimask = gr.Checkbox(
label="Multiple Masks",
value=True
)
simple_button = gr.Button("Segment", variant="primary")
with gr.Column():
simple_mask = gr.Image(label="Output Mask")
simple_info = gr.Textbox(label="Info")
simple_button.click(
fn=segment_simple,
inputs=[simple_image, simple_x, simple_y, simple_label, simple_multimask],
outputs=[simple_mask, simple_info]
)
gr.Markdown("""
---
### 📡 API Usage from Python (Backend-Compatible)
```python
from gradio_client import Client, handle_file
import json
client = Client("Aniketg6/medsam-inference")
# Point-based segmentation (matches backend format)
result = client.predict(
image=handle_file("image.jpg"),
request_json=json.dumps({
"points": [[150, 200], [300, 400]],
"labels": [1, 1]
}),
api_name="/segment_points"
)
# Multiple box segmentation (main frontend API)
result = client.predict(
image=handle_file("image.jpg"),
request_json=json.dumps({
"bboxes": [[100, 100, 300, 300], [400, 400, 600, 600]]
}),
api_name="/segment_multiple_boxes"
)
# Parse response
data = json.loads(result)
print(f"Success: {data['success']}")
print(f"Masks: {len(data['masks'])}")
print(f"Confidences: {data['confidences']}")
print(f"Method: {data['method']}")
```
""")
# Launch
if __name__ == "__main__":
demo.launch(
server_name="0.0.0.0",
server_port=7860,
share=False,
show_error=True
)
|