LuGot16 commited on
Commit
577e47e
·
verified ·
1 Parent(s): e3dc764

Upload inference_v1.0.0.py

Browse files
Files changed (1) hide show
  1. inference_v1.0.0.py +235 -0
inference_v1.0.0.py ADDED
@@ -0,0 +1,235 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ =================================================================
3
+ INFERENCE v1.0.0 — Full-image + SCSE attention + TTA
4
+ =================================================================
5
+ """
6
+
7
+ import os, cv2, torch, numpy as np, json, time, csv
8
+ import torch.nn as nn
9
+ import torch.nn.functional as F
10
+ import segmentation_models_pytorch as smp
11
+ from segmentation_models_pytorch.decoders.unet.decoder import UnetDecoder
12
+ from pathlib import Path
13
+
14
+ # ============================================================
15
+ # MODEL v1.0.0
16
+ # ============================================================
17
+ class TubuleSegModel(nn.Module):
18
+ def __init__(self, attention_type='scse'):
19
+ super().__init__()
20
+ self.encoder = smp.encoders.get_encoder(
21
+ 'timm-efficientnet-b4', in_channels=3, depth=5, weights=None
22
+ )
23
+ ec = self.encoder.out_channels
24
+ self.seg_decoder = UnetDecoder(
25
+ encoder_channels=ec, decoder_channels=(256,128,64,32,16),
26
+ n_blocks=5, use_norm='batchnorm', attention_type=attention_type
27
+ )
28
+ self.seg_head = nn.Conv2d(16, 3, kernel_size=1)
29
+ self.border_decoder = UnetDecoder(
30
+ encoder_channels=ec, decoder_channels=(256,128,64,32,16),
31
+ n_blocks=5, use_norm='batchnorm', attention_type=attention_type
32
+ )
33
+ self.border_head = nn.Conv2d(16, 2, kernel_size=1)
34
+
35
+ def forward(self, x):
36
+ features = self.encoder(x)
37
+ return (
38
+ self.seg_head(self.seg_decoder(features)),
39
+ self.border_head(self.border_decoder(features))
40
+ )
41
+
42
+
43
+ # ============================================================
44
+ # TTA — Test-Time Augmentation
45
+ # ============================================================
46
+ def predict_with_tta(model, img_t, device):
47
+ """
48
+ 8 pasadas: 4 rotaciones (0, 90, 180, 270) x 2 flips (original, horizontal).
49
+ Promedia probabilidades softmax en el espacio original.
50
+ img_t: [1, 3, H, W] tensor normalizado en device.
51
+ Retorna: [3, H, W] numpy array de probabilidades promediadas.
52
+ """
53
+ preds = []
54
+ with torch.no_grad():
55
+ for flip in [False, True]:
56
+ x = img_t.flip(-1) if flip else img_t
57
+ for k in range(4): # rotaciones 0, 90, 180, 270
58
+ x_rot = torch.rot90(x, k, dims=[-2, -1])
59
+ seg_logits, _ = model(x_rot)
60
+ prob = F.softmax(seg_logits, dim=1) # [1, 3, H, W]
61
+ # Revertir la rotacion
62
+ prob = torch.rot90(prob, -k, dims=[-2, -1])
63
+ # Revertir el flip
64
+ if flip:
65
+ prob = prob.flip(-1)
66
+ preds.append(prob.cpu().numpy()[0]) # [3, H, W]
67
+
68
+ return np.mean(preds, axis=0) # [3, H, W]
69
+
70
+
71
+ # ============================================================
72
+ # POST-PROCESSING (identico a inference_v22.py)
73
+ # ============================================================
74
+ def postprocess(pred, h, w):
75
+ epi_mask = (pred == 1).astype(np.uint8)
76
+ n, labels = cv2.connectedComponents(epi_mask)
77
+ if n > 1:
78
+ areas = [(lid, (labels==lid).sum()) for lid in range(1, n)]
79
+ areas.sort(key=lambda x: x[1], reverse=True)
80
+ epi_clean = (labels == areas[0][0]).astype(np.uint8)
81
+ else:
82
+ epi_clean = epi_mask
83
+
84
+ k_close = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (11, 11))
85
+ epi_closed = cv2.morphologyEx(epi_clean, cv2.MORPH_CLOSE, k_close)
86
+
87
+ flood = epi_closed.copy()
88
+ fm = np.zeros((h+2, w+2), dtype=np.uint8)
89
+ for y in range(0, h, 3):
90
+ if flood[y, 0] == 0: cv2.floodFill(flood, fm, (0, y), 2)
91
+ if flood[y, w-1] == 0: cv2.floodFill(flood, fm, (w-1, y), 2)
92
+ for x in range(0, w, 3):
93
+ if flood[0, x] == 0: cv2.floodFill(flood, fm, (x, 0), 2)
94
+ if flood[h-1, x] == 0: cv2.floodFill(flood, fm, (x, h-1), 2)
95
+ holes = (flood == 0).astype(np.uint8)
96
+ epi_filled = epi_closed | holes
97
+
98
+ final = np.zeros((h, w), dtype=np.uint8)
99
+ final[epi_filled > 0] = 1
100
+
101
+ lum_mask = (pred == 2).astype(np.uint8)
102
+ lum_inside = lum_mask & epi_filled
103
+ n, labels = cv2.connectedComponents(lum_inside)
104
+ if n > 1:
105
+ lum_clean = np.zeros_like(lum_inside)
106
+ for lid in range(1, n):
107
+ comp = (labels == lid).astype(np.uint8)
108
+ if comp.sum() < 300: continue
109
+ dilated = cv2.dilate(comp, cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5,5)))
110
+ border_px = dilated & (~epi_filled.astype(bool)).astype(np.uint8)
111
+ if border_px.sum() < comp.sum() * 0.05:
112
+ lum_clean[labels == lid] = 1
113
+ final[lum_clean > 0] = 2
114
+
115
+ return final
116
+
117
+
118
+ # ============================================================
119
+ # MAIN
120
+ # ============================================================
121
+ def main():
122
+ base_dir = Path(r"D:\Lu\AI\Tubules"); os.chdir(str(base_dir))
123
+ device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
124
+ print(f"Device: {device}")
125
+
126
+ model_file = 'best_model_segmentation_v1.0.0.pt'
127
+ attention = 'scse'
128
+
129
+ model = TubuleSegModel(attention_type=attention)
130
+ ckpt = torch.load(str(base_dir / model_file), map_location='cpu', weights_only=False)
131
+ model.load_state_dict(ckpt['model_state_dict'])
132
+ model.to(device); model.eval()
133
+
134
+ from huggingface_hub import snapshot_download
135
+ data_repo = base_dir / 'data_repo'
136
+ if not data_repo.exists():
137
+ snapshot_download('LuGot16/tubules', repo_type='dataset', local_dir=str(data_repo))
138
+
139
+ test_dir = data_repo / 'area_test'
140
+ out_dir = base_dir / f'inference_results_{model_file.replace("best_model_","").replace(".pt","")}'
141
+ out_dir.mkdir(exist_ok=True)
142
+
143
+ SCALE = 0.32 # um/pixel
144
+ results = []
145
+ files = sorted(test_dir.glob('*.tif'))
146
+ t0 = time.time()
147
+
148
+ for i, fp in enumerate(files):
149
+ img = cv2.imread(str(fp))
150
+ if img is None: continue
151
+ h, w = img.shape[:2]
152
+ img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
153
+
154
+ img_resized = cv2.resize(img_rgb, (512, 512), interpolation=cv2.INTER_LINEAR)
155
+ img_norm = (img_resized.astype(np.float32)/255.0 - [0.485,0.456,0.406]) / [0.229,0.224,0.225]
156
+ img_t = torch.from_numpy(img_norm.transpose(2,0,1)).float().unsqueeze(0).to(device)
157
+
158
+ # TTA
159
+ seg_probs = predict_with_tta(model, img_t, device) # [3, 512, 512]
160
+
161
+ pred_512 = seg_probs.argmax(0).astype(np.uint8)
162
+ pred = cv2.resize(pred_512, (w, h), interpolation=cv2.INTER_NEAREST)
163
+ pred = postprocess(pred, h, w)
164
+
165
+ total = h * w
166
+ epi_pct = (pred==1).sum() / total * 100
167
+ lum_pct = (pred==2).sum() / total * 100
168
+ tub_pct = epi_pct + lum_pct
169
+ epi_um2 = (pred==1).sum() * SCALE * SCALE
170
+ lum_um2 = (pred==2).sum() * SCALE * SCALE
171
+ tub_um2 = epi_um2 + lum_um2
172
+ status = "OK" if tub_pct >= 30 and lum_pct > 1 else "CHECK"
173
+
174
+ results.append({
175
+ 'image': fp.name,
176
+ 'tubule_pct': round(tub_pct, 1),
177
+ 'epithelium_pct': round(epi_pct, 1),
178
+ 'lumen_pct': round(lum_pct, 1),
179
+ 'tubule_um2': round(tub_um2, 1),
180
+ 'epithelium_um2': round(epi_um2, 1),
181
+ 'lumen_um2': round(lum_um2, 1),
182
+ 'lumen_epi_ratio': round(lum_um2 / (epi_um2 + 1e-6), 4),
183
+ 'status': status
184
+ })
185
+
186
+ # Overlay
187
+ overlay = img.copy()
188
+ overlay[pred==1] = (overlay[pred==1]*0.5 + np.array([0,180,0])*0.5).astype(np.uint8)
189
+ overlay[pred==2] = (overlay[pred==2]*0.5 + np.array([255,100,0])*0.5).astype(np.uint8)
190
+
191
+ tubule_mask = (pred >= 1).astype(np.uint8)
192
+ contours_outer, _ = cv2.findContours(tubule_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
193
+ cv2.drawContours(overlay, contours_outer, -1, (0, 255, 255), 2)
194
+
195
+ lum_mask_vis = (pred == 2).astype(np.uint8)
196
+ contours_lumen, _ = cv2.findContours(lum_mask_vis, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
197
+ cv2.drawContours(overlay, contours_lumen, -1, (0, 0, 255), 2)
198
+
199
+ cv2.imwrite(str(out_dir / f"{fp.stem}_overlay.png"), overlay)
200
+ cv2.imwrite(str(out_dir / f"{fp.stem}_mask.png"), pred * 127)
201
+
202
+ if (i+1) % 10 == 0 or status == "CHECK":
203
+ print(f" [{i+1}/{len(files)}] {fp.name}: "
204
+ f"tub={tub_pct:.1f}% epi={epi_pct:.1f}% lum={lum_pct:.1f}% [{status}]")
205
+
206
+ ok = sum(1 for r in results if r['status'] == 'OK')
207
+ tub_vals = [r['tubule_pct'] for r in results]
208
+ epi_vals = [r['epithelium_pct'] for r in results]
209
+ lum_vals = [r['lumen_pct'] for r in results]
210
+
211
+ print(f"\n{'='*60}")
212
+ print(f" RESULTS {model_file} + TTA — {ok}/{len(results)} OK")
213
+ print(f" Tubule: {np.mean(tub_vals):.1f} +/- {np.std(tub_vals):.1f}%")
214
+ print(f" Epithelium: {np.mean(epi_vals):.1f} +/- {np.std(epi_vals):.1f}%")
215
+ print(f" Lumen: {np.mean(lum_vals):.1f} +/- {np.std(lum_vals):.1f}%")
216
+ print(f" Time: {(time.time()-t0)/60:.1f} min")
217
+ print(f"{'='*60}")
218
+
219
+ with open(out_dir / 'results.csv', 'w', newline='') as f:
220
+ w = csv.DictWriter(f, fieldnames=results[0].keys())
221
+ w.writeheader(); w.writerows(results)
222
+ with open(out_dir / 'results.json', 'w') as f:
223
+ json.dump({
224
+ 'model': model_file, 'tta': True, 'n_tta': 8,
225
+ 'mIoU': float(ckpt['best_val_iou']),
226
+ 'n_ok': ok, 'n_total': len(results),
227
+ 'mean_tubule_pct': round(np.mean(tub_vals), 1),
228
+ 'mean_lumen_pct': round(np.mean(lum_vals), 1),
229
+ 'results': results
230
+ }, f, indent=2)
231
+
232
+
233
+
234
+ if __name__ == '__main__':
235
+ main()