Add mold detection FastAPI backend v2

Dockerfile

@@ -28,3 +28,4 @@ EXPOSE 7860
 CMD uvicorn app:app --host 0.0.0.0 --port 7860
 
 
+

advanced_decision.py

@@ -0,0 +1,55 @@
+import torch
+import torch.nn.functional as F
+import numpy as np
+
+MOLD_HIGH = 0.85
+MOLD_LOW  = 0.50
+BIO_TH    = 0.70
+UNCERT_TH = 0.15
+PATCH_TH  = 0.50
+DINO_TH   = 0.75
+
+def enable_dropout(model):
+    for m in model.modules():
+        if m.__class__.__name__.startswith("Dropout"):
+            m.train()
+
+def mc_uncertainty(model, img_tensor, mold_idx, T=15):
+    enable_dropout(model)
+    probs = []
+    for _ in range(T):
+        out = model(img_tensor.unsqueeze(0))
+        cp = F.softmax(out["class"],1)[0]
+        probs.append(cp[mold_idx].item())
+    model.eval()
+    return float(np.mean(probs)), float(np.std(probs))
+
+def patch_consistency(model, image, transform, mold_idx, device, stride=112):
+    w,h = image.size
+    votes = []
+    for y in range(0, h-224+1, stride):
+        for x in range(0, w-224+1, stride):
+            patch = image.crop((x,y,x+224,y+224))
+            pt = transform(patch).unsqueeze(0).to(device)
+            with torch.no_grad():
+                out = model(pt)
+                cp = F.softmax(out["class"],1)[0]
+            votes.append(cp[mold_idx].item())
+    return float(np.mean(np.array(votes) > 0.7)) if votes else 0.0
+
+def final_decision_v2(
+    mold_p, bio_p, uncert, patch_ratio, dino_sim
+):
+    if (
+        mold_p > MOLD_HIGH and
+        bio_p > BIO_TH and
+        uncert < UNCERT_TH and
+        patch_ratio > PATCH_TH and
+        dino_sim > DINO_TH
+    ):
+        return "Mold"
+
+    if mold_p > MOLD_LOW and bio_p > BIO_TH:
+        return "Possible Mold"
+
+    return "Not Mold"

app.py

@@ -27,7 +27,7 @@ app.add_middleware(
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
 # Model path for HuggingFace Spaces (flat structure)
-model_path = Path("resnet50_multitask_bio.pth")
+model_path = Path("resnet50_multitask_mold.pth")
 
 print(f"Loading model from: {model_path.absolute()}")
 print(f"Model exists: {model_path.exists()}")

dino.py

@@ -0,0 +1,28 @@
+import os
+import numpy as np
+import torch
+import torch.hub
+from PIL import Image
+from sklearn.metrics.pairwise import cosine_similarity
+
+def load_dino(device):
+    model = torch.hub.load("facebookresearch/dinov2", "dinov2_vits14")
+    model.eval().to(device)
+    return model
+
+def build_embeddings(dino, transform, image_dir, device):
+    embs = []
+    for f in os.listdir(image_dir):
+        if f.lower().endswith((".jpg",".png",".jpeg")):
+            img = Image.open(os.path.join(image_dir,f)).convert("RGB")
+            t = transform(img).unsqueeze(0).to(device)
+            with torch.no_grad():
+                e = dino(t)
+            embs.append(e.squeeze().cpu().numpy())
+    return np.vstack(embs)
+
+def similarity(dino, mold_embs, image, transform, device):
+    t = transform(image).unsqueeze(0).to(device)
+    with torch.no_grad():
+        e = dino(t).cpu().numpy()
+    return float(cosine_similarity(e, mold_embs).max())

gradcam.py

@@ -0,0 +1,30 @@
+import torch
+import numpy as np
+
+class GradCAM:
+    def __init__(self, model, target_layer):
+        self.model = model
+        self.gradients = None
+        self.activations = None
+
+        target_layer.register_forward_hook(self._fwd)
+        target_layer.register_backward_hook(self._bwd)
+
+    def _fwd(self, m, i, o):
+        self.activations = o
+
+    def _bwd(self, m, gi, go):
+        self.gradients = go[0]
+
+    def generate(self, img_tensor, class_idx):
+        out = self.model(img_tensor.unsqueeze(0))
+        score = out["class"][:, class_idx].sum()
+        self.model.zero_grad()
+        score.backward()
+
+        w = self.gradients.mean(dim=(2,3), keepdim=True)
+        cam = (w * self.activations).sum(dim=1)
+        cam = torch.relu(cam)
+        cam = cam / (cam.max() + 1e-8)
+
+        return cam.detach().cpu().numpy()[0]

requirements.txt

@@ -7,3 +7,4 @@ numpy<2
 python-multipart
 
 
+