Add ResNet baseline and ConvNeXt v2 backend

app.py

@@ -14,7 +14,7 @@ from advanced_decision import (
 from gradcam import GradCAM
 from typing import Optional
 
-app = FastAPI(title="Mold Detection API v2 (ConvNeXt)")
+app = FastAPI(title="Mold Detection API (ResNet + ConvNeXt)")
 
 app.add_middleware(
     CORSMiddleware,
@@ -118,7 +118,22 @@ def ensure_dino():
 async def predict_v1(file: UploadFile):
     img = Image.open(io.BytesIO(await file.read())).convert("RGB")
     img_t = transform(img).to(device)
-    return final_decision(model, img_t)
+
+    with torch.no_grad():
+        out = model(img_t.unsqueeze(0))
+        cp = torch.softmax(out["class"], 1)[0]
+        bp = torch.softmax(out["bio"], 1)[0]
+
+    mold_p = cp[mold_idx].item()
+    bio_p = bp[1].item()
+
+    decision = final_decision(mold_p, bio_p)
+
+    return {
+        "decision": decision,
+        "mold_probability": round(mold_p, 3),
+        "biological_probability": round(bio_p, 3),
+    }
 
 
 @app.post("/predict/v2")

dino.py

@@ -1,7 +1,10 @@
 import torch
 import numpy as np
 from PIL import Image
-from datasets import load_dataset
+
+# NOTE:
+# We intentionally avoid importing `datasets` at module import time so the API can
+# start even if the optional DINO dependencies are not installed locally.
 from sklearn.metrics.pairwise import cosine_similarity
 
 
@@ -15,6 +18,9 @@ def load_dino(device):
 
 
 def build_embeddings(dino, transform, device):
+    # Lazy import to keep DINO optional.
+    from datasets import load_dataset
+
     dataset = load_dataset(
         "AdarshDS/mold-reference-images",
         split="train"

requirements.txt

@@ -6,7 +6,6 @@ pillow
 numpy<2
 python-multipart
 scikit-learn
-scikit-learn
 datasets