perf: major build optimizations - remove matplotlib, fix duplicate endpoint, optimize dependencies

app.py

@@ -32,18 +32,6 @@ DEVICE       = os.getenv("DEVICE",       "cpu")
 # This prevents timeout on HuggingFace free tier during container startup
 _model_loaded = False
 
-@app.get("/health")
-def health():
-    global _model_loaded
-    if not _model_loaded:
-        try:
-            load_model(WEIGHTS_PATH, DEVICE, meta_path=META_PATH)
-            _model_loaded = True
-            print("Model loaded and cached on first request.")
-        except Exception as e:
-            return {"status": "error", "device": DEVICE, "model": "DenseNet121-CBAM", "error": str(e)}
-    return {"status": "ok", "device": DEVICE, "model": "DenseNet121-CBAM"}
-
 
 # ─── Routes ───────────────────────────────────────────────────────────────────
 from fastapi.responses import HTMLResponse
@@ -63,6 +51,14 @@ def root():
 
 @app.get("/health")
 def health():
+    global _model_loaded
+    if not _model_loaded:
+        try:
+            load_model(WEIGHTS_PATH, DEVICE, meta_path=META_PATH)
+            _model_loaded = True
+            print("Model loaded and cached on first request.")
+        except Exception as e:
+            return {"status": "error", "device": DEVICE, "model": "DenseNet121-CBAM", "error": str(e)}
     return {"status": "ok", "device": DEVICE, "model": "DenseNet121-CBAM"}
 
 

inference.py

@@ -156,6 +156,15 @@ class GradCAMPlusPlus:
 
 # ─── Preprocessing ─────────────────────────────────────────────────────────────
 
+def apply_jet_colormap(gray_img):
+    """Apply jet colormap manually without matplotlib (0=blue, 1=red)."""
+    gray_img = np.clip(gray_img, 0, 1)
+    r = np.clip(1.5 - np.abs(gray_img * 2 - 3), 0, 1)
+    g = np.clip(1.5 - np.abs(gray_img * 2 - 2), 0, 1)
+    b = np.clip(1.5 - np.abs(gray_img * 2 - 1), 0, 1)
+    return (np.stack([r, g, b], axis=-1) * 255).astype(np.uint8)
+
+
 def get_transform(image_size=512):
     return A.Compose([
         A.CLAHE(clip_limit=2.0, tile_grid_size=(8, 8), p=1.0),
@@ -257,8 +266,8 @@ def run_inference(image_path: str, weights_path: str, device: str = "cpu",
             orig_rgb  = cv2.cvtColor(orig, cv2.COLOR_GRAY2RGB)
             cam_resized = cv2.resize(cam_map, (orig_rgb.shape[1], orig_rgb.shape[0]))
 
-            import matplotlib.cm as mpl_cm
-            heatmap = (mpl_cm.jet(cam_resized)[:, :, :3] * 255).astype(np.uint8)
+            # Manual jet colormap (replaces matplotlib.cm.jet) - saves ~100MB in build
+            heatmap = apply_jet_colormap(cam_resized)
             overlay = cv2.addWeighted(orig_rgb, 0.6, heatmap, 0.4, 0)
             result["gradcam_overlay"] = overlay   # numpy array, encode downstream
         finally:

requirements.txt

@@ -7,9 +7,6 @@ fastapi==0.111.0
 uvicorn[standard]==0.29.0
 python-multipart==0.0.9
 opencv-python-headless==4.9.0.80
+albumentations==1.4.2
 numpy==1.26.4
 Pillow==10.3.0
-
-# Optional: Remove these if not used in inference
-# albumentations==1.4.2
-# matplotlib==3.8.4