Update app.py

app.py

@@ -1,17 +1,18 @@
-import keras
 import numpy as np
-import gradio as gr
+import keras
 from PIL import Image
+import gradio as gr
 
 MODEL_PATH = "small32cnn_mlbt_mmat.keras"
-STATS_PATH = "preproc_stats_smallcnn.npz"
+STATS_PATH = "jet_image_scale_and_stats.npz"
 
-# Load with standalone Keras 3 (matches the saver)
 model = keras.models.load_model(MODEL_PATH, compile=False)
 
 stats = np.load(STATS_PATH)
-MEAN = float(stats["MEAN"])
-STD  = float(stats["STD"])
+SCALE = float(stats["SCALE"])
+MEAN  = float(stats["MEAN"])
+STD   = float(stats["STD"])
+print("Loaded SCALE/MEAN/STD:", SCALE, MEAN, STD, flush=True)
 
 CLASS_NAMES = ["MMAT", "MLBT"]
 
@@ -19,34 +20,28 @@ CLASS_NAMES = ["MMAT", "MLBT"]
 
 # ---- Preprocessing: Image -> (1, 32, 32, 1) normalized ----
 def preprocess(img: np.ndarray) -> np.ndarray:
-    """
-    img: H x W x C in [0, 255] from Gradio (numpy)
-    returns: (1, 32, 32, 1) float32, z-score normalized
-    """
-
-    # Convert to grayscale
     if img.ndim == 3 and img.shape[2] == 3:
-        # RGB -> grayscale
         img_gray = np.dot(img[..., :3], [0.2989, 0.5870, 0.1140])
     elif img.ndim == 3 and img.shape[2] == 1:
         img_gray = img[..., 0]
     elif img.ndim == 2:
         img_gray = img
     else:
-        # fallback: take first channel
         img_gray = img[..., 0]
 
-    # Resize to 32x32
     pil_img = Image.fromarray(img_gray.astype("uint8"))
     pil_img = pil_img.resize((32, 32), Image.BILINEAR)
     arr = np.array(pil_img).astype("float32")
 
-    # z-score normalization using training stats
-    arr = (arr - MEAN) / STD
+    # Invert the global scaling to approximate original X
+    arr_unscaled = arr / SCALE
+
+    # Now apply the same normalization as during training
+    arr_norm = (arr_unscaled - MEAN) / (STD + 1e-8)
+
+    arr_norm = arr_norm[None, ..., None]
+    return arr_norm
 
-    # Add batch and channel dims: (1, 32, 32, 1)
-    arr = arr[None, ..., None]
-    return arr
 
 
 # ---- Prediction function for Gradio ----
@@ -55,15 +50,16 @@ def predict(img: np.ndarray):
     raw = float(model.predict(x, verbose=0)[0, 0])
     print("Raw model output:", raw, flush=True)
 
-    # raw ≈ P(MMAT)
-    prob_mmat = raw
-    prob_mlbt = 1.0 - prob_mmat
+    # raw ≈ P(MLBT) as in training
+    prob_mlbt = raw
+    prob_mmat = 1.0 - prob_mlbt
 
     return {"MLBT": prob_mlbt, "MMAT": prob_mmat}
 
 
 
 
+
 # ---- Gradio interface ----
 input_component = gr.Image(
     type="numpy",