2026001

scripts/analysis/verify_feature_dims.py

@@ -0,0 +1,50 @@
+#!/usr/bin/env python3
+"""
+verify_feature_dims.py
+
+Quick check: what does TF/Keras actually output for each model
+with include_top=False?  Compare against manuscript claims.
+"""
+
+import numpy as np
+from tensorflow.keras.applications import (
+    VGG16, DenseNet121, EfficientNetB1, MobileNetV2, ResNet50
+)
+
+# Match the manuscript's input sizes
+models = {
+    "VGG16":         (VGG16,         224),
+    "DenseNet121":   (DenseNet121,   224),
+    "EfficientNetB1":(EfficientNetB1, 240),
+    "MobileNetV2":   (MobileNetV2,   224),
+    "ResNet50":      (ResNet50,      224),
+}
+
+manuscript_claims = {
+    "VGG16":         25088,
+    "DenseNet121":   50176,
+    "EfficientNetB1":62720,
+    "MobileNetV2":   62720,
+    "ResNet50":      100352,
+}
+
+print(f"{'Model':<18s} {'Input':>5s} {'Spatial':>10s} {'Flattened':>10s} {'Manuscript':>12s} {'Match?':>7s}")
+print("-" * 70)
+
+for name, (model_fn, input_size) in models.items():
+    model = model_fn(weights="imagenet", include_top=False,
+                     input_shape=(input_size, input_size, 3))
+
+    # Pass a dummy batch through
+    dummy = np.random.randn(1, input_size, input_size, 3)
+    # Need to preprocess correctly — but shape doesn't depend on values
+    output = model.predict(dummy, verbose=0)
+
+    spatial = output.shape[1:4]  # (H, W, C) for channels_last
+    flattened = int(np.prod(spatial))
+    claimed = manuscript_claims[name]
+    match = "✅" if flattened == claimed else "❌"
+
+    print(f"{name:<18s} {input_size:>4}d  {str(spatial):>10s} {flattened:>10d} {claimed:>12d} {match:>7s}")
+
+print("\nNote: TF uses channels_last (NHWC) format.")