Vision Defect Inspection Results Dashboard
A CNN-based defect classifier built independently by Bharadwaj Rachuri using transfer learning on
MobileNetV2. Built, trained, and tested locally — not deployed to a live endpoint.
TRAINING RESULTS — VERIFIED FROM TRAINING LOGS
Final Train Accuracy
0%
Epoch 8 of 8
Final Validation Accuracy
0%
Epoch 8 of 8
Test Suite
0/26
All tests passing
Total Parameters
2.42M
1.37M trainable
⚠️
FINE-TUNING INSTABILITY AT EPOCH 6 — SHOWN, NOT HIDDEN
Training ran in two phases: Phase 1 (epochs 1–5) trained only the classification head
with the MobileNetV2 base frozen, reaching 98.8% train / 95.0% val accuracy. Phase 2
(epochs 6–8) unfroze the last 20 MobileNetV2 layers for fine-tuning — and train accuracy immediately
dropped to 78.3% at epoch 6, before recovering to 98.3% by epoch 7–8. This is the expected
signature of unfreezing pretrained layers with an unadjusted learning rate, not a bug. Final validation
accuracy (91.7%) is lower than the Phase 1 peak (98.3% at epoch 2) — a real trade-off worth being upfront about.
Training & Validation Accuracy — 8 Epochs
Two-phase transfer learning: frozen-base head training (epochs 1–5) then fine-tuning the last 20 MobileNetV2 layers (epochs 6–8). Dotted orange line marks the phase transition.
MODEL ARCHITECTURE — VERIFIED FROM SAVED MODEL FILE
| Layer | Output Shape | Params |
|---|
| Input | (224, 224, 3) | 0 |
| MobileNetV2 (backbone) | (7, 7, 1280) | 2,257,984 |
| GlobalAveragePooling2D | (1280) | 0 |
| Dropout | (1280) | 0 |
| Dense (128, ReLU) | (128) | 163,968 |
| Dropout | (128) | 0 |
| Dense (1, Sigmoid) | (1) | 129 |
What's verified vs. not shown here: Train/validation accuracy per epoch, test suite result,
and full model architecture are confirmed directly from training logs and the saved .keras model
file. Validation AUC, exact training time, and dataset composition were not available in the source material
used to build this page, so they are intentionally omitted rather than estimated.