Update README.md

README.md

@@ -13,14 +13,11 @@ tags:
   - calibration
   - temperature-scaling
   - computer-vision
-  - fgvc
+  - tensorflow.js
 library_name: keras
 language: en
 datasets:
   - 0xgr3y/arch-building-dataset
-widget:
-  - src: results/greyscope-labs-architecture-classification-efficientnetv2.jpg
-    example_title: Bridge Classification
 model-index:
   - name: Architectural Building Image Classifier
     results:
@@ -82,10 +79,14 @@ A fine-grained image classification model for world architectural buildings. Bui
 
 - **GeM Pooling** (Radenovic et al., CVPR 2018) — replaces global average pooling with a learnable power parameter (p=3.0) that emphasizes high-activation features, yielding stronger discriminative representations for FGIC tasks
 - **Focal Loss** (Lin et al., ICCV 2017, gamma=2.0) — down-weights well-classified examples to focus gradient updates on hard-to-classify building pairs
-- **DiscriminativeAdamW LR** — extends AdamW with per-variable LR scaling on block6 (×0.1) via update_step override, combined with selective fine-tuning (block6+top_conv unfrozen, BN frozen). LR scaling produces truly discriminative updates — block6 variables receive 10× smaller learning rate than head variables
-- **SWA with BN re-estimation** (Izmailov et al., UAI 2018) — 10-epoch post-training weight averaging with constant LR 1e-4, followed by 100-step batch normalization statistics re-estimation
-- **Grad-CAM** (Selvaraju et al., ICCV 2017) — gradient-weighted class activation mapping for explainability, targeting *top_conv* (last Conv2D layer)
+- **DiscriminativeAdamW LR** — extends AdamW with per-variable LR scaling on block6 (×0.1) via (update_step) override, combined with selective fine-tuning (block6+top_conv unfrozen, BN frozen). LR scaling produces truly discriminative updates — block6 variables receive 10× smaller learning rate than head variables (117 total: 105 block6 + 12 head)
+- **Mixup + CutMix** (Zhang et al., ICLR 2018. Yun et al., ICCV 2019) — alternating per-batch (50/50): Mixup (alpha=0.2, linear interpolation) and CutMix (alpha=1.0, spatial patch). Applied only in Phase 1 training to regularize head learning
+- **Selective Unfreeze** (Yosinski et al., 2014) — Phase 2 unfreezes block6+top_conv layers (180/513 EfficientNetV2-S layers) while keeping BatchNormalization frozen to preserve pretrained statistics
+- **SWA with BN re-estimation** (Izmailov et al., UAI 2018) — 10-epoch post-training weight averaging with constant LR 1e-4, followed by 100-step batch normalization statistics re-estimation (3,200 images)
+- **Test-Time Augmentation** — 6 variations averaged at inference: original, horizontal flip, center crop 85%, center crop 70%, corner crop top-left 80%, corner crop bottom-right 80%. Yields +0.22% accuracy improvement (97.77% → 97.99%)
+- **Grad-CAM** (Selvaraju et al., ICCV 2017) — gradient-weighted class activation mapping for explainability, targeting *top_conv* (last Conv2D layer of EfficientNetV2-S)
 - **ECE Calibration** (Guo et al., ICML 2017) — Expected Calibration Error with 15-bin reliability diagram to assess prediction confidence reliability
+- **Temperature Scaling** (Guo et al., ICML 2017) — post-hoc calibration via scalar temperature parameter T optimized on validation set (NLL minimization). T=0.54 reduces ECE from 12.04% (underconfident due to Label Smoothing) to 0.53% — applied at inference via (softmax(log(probs) / T)) trick
 
 ## Architecture
 
@@ -142,7 +143,7 @@ Output (8 classes)
 | Top-2 Accuracy | 99.26% |
 | Top-3 Accuracy | 99.70% |
 | Macro ROC-AUC (OvR) | 0.9985 |
-| ECE (15 bins) | 0.1204 (pre-T-scaling; post-T-scaling: 0.0053, T=0.54) |
+| ECE (15 bins) | 0.1204 (pre-T-scaling. post-T-scaling: 0.0053, T=0.54) |
 
 ### Per-Class Results
 
@@ -169,21 +170,25 @@ Four candidate models were evaluated on the validation set:
 | `fine_tuning_ema.keras` | 93.53% | 0.6007 | Phase 2 EMA (per-step Polyak averaging) |
 | **`fine_tuning_swa.keras`** | **98.36%** | **0.4109** | **SWA averaged weights ← SELECTED** |
 
-### SWA Progression
-
-| SWA Epoch | Val Accuracy | Val Loss |
-|-----------|-------------|----------|
-| 1 | 95.76% | 0.5831 |
-| 2 | 97.62% | 0.5116 |
-| 3 | 97.69% | 0.4748 |
-| 4 | 96.95% | 0.4390 |
-| 5 | 97.47% | 0.4490 |
-| 6 | 97.84% | 0.4416 |
-| 7 | 98.14% | 0.4055 |
-| 8 | 97.32% | 0.4359 |
-| 9 | 97.02% | 0.4519 |
-| 10 | 97.54% | 0.4226 |
-| **SWA + BN (final)** | **98.36%** | **0.4109** |
+### Training Progression
+
+| Phase | Epoch | Train Acc | Val Accuracy | Val Loss |
+|-------|-------|-----------|-------------|----------|
+| Phase 1 (Head Training) | 1 | 56.96% | 92.19% | 1.0079 |
+| Phase 2 (Selective Fine-Tuning) | 1 | 84.96% | 96.21% | 0.5656 |
+| SWA | 1 | 90.83% | 95.76% | 0.5831 |
+| SWA | 2 | 94.07% | 97.62% | 0.5116 |
+| SWA | 3 | 95.36% | 97.69% | 0.4748 |
+| SWA | 4 | 96.56% | 96.95% | 0.4390 |
+| SWA | 5 | 97.18% | 97.47% | 0.4490 |
+| SWA | 6 | 97.76% | 97.84% | 0.4416 |
+| SWA | 7 | 97.91% | 98.14% | 0.4055 |
+| SWA | 8 | 98.19% | 97.32% | 0.4359 |
+| SWA | 9 | 98.14% | 97.02% | 0.4519 |
+| SWA | 10 | 98.59% | 97.54% | 0.4226 |
+| **SWA + BN (final)** | — | — | **98.36%** | **0.4109** |
+
+> Phase 1 and Phase 2 each stopped after 1 epoch via `myCallback` (custom early stopping at target accuracy: 85% Phase 1, 92% Phase 2). SWA ran 10 epochs with constant LR 1e-4, followed by BN re-estimation (100 steps, 3,200 images). Values shown are training-time metrics from progress bar. checkpoint evaluation values may differ slightly (see Model Selection table above).
 
 ![Training Curves](results/training_curves.png)
 
@@ -191,16 +196,12 @@ Four candidate models were evaluated on the validation set:
 
 ![Per-Class Accuracy](results/per_class_accuracy.png)
 
-![Reliability Diagram](results/reliability_diagram.png)
-
-![ROC Curves](results/roc_curves.png)
+![Confidence Per Class](results/confidence_per_class.png)
 
 ![t-SNE Embedding](results/tsne_embedding.png)
 
 ![Grad-CAM Heatmaps](results/gradcam_heatmaps.png)
 
-![Confidence Per Class](results/confidence_per_class.png)
-
 ## Training Details
 
 ### Training Strategy
@@ -245,12 +246,12 @@ Two-phase progressive training with SWA post-processing:
 | Transfer Learning | EfficientNetV2-S backbone frozen in Phase 1 | Yosinski et al., NeurIPS 2014 |
 | Selective Fine-Tuning | Unfreeze block6+top_conv only, BN stays frozen | Howard & Ruder, ACL 2018 |
 | Discriminative LR Scaling | block6 LR×0.1 via update_step (truly discriminative — 10× smaller updates for pretrained features) | Howard & Ruder, ACL 2018 |
-| CutMix + Mixup | Alternation per batch (50/50), Phase 1 only | Yun et al., ICCV 2019; Zhang et al., ICLR 2018 |
+| CutMix + Mixup | Alternation per batch (50/50), Phase 1 only | Yun et al., ICCV 2019. Zhang et al., ICLR 2018 |
 | Focal Loss | gamma=2.0, down-weights easy examples | Lin et al., ICCV 2017 |
 | Label Smoothing | 0.1 (Phase 1) → 0.05 (Phase 2) | Szegedy et al., CVPR 2016 |
 | GeM Pooling | p=3.0 learnable, replaces GAP | Radenovic et al., CVPR 2018 |
 | Dropout | 0.4 after Dense(256)+BN | Srivastava et al., JMLR 2014 |
-| Batch Normalization | After Conv2D and Dense; frozen during fine-tuning | Ioffe & Szegedy, arXiv 2015 |
+| Batch Normalization | After Conv2D and Dense. frozen during fine-tuning | Ioffe & Szegedy, arXiv 2015 |
 | EMA (per-step) | Shadow weights, decay=0.999, Polyak averaging | Tarvainen & Valpola, NeurIPS 2017 |
 | SWA | 10-epoch post-training, constant LR 1e-4 | Izmailov et al., UAI 2018 |
 | Data Augmentation | Rotation ±15°, shift ±10%, shear ±0.1 rad, zoom ±20%, brightness 0.75–1.15, channel shift ±10.0, horizontal flip | Perez & Wang, arXiv 2017 |
@@ -273,7 +274,7 @@ See the dataset curation page for [World Architectural Buildings Dataset for Mul
 
 - **Normalization:** `preprocess_input` from `tf.keras.applications.efficientnet_v2` (ImageNet distribution)
 - **Input resolution:** 320×320 (higher than ImageNet default 224×224 to capture fine-grained architectural details — textures, ornaments, facade patterns)
-- **Augmentation:** Applied to training set only; validation and test sets use clean preprocessing
+- **Augmentation:** Applied to training set only. validation and test sets use clean preprocessing
 - **Split method:** `splitfolders.ratio` from `dataset/`, seed=42
 
 ## Files
@@ -285,14 +286,14 @@ See the dataset curation page for [World Architectural Buildings Dataset for Mul
 | **Config** | `config.json` · `label_mapping.json` · `preprocessor_config.json` |
 | **Evaluation** | `calibration_data.json` · `model_benchmark.json` · `confusion_pairs.json` · `class_confidence_stats.json` · `temperature_config.json` |
 | **Deployment** | `saved_model/` (183 MB) · `tflite/` (88 MB) · `tfjs_model/` (90 MB, 23 shards) |
-| **Results** | `results/` — 12 PNG (training curves, confusion matrix, ROC, t-SNE, Grad-CAM, etc.) |
+| **Results** | `results/` — 12 PNG (augmentation, reliability-diagram, training curves, confusion matrix, ROC, t-SNE, Grad-CAM, etc.) |
 | **Archive** | `models_keras/` — 3 checkpoints (head_training, fine_tuning, fine_tuning_ema) |
 
 ## Usage
 
 ### Gradio Space
 
-Try the live building classifier: [Architecture Building Image Classifier with Space](https://huggingface.co/spaces/0xgr3y/arch-building-classifier)
+Try the live building classify: [Architecture Building Image Classifier with Space](https://huggingface.co/spaces/0xgr3y/arch-building-classifier)
 
 ### Python — build_model.py (recommended)
 
@@ -459,6 +460,8 @@ Keras vs TFLite consistency was verified on 8 random test samples (1 per class):
 
 > The 1 misclassification (castle→barn, 65% confidence) is consistent with the 97.77% test accuracy. The 8/8 match confirms TFLite conversion preserves model behavior exactly.
 
+![TFLite Inference](results/inference_tflite.png)
+
 ## Security Notice (PAIT-KERAS-301)
 
 The `.keras` files in this repository are flagged **"Unsafe"** by [Protect AI Guardian](https://protectai.com/insights/models/0xgr3y/Arch-Building-Image-Classification) (threat: PAIT-KERAS-301). This is a **structural false positive**, not a malware detection:
@@ -487,7 +490,7 @@ With model is provided in multiple formats to suit different deployment scenario
 | **Weights H5** ✓ | `fine_tuning_swa.weights.h5` | ~158 MB | ✓ Safe | — | Programmatic load via `build_model.py` |
 | **safetensors** ✓ | `fine_tuning_swa.safetensors` | ~157 MB | ✓ Safe | — | HF standard, cross-framework |
 | **Build Script** ✓ | `build_model.py` | ~21 KB | ✓ Safe | — | Architecture reconstruction + `load_weights()` |
-| **Keras** ⚠️ | `fine_tuning_swa.keras` | ~227 MB | ⚠️ Flagged | 358.0 ms | Developer reference, fine-tuning |
+| **Keras** ℹ | `fine_tuning_swa.keras` | ~227 MB | ℹ Flagged | 358.0 ms | Developer reference, fine-tuning |
 
 ### Load Examples
 
@@ -504,7 +507,7 @@ See **Usage** section above for complete load + inference examples for each form
 
 - Trained on Pexels stock photography — performance may differ on user-generated or field photographs
 - Limited to 8 architectural classes (barn, bridge, castle, mosque, skyscraper, stadium, temple, windmill)
-- Confusion pair analysis found **0 significant pairs** (threshold >5%) — all 8 classes are well-distinguished by the model; see `confusion_pairs.json` for details
+- Confusion pair analysis found **0 significant pairs** (threshold >5%) — all 8 classes are well-distinguished by the model. see `confusion_pairs.json` for details
 - Barn and windmill share 3 cross-class duplicates (0.02% of dataset) — left as-is due to negligible impact
 - Inference confidence can be low on atypical examples
 
@@ -522,7 +525,7 @@ See **Usage** section above for complete load + inference examples for each form
 
 - **Gradio Space (Live):** [arch-building-classifier Space](https://huggingface.co/spaces/0xgr3y/arch-building-classifier)
 - **Dataset Studio:** [0xgr3y/arch-building-dataset](https://huggingface.co/datasets/0xgr3y/arch-building-dataset)
-- **GitHub Repo:** [arcxteam/building-architectural-image-classifier](https://github.com/arcxteam/building-architectural-image-classifier)
+- **GitHub Repository:** [arcxteam/building-architectural-image-classifier](https://github.com/arcxteam/building-architectural-image-classifier)
 
 ## References