Update model README

README.md

@@ -114,9 +114,9 @@ Output (6 classes)
 
 | Metric | Value |
 |--------|-------|
-| Test Accuracy | **96.23%** (970/1,008) |
-| Validation Accuracy (SWA) | **95.93%** |
-| Test-Time Augmentation | **96.33%** (+0.10%) |
+| Test Accuracy | 96.23% (970/1,008) |
+| Validation Accuracy (SWA) | 95.93% |
+| Test-Time Augmentation | 96.33% (+0.10%) |
 | Test Loss | 0.3974 |
 | Overfitting Gap (Train − Test) | 3.22% |
 | Macro Avg Precision | 96.29% |
@@ -135,7 +135,7 @@ Output (6 classes)
 | Temple | 97.55% | 94.64% | 96.07% | 168 |
 | **Macro Avg** | **96.29%** | **96.23%** | **96.21%** | **1,008** |
 
-**Highest performing classes:** Skyscraper (recall=99.40%), Castle (F1=97.92%)
+**Highest performing classes:** Skyscraper (recall=99.40%), Castle (F1=97.92%)  
 **Most challenging classes:** Stadium (recall=90.48%), Bridge (precision=94.19%)
 
 ### Model Selection
@@ -174,8 +174,8 @@ Two-phase progressive training with SWA post-processing:
 
 | Phase | Description | Backbone | Optimizer | LR | Max Epochs | Actual Epochs | CutMix+Mixup | FocalLoss LS |
 |-------|-------------|----------|-----------|-----|-----------|---------------|---------------|-------------|
-| **Phase 1** — Feature Extraction | Train custom head only | Frozen (all) | AdamW (wd=2e-5) | 0.001 + CosineDecay + Warmup 3ep | 25 | **1** ¹ | Yes (50/50 alternation) | 0.1 |
-| **Phase 2** — Selective Fine-Tuning | Load best_phase1 → fine-tune | conv4_block + conv5_block unfrozen (BN frozen) | DiscriminativeAdamW (conv4=0.1×) | 3e-4 + CosineDecay + Warmup 5ep | 50 | **6 + 5 SWA** ² | No | 0.05 |
+| **Phase 1** — Feature Extraction | Train custom head only | Frozen (all) | AdamW (wd=2e-5) | 0.001 + CosineDecay + Warmup 3ep | 25 | 1 ¹ | Yes (50/50 alternation) | 0.1 |
+| **Phase 2** — Selective Fine-Tuning | Load best_phase1 → fine-tune | conv4_block + conv5_block unfrozen (BN frozen) | DiscriminativeAdamW (conv4=0.1×) | 3e-4 + CosineDecay + Warmup 5ep | 50 | 6 + 5 SWA ² | No | 0.05 |
 
 > ¹ Phase 1 stopped at epoch 1 because `val_accuracy = 86.71% ≥ 85%` threshold (myCallback). This demonstrates the effectiveness of ImageNet transfer learning — a single epoch of head training exceeds the target.
 
@@ -209,19 +209,19 @@ Two-phase progressive training with SWA post-processing:
 |-----------|---------------|-----------|
 | Transfer Learning | DenseNet121 backbone frozen in Phase 1 | Yosinski et al., NeurIPS 2014 |
 | Selective Fine-Tuning | Unfreeze conv4+conv5 only, BN stays frozen | Howard & Ruder, ACL 2018 |
-| Discriminative LR | conv4=0.1×, conv5+head=1.0× | — |
+| Discriminative LR | conv4=0.1×, conv5+head=1.0× | Howard & Ruder, ACL 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 | — |
-| Batch Normalization | After Conv2D and Dense; frozen during fine-tuning | — |
-| EMA | Shadow weights, decay=0.999 | — |
+| 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 |
+| EMA | Shadow weights, decay=0.999 | Tarvainen & Valpola, NeurIPS 2017 |
 | SWA | 5-epoch post-training, constant LR 1e-4 | Izmailov et al., UAI 2018 |
-| Data Augmentation | Rotation ±15°, shift ±10%, zoom ±20%, brightness 0.75–1.15, horizontal flip | — |
-| Test-Time Augmentation | 6 augmentation variants, averaged | — |
-| WarmupCosineDecay | Linear warmup + cosine annealing | — |
-| Early Stopping | Patience 7 (Phase 1) / 12 (Phase 2) | — |
+| Data Augmentation | Rotation ±15°, shift ±10%, zoom ±20%, brightness 0.75–1.15, horizontal flip | Perez & Wang, arXiv 2017 |
+| Test-Time Augmentation | 6 augmentation variants, averaged | Shanmugam et al., ICML 2020 |
+| WarmupCosineDecay | Linear warmup + cosine annealing | Loshchilov & Hutter, ICLR 2017 (SGDR) |
+| Early Stopping | Patience 7 (Phase 1) / 12 (Phase 2) | Prechelt, Neural Networks 1998 |
 
 ### Dataset
 
@@ -238,7 +238,7 @@ Two-phase progressive training with SWA post-processing:
 - **Normalization:** `preprocess_input` from `tf.keras.applications.densenet` (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
-- **Split method:** `splitfolders.ratio` from `dataset-raw/`, seed=42
+- **Split method:** `splitfolders.ratio` from `dataset/`, seed=42
 
 ## Files
 
@@ -270,7 +270,7 @@ from tensorflow.keras.layers import Layer
 from PIL import Image
 import numpy as np
 
-# --- Custom Layers (must match training definition) ---
+# ========---Custom Layers (must match training definition)---========
 
 class GeMPooling(Layer):
     def __init__(self, p=3.0, eps=1e-6, **kwargs):
@@ -345,7 +345,7 @@ class DiscriminativeAdamW(tf.keras.optimizers.AdamW):
         return {**super().get_config(), "lr_multipliers": self.lr_multipliers,
                 "backbone_layer_idx": self.backbone_layer_idx}
 
-# --- Load Model ---
+# =====================---Load Model---==========================
 
 LABELS = ["bridge", "castle", "mosque", "skyscraper", "stadium", "temple"]
 custom_objects = {
@@ -357,7 +357,7 @@ custom_objects = {
 model_path = hf_hub_download("0xgr3y/Arch-Building-Image-Classification", "best_phase2_swa.keras")
 model = tf.keras.models.load_model(model_path, custom_objects=custom_objects, compile=False)
 
-# --- Inference ---
+# =======================---Inference---==========================
 
 img = Image.open("building.jpg").convert("RGB").resize((320, 320))
 arr = np.expand_dims(preprocess_input(np.array(img, dtype=np.float32)), axis=0)
@@ -365,7 +365,7 @@ preds = model.predict(arr, verbose=0)[0]
 print(f"Predicted: {LABELS[np.argmax(preds)]} ({np.max(preds)*100:.1f}%)")
 ```
 
-### Python — TF-Lite
+### Python — TensorFlow Lite
 
 Download the model first:
 ```python
@@ -448,7 +448,7 @@ print(f"Predicted: {LABELS[np.argmax(preds)]} ({np.max(preds)*100:.1f}%)")
 
 ## Ethical Considerations
 
-- All training images sourced from [Pexels](https://www.pexels.com) under the Pexels License (free for commercial use, no attribution required). No copyrighted or personally identifiable images were used.
+- All training images sourced from [Pexels.com](https://www.pexels.com) under the Pexels License (free for commercial use, no attribution required). No copyrighted or personally identifiable images were used.
 - The dataset contains only photographs of buildings and structures — no people, faces, or private property are the subject of classification.
 - The model reflects the visual distribution of Pexels stock photography, which may over-represent Western and iconic architectural styles and under-represent vernacular or regional architecture.
 - The 6 class categories are broad and do not capture the full diversity of world architecture. Results should not be used to make definitive claims about architectural categorization.
@@ -471,6 +471,13 @@ print(f"Predicted: {LABELS[np.argmax(preds)]} ({np.max(preds)*100:.1f}%)")
 7. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the Inception Architecture for Computer Vision. *CVPR 2016*. [arXiv:1512.00567](https://arxiv.org/abs/1512.00567)
 8. Yosinski, J., Clune, J., Bengio, Y., & Lipson, H. (2014). How Transferable Are Features in Deep Neural Networks? *NeurIPS 2014*. [arXiv:1411.1792](https://arxiv.org/abs/1411.1792)
 9. Howard, J., & Ruder, S. (2018). Universal Language Model Fine-tuning for Text Classification. *ACL 2018*. [arXiv:1801.06146](https://arxiv.org/abs/1801.06146)
+10. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2014). Dropout: A Simple Way to Prevent Neural Networks from Overfitting. *JMLR*, 15(56), 1929–1958. [http://jmlr.org/papers/v15/srivastava14a.html](http://jmlr.org/papers/v15/srivastava14a.html)
+11. Ioffe, S., & Szegedy, C. (2015). Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. *arXiv preprint*. [arXiv:1502.03167](https://arxiv.org/abs/1502.03167)
+12. Tarvainen, A., & Valpola, H. (2017). Mean Teachers are Better Role Models: Weight-averaged Consistency Targets Improve Semi-supervised Deep Learning Results. *NeurIPS 2017*. [arXiv:1703.01780](https://arxiv.org/abs/1703.01780)
+13. Perez, L., & Wang, J. (2017). The Effectiveness of Data Augmentation in Image Classification using Deep Learning. *arXiv preprint*. [arXiv:1712.04621](https://arxiv.org/abs/1712.04621)
+14. Shanmugam, D., Blalock, D., Balakrishnan, G., Guttag, J., & Sarma, A. (2020). Towards Principled Test-Time Augmentation. *ICML 2020*. [PDF](https://dmshanmugam.github.io/pdfs/icml_2020_testaug.pdf)
+15. Loshchilov, I., & Hutter, F. (2017). SGDR: Stochastic Gradient Descent with Warm Restarts. *ICLR 2017*. [arXiv:1608.03983](https://arxiv.org/abs/1608.03983)
+16. Prechelt, L. (1998). Automatic Early Stopping Using Cross Validation: Quantifying the Criteria. *Neural Networks*, 11(4), 761–767. [https://doi.org/10.1016/S0893-6080(98)00010-0](https://doi.org/10.1016/S0893-6080(98)00010-0)
 
 ## Citation