Add Quick Inference, machine-readable artefacts, and compute disclosure

README.md

@@ -363,3 +363,46 @@ Per-class Grad-CAM saliency maps (one PNG per CNN-class model, 10 classes per fi
 | VGG19              | 79.31 ± 1.89% |
 | ResNet101          | 79.27 ± 1.07% |
 | CLIP Transformer   | 63.92 ± 1.79% |
+
+
+## Quick Inference
+
+```bash
+pip install torch torchvision timm open_clip_torch huggingface_hub pillow
+python code/inference_example.py path/to/fundus.jpg --model densenet121
+```
+
+The script downloads the requested checkpoint from this repo via the HF Hub, applies the
+same 224x224 + ImageNet-normalisation preprocessing used during evaluation, and prints
+the top-3 predicted classes with probabilities.
+
+## Dataset Figures
+
+- `dataset/class_distribution.png` + `.csv` — per-class counts, original vs augmented
+- `dataset/sample_grid.png` — one representative image per class
+
+## Machine-Readable Artefacts
+
+- `code/hparams.json` — full hyperparameter table (optimizer, schedule, augmentations, TTA, ensemble, per-model LR, GPU hours)
+- `analysis/per_class_metrics.csv` — precision/recall/F1/support per model per class
+- `analysis/confusion_matrices_csv/cm_<model>.csv` — numeric confusion matrices
+- `CITATION.cff` — Citation File Format
+
+## Compute Disclosure
+
+All 9 models + ensemble were trained on a single **NVIDIA Tesla T4 (16 GB)** under
+PyTorch 2.11.0+cu128. Approximate wall-clock GPU-hours:
+
+| Model | GPU-hours |
+|---|---:|
+| VGG19 | 2.5 |
+| ResNet50 | 2.3 |
+| ResNet101 | 3.1 |
+| DenseNet121 | 2.8 |
+| InceptionV3 | 2.2 |
+| Swin-B | 4.5 |
+| CLIP-ViT-B/16 | 3.8 |
+| DINOv2-L | 11.2 |
+| RETFound | 9.4 |
+| Ensemble + stats | 0.3 |
+| **Total** | **~42 GPU-hours** |