Upload folder using huggingface_hub

README.md

@@ -158,11 +158,13 @@ python scripts/train_model.py
 
 ## Technical Details
 
-- **Architecture**: 3-layer CNN with BatchNormalization, GlobalAveragePooling, and data augmentation
-- **Input**: 28x28 grayscale images
-- **Dataset**: Google Quick Draw! numpy bitmap format
+- **Architecture**: 4-layer CNN (32→64→128→256 filters) with BatchNormalization, GlobalAveragePooling2D, and online data augmentation (rotation, translation, zoom)
+- **Input**: 28×28 grayscale images (resized from canvas, colors inverted to match training format)
+- **Dataset**: Google Quick Draw! numpy bitmap format, 15,000 samples per class
+- **Training split**: 70% train / 12.5% val / 17.5% test (stratified)
+- **Validation accuracy**: ~72% across 120 classes (random baseline: ~0.8%)
 - **Framework**: TensorFlow/Keras
-- **Model size**: ~6.4 MB
+- **Model size**: ~5.9 MB
 - **Inference time**: <100ms
 - **API**: Flask with rate limiting and CORS
 

model/best_model.keras

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5d78526921aecdead462488e540a7e2214a00f00e2dbc2fe48e2f7ed93b3b691
-size 6639665
+oid sha256:ff10525980803a6b0115c58ef8274b44c770ac16ddf8c49174e90c4512d70eae
+size 6139073