---
license: gpl-3.0
tags:
- classifier
- CNN
- convolutions
---

# Box Classifier

A bunch of CNNs (in parameter-sizes) trained to classify whether a checkbox is filled or not.

The training code, data generator can be found here: https://github.com/papaj2139/box_classifier

# Models

Every model was trained on 10 epochs with 500 batches per epoch on 20000 images.
Every image is 32x32.

**model-5k.bin (5017 parameters)**
Architecture:
- Conv2D(1→8, 3×3) + ReLU + MaxPool(2)
- Conv2D(8→40, 3×3) + ReLU + MaxPool(2)
- GlobalAvgPool
- Dense(40→48) + ReLU + Dropout(0.2)
- Dense(48→1)


Loss trajectory:

![loss-5k](https://cdn-uploads.huggingface.co/production/uploads/69804655d742adcb7e7d6df2/jFjwOkBKOj9Vuy3GI4t_5.png)


Final test Acc: **99.72%**
Final train Acc: **99.57%**

**model-20k.bin (20373 parameters)**
Architecture:
- Conv2D(1→16, 3×3, pad=1), ReLU, MaxPool(2)
- Conv2D(16→100, 3×3, pad=1), ReLU, MaxPool(2)
- GlobalAvgPool2D
- Dense(100→56), ReLU, Dropout(0.2)
- Dense(56→1)

Loss trajectory:

![loss](https://cdn-uploads.huggingface.co/production/uploads/69804655d742adcb7e7d6df2/iqCkMV7G97cWNGveG3-gq.png)


Final test Acc: **99.85%**
Final train Acc: **99.68%**

**model-100k.bin (99713 parameters)**
Architecture:
- Conv2D(1→32, 3×3, pad=1), ReLU, MaxPool(2)
- Conv2D(32→128, 3×3, pad=1), ReLU, MaxPool(2)
- GlobalAvgPool2D
- Dense(128→480), ReLU, Dropout(0.3)
- Dense(480→1)

Loss trajectory:

![loss](https://cdn-uploads.huggingface.co/production/uploads/69804655d742adcb7e7d6df2/l97Ok7nfLVcr20gSmkhpF.png)


Final test Acc: **99.65%**
Final train Acc: **99.81%**

# License
GPLv3