Deep CNN with PyTorch on mini MIO-TCD dataset

The MIO-TCD dataset consists of total 786,702 images with 648,959 in the classification dataset and 137,743 in the localization dataset acquired at different times of the day and different periods of the year by thousands of traffic cameras deployed all over Canada and the United States.

This dataset is a mini version of MIO-TCD, consisting of 25000 images in 5 classes, and could be downloaded here.

Class	Train	Valid	Test
articulated_truck	3000	1000	1000
background	3000	1000	1000
bus	3000	1000	1000
car	3000	1000	1000
work-van	3000	1000	1000
Total	15000	5000	5000

This CNN was implemented in PyTorch with logging and hyperparameter tuning in W&B and consists of:

1. An underfit model
2. An Overfit model
3. Best fit model (Grid search for learning_rate with Skorch and W&B)
4. Transfer learning with ResNet18 (Freezing weights)
5. Transfer learning with ResNet18 (Fine tune the full CNN)
6. Evaluation metrics

Evaluation metrics

	Train	Valid	Test
Cohen kappa	99.23	87.70	87.42
Precision	99.39	90.19	90.05
Recall	99.38	90.16	89.94
F1	99.38	90.00	89.84
Accuracy	99.38	90.16	89.94