UpsideDownClassifier

This classifier was trained using the auto-cats-and-dogs dataset. It was trained over 5 epochs using a pretrained resent18.

The configuration for the model was

config = {
    "batch_size": 64,
    "num_epochs": 5,
    "lr": 0.005,
    "betas": (0.9, 0.999),
    "eps": 1e-6,
    "lr": 8e-3,
    "do_eval": True
}

Traning Plots

We can see in the figures below the training plots for accuracy and the loss in both, training and validation sets.

Accuracy Plot

Loss Plot

Some Results

Evaluating on the Test Set, we obtain:

• Accuracy = 0.9696

A batch with some missclassifications can be seen in the picture below.