Model

Pretrained model

The pretrained model from which the fine-tuning will starts.

Number of fine-tuned layers

The number of backbone layers from the pretrained model to fine-tune.

Note: the last fully-connected layer is trained from scratch.

Optimization

Optimizer

The optimizer to use for parameters optimization.

Learning rate scheduler

The method to adjust the learning rate based on the number of epochs.

Initial learning rate

The initial learning rate for gradient descent. Smaller values lead to better convergence, but take more time.

Weight decay

Weight decay used for L2 penalty

Batch size

The number of samples to include in each mini-batch.

Number of epochs

The number of epochs for learning. Higher values lead to better convergence, but take more time.

Train/Validation split ratio

Ratio for splitting train & validation sets to compute metrics for best epoch & early stopping.

Early stopping

Stop the training when no improvement is observed on the loss for X epochs

Early stopping min delta

Minimal value for a score change to be considered significant

Early stopping patience

The number of epochs waited before stopping the training if no significant improvement in LR is found