finetuner.callback module#

class finetuner.callback.BestModelCheckpoint(monitor='val_loss', mode='auto')[source]#

Bases: object

Callback to save the best model across all epochs

An option this callback provides include: - Definition of ‘best’; which quantity to monitor and whether it should be

maximized or minimized.

Parameters
  • monitor (str) – if monitor=’train_loss’ best model saved will be according to the training loss, while if monitor=’val_loss’ best model saved will be according to the validation loss.

  • mode (str) – one of {‘auto’, ‘min’, ‘max’}. The decision to overwrite the currently saved model is made based on either the maximization or the minimization of the monitored quantity. For an evaluation metric, this should be max, for val_loss this should be min, etc. In auto mode, the mode is set to min if monitor=’loss’ or monitor=’val_loss’ and to max otherwise.

monitor: str = 'val_loss'#
mode: str = 'auto'#
class finetuner.callback.TrainingCheckpoint(last_k_epochs=1)[source]#

Bases: object

Callback that saves the tuner state at every epoch or the last k epochs.

Parameters

last_k_epochs (int) – This parameter is an integer. Only the most recent k checkpoints will be kept. Older checkpoints are deleted.

last_k_epochs: int = 1#
class finetuner.callback.WandBLogger(token, wandb_args=<factory>)[source]#

Bases: object

Weights & Biases logger to log metrics for training and validation. To use this logger, make sure to have a WandB account created, install the WandB client (which you can do using pip install wandb)

Parameters
  • token (str) – weights and biases authentication key.

  • wandb_args (dict) – Keyword arguments that are passed to wandb.init function.

token: str#
wandb_args: dict#
class finetuner.callback.EarlyStopping(monitor='val_loss', mode='auto', patience=2, min_delta=0, baseline=None)[source]#

Bases: object

Callback to stop training when a monitored metric has stopped improving. A finetuner.fit() training loop will check at the end of every epoch whether the monitored metric is still improving or not.

Parameters
  • monitor (str) – if monitor=’train_loss’ best model saved will be according to the training loss, while if monitor=’val_loss’ best model saved will be according to the validation loss.

  • mode (str) – one of {‘auto’, ‘min’, ‘max’}. The decision to overwrite the current best monitor value is made based on either the maximization or the minimization of the monitored quantity. For an evaluation metric, this should be max, for val_loss this should be min, etc. In auto mode, the mode is set to min if monitor=’loss’ or monitor=’val_loss’ and to max otherwise.

  • patience (int) – integer, the number of epochs after which the training is stopped if there is no improvement. For example for patience = 2’, if the model doesn’t improve for 2 consecutive epochs, the training is stopped.

  • min_delta (int) – Minimum change in the monitored quantity to qualify as an improvement, i.e. an absolute change of less than min_delta, will count as no improvement.

  • baseline (Optional[float]) – Baseline value for the monitored quantity. Training will stop if the model doesn’t show improvement over the baseline.

monitor: str = 'val_loss'#
mode: str = 'auto'#
patience: int = 2#
min_delta: int = 0#
baseline: Optional[float] = None#
class finetuner.callback.EvaluationCallback(query_data, index_data=None, batch_size=8, exclude_self=True, limit=20, distance='cosine')[source]#

Bases: object

A callback that uses the Evaluator to calculate IR metrics at the end of each epoch. When used with other callbacks that rely on metrics, like checkpoints and logging, this callback should be defined first, so that it precedes in execution.

Parameters
  • query_data (Union[DocumentArray, str]) – Search data used by the evaluator at the end of each epoch, to evaluate the model.

  • index_data (Union[DocumentArray, str, None]) – Index data or catalog used by the evaluator at the end of each epoch, to evaluate the model.

  • batch_size (int) – Batch size for computing embeddings.

  • metrics – A List of the metrics to calculate. If set to None, default metrics are computed.

  • exclude_self (bool) – Whether to exclude self when matching.

  • limit (int) – The number of top search results to consider when computing the evaluation metrics.

  • distance (str) – The type of distance metric to use when matching query and index docs, available options are 'cosine', 'euclidean' and 'sqeuclidean'.

query_data: Union[docarray.DocumentArray, str]#
index_data: Optional[Union[docarray.DocumentArray, str]] = None#
batch_size: int = 8#
exclude_self: bool = True#
limit: int = 20#
distance: str = 'cosine'#