crested.tl.Crested

crested.tl.Crested#

class crested.tl.Crested(data, model=None, config=None, project_name=None, run_name=None, logger=None, seed=None)#

Main class to handle training, testing, predicting and calculation of contribution scores.

Parameters:

  • data (AnnDataModule) – AnndataModule object containing the data.

  • model (Model | None (default: None)) – Model architecture to use for training.

  • config (TaskConfig | None (default: None)) – Task configuration (optimizer, loss, and metrics) for use in tl.Crested.

  • project_name (str | None (default: None)) – Name of the project. Used for logging and creating output directories. If not provided, the default project name “CREsted” will be used.

  • run_name (str | None (default: None)) – Name of the run. Used for wandb logging and creating output directories. If not provided, the current date and time will be used.

  • logger (str | None (default: None)) – Logger to use for logging. Can be “wandb”, “tensorboard”, or “dvc” (tensorboard not implemented yet) If not provided, no additional logging will be done.

  • seed (int (default: None)) – Seed to use for reproducibility. WARNING: this doesn’t make everything fully reproducible, especially on GPU. Some (GPU) operations are non-deterministic and simply can’t be controlled by the seed.

Examples

>>> from crested.tl import Crested
>>> from crested.tl import default_configs
>>> from crested.tl.data import AnnDataModule
>>> from crested.tl.zoo import deeptopic_cnn

>>> # Load data
>>> anndatamodule = AnnDataModule(anndata, genome="path/to/genome.fa")
>>> model_architecture = deeptopic_cnn(seq_len=1000, n_classes=10)
>>> configs = default_configs("topic_classification")

>>> # Initialize trainer
>>> trainer = Crested(
...     data=anndatamodule,
...     model=model_architecture,
...     config=configs,
...     project_name="test",
... )

>>> # Fit the model
>>> trainer.fit(epochs=100)

>>> # Evaluate the model
>>> trainer.test()

Methods table#

fit([epochs, mixed_precision, ...])

Fit the model on the training and validation set.

load_model(model_path[, compile])

Load a (pretrained) model from a file.

test([return_metrics])

Evaluate the model on the test set.

transferlearn([epochs_first_phase, ...])

Perform transfer learning on the model.

Methods#

Crested.fit(epochs=100, mixed_precision=False, model_checkpointing=True, model_checkpointing_best_only=True, model_checkpointing_metric='val_loss', model_checkpointing_mode='min', early_stopping=True, early_stopping_patience=10, early_stopping_metric='val_loss', early_stopping_mode='min', learning_rate_reduce=True, learning_rate_reduce_patience=5, learning_rate_reduce_metric='val_loss', learning_rate_reduce_mode='min', save_dir=None, custom_callbacks=None)#

Fit the model on the training and validation set.

Parameters:

  • epochs (int (default: 100)) – Number of epochs to train the model.

  • mixed_precision (bool (default: False)) – Enable mixed precision training.

  • model_checkpointing (bool (default: True)) – Save model checkpoints.

  • model_checkpointing_best_only (bool (default: True)) – Save only the best model checkpoint.

  • model_checkpointing_metric (str (default: 'val_loss')) – Metric to monitor to choose best models.

  • model_checkpointing_mode (str (default: 'min')) – ‘max’ if a high metric is better, ‘min’ if a low metric is better

  • early_stopping (bool (default: True)) – Enable early stopping.

  • early_stopping_patience (int (default: 10)) – Number of epochs with no improvement after which training will be stopped.

  • early_stopping_metric (str (default: 'val_loss')) – Metric to monitor for early stopping.

  • early_stopping_mode (str (default: 'min')) – ‘max’ if a high metric is better, ‘min’ if a low metric is better

  • learning_rate_reduce (bool (default: True)) – Enable learning rate reduction.

  • learning_rate_reduce_patience (int (default: 5)) – Number of epochs with no improvement after which learning rate will be reduced.

  • learning_rate_reduce_metric (str (default: 'val_loss')) – Metric to monitor for reducing the learning rate.

  • learning_rate_reduce_mode (str (default: 'min')) – ‘max’ if a high metric is better, ‘min’ if a low metric is better

  • save_dir (str | None (default: None)) – Directory for saving model to. Default to project name.

  • custom_callbacks (list | None (default: None)) – List of custom callbacks to use during training.

Return type:

None

Crested.load_model(model_path, compile=True)#

Load a (pretrained) model from a file.

Parameters:

  • model_path (str | PathLike) – Path to the model file.

  • compile (bool (default: True)) – Compile the model after loading. Set to False if you only want to load the model weights (e.g. when finetuning a model). If False, you should provide a TaskConfig to the Crested object before calling fit.

Return type:

None

Crested.test(return_metrics=False)#

Evaluate the model on the test set.

Make sure to load a model first using Crested.load_model() before calling this function. Make sure the model is compiled before calling this function.

Parameters:

return_metrics (bool (default: False)) – Return the evaluation metrics as a dictionary.

Return type:

dict | None

Returns:

Evaluation metrics as a dictionary or None if return_metrics is False.

Crested.transferlearn(epochs_first_phase=50, epochs_second_phase=50, learning_rate_first_phase=0.0001, learning_rate_second_phase=1e-06, freeze_until_layer_name=None, freeze_until_layer_index=None, set_output_activation=None, **kwargs)#

Perform transfer learning on the model.

The first phase freezes layers up to a specified layer (if provided), removes the later layers, adds a dense output layer, and trains with a low learning rate. The second phase unfreezes all layers and continues training with an even lower learning rate.

Ensure that you load a model first using Crested.load_model() before calling this function and have a datamodule and config loaded in your Crested object.

One of freeze_until_layer_name or freeze_until_layer_index must be provided.

Parameters:

  • epochs_first_phase (int (default: 50)) – Number of epochs to train in the first phase.

  • epochs_second_phase (int (default: 50)) – Number of epochs to train in the second phase.

  • learning_rate_first_phase (float (default: 0.0001)) – Learning rate for the first phase.

  • learning_rate_second_phase (float (default: 1e-06)) – Learning rate for the second phase.

  • freeze_until_layer_name (str | None (default: None)) – Name of the layer up to which to freeze layers. If None, defaults to freezing all layers except the last layer.

  • freeze_until_layer_index (int | None (default: None)) – Index of the layer up to which to freeze layers. If None, defaults to freezing all layers except the last layer.

  • set_output_activation (str | None (default: None)) – Set output activation if different from the previous model.

  • kwargs – Additional keyword arguments to pass to the fit method.

See also

crested.tl.Crested.fit

Contents