ebes package

Subpackages

Submodules

ebes.trainer module

class ebes.trainer.Trainer(*, model=None, loss=None, optimizer=None, lr_scheduler=None, train_loader=None, val_loader=None, metrics=None, run_name=None, total_iters=None, total_epochs=None, patience=-1, iters_per_epoch=10000, ckpt_dir=None, ckpt_replace=True, ckpt_track_metric='epoch', ckpt_resume=None, device='cpu', metrics_on_train=False, verbose=True)

Bases: object

A base class for all trainers.

best_checkpoint()

Return the path to the best checkpoint

Return type:: Path

compute_metrics(phase)

Compute and log metrics.

The metrics are computed based on the whole epoch data, so the granularity of metrics is epoch, so when the metrics are not None, the epoch is not None to.

Parameters:: phase (Literal['train', 'val']) – wether the metrics were collected during train or validatoin.
Return type:: dict[str, Any]

property device: str

load_best_model()

Loads the best model to self._model according to the track metric.

Return type:: None

load_ckpt(ckpt_fname, strict=True)

Load model, optimizer and scheduler states.

Parameters:: ckpt_fname (str | PathLike) – path to checkpoint.
Return type:: None

property lr_scheduler: _LRScheduler | None

property model: Module | None

property optimizer: Optimizer | None

run()

Train and validate model.

Return type:: None

property run_name

save_ckpt(ckpt_path=None)

Save model, optimizer and scheduler states.

Parameters:: ckpt_path (Union[str, PathLike, None]) – path to checkpoints. If ckpt_path is a directory, the checkpoint will be saved there with epoch, loss an metrics in the filename. All scalar metrics returned from compute_metrics are used to construct a filename. If full path is specified, the checkpoint will be saved exectly there. If None ckpt_dir from construct is used with subfolder named run_name from Trainer’s constructor.
Return type:: None

train(iters)

Return type:: dict[str, Any]

property train_loader: Iterable[Batch] | None

property val_loader: Iterable[Batch] | None

validate(loader=None)

Return type:: dict[str, Any]

ebes.types module

class ebes.types.Batch(*, lengths, time, index=None, num_features=None, cat_features=None, target=None, cat_features_names=None, num_features_names=None, cat_mask=None, num_mask=None)

Bases: object

cat_features: Tensor | None = None

cat_features_names: list[str] | None = None

cat_mask: Tensor | None = None

index: Tensor | ndarray | None = None

lengths: Tensor

num_features: Tensor | None = None

num_features_names: list[str] | None = None

num_mask: Tensor | None = None

pop_target()

Return type:: Tensor | None

target: Tensor | None = None

time: ndarray | Tensor

to(device)

class ebes.types.NHReturn(pre_event_intensities_of_gt, non_event_intensity, clustering_loss, lengths, clus_labels, pred_labels)

Bases: object

clus_labels: Tensor

clustering_loss: Tensor

lengths: Tensor

non_event_intensity: Tensor

pre_event_intensities_of_gt: Tensor

pred_labels: Tensor

to(device)

class ebes.types.NHSeq(*, tokens, lengths, time, masks=None, clustering_loss, clus_labels)

Bases: Seq

clus_labels: Tensor

clustering_loss: Tensor

class ebes.types.PrimeNetReturn(loss, cl_loss, mse_loss, correct_num, total_num)

Bases: object

cl_loss: Tensor

correct_num: int

loss: Tensor

mse_loss: Tensor

to(device)

total_num: int

class ebes.types.Seq(*, tokens, lengths, time, masks=None)

Bases: object

lengths: Tensor

masks: Tensor | None = None

time: Tensor

to(device)

tokens: Tensor

ebes package

Subpackages

Submodules

ebes.trainer module

ebes.types module

Module contents