synicix_ml_pipeline.trainers package

Submodules

synicix_ml_pipeline.trainers.NNTrainer module

class synicix_ml_pipeline.trainers.NNTrainer.NNTrainer(train_dataloader, validation_dataloader, test_dataloader, device, model_class, model_class_params, optimizer_class, optimizer_class_params, criterion_class, criterion_class_params, model_save_path, max_epochs, training_strip_length, sucessive_validation_score_decline_tolerance, fp16=False, fp16_opt_level=None)

Bases: object

NNTrainer is a trainer class provided with synicix_ml_pipeline. It handles training and saving/loading of models with some additional useful features.

One of the most notable features it provide is integrated fp16 or mixed precision training which can drastically improve your network training speed at little to no loss in performance This is provided via Nvidia’s the apex libaray: https://github.com/NVIDIA/apex

The other feature is a built in early stoping algorithm which was implemented base off this paper: https://page.mi.fu-berlin.de/prechelt/Biblio/stop_tricks1997.pdf

eval_step(data, return_outputs_targets_and_loss=False)

Evluate step function to handle training of one batch

Parameters:

data (tuple): Data point sample obtain from pytorch dataloader return_outputs_targets_and_loss If true then it will return outputs_targets_and_losses as a dict

Returns:

float: Model’s test dataset loss if return_outputs_targets_and_loss False dict(outputs, target, loss): If return_outputs_targets_and_loss True

evaluate(return_outputs_targets_and_loss=False)

Run the test dataset through the model and return the average loss

Parameters:

return_outputs_targets_and_loss (bool): If true then it will return outputs_targets_and_losses as a dict

Returns:

float: Model’s test dataset loss if return_outputs_targets_and_loss False dict(outputs, target, loss): If return_outputs_targets_and_loss True

load_best_performing_model()

Load the best performing model which is the previous save file of the lowest validation score form model_save_path

Parameters: Returns:

None

load_model(path)

Load model, optmizer, and amp from path

Parameters:

path (str): path of where to load model file from

Returns:

None

save_model()

Save model optmizer, and amp to model_save_path

Parameters:

None

Returns:

None

should_early_stop(current_epoch)

Checks wheter to early stop or not based on the the previous validation score and training strip performance

Parameters:

current_epoch (int): Current epoch, mainly use for printing

Returns:

bool: Whether to early stop or not

train()

Function to start the training

Parameters:

None

Returns:

None

train_step(data)

Training step function to handle training of one batch

Parameters:

data (tuple): Data point sample obtain from pytorch dataloader

Returns:

pytorch.loss: Model’s loss for the given batch Tensor: Regularlization loss for the given batch

validate()

Run the validation dataset through the model and return the average loss

Parameters:

None

Returns:

float: model’s validation dataset loss

Module contents