Hydra pipelines#

Pipelines and utils#

batorch.utils.pipelines.mcmc

Pipeline for running SGMCMC algorithms.

batorch.utils.pipelines.train_loop

Function that performs the a basic training loop for estimating the maximum likelihood (MLE) or maximum a posteriori (MAP).

batorch.utils.pipelines.test_loop

Function that sweeps through a test dataset and outputs the model predictions.
batorch.utils.pipelines.train_loop(num_epoch, model, train_loader, val_loader, optimizer, scheduler, loglike, writer, device, fold=1)[source]#

Function that performs the a basic training loop for estimating the maximum likelihood (MLE) or maximum a posteriori (MAP).

Parameters

  • num_epoch – Number of sweeps through the dataset.

  • model – Neural network.

  • train_loader – Dataloader that sweeps through the training dataset.

  • val_loader – Dataloader that sweeps through the validation dataset.

  • optimizer – Optimizer.

  • scheduler – Step size scheduler

  • loglike – Negative log likelihood function

  • writer – Tensorboard summary writer

  • device – Device to be used (cpu or cuda)

  • fold – The fold number, defaults to 1

Returns

Dictionnary with losses

batorch.utils.pipelines.test_loop(model, test_loader, loglike, writer, device, fold=1, enable_dropout=False)[source]#

Function that sweeps through a test dataset and outputs the model predictions.

Parameters

  • model – Neural network

  • test_loader – Dataloader that sweeps through a test dataset

  • loglike – Negative log likelihood function

  • writer – Tensorboard SummaryWriter

  • device – Device to be used (cpu or cuda)

Returns

Dictionnary with losses and predictions

batorch.utils.pipelines.mcmc(sampler: Sampler, datamodule: DataModule, max_burnin_iterations: int, max_sampling_iterations: int, device: torch.device, quantization=None)[source]#

Pipeline for running SGMCMC algorithms.

Parameters

  • sampler – SGMCMC sampler.

  • datamodule – DataModule holding at least a training dataset and a test dataset.

  • max_burnin_iterations – Number of burnin iterations.

  • max_sampling_iterations – Number of iterations after burnin.

  • device – Device (cuda or cpu).

  • quantization – Quantization object, default is None.

Entry points#

hydra_entries.mcmc_pipeline(config: omegaconf.DictConfig)#

Pipeline for Bayesian regression with SGMCMC methods.

The input configuration file is parsed by hydra and passed to mcmc().

Parameters

config – Configuration given by a omegaconf.DictConfig.

hydra_entries.map_pipeline(config: omegaconf.DictConfig)#

Pipeline for maximum a posteriori or maximum likelihood estimation.

The input configuration file is parsed by hydra. The functions train_loop() and batorch.utils.pipelines.test_loop() are used to train the model.

Parameters

config – Configuration given by a omegaconf.DictConfig.

hydra_entries.ensemble_pipeline(config: omegaconf.DictConfig)#

Pipeline for Bayesian regression with ensembles.

The input configuration file is parsed by hydra. The functions train_loop() and batorch.utils.pipelines.test_loop() are used to train an ensemble of models.

Parameters

config – Configuration given by a omegaconf.DictConfig.

hydra_entries.mcdropout_pipeline(config: omegaconf.DictConfig)#

Pipeline for Bayesian regression Monte Carlo dropout.

The input configuration file is parsed by hydra. The input configuration file is parsed by hydra. The functions train_loop() and batorch.utils.pipelines.test_loop() are used to perform Monte Carlo dropout.

Parameters

config – Configuration given by a omegaconf.DictConfig.