padertorch

Padertorch is designed to simplify the training of deep learning models written with PyTorch. While focusing on speech and audio processing, it is not limited to these application areas.

Highlights

• Easily extensible: Write your own network modules and models based on padertorch.Module and padertorch.Model.

• Seamless integration: You provide your own data and model. We provide the trainer to wrap around your model and data. Calling train starts the training.

• Fast prototyping: The trainer and models are all compatible with sacred and easily allow hyperparameter changes over the command line. Check out the configurable module and the examples for how it works.

• Trainer: Our padertorch.Trainer takes care of the repetitive training loop and allows you to focus on network tuning. Included are features such as:

  • Periodically executed validation runs

  • Automatic checkpointing

    The parameters of the model and the state of the trainer are periodically saved. The checkpoint interval and number of total checkpoints are customizable: Keep one, some or all checkpoints. We also keep track of the best checkpoint on the validation data given some metric.

    • Resume from the latest checkpoint if the training was interrupted.

  • Learning rate scheduling

  • Backoff: Restore the best checkpoint and change the learning rate if the loss is not decreasing.

  • Hooks: Extend the basic features of the trainer with your own functionality.

  • Test run

    The trainer has a test_run function to train the model for few iterations and test if

    • the model is executable (burn test),
    • the validation is deterministic/reproducible, and
    • the model changes the parameter during training.

  • Virtual minibatch

    • The Trainer usually does not know if the model is trained with a single example or multiple examples (minibatch), because the examples that are yielded from the dataset are directly forwarded to the model.
    • When the virtual_minibatch_size option is larger than one, the trainer calls the forward and backward step virtual_minibatch_size times before applying the gradients. This increases the number of examples over which the gradient is calculated while the memory consumption stays similar, e.g., with virtual_minibatch_size=2 the gradient is accumulated over two (batched) examples before calling optimizer.step(); optimizer.zero_grad(). See here for a more thorough explanation.

• Logging: As logging backend, we use tensorboardX to generate a tfevents file that can be visualized from a tensorboard. Custom values to be logged can be defined in subclasses of padertorch.Model.

• Multi-GPU training: Easily deploy your model onto multiple GPUs to increase the total batch size and speed up the training. See here for implementation details and the example for how to enable it.

• Support for lazy data preparation: Ever tired of pre-computing features, taking up time and space on the hard disk? Padertorch works with lazy dataloaders (e.g., lazy_dataset) which extract the features on the fly!

• Hands-on examples: We provide models and training scripts which help you getting started with padertorch.

Installation

$ git clone https://github.com/fgnt/padertorch.git
$ cd padertorch && pip install -e .[all]

This will install all dependencies. For a light installation, you can drop [all].

Requirements

• Python 3
• matplotlib
• Installed by padertorch (core dependencies):
  • torch
  • tensorboardX
  • einops
  • tqdm
  • natsort
  • lazy_dataset
  • IPython
  • paderbox
• To run the examples (included in [all]):
  • sacred
  • torchvision
  • pb_bss

Getting Started

A Short Explanation of padertorch.Module and padertorch.Model

You can build your models upon padertorch.Module and padertorch.Model. Both expect a forward method which has the same functionality as the forward call of torch.nn.Module: It takes some data as input, applies some transformations, and returns the network output:

class MyModel(pt.Module):

  def forward(self, example):
      x = example['x']
      out = transform(x)
      return out

Additionally, padertorch.Model expects a review method to be implemented which takes the input and output of the forward call as its inputs from which it computes the training loss and metrics for logging in tensorboard. The following is an example for a classification problem using the cross-entropy loss:

import torch

class MyModel(pt.Model):

  def forward(self, example):
      output = ...  # e.g., output has shape (N, C), where C is the number of classes
      return output

  def review(self, example, output):
      # loss computation, where example['label'] has shape (N,)
      ce_loss = torch.nn.CrossEntropyLoss()(output, example['label'])
      # compute additional metrics
      with torch.no_grad():
          prediction = torch.argmax(output, dim=1)
          accuracy = (prediction == example['label']).float().mean()
      return {
          'loss': ce_loss,
          'scalars': {'accuracy': accuracy}
    }

See padertorch.summary.tbx_utils.review_dict for how the review dictionary should be constructed. For each training step, the trainer calls forward, passes its output to review and performs a backpropagation step on the loss. Typically, the input to the forward of a Module is a Tensor, while for a Model, it is a dictionary which contains additional entries, e.g., labels, which are needed in the review. This is only a recommendation and there is no restriction for the input type.

While these two methods are mandatory, you are free to add any further methods to your models. Since a Module does not need a review method, it can be used as a component of a Model.

How to Integrate your Data and Model with the Trainer

The trainer works with any kind of iterable, e.g., list, torch.utils.data.DataLoader or lazy_dataset.Dataset. The train method expects an iterable as input which yields training examples or minibatches of examples that are forwarded to the model without being interpreted by the trainer, i.e., the yielded entries can have any data type and only the model has to be designed to work with them. In our examples, the iterables always yield a dict.

The Model implements an example_to_device which is called by the trainer to move the data to a CPU or GPU. Per default, example_to_device uses padertorch.data.example_to_device which recursively converts numpy arrays to Tensors and moves all Tensors to the available device. The training device can be directly provided to the call of Trainer.train. Otherwise, it is automatically set by the trainer according to torch.cuda.is_available.

Optionally, you can add an iterable with validation examples by using Trainer.register_validation_hook. Some functionalities (e.g., keeping track of the best checkpoint) are then performed on the validation data.

A simple sketch for the trainer setup is given below:

import torch
import padertorch as pt

train_dataset = ...
validation_dataset = ...

class MyModel(pt.Model):
    def __init__(self):
        super().__init__()
        self.net = torch.nn.Sequential(...)

    def forward(self, example):
        output = self.net(example['observation'])
        return output

    def review(self, example, output):
        loss = ...  # calculate loss
        with torch.no_grad():
            ...  # calculate general metrics
            if self.training:
                ...  # calculate training specific metrics
            else:
                ...  # calculate validation specific metrics
        return {
            'loss': loss,
            'scalars': {
                'accuracy': ...,
                ...
            },
        }  # Furthers keys: 'images', 'audios', 'histograms', 'texts', 'figures'


trainer = padertorch.Trainer(
    model=MyModel(),
    storage_dir=pt.io.get_new_storage_dir('my_experiment'),  # checkpoints of the trained model are stored here
    optimizer=pt.train.optimizer.Adam(),
    loss_weights=None,
    summary_trigger=(1, 'epoch'),
    checkpoint_trigger=(1, 'epoch'),
    stop_trigger=(1, 'epoch'),
    virtual_minibatch_size=1,
)
trainer.test_run(train_dataset, validation_dataset)
trainer.register_validation_hook(validation_dataset)
trainer.train(train_dataset)

See the trainer for an explanation of its signature. If you want to use pt.io.get_new_storage_dir to manage your experiments, you have to define an environment variable STORAGE_ROOT which points to the path where all your experiments will be stored, i.e., in the example above, a new directory under $STORAGE_ROOT/my_experiment/1 will be created. Otherwise, you can use pt.io.get_new_subdir where you can directly input the path to store your model without defining an environment variable.

Features for Application in Deep Learning

Padertorch provides a selection of frequently used network architectures and functionalities such as activation and normalization, ready for you to integrate into your own models.

• Multi-layer Feed-Forward: Multiple fully-connected layers with non-linearity and dropout.
• CNN (currently subject to breaking changes and hardly any documentation): 1d- and 2d-CNNs with residual connections, dropout, gated activations, batch and sequence norm and correct handling of down- and upsampling.
• Normalization: Perform normalization of arbitrary axes/dimensions of your features, keep track of running statistics and apply learnable affine transformation.
• Collection of activation functions: Fast access of various activation functions with just a string.
• Losses: We provide an implementation of the Kullback-Leibler divergence and different regression objectives (SI-SNR, log-MSE and others).

Support for Sequential and Speech Data

Padertorch especially offers support for training with sequential data such as:

• Masking: Calculate a mask which has non-zero entries for non-padded positions and zero entries for padded positions in the sequence.
• Segmenting: Segment a sequence into smaller chunks of the same length.
• Dual-Path RNN (DPRNN): See the paper.
• WaveNet: Synthesize waveforms from spectrograms (supports fast inference with nv_wavenet).
• Visualization in tensorboard: Prepare spectrograms and speech masks for visualization and audio for playback in tensorboard.

We also provide a loss wrapper for permutation-invariant training (PIT) criteria which are, e.g., commonly used in (speech) source separation.

Further Reading

Have a look at the following links to get the most out of your experience with padertorch:

• Configurable: Provides a thorough explanation of our configurable module which allows to create model instances from a config dict.
• Sacred: Explains how to use sacred in combination with pt.Configurable.
• contrib/: Unordered collection of advanced and experimental features. Subject to breaking changes so be careful with relying on it too much. But it might provide ideas for your own implementations.
• contrib/examples/: Shows advanced usage of padertorch with actual data and models. A good starting point to get ideas for writing own models and experiments. See here for a guide to navigate through the examples and recommendations for which examples to start with.