Tfignite

Tfignite is a project that stems from ignite. Ignite is a high-level library to help with training neural networks in PyTorch, while Tfignite bares similar design / api and is designed dedicated to tensorflow 2.0.

For the full API reference, read the online documentation. For example usage, check the examples directory.

Why tfignite?

Tfignite separates training/evaluation loop from model/dataset computation graph. This makes a single training/evaluation script highly portable to different project and developers only need to focus on how to build the model and dataset for their tasks. The difference from Keras is that the training/evaluation loop is not part of Model's APIs; instead, developers define model forward pass function, which is then injected into the loop defined by an Engine. Users can also register event handlers in different phases of a training/evaluation loop (for e.g. ITERATION_STARTED).

Apart from the aforementioned separation of model forward pass function and boilerplate loop in Engine (ignite has full credits for this). Tfignite further reduce boilerplate code by defining the Model, Dataset, Callback, ArgumentParser interfaces:

1. Model: Defines the create_trainer and the create_evaluator function, both of which injects a forward pass function to an Engine and returns it to the training/evaluation script.
2. Dataset: Defines a unified interface Dataset.create wrapping over tf.data.Dataset.
3. Callback: Defines a interface to group related Engine event handlers in different phases. For example, Checkpointer loads the checkpoint at the beginning of training and stores the checkpoint on EPOCH_COMPLETED.
4. ArgumentParser: Inherited from argparse.ArgumentParser, the parser pass itself to the Model and Dataset classes for parsing Model-specific and Dataset-specific arguments. This further separates the Model development and the training/evaluation script.

Installation

pip install tfignite