This is a first stable version for Transfer NLP, allowing users to:
Keep track of experiments and enforce reproducible research
Combine custom and open-source code into controlled experiments
Here are a few features available in the release:
Configuring all objects from an experiment using a json file
Running sequential jobs for the same experiment using different sets of parameters (parameter tuning, ablation studies...)
Keep track of your experiments and make them reproducible / incrementally improvable
Allow dynamic re-creation of any instantiated object during training through object factories
Use several basic building blocks: Vocabulary class, PyTorch optimizer, Predictors...
Transfer Learning: use the BasicTrainer to fine-tune pre-trained models to your custom downstream tasks.