Releases · facebookresearch/mmf

Features

Multi-tasking support: Multitasking over various datasets available in Pythia
Distributed training support
Better Customization Support: Use your custom losses, metrics, optimizers and lot of more
Standardized Trainer API: A standard trainer API to fit most of your use-cases, if not inherit and build your own trainer
Processors: Use processors to build out your datasets easily and without pain
SampleList and Sample: Use SampleList and Sample to have more granular control over what you pass and a single unified API for accessing attributes whether inside a dataset, a single sample or a batch
Feature Extraction: A new simple script to extract out features and related information from VQA MaskRCNN Benchmark
Registry: No need to manually load datasets and models anymore, registry takes care of loading your models, datasets and other classes at the fly. Think of registry as a singleton containing all that you need.
Tensorboard Logging: Tensorboard logging is now provided by default.
Configuration: Better hierarchal configuration system for better separation of concerns.
Checkpointing: Better control checkpointing and resuming
Logging: Better logging is provided now with eta, individual val losses and metrics. Just pass them back from your model and everything logs automatically
EvalAI Evaluation: Now, directly output JSON files that can be uploaded to EvalAI
Early Stopping
New Embeddings Support: FastText, GloVe, BERT etc.

Following new datasets were added:

Note: There are a lot of breaking changes in the API from v0.1. Refer to the documentation to learn more on how to work with Pythia v0.3.