Stable release (stronger baseline) for VisDial challenge 2019.

Summarizing changes with PR #7:

A few bug fixes and tweaks for a stronger baseline.

This improves MRR from 0.5845 to 0.6155 and NDCG from 0.5070 to 0.5315 on val.

Changes:

• Switched off dropout during evaluation on val in train.py.
• Shuffling batches during training (shuffle=True to DataLoader).
• Explicitly clearing GPU memory cache with torch.cuda.empty_cache(). Negligible time hit on single GPU, and fits batch sizes of up to 32 x no. of GPUs. There's some time gain when training with larger batch sizes.
• Added a linear learning rate warm up (https://arxiv.org/abs/1706.02677), followed by multi-step decaying.
• Using a multi-layer LSTM + dropout for the decoder.
• Switched from dot-product attention to a richer element-wise multiplication + fc layer attention. (The network can learn dot-product attention if it needs to.)

Stable release for VisDial challenge 2019.

This tag links to the official released code for VisDial challenge 2019. It does not preserve backward compatibility with v2018.

What's new with v2019?

1. Almost a complete rewrite of v2018, which increased speed, readability, modularity, and extensibility.
2. Multi-GPU support - try out specifying GPU ids to train/evaluate scripts as: --gpu-ids 0 1 2 3
3. Docker support - we provide a Dockerfile which can help you set up all the dependencies with ease.
4. Stronger baseline - our Late Fusion Encoder is equipped with Bottom-up Top-Down attention. We also provide pre-extracted image features (links below).
5. Minimal pre-processed data - no requirement to download tens of pre-processed data files anymore (were typically referred to as visdial_data.h5 and visdial_params.json).

Stable release for VisDial challenge 2018.

This tag links to the official released code for VisDial challenge 2018.

Code for VisDial challenge 2019 is subject to change and may break backwards compatibility.