Bi-directional Attention Flow for Machine Comprehension

• This the original implementation of Bi-directional Attention Flow for Machine Comprehension.
• The CodaLab worksheet for the SQuAD Leaderboard submission is available here.
• For TensorFlow v1.2 compatible version, see the dev branch.
• Please contact Minjoon Seo (@seominjoon) for questions and suggestions.

0. Requirements

General

• Python (verified on 3.5.2. Issues have been reported with Python 2!)
• unzip, wget (for running download.sh only)

Python Packages

• tensorflow (deep learning library, only works on r0.11)
• nltk (NLP tools, verified on 3.2.1)
• tqdm (progress bar, verified on 4.7.4)
• jinja2 (for visaulization; if you only train and test, not needed)

1. Pre-processing

First, prepare data. Donwload SQuAD data and GloVe and nltk corpus (~850 MB, this will download files to $HOME/data):

chmod +x download.sh; ./download.sh

Second, Preprocess Stanford QA dataset (along with GloVe vectors) and save them in $PWD/data/squad (~5 minutes):

python -m squad.prepro

2. Training

The model has ~2.5M parameters. The model was trained with NVidia Titan X (Pascal Architecture, 2016). The model requires at least 12GB of GPU RAM. If your GPU RAM is smaller than 12GB, you can either decrease batch size (performance might degrade), or you can use multi GPU (see below). The training converges at ~18k steps, and it took ~4s per step (i.e. ~20 hours).

Before training, it is recommended to first try the following code to verify everything is okay and memory is sufficient:

python -m basic.cli --mode train --noload --debug

Then to fully train, run:

python -m basic.cli --mode train --noload

You can speed up the training process with optimization flags:

python -m basic.cli --mode train --noload --len_opt --cluster

You can still omit them, but training will be much slower.

Note that during the training, the EM and F1 scores from the occasional evaluation are not the same with the score from official squad evaluation script. The printed scores are not official (our scoring scheme is a bit harsher). To obtain the official number, use the official evaluator (copied in squad folder, squad/evaluate-v1.1.py). For more information See 3.Test.

3. Test on the SQuAD dataset

To test, run:

python -m basic.cli

Similarly to training, you can give the optimization flags to speed up test (5 minutes on dev data):

python -m basic.cli --len_opt --cluster

This command loads the most recently saved model during training and begins testing on the test data. After the process ends, it prints F1 and EM scores, and also outputs a json file ($PWD/out/basic/00/answer/test-####.json, where #### is the step # that the model was saved). Note that the printed scores are not official (our scoring scheme is a bit harsher). To obtain the official number, use the official evaluator (copied in squad folder) and the output json file:

python squad/evaluate-v1.1.py $HOME/data/squad/dev-v1.1.json out/basic/00/answer/test-####.json

3.1 Loading from pre-trained weights

Instead of training the model yourself, you can choose to use pre-trained weights that were used for SQuAD Leaderboard submission. Refer to this worksheet in CodaLab to reproduce the results. If you are unfamiliar with CodaLab, follow these simple steps (given that you met all prereqs above):

1. Download save.zip from the worksheet and unzip it in the current directory.
2. Copy glove.6B.100d.txt from your glove data folder ($HOME/data/glove/) to the current directory.
3. To reproduce single model:

basic/run_single.sh $HOME/data/squad/dev-v1.1.json single.json

This writes the answers to single.json in the current directory. You can then use the official evaluator to obtain EM and F1 scores. If you want to run on GPU (~5 mins), change the value of batch_size flag in the shell file to a higher number (60 for 12GB GPU RAM). 4. Similarly, to reproduce ensemble method:

basic/run_ensemble.sh $HOME/data/squad/dev-v1.1.json ensemble.json 

If you want to run on GPU, you should run the script sequentially by removing '&' in the forloop, or you will need to specify different GPUs for each run of the for loop.

4. Train over MS-MARCO

To train over MS-MARCO, copy the documents in marco-data within the repository to $HOME/data/marco. These documents were created from running the MArcoToSquadConverter tool under tools upon the original downloaded MS-MARCO dataset. This filters the questions that we want to study, where the answer is a subspan of the passage, and furthermore modifies the format so it matches that of the SQuAD dataset, and fits the input of the bi-directional attention flow implementation.

Preprocess the MS-Marco data:

python -m marco.prepro

Before training, it is recommended to first try the following code to verify everything is okay and memory is sufficient:

python -m marco.cli --mode train --debug

Then, train the existing model. It is important to not specify --noload so that it will load the network we trained with the SQuAD dataset:

python -m marco.cli --mode train --debug

To test on MS-MARCO, you can run:

python -m marco.cli --len_opt --cluster

5. Multi-GPU Training & Testing

Our model supports multi-GPU training. We follow the parallelization paradigm described in TensorFlow Tutorial. In short, if you want to use batch size of 60 (default) but if you have 3 GPUs with 4GB of RAM, then you initialize each GPU with batch size of 20, and combine the gradients on CPU. This can be easily done by running:

python -m basic.cli --mode train --noload --num_gpus 3 --batch_size 20

Similarly, you can speed up your testing by:

python -m basic.cli --num_gpus 3 --batch_size 20 

Demo

For now, please refer to the demo branch of this repository.