motherofunicorns / fact_ai_project

FACT AI project - University of Amsterdam 2021

This is a reproduction study for the paper Towards Transparent and Explainable Attention Models (ACL 2020)

This code is built on the original authors' repository.

Structure and overview

• The job_scripts directory contains some of the scripts we used to train the models on the lisa cluster.
• The Notes directory contains our notes while working on this project.
• The Transparency directory on the master branch of this repository contains a duplication of the original code with some minor additions:
  • We added a seeding mechanism
  • We added the flexibility to separate the train phase and the experiment phase. This makes it easier to handle large datasets that take a long compute time.

other branches

In the other branches we tested out some hypothesis and extensions of the models. Here we provide an overview of the different branches, and details for running scripts in different branches are in the sections below:

• biLSTM contains our extension of using biLSTM as encoder instead of uni-directional LSTM
• const_attention enables forcing the attention weights of the model to be 1) equal on all hidden representations, or 2) all zeros except for the first hidden representation, or 3) all zeros except for the last hidden representation
• embedding_params do not fine-tune the pre-trained embeddings if they are used
• lime contains our experiments that compare attention weights with LIME scores
• Q_route_fix contains our investigation about whether orthogonalisation should also be applied on the Q-route when training models
• dataset_analysis was use to run additional experiments.

Installation

• Clone this repository git clone [email protected]:MotherOfUnicorns/FACT_AI_project.git
• Move to the project directory cd FACT_AI_project
• Add the parent directory of Transparency directory (which should be your current directory) to your python path export PYTHONPATH=$PYTHONPATH:$(pwd)

Requirements

• Python 3.6 or 3.7
• Install all required packages as specified:

  pip install -r requirements.txt
  

• To run the LIME experiments, also

  pip install lime
  

• Install the required packages and download the spacy en model:

  python -m spacy download en
  

• Download some nltk taggers needed for the experiments:

  python -c "import nltk; nltk.download('averaged_perceptron_tagger'); nltk.download('universal_tagset')"
  

Preparing the Datasets

Each dataset has a separate jupyter notebook in the Transparency/preprocess folder. Follow the instructions in the jupyter notebooks to download and preprocess the datasets.

Alternatively, we have made our pre-processed datasets available on the lisa cluster at /home/lgpu0136/project/Transparency/preprocess

Training & Running Experiments (on the master branch)

How to train and experiment with models using different encoders, including vanilla_lstm, ortho_lstm, diversity_lstm:

single input sequence tasks

Datasets available are sst, imdb, yelp, amazon, 20News_sports, tweet.

An example to train and test the orthogonal LSTM model on imdb dataset:

dataset_name=imdb
model_name=ortho_lstm
output_path=./experiments
python Transparency/train_and_run_experiments_bc.py --dataset ${dataset_name} --data_dir Transparency --output_dir ${output_path} --encoder ${model_name}

To use the diversity_lstm model, an additional --diversity flag is needed to specify the diversity weight:

python Transparency/train_and_run_experiments_bc.py --dataset ${dataset_name} --data_dir Transparency --output_dir ${output_path} --encoder diversity_lstm --diversity 0.5

dual input sequences tasks

Datasets available are snli, qqp, babi_1, babi_2, babi_3, cnn.

An example to train and test the orthogonal LSTM model on babi_1 dataset:

dataset_name=babi_1
model_name=ortho_lstm
output_path=./experiments
python Transparency/train_and_run_experiments_qa.py --dataset ${dataset_name} --data_dir Transparency --output_dir ${output_path} --encoder ${model_name}

Similarly, to use the diversity_lstm model, an additional --diversity flag is needed to specify the diversity weight:

python Transparency/train_and_run_experiments_qa.py --dataset ${dataset_name} --data_dir Transparency --output_dir ${output_path} --encoder ${model_name} --diversity 0.5

other arguments

Using the --job_type flag, you can specify whether you want to only train the model:

--job_type train

or to only run the experiments when you already have a trained model:

--job_type experiment

The default behaviour if you don't specify this flag is performing both train and experiment, i.e. equivalent to

--job_type both

Using the --seed flag, you can manually set a random seed. If unspecified, the default seed is zero, i.e. equivalent to

--seed 0

branch-specific tasks

The following are only available in the specified branches. To use them, first move to the correct branch:

git checkout [branchname]

biLSTM branch

The --encoder flag now can accept either one of these arguments:

• vanilla_lstm
• ortho_lstm
• diversity_lstm
• bi_lstm
• ortho_bi_lstm
• diversity_bi_lstm

const_attention branch

Using the --attention flag, you can switch between different attention weight distributions:

• tanh: normal unconstrained attention weight with tanh activation
• equal: equal attention weights to all words in the sentence
• first_only: all attention weights are concentrated on the first word
• last_only: all attention weights are concentrated on the last word (equivalent to LSTM only without attention)

embedding_params branch

The utilities in the const_attention branch are also available here.

lime branch

First make sure you have a trained model, then use the Transparency/lime_explanation_bc.py or Transparency/lime_explanation_qa.py script to compare attention weights with LIME scores, depending on the specific dataset.

single input sequence tasks

Datasets available are sst, imdb, yelp, amazon, 20News_sports, tweet.

An example to run the LIME experiment on the imdb dataset with ortho_lstm encoder:

dataset_name=imdb
model_name=ortho_lstm
output_path=./experiments # make sure this the same directory as when you trained your model
python Transparency/lime_explanation_bc.py --dataset ${dataset_name} --data_dir Transparency --output_dir ${output_path} --encoder ${model_name}

The LIME outputs will also be found in ${output_path}.

When using diversity_lstm as the encoder, an additional --diversity flag is needed to specify the diversity weight.

dual input sequences tasks

Datasets available are snli, qqp, babi_1, babi_2, babi_3, cnn.

An example to run the LIME experiment on the babi_1 dataset with ortho_lstm encoder:

dataset_name=babi_1
model_name=ortho_lstm
output_path=./experiments # make sure this the same directory as when you trained your model
python Transparency/lime_explanation_qa.py --dataset ${dataset_name} --data_dir Transparency --output_dir ${output_path} --encoder ${model_name}

The LIME outputs will also be found in ${output_path}.

When using diversity_lstm as the encoder, an additional --diversity flag is needed to specify the diversity weight.

Q_route_fix branch

Same arguments as the master branch are accepted here. But when the ortho_lstm encoder is used for any dual input sequence tasks, only the P-path is orthogonalised, and not the Q-path.

Our Results

For as long as they persist on the lisa cluster, our results (trained models + various experiments) are available to view at:

• /home/lgpu0136/experiments
• /home/lgpu0136/experiments_attentions
• /home/lgpu0136/experiments_bilstm
• /home/lgpu0136/experiments_fixed_embeddings
• /home/lgpu0136/experiments_q_route

fact_ai_project's Issues

fill in the excel sheets for results comparison: Question Answering tasks

• snli
• qqp
• babi1
• babi2
• babi3

rerun SNLI + diversity LSTM with correct diversity weight

should use 0.1 and not 0.5

Amazon dataset runs out of memory

• is it easy to run the post-training analysis with trained model?
• skip if too complicated?

redundant random seed

should remove redundant random seeds set in Binary_Classification.py and Question_Answering.py

Implement attention with equal weights

• interface with the current code base
• visualise magnitude of hidden state vectors (after each training epoch)

(optional) different initialisation for hidden states and cell states

(maybe, depends on time constraints?) do we observe different results if we initialise with zero instead of Xavier init?

fill in the excel sheets for results comparison: Binary Classification tasks

• sst
• imdb
• yelp
• amazon
• 20news

(optional) dive into code and report other metrics (numerical values) after training

Instead of only outputting graphs, also dump the numerical values for e.g. Figure 3, 4, 5, in paper

especially figure 5 where we get a pie chart without the PUNC category

(optional) retrain smaller datasets with multiple seeds

Train all BC models with 'equal' and 'last_only' attention and log in spreadsheet

Goal: determine impact on accuracy of voiding the attention mechanism

Yelp diversity lstm: different conicity values

use the changed made for Amazon dataset to rerun training for Yelp+diversity and skip other experiments