Siamese Deep Neural Networks for semantic similarity.

This repository contains implementation of Siamese Neural Networks in Tensorflow built based on 3 different and major deep learning architectures:

• Convolutional Neural Networks
• Recurrent Neural Networks
• Multihead Attention Networks

The main reason of creating this repository is to compare well-known implementaions of Siamese Neural Networks available on GitHub mainly built upon CNN and RNN architectures with Siamese Neural Network built based on multihead attention mechanism originally proposed in Transformer model from Attention is all you need paper.

Supported datasets

Current version of pipeline supports working with 3 datasets:

• The Stanford Natural Language Inference (SNLI) Corpus
• Quora Question Pairs
• 🆕 Adversarial Natural Language Inference (ANLI) benchmark: GitHub, arXiv

Installation

Data preparation

In order to download data, execute the following commands (this process can take a while depending on your network throughput):

cd bin
chmod a+x prepare_data.sh
./prepare_data.sh

As as result of executing above script, corpora directory will be created with QQP, SNLI and ANLI data.

Dependency installation

This project was developed in and has been tested on Python 3.6. The package requirements are stored in requirements folder.

To install the requirements, execute the following command:

For GPU usage, execute:

pip install -r requirements/requirements-gpu.txt

and for CPU usage:

pip install -r requirements/requirements-cpu.txt

Training models

To train model run the following command:

python3 run.py train SELECTED_MODEL SELECTED_DATASET --experiment_name NAME --gpu GPU_NUMBER

where SELECTED_MODEL represents one of the selected model among:

• cnn
• rnn
• multihead

and SELECTED_DATASET is represented by:

• SNLI
• QQP
• ANLI

--experiment_name is an optional argument used for indicating experiment name. Default value {SELECTED_MODEL}_{EMBEDDING_SIZE}.

--gpu is an optional argument, use it in order to indicate specific GPU on your machine (the default value is '0').

Example (GPU usage): Run the following command to train Siamese Neural Network based on CNN and trained on SNLI corpus:

python3 run.py train cnn SNLI --gpu 1

Example (CPU usage): Run the following command to train Siamese Neural Network based on CNN:

python3 run.py train cnn SNLI

Training configuration

This repository contains main configuration training file placed in 'config/main.ini'.

[TRAINING]
num_epochs = 10
batch_size = 512
eval_every = 20
learning_rate = 0.001
checkpoints_to_keep = 5
save_every = 100
log_device_placement = false

[DATA]
logs_path = logs
model_dir = model_dir

[PARAMS]
embedding_size = 64
loss_function = mse

Model configuration

Additionally each model contains its own specific configuration file in which changing hyperparameters is possible.

Multihead Attention Network configuration file

[PARAMS]
num_blocks = 2
num_heads = 8
use_residual = False
dropout_rate = 0.0

Convolutional Neural Network configuration file

[PARAMS]
num_filters = 50,50,50
filter_sizes = 2,3,4
dropout_rate = 0.0

Recurrent Neural Network configuration file

[PARAMS]
hidden_size = 128
cell_type = GRU
bidirectional = True

Training models with GPU support on Google Colaboratory

If you don't have an access to workstation with GPU, you can use the below exemplary Google Colaboratory notebook for training your models (CNN, RNN or Multihead) on SNLI or QQP datasets with usage of NVIDIA Tesla T4 16GB GPU available within Google Colaboratory backend: Multihead Siamese Nets in Google Colab

Testing models

Download pretrained models from the following link: pretrained Siamese Nets models, unzip and put them into ./model_dir directory. After that, you can test models either using predict mode of pipeline:

python3 run.py predict cnn

or using GUI demo:

python3 gui_demo.py

The below pictures presents Multihead Siamese Nets GUI for:

1. Positive example:

2. Negative example:

Attention weights visualization

In order to visualize multihead attention weights for compared sentences use GUI demo - check 'Visualize attention weights' checkbox which is visible after choosing model based on multihead attention mechanism.

The example of attention weights visualization looks as follows (4 attention heads):

Comparison of models

Experiments performed on GPU Nvidia GeForce GTX 1080Ti.

> SNLI dataset.

Experiment parameters:

Number of epochs : 10
Batch size : 512
Learning rate : 0.001

Number of training instances : 326959
Number of dev instances : 3674
Number of test instances : 36736

Embedding size : 64
Loss function: mean squared error (MSE)

Specific hyperparameters of models:

Evaluation results:

*Mean-Dev-Acc: the mean development set accuaracy over all epochs.

**Last-Dev-Acc: the development set accuaracy for the last epoch.

Training curves (Accuracy & Loss):

> QQP dataset.

Experiment parameters:

Number of epochs : 10
Batch size : 512
Learning rate : 0.001

Number of training instances : 362646
Number of dev instances : 1213
Number of test instances : 40428

Embedding size : 64
Loss function: mean squared error (MSE)

Specific hyperparameters of models:

Evaluation results:

*Mean-Dev-Acc: the mean development set accuracy over all epochs.

**Last-Dev-Acc: the development set accuracy for the last epoch.

Training curves (Accuracy & Loss):