dcnhan / SMC-T-v2

Code for the paper: The Monte Carlo Transformer: a stochastic self-attention model for sequence prediction

Requirements

The libraries needed to run the code are provided in the file requirements.txt.

• To run all the scripts from the origin repo (SMC-T-v2), run first the following command line: export PYTHONPATH=src:${PYTHONPATH}

Preprocessing the datasets

Synthetic datasets

• python src/preprocessing/create_synthetic_dataset.py -model 1 -num_samples 1000 -seq_len 24 -num_features 1
• python src/preprocessing/create_synthetic_dataset.py -model 2 -alpha 0.9 -variance 0.3 -num_samples 1000 -seq_len 24 -num_features 1

Real-world datasets

• To preprocess all the real-world datasets, please run: sh preprocess_datasets.sh

Training the Baselines

LSTM with Dropout

• python src/scripts/run.py -dataset "synthetic" -dataset "synthetic" -dataset_model 1 -data_path "../../data/synthetic_model_1" -algo "lstm" -rnn_units 32 -p_drop 0.1 -rnn_drop 0.1 -bs 32 -ep 50 -output_path "exp_synthetic_model_1"

Classic Transformer

• python src/scripts/run.py -dataset "synthetic" -dataset_model 1 -data_path "../../data/synthetic_model_1" -algo "baseline_t" -d_model 8 -dff 8 -pe 50 -bs 32 -ep 50 -output_path "exp_synthetic_model_1"`

Bayesian LSTM

• python src/scripts/run.py -dataset "synthetic" -dataset_model 1 -data_path "data/synthetic_model_1" -algo "bayesian_lstm" -rnn_units 32 -bs 32 -ep 150 -output_path "output/exp_synthetic_model_1" -mc_samples 1000 -particles 3 -cv 1

Training the SMC-Transformer

Recurrent Transformer without noise and smc algo

• python src/scripts/run.py -dataset "synthetic" -dataset_model 1 -data_path "../../data/synthetic_model_1" -algo "smc_t" -d_model 8 -dff 8 -bs 32 -ep 50 -smc False -output_path "exp_synthetic_model_1"

SMC Transformer

• python src/scripts/run.py -dataset "synthetic" -dataset_model 1 -data_path "../../data/synthetic_model_1" -algo "smc_t" -d_model 8 -dff 8 -bs 32 -ep 50 -particles 10 -smc True -output_path "exp_synthetic_model_1"

Examples of sh scripts are provided in the scripts/sh folder.