Awesome

CausalTransformer

Causal Transformer for estimating counterfactual outcomes over time.

The project is built with following Python libraries:

1. Pytorch-Lightning - deep learning models
2. Hydra - simplified command line arguments management
3. MlFlow - experiments tracking

Installations

First one needs to make the virtual environment and install all the requirements:

pip3 install virtualenv
python3 -m virtualenv -p python3 --always-copy venv
source venv/bin/activate
pip3 install -r requirements.txt

MlFlow Setup / Connection

To start an experiments server, run:

mlflow server --port=5000

To access MlFLow web UI with all the experiments, connect via ssh:

ssh -N -f -L localhost:5000:localhost:5000 <username>@<server-link>

Then, one can go to local browser http://localhost:5000.

Experiments

Main training script is universal for different models and datasets. For details on mandatory arguments - see the main configuration file config/config.yaml and other files in configs/ folder.

Generic script with logging and fixed random seed is following (with training-type enc_dec, gnet, rmsn and multi):

PYTHONPATH=. CUDA_VISIBLE_DEVICES=<devices> 
python3 runnables/train_<training-type>.py +dataset=<dataset> +backbone=<backbone> exp.seed=10 exp.logging=True

Backbones (baselines)

One needs to choose a backbone and then fill the specific hyperparameters (they are left blank in the configs):

• Causal Transformer (this paper): runnables/train_multi.py +backbone=ct
• Encoder-Decoder Causal Transformer (this paper): runnables/train_enc_dec.py +backbone=edct
• Marginal Structural Models (MSMs): runnables/train_msm.py +backbone=msm
• Recurrent Marginal Structural Networks (RMSNs): runnables/train_rmsn.py +backbone=rmsn
• Counterfactual Recurrent Network (CRN): runnables/train_enc_dec.py +backbone=crn
• G-Net: runnables/train_gnet.py +backbone=gnet

Models already have best hyperparameters saved (for each model and dataset), one can access them via: +backbone/<backbone>_hparams/cancer_sim_<balancing_objective>=<coeff_value> or +backbone/<backbone>_hparams/mimic3_real=diastolic_blood_pressure.

For CT, EDCT, and CT, several adversarial balancing objectives are available:

• counterfactual domain confusion loss (this paper): exp.balancing=domain_confusion
• gradient reversal (originally in CRN, but can be used for all the methods): exp.balancing=grad_reverse

To train a decoder (for CRN and RMSNs), use the flag model.train_decoder=True.

To perform a manual hyperparameter tuning use the flags model.<sub_model>.tune_hparams=True, and then see model.<sub_model>.hparams_grid. Use model.<sub_model>.tune_range to specify the number of trials for random search.

Datasets

One needs to specify a dataset / dataset generator (and some additional parameters, e.g. set gamma for cancer_sim with dataset.coeff=1.0):

• Synthetic Tumor Growth Simulator: +dataset=cancer_sim
• MIMIC III Semi-synthetic Simulator (multiple treatments and outcomes): +dataset=mimic3_synthetic
• MIMIC III Real-world dataset: +dataset=mimic3_real

Before running MIMIC III experiments, place MIMIC-III-extract dataset (all_hourly_data.h5) to data/processed/

Example of running Causal Transformer on Synthetic Tumor Growth Generator with gamma = [1.0, 2.0, 3.0] and different random seeds (total of 30 subruns), using hyperparameters:

PYTHONPATH=. CUDA_VISIBLE_DEVICES=<devices> 
python3 runnables/train_multi.py -m +dataset=cancer_sim +backbone=ct +backbone/ct_hparams/cancer_sim_domain_conf=\'0\',\'1\',\'2\' exp.seed=10,101,1010,10101,101010

Updated results

Self- and cross-attention bug

New results for semi-synthetic and real-world experiments after fixing a bug with self- and cross-attentions (https://github.com/Valentyn1997/CausalTransformer/issues/7). Therein, the bug affected only Tables 1 and 2, and Figure 5 (https://arxiv.org/pdf/2204.07258.pdf). Nevertheless, the performance of the CT with the bug fixed did not change drastically.

Table 1 (updated). Results for semi-synthetic data for $\tau$-step-ahead prediction based on real-world medical data (MIMIC-III). Shown: RMSE as mean ± standard deviation over five runs.

Table 2 (updated). Results for experiments with real-world medical data (MIMIC-III). Shown: RMSE as mean ± standard deviation over five runs.

Figure 6 (updated). Subnetworks importance scores based on semi-synthetic benchmark (higher values correspond to higher importance of subnetwork connectivity via cross-attentions). Shown: RMSE differences between model with isolated subnetwork and full CT, means ± standard errors.

Last active entry zeroing bug

New results after fixing a bug with the synthetic tumor-growth simulator: outcome corresponding to the last entry for every time series was zeroed.

Table 9 (updated). Normalized RMSE for one-step-ahead prediction. Shown: mean and standard deviation over five runs (lower is better). Parameter $\gamma$ is the the amount of time-varying confounding: higher values mean larger treatment assignment bias.

Table 10 (updated). Normalized RMSE for $\tau$-step-ahead prediction (here: random trajectories setting). Shown: mean and standard deviation over five runs (lower is better). Parameter $\gamma$ is the amount of time-varying confounding: higher values mean larger treatment assignment bias.

Table 11 (updated). Normalized RMSE for $\tau$-step-ahead prediction (here: single sliding treatment setting). Shown: mean and standard deviation over five runs (lower is better). Parameter $\gamma$ is the amount of time-varying confounding: higher values mean larger treatment assignment bias.