Counterfactual regression (CFR) by learning balanced representations, as developed by Johansson, Shalit & Sontag (2016) and Shalit, Johansson & Sontag (2016). cfrnet is implemented in Python using TensorFlow 0.12.0-rc1 and NumPy 1.11.3. The code has not been tested with TensorFlow 1.0.
The core components of cfrnet, i.e. the TensorFlow graph, is contained in cfr/cfr_net.py. The training is performed by cfr_net_train.py. The file cfr_param_search.py takes a configuration file as input and allows the user to randomly sample from the supplied parameters (given that there are multiple values given in a list. See configs/example_ihdp.txt for an example.
A typical experiment uses cfr_param_search.py and evaluate.py as sub-routines. cfr_param_search is best used to randomly search the parameter space. In the output directory, it creates a log of which configurations have been used so far, so that the same experiment is not repeated. evaluate.py goes through the predictions produced by the model and evaluates the error.
The script cfr_param_search.py runs a random hyperparameter search given a configuration file.
Usage:
python cfr_param_search.py <config_file> <num_runs>
The config_file argument should contain the path to a text file where each line is a key-value pair for a CFR parameter.
The num_run argument should contain an integer to indicate how many parameter settings should be sampled. If all possible configurations should be used, this can be set arbitrarily high as the script will terminate when all have been used. If the number of possible settings is vast, a smaller value for num_runs may be appropriate.
Example:
python evaluate.py configs/example_ihdp.txt 10
Example configuration file (from configs/example_ihdp.txt):
p_alpha=[0, 1e-1]
p_lambda=[1e-3]
n_in=[2]
n_out=[2]
dropout_in=1.0
...
Note that some of the lines have square brackets to indicate lists. If a parameter list contains more than a single element, cfr_param_search will sample uniformly from these values. In this way, random parameter search can be performed.
The script evaluate.py performs an evaluation of a trained model based on the predictions made for the training and test sets.
Usage:
python evaluate.py <config_file> [overwrite] [filters]
The parameter config_file should be the same as the one used in cfr_param_search. (Note: evaluate only uses the settings for dataform, data_test, datadir and outdir, the rest can be changed without affecting the evaluation.)
If the overwrite parameter is set to "1", the script re-computes all error estimates. If it is set to "0" it re-uses stored values, but re-prints and re-plots all results.
The argument filters accepts a string in the form of a python dict containing values of the parameters the used wishes to filter. This produces plots and text summaries only of results corresponding to configuration that matches the filter.
Example:
python evaluate.py configs/example_ihdp.txt 0 "{p_alpha: 0}"
A simple experiment example is contained in example_ihdp.sh. This file runs the model on (a subset of) the IHDP data with parameters supplied by configs/example_ihdp.txt. The data for this example can be downloaded from http://www.fredjo.com/files/ihdp_npci_1-100.train.npz (train) and http://www.fredjo.com/files/ihdp_npci_1-100.test.npz (test). For the full data (of 1000 replications) used in the ICML 2017 paper, please visit https://www.fredjo.com/.
- Q: What are the hyperparameters used on IHDP in the ICML 2017 paper? A: The parameters were those given in example_ihdp.txt but with p_alpha = 0.3
- Q: I don't get the same IHDP results as in the paper when I try to replicate with the IHDP example from Github. A: The ICML 2017 results were computed over the full set of 1000 replications. The Github IHDP example uses only 100 examples as it is meant to serve as a quick demo. Please find the 1000 replications at https://www.fredjo.com/.
Uri Shalit, Fredrik D. Johansson & David Sontag. Estimating individual treatment effect: generalization bounds and algorithms, 34th International Conference on Machine Learning (ICML), August 2017.
Fredrik D. Johansson, Uri Shalit & David Sontag. Learning Representations for Counterfactual Inference. 33rd International Conference on Machine Learning (ICML), June 2016.