Interventional Causal Representation Learning

Reproduce Results for the Polynomial Decoder Datasets

Generate Data

• python scripts/main_exps.py --case data --target_latent uniform
• python scripts/main_exps.py --case data --target_latent uniform_corr
• python scripts/main_exps.py --case data --target_latent gaussian_mixture
• python scripts/main_exps.py --case data --target_latent scm_sparse
• python scripts/main_exps.py --case data --target_latent scm_dense

Table 2: Observational Data Case

Train Models

• bash main_slurm_launcher 0 ae_poly 1e-3
• bash main_slurm_launcher 0 ae_poly 5e-4
• bash main_slurm_launcher 0 ae_poly 1e-4

Evaluate Models

• python scripts/main_exps.py --case test --method_type ae_poly --lr 1e-3 --intervention_case 0
• python scripts/main_exps.py --case test --method_type ae_poly --lr 5e-4 --intervention_case 0
• python scripts/main_exps.py --case test --method_type ae_poly --lr 1e-4 --intervention_case 0

Log Results

• python scripts/main_exps.py --case log --method_type ae_poly --intervention_case 0

Table 3: Interventional Data Case

Run all the commands stated above for the observational case (Table 2) with the flag --intervention_case set as 1

Running experiments for Neural Network Decoder

Run all the commands stated above for the observational case (Table 2) with the flag --method_type set as 'ae'.

Reproduce Results for the Image Dataset

Table 4: Image Dataset

Train Models

• bash main_slurm_launcher_image.sh balls_uniform_none
• bash main_slurm_launcher_image.sh balls_scm_linear
• bash main_slurm_launcher_image.sh balls_scm_non_linear

Evaluate Models

• python scripts/main_exps_images.py --case test --target_latent balls_uniform_none
• python scripts/main_exps_images.py --case test --target_latent balls_scm_linear
• python scripts/main_exps_images.py --case test --target_latent balls_scm_non_linear

Log Results

• python scripts/main_exps_images.py --case log

License

This source code is released under the MIT license, included here.