ACFlow

Code to reproduce most of the results in our work Flow Models for Arbitrary Conditional Likelihoods.

Download datasets

• NOTE: you need to change the path in the DataLoaders (inside /datasets/ folder)

Training scripts

python scripts/train.py --cfg_file config_file_path &

• One example configuration for training on celeba is provided in /exp/celeba/test/param.json

• for training with 'Best Guess Penalty', set 'lambda_mse' > 0

Quantitative evaluation

• for models without 'BG Penalty'

python scripts/test.py --cfg_file config_file_path &

• for models with 'BG Penalty'

python scripts/test_single.py --cfg_file config_file_path &

Sampling

• for models without 'BG Penalty'

python scripts/sample.py --cfg_file config_file_path &

• for models with 'BG Penalty'

python scripts/sample_single.py --cfg_file config_file_path &