Skip to content

Latest commit

 

History

History
134 lines (121 loc) · 4.73 KB

README.md

File metadata and controls

134 lines (121 loc) · 4.73 KB
Raw

DisCo: Improving Compositional Generalization in Visual Reasoning through Distribution Coverage

demonstrative figure

Joy Hsu, Jiayuan Mao, Jiajun Wu

In Transactions on Machine Learning Research (TMLR)

Setup and Commands

This code is based on the FiLM: Visual Reasoning with a General Conditioning Layer repository, with FiLM as the base VQA model. Please refer to the FiLM repository to install prerequisites and process CLEVR data in the specified format, then set the paths of train_questions, train_features, test_questions, test_features, vocab.

Our framework, DisCo, can be found in the disco_scripts folder. The path to your current run's checkpoints should be set in checkpoint_path. Please first train a base FiLM model with the --training_film flag, and load in its checkpoint in film_checkpoint_path. We release scripts to run DisCo with two image proposal strategies, direct sampling and GAN generation sampling; for both stategies, the path to the directory of unlabelled images should be at real_dataset_dir.

To run DisCo with direct sampling:

python disco_scripts/train_sampler.py \
  --entropy_threshold XXX \
  --execution_engine_start_from $film_checkpoint_path \
  --program_generator_start_from $film_checkpoint_path \
  --checkpoint_path $checkpoint_path \
  --model_type FiLMReconContrastiveIntermediate \
  --r1_regularization_gamma 15.0 \
  --train_question_h5 $train_questions \
  --train_features_h5 $train_features \
  --val_question_h5 $test_questions \
  --val_features_h5 $test_features \
  --vocab_json $vocab \
  --real_dataset_dir $real_dataset_dir \
  --num_iterations 20000000 \
  --print_verbose_every 20000000 \
  --checkpoint_every 1000 \
  --record_loss_every 100 \
  --num_val_samples 149991 \
  --optimizer Adam \
  --learning_rate 1e-4 \
  --batch_size 128 \
  --use_coords 1 \
  --module_stem_batchnorm 1 \
  --module_stem_num_layers 1 \
  --module_batchnorm 1 \
  --classifier_batchnorm 1 \
  --bidirectional 0 \
  --decoder_type linear \
  --encoder_type gru \
  --weight_decay 1e-5 \
  --rnn_num_layers 1 \
  --rnn_wordvec_dim 200 \
  --rnn_hidden_dim 4096 \
  --rnn_output_batchnorm 0 \
  --classifier_downsample maxpoolfull \
  --classifier_proj_dim 512 \
  --classifier_fc_dims 1024 \
  --module_input_proj 1 \
  --module_residual 1 \
  --module_dim 128 \
  --module_dropout 0e-2 \
  --module_stem_kernel_size 3 \
  --module_kernel_size 3 \
  --module_batchnorm_affine 0 \
  --module_num_layers 1 \
  --num_modules 4 \
  --condition_pattern 1,1,1,1 \
  --gamma_option linear \
  --gamma_baseline 1 \
  --use_gamma 1 \
  --use_beta 1 \
  --condition_method bn-film \
  --program_generator_parameter_efficient 1

To run DisCo with GAN generation sampling:

python disco_scripts/train_gan_sampler.py \
  --entropy_threshold XXX \
  --execution_engine_start_from $film_checkpoint_path \
  --program_generator_start_from $film_checkpoint_path \
  --stylegan_start_from $stylegan_checkpoint_path \
  --checkpoint_path $checkpoint_path \
  --model_type FiLMReconContrastiveIntermediate \
  --r1_regularization_gamma 15.0 \
  --train_question_h5 $train_questions \
  --train_features_h5 $train_features \
  --val_question_h5 $test_questions \
  --val_features_h5 $test_features \
  --vocab_json $vocab \
  --real_dataset_dir $real_dataset_dir \
  --num_iterations 20000000 \
  --print_verbose_every 20000000 \
  --checkpoint_every 1000 \
  --record_loss_every 100 \
  --num_val_samples 149991 \
  --optimizer Adam \
  --learning_rate 1e-4 \
  --batch_size 8 \
  --use_coords 1 \
  --module_stem_batchnorm 1 \
  --module_stem_num_layers 1 \
  --module_batchnorm 1 \
  --classifier_batchnorm 1 \
  --bidirectional 0 \
  --decoder_type linear \
  --encoder_type gru \
  --weight_decay 1e-5 \
  --rnn_num_layers 1 \
  --rnn_wordvec_dim 200 \
  --rnn_hidden_dim 4096 \
  --rnn_output_batchnorm 0 \
  --classifier_downsample maxpoolfull \
  --classifier_proj_dim 512 \
  --classifier_fc_dims 1024 \
  --module_input_proj 1 \
  --module_residual 1 \
  --module_dim 128 \
  --module_dropout 0e-2 \
  --module_stem_kernel_size 3 \
  --module_kernel_size 3 \
  --module_batchnorm_affine 0 \
  --module_num_layers 1 \
  --num_modules 4 \
  --condition_pattern 1,1,1,1 \
  --gamma_option linear \
  --gamma_baseline 1 \
  --use_gamma 1 \
  --use_beta 1 \
  --condition_method bn-film \
  --program_generator_parameter_efficient 1

To pretrain StyleGAN2, use the --training_gan flag. We use the StyleGAN repository as our StyleGAN2 backbone; after pretraining, the StyleGAN2 checkpoint should be set in stylegan_checkpoint_path.