This repository contains the codebase for the paper Continual Learning via Local Module Composition.
Setting up the environemnt
Create a new conda environment and install the requirements.
conda create --name ENV python=3.7
conda activate ENV
pip install -r requirements.txt
pip install -e Utils/ctrl/
pip install Utils/nngeometry/
CTrL Benchmark
All experiments were run on Nvidia Quadro RTX 8000 GPUs. To run CTrL experiments use the following comands for different streams:
Stream S-
LMC (task agnostic)
python main_transfer.py --activate_after_str_oh=0 --momentum_bn 0.1 --track_running_stats_bn 1 --pr_name lmc_cr --shuffle_test 0 --init_oh=none --task_sequence s_minus --momentum_bn_decoder=0.1 --activation_structural=sigmoid --deviation_threshold=4 --depth=4 --epochs=100 --fix_layers_below_on_addition=0 --hidden_size=64 --lr=0.001 --mask_str_loss=1 --module_init=mean --multihead=gated_linear --normalize_oh=1 --optmize_structure_only_free_modules=1 --projection_layer_oh=0 --projection_phase_length=20 --reg_factor=10 --running_stats_steps=100 --str_prior_factor=1 --str_prior_temp=0.1 --structure_inv=ae --structure_inv_oh=linear_no_act --task_agnostic_test=1 --temp=0.1 --wdecay=0.001
(test acc. 0.6863, 15 modules)
MNTDP (task aware)
python main_transfer_mntdp.py --momentum_bn 0.1 --pr_name lmc_cr --copy_batchstats 1 --track_running_stats_bn 1 --task_sequence s_minus --gating MNTDP --shuffle_test 0 --epochs 100 --lr 1e-3 --wdecay 1e-3
(test acc. 0.667, 12 modules)
Stream S+
LMC
python main_transfer.py --activate_after_str_oh=0 --activation_structural=sigmoid --deviation_threshold=1.5 --early_stop_complete=0 --pr_name lmc_cr --epochs=100 --epochs_str_only_after_addition=1 --hidden_size=64 --init_oh=none --init_runingstats_on_addition=1 --keep_bn_in_eval_after_freeze=1 --lr=0.001 --module_init=most_likely --momentum_bn=0.1 --momentum_bn_decoder=0.1 --multihead=gated_linear --normalize_oh=1 --optmize_structure_only_free_modules=1 --projection_layer_oh=0 --projection_phase_length=5 --reg_factor=10 --running_stats_steps=100 --str_prior_factor=1 --str_prior_temp=0.1 --structure_inv=ae --structure_inv_oh=linear_no_act --task_agnostic_test=1 --task_sequence=s_plus --temp=1 --wdecay=0.001
(test acc. 0.6244, 22 modules)
MNTDP (task aware)
python main_transfer_mntdp.py --momentum_bn 0.1 --pr_name lmc_cr --copy_batchstats 1 --track_running_stats_bn 1 --task_sequence s_plus --gating MNTDP --shuffle_test 0 --epochs 100 --lr 1e-3 --wdecay 1e-3 --regenerate_seed 0
(test acc. 0.609, 18 modules)
Stream Sin
LMC
python main_transfer.py --activate_after_str_oh=0 --momentum_bn 0.1 --track_running_stats_bn 1 --pr_name lmc_cr --shuffle_test 0 --init_oh=none --task_sequence s_in --momentum_bn_decoder=0.1 --activation_structural=sigmoid --deviation_threshold=4 --depth=4 --epochs=100 --fix_layers_below_on_addition=0 --hidden_size=64 --lr=0.001 --mask_str_loss=1 --module_init=most_likely --multihead=gated_linear --normalize_oh=1 --optmize_structure_only_free_modules=1 --projection_layer_oh=0 --projection_phase_length=20 --reg_factor=10 --running_stats_steps=100 --str_prior_factor=1 --str_prior_temp=0.1 --structure_inv=ae --structure_inv_oh=linear_no_act --task_agnostic_test=1 --temp=0.1 --wdecay=0.001
(test acc. 0.7081, 21 modules)
MNTDP (task aware)
python main_transfer_mntdp.py --momentum_bn 0.1 --pr_name lmc_cr --copy_batchstats 1 --track_running_stats_bn 1 --task_sequence s_in --gating MNTDP --shuffle_test 0 --epochs 100 --lr 1e-3 --wdecay 1e-3 --regenerate_seed 0
(test acc. 0.6646, 15 modules)
Stream Sout
LMC
python main_transfer.py --activate_after_str_oh=0 --momentum_bn 0.1 --track_running_stats_bn 1 --pr_name lmc_cr --shuffle_test 0 --init_oh=none --task_sequence s_out --momentum_bn_decoder=0.1 --activation_structural=sigmoid --deviation_threshold=4 --depth=4 --epochs=100 --fix_layers_below_on_addition=0 --hidden_size=64 --lr=0.001 --mask_str_loss=1 --module_init=mean --multihead=gated_linear --normalize_oh=1 --optmize_structure_only_free_modules=1 --projection_layer_oh=0 --projection_phase_length=20 --reg_factor=10 --running_stats_steps=100 --str_prior_factor=1 --str_prior_temp=0.1 --structure_inv=ae --structure_inv_oh=linear_no_act --task_agnostic_test=1 --temp=0.1 --wdecay=0.001
(test acc. 0.5849, 15 modules)
MNTDP (task aware)
python main_transfer_mntdp.py --momentum_bn 0.1 --pr_name lmc_cr --copy_batchstats 1 --track_running_stats_bn 1 --task_sequence s_out --gating MNTDP --shuffle_test 0 --epochs 100 --lr 1e-3 --wdecay 0 --regenerate_seed 0
(test acc. 0.6567, 11 modules)
Stream Spl
LMC
python main_transfer.py --activate_after_str_oh=0 --activation_structural=sigmoid --pr_name lmc_cr --deviation_threshold=1.5 --early_stop_complete=0 --epochs=100 --hidden_size=64 --init_oh=none --init_runingstats_on_addition=0 --keep_bn_in_eval_after_freeze=1 --lr=0.001 --module_init=most_likely --momentum_bn=0.1 --momentum_bn_decoder=0.1 --multihead=gated_linear --normalize_oh=1 --optmize_structure_only_free_modules=1 --projection_layer_oh=0 --projection_phase_length=10 --reg_factor=10 --running_stats_steps=100 --str_prior_factor=1 --str_prior_temp=0.1 --structure_inv=ae --structure_inv_oh=linear_no_act --task_agnostic_test=1 --task_sequence=s_pl --temp=1 --regenerate_seed 0 --wdecay=0.001
(test acc. 0.6241, 19 modules)
MNTDP (task aware)
python main_transfer_mntdp.py --momentum_bn 0.1 --pr_name lmc_cr --copy_batchstats 1 --track_running_stats_bn 1 --task_sequence s_pl --gating MNTDP --shuffle_test 0 --epochs 100 --lr 1e-3 --wdecay 1e-4 --regenerate_seed 0
(test acc. 0.6391, 18 modules)
Stream Slong30 -- 30 tasks
LMC (task aware)
python main_transfer.py --activate_after_str_oh=0 --activation_structural=sigmoid --deviation_threshold=1.5 --epochs=50 --hidden_size=64 --init_oh=none --keep_bn_in_eval_after_freeze=1 --lr=0.001 --module_init=most_likely --momentum_bn_decoder=0.1 --multihead=gated_linear --n_tasks=100 --normalize_oh=1 --optmize_structure_only_free_modules=1 --projection_layer_oh=0 --projection_phase_length=5 --reg_factor=1 --running_stats_steps=50 --seed=180 --str_prior_factor=1 --str_prior_temp=0.01 --structure_inv=ae --structure_inv_oh=linear_no_act --task_agnostic_test=0 --task_sequence=s_long30 --temp=1 --wdecay=0.001
(test acc. 62.44, 50 modules)
MNTDP (task aware)
python main_transfer_mntdp.py --epochs=50 --hidden_size=64 --lr=0.001 --module_init=most_likely --multihead=gated_linear --n_tasks=100 --seed=180 --task_sequence=s_long30 --wdecay=0.001
(test acc. 64.58, 64 modules)
Stream Slong -- 100 tasks
LMC (task aware)
python main_transfer.py --activate_after_str_oh=0 --activation_structural=sigmoid --deviation_threshold=4 --epochs=100 --hidden_size=64 --init_oh=none --keep_bn_in_eval_after_freeze=1 --lr=0.001 --module_init=most_likely --momentum_bn_decoder=0.1 --multihead=gated_linear --n_tasks=100 --normalize_oh=1 --optmize_structure_only_free_modules=1 --projection_layer_oh=0 --projection_phase_length=5 --reg_factor=1 --running_stats_steps=50 --seed=180 --str_prior_factor=1 --str_prior_temp=0.01 --structure_inv=ae --structure_inv_oh=linear_no_act --task_agnostic_test=0 --task_sequence=s_long --temp=1 --pr_name s_long_cr --wdecay=0
(test acc. 63.88, 32 modules)
MNTDP (task aware)
python main_transfer_mntdp.py --momentum_bn 0.1 --n_tasks 100 --hidden_size 64 --searchspace topdown --keep_bn_in_eval_after_freeze 1 --pr_name s_long_cr --copy_batchstats 1 --track_running_stats_bn 1 --wand_notes correct_MNTDP --task_sequence s_long --gating MNTDP --shuffle_test 0 --epochs 50 --lr 1e-3 --wdecay 1e-3
(test acc. 68.92, 142 modules)
OOD generalization experiments
LMC
python main_transfer.py --regenerate_seed 0 --deviation_threshold=8 --epochs=50 --pr_name lmc_cr --hidden_size=64 --keep_bn_in_eval_after_freeze=0 --lr=0.001 --module_init=none --momentum_bn_decoder=0.1 --normalize_data=1 --optmize_structure_only_free_modules=0 --projection_phase_length=10 --no_projection_phase 0 --reg_factor=10 --running_stats_steps=1000 --str_prior_factor=1 --str_prior_temp=0.1 --structure_inv=linear_no_act --task_sequence=s_ood --temp=1 --wdecay=0 --task_agnostic_test=0
EWC
python main_transfer.py --epochs=50 --ewc=1000 --hidden_size=256 --keep_bn_in_eval_after_freeze=0 --lr=0.001 --module_init=none --pr_name lmc_cr --multihead=usual --normalize_data=1 --task_sequence=s_ood --use_structural=0 --wdecay=0 --projection_phase_length=0
MNTDP
python main_transfer_mntdp.py --epochs=50 --regenerate_seed 0 --hidden_size=64 --keep_bn_in_eval_after_freeze=0 --pr_name lmc_cr --lr=0.01 --module_init=none --multihead=usual --normalize_data=1 --task_sequence=s_ood --use_structural=0 --wdecay=0
LMC (no projetion)
python main_transfer.py --regenerate_seed 0 --deviation_threshold=8 --epochs=50 --pr_name lmc_cr --hidden_size=64 --keep_bn_in_eval_after_freeze=0 --lr=0.001 --module_init=none --momentum_bn_decoder=0.1 --normalize_data=1 --optmize_structure_only_free_modules=0 --projection_phase_length=0 --no_projection_phase 1 --reg_factor=10 --running_stats_steps=1000 --str_prior_factor=1 --str_prior_temp=0.1 --structure_inv=linear_no_act --task_sequence=s_ood --temp=1 --wdecay=0
Plug and play (combining independently trained modular learners)
python main_plug_and_play.py --activate_after_str_oh=0 --activation_structural=sigmoid --deviation_threshold=1.5 --early_stop_complete=0 --epochs=100 --epochs_str_only_after_addition=1 --pr_name lmc_cr --hidden_size=64 --init_oh=none --init_runingstats_on_addition=1 --keep_bn_in_eval_after_freeze=1 --lr=0.001 --module_init=mean --momentum_bn=0.1 --momentum_bn_decoder=0.1 --multihead=gated_linear --n_tasks=3 --normalize_oh=1 --optmize_structure_only_free_modules=1 --projection_layer_oh=0 --projection_phase_length=5 --reg_factor=10 --running_stats_steps=10 --str_prior_factor=1 --str_prior_temp=0.1 --structure_inv=ae --structure_inv_oh=linear_no_act --task_agnostic_test=1 --task_sequence=s_pnp_comp --temp=1 --wdecay=0.001
A list of hyperparameters used for other baselines can be found in the baselines.txt file.
1 Jan 04, 2022
229 Dec 29, 2022
56 Nov 29, 2022
2 Aug 21, 2022
24 Nov 29, 2022
17 Jun 19, 2022
5 Oct 15, 2022
11 Nov 27, 2022
38 Dec 06, 2022
46 Jan 07, 2023
2 Jul 12, 2022
317 Dec 26, 2022
97 Jan 07, 2023
9 May 30, 2022
61 Dec 27, 2022
2 Oct 02, 2022
1 Feb 23, 2022
80 Sep 17, 2022
126 Dec 20, 2022
45 Feb 06, 2022