├── arlc ├── __init__.py ├── datasets │ ├── __init__.py │ └── iraven.py ├── utils │ ├── const.py │ ├── checkpath.py │ ├── averagemeter.py │ ├── general.py │ ├── vsa.py │ ├── raven │ │ ├── scene.py │ │ ├── env.py │ │ ├── raven_one_hot.py │ │ └── extraction.py │ └── parsing.py ├── losses.py ├── selection.py ├── execution.py └── rule_templates.py ├── requirements.txt ├── figs └── arlc_preview.png ├── .gitignore ├── .pre-commit-config.yaml ├── experiments ├── arlc_learn.sh ├── arlc_progr_to_learn.sh ├── ablations │ ├── context.sh │ └── learnvrf_nopn_2x2.sh ├── uncertainty │ ├── train_noisy.sh │ ├── arlc_learn_noisy.sh │ ├── train_confounders.sh │ ├── arlc_eval_noisy.sh │ ├── exp_dist_inference.sh │ ├── exp_confounders_inference.sh │ └── eval_noisy.sh ├── arlc_progr.sh ├── arlc_ood.sh └── iravenx │ ├── arlc_program_eval.sh │ └── arlc_learn_iravenx_50.sh ├── setup.py ├── data └── README.md ├── results.py ├── README.md ├── main.py └── LICENSE /arlc/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | tqdm 2 | matplotlib 3 | tensorboard 4 | progressbar 5 | scikit-learn 6 | pre-commit 7 | -------------------------------------------------------------------------------- /arlc/datasets/__init__.py: -------------------------------------------------------------------------------- 1 | from .iraven import GeneralIRAVENDataset 2 | 3 | __all__ = ["GeneralIRAVENDataset"] 4 | -------------------------------------------------------------------------------- /figs/arlc_preview.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IBM/abductive-rule-learner-with-context-awareness/main/figs/arlc_preview.png -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # MacOS related files 2 | *.DS_Store 3 | *.icloud 4 | 5 | # Python files 6 | **__pycache__** 7 | **.pyc 8 | *.out 9 | *.vscode 10 | **.egg-info 11 | *.py[cod] 12 | *$py.class 13 | .ruff_cache 14 | *.log 15 | .pytest_cache 16 | results 17 | 18 | # Jupyter notebooks 19 | *.ipynb_checkpoints 20 | 21 | # Credentials 22 | **/.config 23 | 24 | **.code-workspace 25 | 26 | models 27 | data 28 | -------------------------------------------------------------------------------- /.pre-commit-config.yaml: -------------------------------------------------------------------------------- 1 | repos: 2 | - repo: https://github.com/psf/black 3 | rev: 22.3.0 4 | hooks: 5 | - id: black 6 | language_version: python3.10 7 | - repo: https://github.com/pre-commit/pre-commit-hooks 8 | rev: v4.4.0 9 | hooks: 10 | - id: trailing-whitespace 11 | - id: end-of-file-fixer 12 | - id: check-yaml 13 | - id: check-json 14 | - id: check-merge-conflict 15 | - id: check-case-conflict 16 | - id: mixed-line-ending 17 | - id: fix-byte-order-marker 18 | -------------------------------------------------------------------------------- /experiments/arlc_learn.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | CONFIG="distribute_four" 4 | RUN="arlc_learn" 5 | EXP_DIR="models" 6 | EPOCHS=25 7 | NTEST=10 8 | NRULES=5 9 | 10 | for SEED in $(seq 1 $NTEST); 11 | do 12 | echo "Running training with seed $SEED" 13 | python main.py --epochs $EPOCHS \ 14 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --data_dir data \ 15 | --batch_size 8 --num_workers 1 --num_rules $NRULES --seed $SEED --run_name $RUN --exp_dir $EXP_DIR 16 | done 17 | python results.py --path $EXP_DIR/$RUN --seeds $NTEST 18 | -------------------------------------------------------------------------------- /experiments/arlc_progr_to_learn.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | CONFIG="distribute_four" 4 | RUN="arlc_progr" 5 | EXP_DIR="models" 6 | EPOCHS=25 7 | NTEST=10 8 | NRULES=5 9 | 10 | 11 | for SEED in $(seq 1 $NTEST); 12 | do 13 | echo "Running training with seed $SEED" 14 | python main.py --epochs $EPOCHS \ 15 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --data_dir data \ 16 | --batch_size 8 --num_workers 1 --num_rules $NRULES --seed $SEED --run_name $RUN --exp_dir $EXP_DIR --program 17 | done 18 | python results.py --path $EXP_DIR/$RUN --seeds $NTEST 19 | -------------------------------------------------------------------------------- /experiments/ablations/context.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | CONFIG="distribute_four" 4 | RUN="ablation_np_pn_superpos" 5 | EXP_DIR="models" 6 | EPOCHS=25 7 | NTEST=10 8 | NRULES=5 9 | 10 | for SEED in $(seq 1 $NTEST); 11 | do 12 | echo "Running training with seed $SEED" 13 | python main.py --epochs $EPOCHS \ 14 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --data_dir /dccstor/saentis/data/I-RAVEN \ 15 | --batch_size 8 --num_workers 1 --num_rules $NRULES --seed $SEED --run_name $RUN --num_term 6 --rule_type arlc \ 16 | --exp_dir $EXP_DIR 17 | done 18 | python results.py --path $EXP_DIR/$RUN --seeds $NTEST 19 | -------------------------------------------------------------------------------- /experiments/ablations/learnvrf_nopn_2x2.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | CONFIG="distribute_four" 4 | RUN="ablation_np_pn_superpos" 5 | EXP_DIR="models" 6 | EPOCHS=25 7 | NTEST=10 8 | NRULES=5 9 | 10 | for SEED in $(seq 1 $NTEST); 11 | do 12 | echo "Running training with seed $SEED" 13 | python main.py --epochs $EPOCHS \ 14 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --data_dir /dccstor/saentis/data/I-RAVEN \ 15 | --batch_size 8 --num_workers 1 --num_rules $NRULES --seed $SEED --run_name $RUN --num_term 6 --rule_type learnvrf \ 16 | --exp_dir $EXP_DIR 17 | done 18 | python results.py --path $EXP_DIR/$RUN --seeds $NTEST 19 | -------------------------------------------------------------------------------- /experiments/uncertainty/train_noisy.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | CONFIG="center_single" 4 | EXP_DIR="models" 5 | NTEST=5 6 | NRULES=5 7 | DATA="iravenx" 8 | EPOCHS=10 9 | SIGMA=0.7 10 | RUN="iravenx_noisy_$SIGMA" 11 | 12 | if [ "$DATA" = "iraven" ]; then 13 | DATA_DIR="data/I-RAVEN" 14 | else 15 | DATA_DIR="data/I-RAVEN-X" 16 | fi 17 | 18 | for SEED in $(seq 1 $NTEST); 19 | do 20 | echo $SEED 21 | python main.py --epochs $EPOCHS --dyn_range 100 --n 10 \ 22 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR --entropy --sigma $SIGMA \ 23 | --batch_size 8 --num_workers 1 --num_rules $NRULES --num_terms 26 --seed $SEED --run_name $RUN --exp_dir $EXP_DIR --partition _shuffle 24 | done 25 | python results.py --path $EXP_DIR/$RUN --seeds $NTEST 26 | -------------------------------------------------------------------------------- /experiments/uncertainty/arlc_learn_noisy.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | CONFIG="center_single" 4 | SIGMA=0.7 5 | RUN="arlc_learn_noisy_$SIGMA" 6 | EXP_DIR="models" 7 | EPOCHS=15 8 | NTEST=5 9 | NRULES=5 10 | DATA="iravenx" 11 | if [ "$DATA" = "iraven" ]; then 12 | DATA_DIR="data/I-RAVEN" 13 | else 14 | DATA_DIR="data/I-RAVEN-X" 15 | fi 16 | 17 | 18 | 19 | for SEED in $(seq 1 $NTEST); 20 | do 21 | echo $SEED 22 | python main.py --epochs $EPOCHS --n 3 --dyn_range 10 --num_terms 12 \ 23 | --vsa_conversion --vsa_selection --config $CONFIG --dataset $DATA --data_dir $DATA_DIR --partition _shuffle --entropy \ 24 | --batch_size 8 --num_workers 1 --num_rules $NRULES --seed $SEED --run_name $RUN --exp_dir $EXP_DIR --orientation-confounder 0 --sigma $SIGMA 25 | done 26 | python results.py --path $EXP_DIR/$RUN --seeds $NTEST 27 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | from setuptools import setup, find_packages 7 | 8 | setup( 9 | name="arlc", 10 | version="1.0.0", 11 | description="Abductive Rule Learner with Context-awareness", 12 | url="https://research.ibm.com/people/giacomo-camposampiero--1", 13 | author="Giacomo Camposampiero", 14 | author_email="giacomo.camposampiero1@ibm.com", 15 | license="GPL-3.0", 16 | packages=find_packages( 17 | where="arlc", 18 | ), 19 | include_package_data=True, 20 | zip_safe=False, 21 | ) 22 | -------------------------------------------------------------------------------- /experiments/uncertainty/train_confounders.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | CONFIG="center_single" 4 | CONF=5 5 | RUN="iravenx_confounders_entropy_$CONF" 6 | EXP_DIR="models" 7 | NTEST=5 8 | NRULES=5 9 | DEBUG=0 10 | DATA="iravenx" 11 | EPOCHS=10 12 | 13 | if [ "$DATA" = "iraven" ]; then 14 | DATA_DIR="data/I-RAVEN" 15 | else 16 | DATA_DIR="data/I-RAVEN-X" 17 | fi 18 | 19 | 20 | for SEED in $(seq 1 $NTEST); 21 | do 22 | echo $SEED 23 | python main.py --epochs $EPOCHS --dyn_range 100 --n 10 \ 24 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR --entropy \ 25 | --batch_size 8 --num_workers 1 --num_rules $NRULES --num_terms 26 --seed $SEED --run_name $RUN --exp_dir $EXP_DIR --partition _shuffle --orientation-confounder $CONF 26 | done 27 | python results.py --path $EXP_DIR/$RUN --seeds $NTEST 28 | -------------------------------------------------------------------------------- /arlc/utils/const.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | DIM_EXIST = 2 7 | DIM_POSITION_2x2 = 15 8 | DIM_POSITION_3x3 = 511 9 | DIM_NUMBER_2x2 = 4 10 | DIM_NUMBER_3x3 = 9 11 | DIM_ONEHOT = 1001 12 | LOG_EPSILON = 1e-39 13 | NORM_SCALE = 1e15 14 | 15 | NUMPOS = { 16 | "center_single": 1, 17 | "distribute_four": 4, 18 | "distribute_nine": 9, 19 | "left_center_single_right_center_single": 2, 20 | "up_center_single_down_center_single": 2, 21 | "in_center_single_out_center_single": 2, 22 | "in_distribute_four_out_center_single": 5, 23 | } 24 | -------------------------------------------------------------------------------- /data/README.md: -------------------------------------------------------------------------------- 1 | ## Get the Data 2 | Generate the I-RAVEN dataset with the instructions proveded [here](https://github.com/husheng12345/SRAN) and save it in this folder. 3 | 4 | ```bash 5 | git clone https://github.com/husheng12345/SRAN 6 | pip2 install --user -r SRAN/I-RAVEN/requirements.txt 7 | python2 SRAN/I-RAVEN/main.py --save-dir . 8 | ``` 9 | 10 | ## Prepare the Data 11 | 12 | Run the rule preprocessing script: 13 | ```bash 14 | python arlc/utils/raven/extraction.py --data_path data 15 | ``` 16 | 17 | In the latest version of the code we migrated from the original numpy-based dataset to a JSON-based following the approach of [Hu et al.](https://github.com/hxiaoyang/lm-raven). 18 | To convert the original dataset to the JSON files required by the new dataloader, use the script provided [here](https://github.com/IBM/raven-large-language-models/blob/main/src/datasets/generation/iraven_task.py). 19 | -------------------------------------------------------------------------------- /experiments/uncertainty/arlc_eval_noisy.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | MODEL=$1 4 | 5 | CONFIG="center_single" 6 | EXP_DIR="models" 7 | EPOCHS=10 8 | NTEST=5 9 | NRULES=5 10 | DEBUG=1 11 | DATA="iraven" 12 | SIGMA=0.3 13 | 14 | if [ "$DATA" = "iraven" ]; then 15 | DATA_DIR="data/I-RAVEN" 16 | else 17 | DATA_DIR="data/I-RAVEN-X" 18 | fi 19 | 20 | for SEED in $(seq 1 $NTEST); 21 | do 22 | python main.py --resume models/$MODEL/$SEED/ckpt --n 3 --dyn_range 10 --num_terms 12 \ 23 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR \ 24 | --batch_size 512 --num_workers 1 --num_rules $NRULES --num_terms 12 --seed $SEED --run_name $MODEL --orientation-confounder 0 \ 25 | --exp_dir $EXP_DIR --partition _shuffle --mode test --evaluate-rule --entropy --sigma $SIGMA 26 | done 27 | python results.py --path models/$MODEL --seeds $NTEST 28 | -------------------------------------------------------------------------------- /experiments/uncertainty/exp_dist_inference.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | RUN=$1 4 | 5 | CONFIG="center_single" 6 | EXP_DIR="models" 7 | EPOCHS=10 8 | NTEST=5 9 | NRULES=5 10 | DEBUG=1 11 | DATA="iravenx" 12 | SIGMAS=(-0.7 -0.51) 13 | 14 | if [ "$DATA" = "iraven" ]; then 15 | DATA_DIR="data/I-RAVEN" 16 | else 17 | DATA_DIR="data/I-RAVEN-X" 18 | fi 19 | 20 | for SIGMA in "${SIGMAS[@]}"; 21 | do 22 | echo "****************** sigma $SIGMAS" 23 | for SEED in $(seq 1 $NTEST); 24 | do 25 | python main.py --dyn_range 100 --n 10 --resume models/$RUN/$SEED/ckpt \ 26 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR \ 27 | --batch_size 128 --num_workers 1 --num_rules $NRULES --num_terms 26 --seed $SEED --run_name $RUN \ 28 | --exp_dir $EXP_DIR --partition _shuffle --mode test --evaluate-rule --orientation-confounder 0 --entropy --sigma $SIGMA 29 | done 30 | python results.py --path $EXP_DIR/$RUN --seeds $NTEST 31 | done 32 | -------------------------------------------------------------------------------- /experiments/uncertainty/exp_confounders_inference.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | RUN=$1 4 | 5 | CONFIG="center_single" 6 | EXP_DIR="models" 7 | EPOCHS=10 8 | NTEST=3 9 | NRULES=5 10 | DEBUG=1 11 | DATA="iravenx" 12 | N_CONFOUNDERS=(0 1 3 5 10 30 300) 13 | 14 | if [ "$DATA" = "iraven" ]; then 15 | DATA_DIR="data/I-RAVEN" 16 | else 17 | DATA_DIR="data/I-RAVEN-X" 18 | fi 19 | 20 | for N_CONFOUNDERS in "${N_CONFOUNDERS[@]}"; 21 | do 22 | echo "****************** N_CONF $N_CONFOUNDERS" 23 | for SEED in $(seq 1 $NTEST); 24 | do 25 | python main.py --dyn_range 1000 --n 10 --resume models/$RUN/$SEED/ckpt \ 26 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR \ 27 | --batch_size 64 --num_workers 1 --num_rules $NRULES --num_terms 26 --seed $SEED --run_name $RUN \ 28 | --exp_dir $EXP_DIR --partition _shuffle --mode test --evaluate-rule --orientation-confounder $N_CONFOUNDERS --entropy 29 | done 30 | python results.py --path $EXP_DIR/$RUN --seeds $NTEST 31 | done 32 | -------------------------------------------------------------------------------- /experiments/arlc_progr.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | python main.py --vsa_conversion --vsa_selection --shared_rules --config center_single --data_dir data --batch_size 256 --num_workers 1 --mode test --program 4 | python main.py --vsa_conversion --vsa_selection --shared_rules --config distribute_four --data_dir data --batch_size 256 --num_workers 1 --mode test --program 5 | python main.py --vsa_conversion --vsa_selection --shared_rules --config distribute_nine --data_dir data --batch_size 256 --num_workers 1 --mode test --program 6 | python main.py --vsa_conversion --vsa_selection --shared_rules --config left_right --data_dir data --batch_size 256 --num_workers 1 --mode test --program 7 | python main.py --vsa_conversion --vsa_selection --shared_rules --config up_down --data_dir data --batch_size 256 --num_workers 1 --mode test --program 8 | python main.py --vsa_conversion --vsa_selection --shared_rules --config in_out_single --data_dir data --batch_size 256 --num_workers 1 --mode test --program 9 | python main.py --vsa_conversion --vsa_selection --shared_rules --config in_out_four --data_dir data --batch_size 256 --num_workers 1 --mode test --program 10 | -------------------------------------------------------------------------------- /arlc/utils/checkpath.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | import os, sys, time, torch, shutil 7 | 8 | 9 | def check_paths(args): 10 | try: 11 | if not os.path.exists(args.save_dir): 12 | os.makedirs(args.save_dir) 13 | if not os.path.exists(args.log_dir): 14 | os.makedirs(args.log_dir) 15 | new_log_dir = os.path.join(args.log_dir, time.ctime().replace(" ", "-")) 16 | args.log_dir = new_log_dir 17 | if not os.path.exists(args.log_dir): 18 | os.makedirs(args.log_dir) 19 | if not os.path.exists(args.checkpoint_dir): 20 | os.makedirs(args.checkpoint_dir) 21 | except OSError as e: 22 | print(e) 23 | sys.exit(1) 24 | 25 | 26 | def save_checkpoint(state, is_best, filename="checkpoint.pth.tar", savedir=""): 27 | save_name = os.path.join(savedir, filename) 28 | torch.save(state, save_name) 29 | if is_best: 30 | save_name = os.path.join(savedir, "model_best.pth.tar") 31 | shutil.copyfile(os.path.join(savedir, filename), save_name) 32 | -------------------------------------------------------------------------------- /results.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | import numpy as np 7 | from collections import defaultdict 8 | import os 9 | from arlc.utils.parsing import eval_parse_args 10 | import json 11 | 12 | 13 | def main(): 14 | print("\n") 15 | args = eval_parse_args() 16 | res = defaultdict(list) 17 | for i in range(1, args.seeds + 1): 18 | with open(os.path.join(args.path, f"{i}/ckpt/eval.json")) as f: 19 | dat = json.load(f) 20 | for k, v in dat.items(): 21 | res[k].append(v) 22 | for k, v in res.items(): 23 | print(f"{k}\t\t{np.mean(v)} ({np.std(v)})") 24 | print(f"{k}\t\t{np.min(v)} {np.max(v)}") 25 | print(f"{k}\t\t{v}") 26 | 27 | mean = np.mean(sum(res.values(), [])) 28 | std = np.mean([np.std(x) for x in res.values()]) 29 | print("\nLaTex table entry:") 30 | print( 31 | " & ".join([f"${np.mean(v):.1f}^{{\pm{np.std(v):.1f}}}$" for v in res.values()]) 32 | + f" & ${mean:.1f}^{{\pm{std:.1f})}}$" 33 | ) 34 | print("\n") 35 | 36 | 37 | if __name__ == "__main__": 38 | main() 39 | -------------------------------------------------------------------------------- /experiments/uncertainty/eval_noisy.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | RUN=$1 4 | SIGMA=$2 5 | 6 | CONFIG="center_single" 7 | EXP_DIR="models" 8 | EPOCHS=10 9 | NTEST=5 10 | NRULES=5 11 | DATA="iravenx" 12 | N_CONFOUNDERS=0 13 | 14 | if [ "$DATA" = "iraven" ]; then 15 | DATA_DIR="data/I-RAVEN" 16 | else 17 | DATA_DIR="data/I-RAVEN-X" 18 | fi 19 | 20 | echo "****************** EVAL 100" 21 | for SEED in $(seq 1 $NTEST); 22 | do 23 | python main.py --dyn_range 100 --n 10 --resume models/$RUN/$SEED/ckpt \ 24 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR \ 25 | --batch_size 128 --num_workers 1 --num_rules $NRULES --num_terms 26 --seed $SEED --run_name $RUN \ 26 | --exp_dir $EXP_DIR --partition _shuffle --mode test --evaluate-rule --entropy --sigma $SIGMA 27 | done 28 | python results.py --path $EXP_DIR/$RUN --seeds $NTEST 29 | 30 | 31 | echo "****************** EVAL 1000" 32 | for SEED in $(seq 1 $NTEST); 33 | do 34 | python main.py --dyn_range 1000 --n 10 --resume models/$RUN/$SEED/ckpt \ 35 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR \ 36 | --batch_size 128 --num_workers 1 --num_rules $NRULES --num_terms 26 --seed $SEED --run_name $RUN \ 37 | --exp_dir $EXP_DIR --partition _shuffle --mode test --evaluate-rule --entropy --sigma $SIGMA 38 | done 39 | python results.py --path $EXP_DIR/$RUN --seeds $NTEST 40 | -------------------------------------------------------------------------------- /experiments/arlc_ood.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | python main.py --data_dir data --vsa_conversion --vsa_selection --shared_rules --gen_attribute 'Type' --gen_rule 'Constant' 4 | python main.py --data_dir data --vsa_conversion --vsa_selection --shared_rules --gen_attribute 'Type' --gen_rule 'Progression' 5 | python main.py --data_dir data --vsa_conversion --vsa_selection --shared_rules --gen_attribute 'Type' --gen_rule 'Distribute_Three' 6 | python main.py --data_dir data --vsa_conversion --vsa_selection --shared_rules --gen_attribute 'Size' --gen_rule 'Constant' 7 | python main.py --data_dir data --vsa_conversion --vsa_selection --shared_rules --gen_attribute 'Size' --gen_rule 'Progression' 8 | python main.py --data_dir data --vsa_conversion --vsa_selection --shared_rules --gen_attribute 'Size' --gen_rule 'Distribute_Three' 9 | python main.py --data_dir data --vsa_conversion --vsa_selection --shared_rules --gen_attribute 'Size' --gen_rule 'Arithmetic' 10 | python main.py --data_dir data --vsa_conversion --vsa_selection --shared_rules --gen_attribute 'Color' --gen_rule 'Constant' 11 | python main.py --data_dir data --vsa_conversion --vsa_selection --shared_rules --gen_attribute 'Color' --gen_rule 'Progression' 12 | python main.py --data_dir data --vsa_conversion --vsa_selection --shared_rules --gen_attribute 'Color' --gen_rule 'Distribute_Three' 13 | python main.py --data_dir data --vsa_conversion --vsa_selection --shared_rules --gen_attribute 'Color' --gen_rule 'Arithmetic' 14 | -------------------------------------------------------------------------------- /arlc/utils/averagemeter.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | 7 | class ProgressMeter(object): 8 | def __init__(self, num_batches, meters, prefix=""): 9 | self.batch_fmtstr = self._get_batch_fmtstr(num_batches) 10 | self.meters = meters 11 | self.prefix = prefix 12 | 13 | def display(self, batch): 14 | entries = [self.prefix + self.batch_fmtstr.format(batch)] 15 | entries += [str(meter) for meter in self.meters] 16 | print("\t".join(entries), flush=True) 17 | 18 | def _get_batch_fmtstr(self, num_batches): 19 | num_digits = len(str(num_batches // 1)) 20 | fmt = "{:" + str(num_digits) + "d}" 21 | return "[" + fmt + "/" + fmt.format(num_batches) + "]" 22 | 23 | 24 | class AverageMeter(object): # Computes and stores the average and current value 25 | def __init__(self, name, fmt=":f"): 26 | self.name = name 27 | self.fmt = fmt 28 | self.reset() 29 | 30 | def reset(self): 31 | self.val = 0 32 | self.avg = 0 33 | self.sum = 0 34 | self.count = 0 35 | 36 | def update(self, val, n=1): 37 | if (self.count + n) != 0: 38 | self.val = val 39 | self.sum += val * n 40 | self.count += n 41 | self.avg = self.sum / self.count 42 | 43 | def __str__(self): 44 | fmtstr = "{name} {val" + self.fmt + "} ({avg" + self.fmt + "})" 45 | return fmtstr.format(**self.__dict__) 46 | -------------------------------------------------------------------------------- /arlc/utils/general.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | import torch 7 | 8 | 9 | iravenx_rule_map = { 10 | "Constant": 0, 11 | "Progression": 1, 12 | "Arithmetic": 2, 13 | "Distribute_Three": 3, 14 | } 15 | iravenx_index_map = { 16 | 0: "Constant", 17 | 1: "Progression", 18 | 2: "Arithmetic", 19 | 3: "Distribute_Three", 20 | } 21 | 22 | 23 | LOG_EPSILON = 1e-39 24 | NORM_SCALE = 1e15 25 | 26 | 27 | def normalize(unnorm_prob, dim=-1): 28 | unnorm_prob = unnorm_prob * NORM_SCALE 29 | sum_dim = torch.sum(unnorm_prob, dim=dim, keepdim=True) 30 | norm_prob = unnorm_prob / sum_dim 31 | return norm_prob, sum_dim 32 | 33 | 34 | def to_n_bit_string(n, number): 35 | format_string = "{" + "0:0{}b".format(n) + "}" 36 | return format_string.format(number) 37 | 38 | 39 | def left_rotate(number, steps, num_bits): 40 | offset = steps % num_bits 41 | index = ((number << offset) | (number >> (num_bits - offset))) & (2**num_bits - 1) 42 | return index 43 | 44 | 45 | def right_rotate(number, steps, num_bits): 46 | offset = steps % num_bits 47 | index = ((number >> offset) | (number << (num_bits - offset))) & (2**num_bits - 1) 48 | return index 49 | 50 | 51 | def count_1(n): 52 | return bin(n).count("1") 53 | 54 | 55 | def sample_action(prob, sample=True): 56 | if sample: 57 | temp = torch.ones_like(prob) * 10 ** (-7) 58 | prob = torch.where(prob < 0, temp, prob) 59 | action = torch.distributions.Categorical(prob).sample() 60 | else: 61 | action = torch.argmax(prob, dim=-1) 62 | logprob = torch.log(torch.gather(prob, -1, action.unsqueeze(-1))).squeeze(-1) 63 | return action, logprob 64 | 65 | 66 | def log(x): 67 | return torch.log(x + LOG_EPSILON) 68 | -------------------------------------------------------------------------------- /experiments/iravenx/arlc_program_eval.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | # All rules 4 | python main.py --epochs 10 --dyn_range 50 --vsa_conversion --vsa_selection --shared_rules --config "center_single" --dataset "iravenx" --data_dir "/dccstor/saentis/data/I-RAVEN-X/shuffle" --batch_size 128 --num_workers 1 --num_rules 5 --num_terms 22 --mode test --program --seed 0 --run_name "iravenx_50_program" --exp_dir "debug" --annealing 8 --partition Arithmetic_shuffle 5 | python main.py --epochs 10 --dyn_range 100 --vsa_conversion --vsa_selection --shared_rules --config "center_single" --dataset "iravenx" --data_dir "/dccstor/saentis/data/I-RAVEN-X/shuffle" --batch_size 128 --num_workers 1 --num_rules 5 --num_terms 22 --mode test --program --seed 0 --run_name "iravenx_50_program" --exp_dir "debug" --annealing 8 --partition Arithmetic_shuffle 6 | python main.py --epochs 10 --dyn_range 1000 --vsa_conversion --vsa_selection --shared_rules --config "center_single" --dataset "iravenx" --data_dir "/dccstor/saentis/data/I-RAVEN-X/shuffle" --batch_size 128 --num_workers 1 --num_rules 5 --num_terms 22 --mode test --program --seed 0 --run_name "iravenx_50_program" --exp_dir "debug" --annealing 8 --partition Arithmetic_shuffle 7 | 8 | # Arithmetic rules 9 | python main.py --epochs 10 --dyn_range 50 --vsa_conversion --vsa_selection --shared_rules --config "center_single" --dataset "iravenx" --data_dir "/dccstor/saentis/data/I-RAVEN-X/shuffle" --batch_size 128 --num_workers 1 --num_rules 5 --num_terms 22 --mode test --program --seed 0 --run_name "iravenx_50_program" --exp_dir "debug" --annealing 8 --partition Arithmetic_shuffle 10 | python main.py --epochs 10 --dyn_range 100 --vsa_conversion --vsa_selection --shared_rules --config "center_single" --dataset "iravenx" --data_dir "/dccstor/saentis/data/I-RAVEN-X/shuffle" --batch_size 128 --num_workers 1 --num_rules 5 --num_terms 22 --mode test --program --seed 0 --run_name "iravenx_50_program" --exp_dir "debug" --annealing 8 --partition Arithmetic_shuffle 11 | python main.py --epochs 10 --dyn_range 1000 --vsa_conversion --vsa_selection --shared_rules --config "center_single" --dataset "iravenx" --data_dir "/dccstor/saentis/data/I-RAVEN-X/shuffle" --batch_size 128 --num_workers 1 --num_rules 5 --num_terms 22 --mode test --program --seed 0 --run_name "iravenx_50_program" --exp_dir "debug" --annealing 8 --partition Arithmetic_shuffle 12 | -------------------------------------------------------------------------------- /arlc/losses.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | import torch 7 | import torch.nn as nn 8 | import torch.nn.functional as F 9 | 10 | 11 | class CosineLoss(nn.Module): 12 | def __init__(self): 13 | super(CosineLoss, self).__init__() 14 | 15 | def loss(self, output, target): 16 | loss = 1 - F.cosine_similarity(output, target, dim=-1) 17 | return loss 18 | 19 | def forward(self, output, target): 20 | loss = self.loss(output, target) 21 | loss = loss.mean(dim=-1) 22 | return loss 23 | 24 | def score(self, output, targets): 25 | losses = self.loss(output, targets) 26 | score = -losses 27 | return score 28 | 29 | 30 | class CrossEntropyLoss(nn.Module): 31 | def __init__(self): 32 | super(CrossEntropyLoss, self).__init__() 33 | 34 | def loss(self, output, target): 35 | loss = torch.sum(-target * torch.log(output), dim=-1) 36 | return loss 37 | 38 | def forward(self, output, target): 39 | loss = self.loss(output, target) 40 | loss = loss.mean(dim=-1) 41 | return loss 42 | 43 | def score(self, output, targets): 44 | losses = self.loss(output, targets) 45 | score = -losses 46 | return score 47 | 48 | 49 | class KLDivLoss(nn.Module): 50 | def __init__(self): 51 | super(KLDivLoss, self).__init__() 52 | 53 | def loss(self, output, target): 54 | output_normalized = output / output.sum(dim=-1, keepdim=True) 55 | target_normalized = target / target.sum(dim=-1, keepdim=True) 56 | epsilon = 1e-10 57 | loss = torch.sum( 58 | output_normalized 59 | * ( 60 | torch.log(output_normalized + epsilon) 61 | - torch.log(target_normalized + epsilon) 62 | ), 63 | dim=-1, 64 | ) 65 | return loss 66 | 67 | def forward(self, output, target): 68 | loss = self.loss(output, target) 69 | loss = loss.mean(dim=-1) 70 | return loss 71 | 72 | def score(self, output, targets): 73 | losses = self.loss(output, targets) 74 | scores = -losses 75 | return scores 76 | -------------------------------------------------------------------------------- /experiments/iravenx/arlc_learn_iravenx_50.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | CONFIG="center_single" 4 | RUN="iravenx_50" 5 | EXP_DIR="models" 6 | EPOCHS=10 7 | NTEST=5 8 | NRULES=5 9 | DATA="iravenx" 10 | 11 | if [ "$DATA" = "iraven" ]; then 12 | DATA_DIR="path_top_standard_raven" 13 | else 14 | DATA_DIR="path_to_iravenx" 15 | fi 16 | 17 | # Train the model 18 | python main.py --epochs $EPOCHS --dyn_range 50 --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR --annealing $EPOCHS --batch_size 256 --num_workers 1 --num_rules $NRULES --num_terms 22 --seed 2 --run_name $RUN --exp_dir $EXP_DIR --partition _shuffle 19 | 20 | # Eval on unseen dynamic ranges 21 | python main.py --epochs $EPOCHS --dyn_range 100 --mode test --resume models/iravenx_50/2/ckpt --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR --annealing $EPOCHS --batch_size 256 --num_workers 1 --num_rules $NRULES --num_terms 22 --seed 0 --run_name $RUN --exp_dir $EXP_DIR --partition _shuffle 22 | 23 | python main.py --epochs $EPOCHS --dyn_range 1000 --mode test --resume models/iravenx_50/2/ckpt \ 24 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR --annealing $EPOCHS \ 25 | --batch_size 256 --num_workers 1 --num_rules $NRULES --num_terms 22 --seed 0 --run_name $RUN --exp_dir $EXP_DIR --partition _shuffle 26 | 27 | 28 | 29 | # Test arithmetic accuracies 30 | python main.py --epochs $EPOCHS --dyn_range 50 --mode test --resume models/iravenx_50/2/ckpt \ 31 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR --annealing $EPOCHS \ 32 | --batch_size 256 --num_workers 1 --num_rules $NRULES --num_terms 22 --seed 0 --run_name $RUN --exp_dir $EXP_DIR --partition Arithmetic_shuffle 33 | 34 | python main.py --epochs $EPOCHS --dyn_range 100 --mode test --resume models/iravenx_50/2/ckpt \ 35 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR --annealing $EPOCHS \ 36 | --batch_size 256 --num_workers 1 --num_rules $NRULES --num_terms 22 --seed 0 --run_name $RUN --exp_dir $EXP_DIR --partition Arithmetic_shuffle 37 | 38 | python main.py --epochs $EPOCHS --dyn_range 1000 --mode test --resume models/iravenx_50/2/ckpt \ 39 | --vsa_conversion --vsa_selection --shared_rules --config $CONFIG --dataset $DATA --data_dir $DATA_DIR --annealing $EPOCHS \ 40 | --batch_size 256 --num_workers 1 --num_rules $NRULES --num_terms 22 --seed 0 --run_name $RUN --exp_dir $EXP_DIR --partition Arithmetic_shuffle -------------------------------------------------------------------------------- /arlc/selection.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | from collections import namedtuple 7 | import torch as t 8 | import torch.nn as nn 9 | import torch.nn.functional as F 10 | import numpy as np 11 | 12 | 13 | class RuleSelector(nn.Module): 14 | def __init__(self, loss_fn, temperature, rule_selector="sample"): 15 | super(RuleSelector, self).__init__() 16 | self.loss_fn = loss_fn 17 | self.temperature = temperature 18 | self.train_mode = True 19 | 20 | def train(self): 21 | self.train_mode = True 22 | 23 | def eval(self): 24 | self.train_mode = False 25 | 26 | def attribute_forward(self, outputs, tests, candidates=None, targets=None): 27 | if self.train_mode: 28 | tests = ( 29 | t.cat( 30 | ( 31 | tests, 32 | candidates[t.arange(candidates.shape[0]), targets].unsqueeze(1), 33 | ), 34 | dim=1, 35 | ) 36 | .unsqueeze(1) 37 | .expand(-1, outputs.shape[1], -1, -1) 38 | ) 39 | scores = self.loss_fn.score(outputs, tests).mean(dim=-1) 40 | weights = F.softmax(scores / self.temperature, dim=-1) 41 | else: 42 | tests = tests.unsqueeze(1).expand(-1, outputs.shape[1], -1, -1) 43 | scores = self.loss_fn.score(outputs[:, :, :2], tests).mean(dim=-1) 44 | weights = F.softmax(scores / self.temperature, dim=-1) 45 | outputs = t.einsum("ijkh,ij->ikh", outputs, weights) 46 | return outputs, weights 47 | 48 | def _entropy(self, dist): 49 | dist = dist.detach().cpu().numpy() 50 | entropy = -(dist * np.log(dist)) 51 | return entropy[~np.isnan(entropy)].sum() / dist.shape[0] 52 | 53 | def forward(self, outputs, tests, candidates=None, targets=None, use_position=True): 54 | res = {} 55 | rules = {} 56 | weights = {} 57 | for attr in outputs._fields: 58 | if attr in ["position", "number"] and ( 59 | not use_position or outputs.position is None 60 | ): 61 | res[attr] = rules[attr] = None 62 | continue 63 | res[attr], weights[attr] = self.attribute_forward( 64 | getattr(outputs, attr), 65 | getattr(tests, attr), 66 | getattr(candidates, attr, None), 67 | targets, 68 | ) 69 | res = type(outputs)(**res) 70 | entropy_attr = {k: self._entropy(v) for k, v in weights.items()} 71 | return res, entropy_attr 72 | -------------------------------------------------------------------------------- /arlc/utils/vsa.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | from collections import namedtuple 7 | from nvsa.reasoning.vsa_block_utils import ( 8 | pmf2vec, 9 | binding_circular, 10 | block_discrete_codebook, 11 | ) 12 | from arlc.utils.const import ( 13 | DIM_POSITION_2x2, 14 | DIM_POSITION_3x3, 15 | DIM_NUMBER_2x2, 16 | DIM_NUMBER_3x3, 17 | DIM_ONEHOT, 18 | ) 19 | import torch.nn as nn 20 | from nvsa.reasoning.vsa_block_utils import ( 21 | block_discrete_codebook, 22 | block_continuous_codebook, 23 | ) 24 | 25 | 26 | def generate_nvsa_codebooks(args, rng): 27 | """ 28 | Generate the codebooks for NVSA frontend and backend. 29 | The codebook can also be loaded if it is stored under args.resume/ 30 | """ 31 | backend_cb_cont, _ = block_continuous_codebook( 32 | device=args.device, 33 | scene_dim=1024, 34 | d=args.nvsa_backend_d, 35 | k=args.nvsa_backend_k, 36 | rng=rng, 37 | fully_orthogonal=False, 38 | ) 39 | backend_cb_discrete, _ = block_discrete_codebook( 40 | device=args.device, d=args.nvsa_backend_d, k=args.nvsa_backend_k, rng=rng 41 | ) 42 | return backend_cb_cont, backend_cb_discrete 43 | 44 | 45 | class VSAConverter(nn.Module): 46 | def __init__( 47 | self, 48 | device, 49 | constellation, 50 | dictionary, 51 | dictionary_type="Discrete", 52 | context_dim=8, 53 | attributes_superposition=False, 54 | ): 55 | super(VSAConverter, self).__init__() 56 | self.device = device 57 | self.constellation = constellation 58 | self.d = dictionary.shape[1] * dictionary.shape[2] 59 | self.k = dictionary.shape[1] 60 | self.dictionary = dictionary 61 | self.dictionary_type = dictionary_type 62 | self.compute_attribute_dicts() 63 | self.context_dim = context_dim 64 | self.attributes_superposition = attributes_superposition 65 | if self.attributes_superposition: 66 | attribute_keys, _ = block_discrete_codebook( 67 | device=device, d=self.d, k=self.k, scene_dim=5 68 | ) 69 | self.attribute_keys = nn.Parameter(attribute_keys) 70 | 71 | def compute_attribute_dicts(self): 72 | if "distribute" in self.constellation or "in_out_four" == self.constellation: 73 | if "four" in self.constellation: 74 | DIM_POSITION = DIM_POSITION_2x2 75 | DIM_NUMBER = DIM_NUMBER_2x2 76 | else: 77 | DIM_POSITION = DIM_POSITION_3x3 78 | DIM_NUMBER = DIM_NUMBER_3x3 79 | self.position_dictionary = self.dictionary[:DIM_POSITION] 80 | self.number_dictionary = self.dictionary[:DIM_NUMBER] 81 | 82 | def compute_values(self, scene_prob): 83 | vsas = {} 84 | for attr in scene_prob._fields: 85 | if attr == "position" and ( 86 | "distribute" in self.constellation 87 | or "in_out_four" == self.constellation 88 | ): 89 | vsas[attr] = pmf2vec(self.position_dictionary, scene_prob.position) 90 | elif attr == "number" and ( 91 | "distribute" in self.constellation 92 | or "in_out_four" == self.constellation 93 | ): 94 | vsas[attr] = pmf2vec(self.number_dictionary, scene_prob.number) 95 | elif attr in ["position", "number"]: 96 | vsas[attr] = None 97 | else: 98 | vsas[attr] = pmf2vec( 99 | self.dictionary[: DIM_ONEHOT + 1], getattr(scene_prob, attr) 100 | ) 101 | return type(scene_prob)(**vsas) 102 | 103 | def forward(self, scene_prob): 104 | return self.compute_values(scene_prob) 105 | -------------------------------------------------------------------------------- /arlc/utils/raven/scene.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | from collections import OrderedDict, namedtuple 7 | from itertools import product 8 | import numpy as np 9 | import torch 10 | import arlc.utils.general as utils 11 | 12 | 13 | class SceneEngine: 14 | def __init__(self, number_slots, device): 15 | self.device = device 16 | self.num_slots = number_slots 17 | self.positions = list(product(range(2), repeat=self.num_slots)) 18 | # assume nonempty 19 | start_index = 1 20 | position2number = np.sum(self.positions[start_index:], axis=1) 21 | # note the correspondence of positions: first digit from the left corresponds to part one 22 | self.positions = torch.tensor(self.positions[start_index:], dtype=torch.int).to( 23 | self.device 24 | ) 25 | self.dim_position = self.positions.shape[0] 26 | self.num_pos_index_map = OrderedDict() 27 | for i in range(start_index, self.num_slots + 1): 28 | self.num_pos_index_map[i] = torch.tensor( 29 | list( 30 | filter( 31 | lambda idx: position2number[idx] == i, 32 | range(len(position2number)), 33 | ) 34 | ), 35 | dtype=torch.long, 36 | ).to(self.device) 37 | 38 | def compute_scene_prob(self, **attribute_logprobs): 39 | position_prob, position_logprob = self.compute_position_prob( 40 | attribute_logprobs.pop("exist") 41 | ) 42 | number_prob, number_logprob = self.compute_number_prob(position_prob) 43 | SceneProb = namedtuple( 44 | "SceneProb", ["position", "number"] + [k for k in attribute_logprobs.keys()] 45 | ) 46 | SceneLogProb = namedtuple( 47 | "SceneLogProb", 48 | ["position", "number"] + [k for k in attribute_logprobs.keys()], 49 | ) 50 | attr_probs = { 51 | k: self.compute_attribute_prob(v, position_logprob) 52 | for k, v in attribute_logprobs.items() 53 | } 54 | att_logprobs = {k: utils.log(v) for k, v in attr_probs.items()} 55 | return ( 56 | SceneProb(position_prob, number_prob, **attr_probs), 57 | SceneLogProb(position_logprob, number_logprob, **att_logprobs), 58 | ) 59 | 60 | def compute_position_prob(self, exist_logprob): 61 | batch_size = exist_logprob.shape[0] 62 | num_panels = exist_logprob.shape[1] 63 | exist_logprob = exist_logprob.unsqueeze(2).expand( 64 | -1, -1, self.dim_position, -1, -1 65 | ) 66 | index = ( 67 | self.positions.unsqueeze(0) 68 | .unsqueeze(0) 69 | .expand(batch_size, num_panels, -1, -1) 70 | .unsqueeze(-1) 71 | .type(torch.long) 72 | ) 73 | position_logprob = torch.gather( 74 | exist_logprob, -1, index 75 | ) # (batch_size, num_panels, self.dim_position, slots, 1) 76 | position_logprob = torch.sum( 77 | position_logprob.squeeze(-1), dim=-1 78 | ) # (batch_size, num_panels, self.dim_position) 79 | position_prob = torch.exp(position_logprob) 80 | # assume nonempty: all zero state is filtered out 81 | position_prob = utils.normalize(position_prob)[0] 82 | position_logprob = utils.log(position_prob) 83 | return position_prob, position_logprob 84 | 85 | def compute_number_prob(self, position_prob): 86 | all_num_prob = [] 87 | for _, indices in self.num_pos_index_map.items(): 88 | num_prob = torch.sum(position_prob[:, :, indices], dim=-1, keepdim=True) 89 | all_num_prob.append(num_prob) 90 | number_prob = torch.cat(all_num_prob, dim=-1) 91 | return number_prob, utils.log(number_prob) 92 | 93 | def compute_attribute_prob(self, logprob, position_logprob): 94 | batch_size = logprob.shape[0] 95 | num_panels = logprob.shape[1] 96 | index = ( 97 | self.positions.unsqueeze(0) 98 | .unsqueeze(0) 99 | .expand(batch_size, num_panels, -1, -1) 100 | .unsqueeze(-1) 101 | .type(torch.float) 102 | ) 103 | logprob = logprob.unsqueeze(2).expand(-1, -1, self.dim_position, -1, -1) 104 | logprob = ( 105 | index * logprob 106 | ) # (batch_size, num_panels, self.dim_position, slots, DIM_TYPE) 107 | logprob = torch.sum(logprob, dim=3) + position_logprob.unsqueeze(-1) 108 | prob = torch.exp(logprob) 109 | prob = torch.sum(prob, dim=2) 110 | inconsist_prob = 1.0 - torch.clamp( 111 | torch.sum(prob, dim=-1, keepdim=True), max=1.0 112 | ) # clamp for numerical stability 113 | prob = torch.cat([prob, inconsist_prob], dim=-1) 114 | return torch.nan_to_num(prob, nan=0.0) 115 | -------------------------------------------------------------------------------- /arlc/datasets/iraven.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | import os 7 | import numpy as np 8 | import torch as t 9 | from torch.utils.data import Dataset 10 | import random 11 | import json 12 | 13 | rule_map = {"Constant": 0, "Progression": 1, "Arithmetic": 2, "Distribute_Three": 3} 14 | 15 | 16 | class GeneralIRAVENDataset(Dataset): 17 | def __init__( 18 | self, 19 | dataset_type, 20 | data_dir, 21 | constellation_filter, 22 | rule_filter="", 23 | attribute_filter="", 24 | n_train=None, 25 | in_memory=False, 26 | partition="", 27 | n=10, 28 | n_show=3, 29 | maxval=1000, 30 | n_confounders=0, 31 | ): 32 | 33 | self.n = n 34 | self.n_show = n_show 35 | self.n_tot = n_show * n - 1 + 8 36 | self.n_confounders = n_confounders 37 | self.maxval = maxval 38 | 39 | if dataset_type == "train": 40 | self.filtered_indeces = np.arange(6000) 41 | elif dataset_type == "val": 42 | self.filtered_indeces = np.arange(6000, 8000) 43 | elif dataset_type == "test": 44 | self.filtered_indeces = np.arange(8000, 10000) 45 | 46 | if rule_filter != "" or attribute_filter != "": 47 | raise ValueError("Rule filtering not implemented") 48 | 49 | if n_train: 50 | self.filtered_indeces = self.filtered_indeces[:n_train] 51 | 52 | self.old_raven = not "I-RAVEN-" in data_dir 53 | data_file = ( 54 | f"{constellation_filter}{partition}_n_{n}_maxval_{maxval}.json" 55 | ) 56 | print(f"Number of confounders: {n_confounders}") 57 | self.constellation = constellation_filter 58 | # load entire dataset from 59 | with open(os.path.join(data_dir, data_file), "r") as f: 60 | self.dataset = json.load(f) 61 | 62 | def __len__(self): 63 | return len(self.filtered_indeces) 64 | 65 | def _get_panel_number(self, x, y): 66 | if not (0 <= x <= 1 and 0 <= y <= 1): 67 | raise ValueError("Point is outside the 1x1 box") 68 | if self.constellation == "distribute_nine": 69 | div = 1 / 3 70 | ppr = 3 71 | elif self.constellation == "distribute_four": 72 | div = 1 / 2 73 | ppr = 2 74 | col = int(x / div) 75 | row = int((1 - y) / div) 76 | panel_number = row * ppr + col 77 | return panel_number 78 | 79 | def __getitem__(self, index): 80 | valid_index = self.filtered_indeces[index % len(self.filtered_indeces)] 81 | data = self.dataset[str(valid_index)] 82 | # dimension panel, slots, attributes 83 | input_tensor = t.ones((self.n_tot, 9, 5 + self.n_confounders)).float() * (-1) 84 | for i in range(self.n_tot): 85 | panels = data["rpm"][i + (self.n - self.n_show) * self.n][0] 86 | if self.constellation == "center_single": 87 | input_tensor[:, 0, 0] = 0 # Fix position in center constellation 88 | panel = data["rpm"][i + (self.n - self.n_show) * self.n][0] 89 | input_tensor[i, 0, 2] = int(panel["Color"]) 90 | input_tensor[i, 0, 3] = int(panel["Size"]) + self.old_raven * 1 91 | input_tensor[i, 0, 4] = int(panel["Type"]) + self.old_raven * 2 92 | input_tensor[i, 0, 1] = int(panel["Angle"]) 93 | for n in range(self.n_confounders): 94 | input_tensor[i, 0, 5 + n] = int(panel[f"Confounder{n}"]) 95 | else: 96 | for pidx, (pos, ent) in enumerate( 97 | zip(panels["positions"], panels["entities"]) 98 | ): 99 | input_tensor[i, pidx, 2] = int(ent["Color"]) 100 | input_tensor[i, pidx, 3] = int(ent["Size"]) 101 | input_tensor[i, pidx, 4] = int(ent["Type"]) 102 | input_tensor[i, pidx, 1] = int(ent["Angle"]) 103 | input_tensor[i, pidx, 0] = int(self._get_panel_number(*pos[:2])) 104 | for n in range(self.n_confounders): 105 | input_tensor[i, 0, 5 + n] = random.randint(0, self.maxval) 106 | 107 | label_tensor = t.tensor(int(data["target"])).long() 108 | rules = data["rules"][0] 109 | if "Number/Position" in rules: 110 | num_pos = "Number/Position" 111 | elif "Number" in rules: 112 | num_pos = "Number" 113 | else: 114 | num_pos = "Position" 115 | pos_num_rule = t.tensor(np.array(rule_map[rules[num_pos]])).float() 116 | color_rule = t.tensor(np.array(rule_map[rules["Color"]])).float() 117 | size_rule = t.tensor(np.array(rule_map[rules["Size"]])).float() 118 | type_rule = t.tensor(np.array(rule_map[rules["Type"]])).float() 119 | rules_tensor = t.stack([pos_num_rule, color_rule, size_rule, type_rule]) 120 | return input_tensor, label_tensor, rules_tensor 121 | 122 | 123 | if __name__ == "__main__": 124 | dataset = GeneralIRAVENDataset( 125 | "train", 126 | "/dccstor/saentis/data/I-RAVEN-X", 127 | "center_single", 128 | n=10, 129 | maxval=100, 130 | partition="_shuffle", 131 | n_confounders=10, 132 | ) 133 | print(dataset.__getitem__(0)) 134 | -------------------------------------------------------------------------------- /arlc/utils/raven/env.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | from arlc.utils.raven.scene import SceneEngine 7 | 8 | 9 | def get_env(env_name, device, **kwargs): 10 | if env_name == "center_single": 11 | return CenterSingle(device, **kwargs) 12 | if env_name == "distribute_four": 13 | return DistributeFour(device, **kwargs) 14 | if env_name == "distribute_nine": 15 | return DistributeNine(device, **kwargs) 16 | if env_name == "in_center_single_out_center_single": 17 | return InCenterSingleOutCenterSingle(device, **kwargs) 18 | if env_name == "in_distribute_four_out_center_single": 19 | return InDistributeFourOutCenterSingle(device, **kwargs) 20 | if env_name == "left_center_single_right_center_single": 21 | return LeftCenterSingleRightCenterSingle(device, **kwargs) 22 | if env_name == "up_center_single_down_center_single": 23 | return UpCenterSingleDownCenterSingle(device, **kwargs) 24 | return None 25 | 26 | 27 | class GeneralEnv(object): 28 | def __init__(self, num_slots, device, **kwargs): 29 | self.num_slots = num_slots 30 | self.device = device 31 | self.scene_engine = SceneEngine(self.num_slots, device) 32 | 33 | def prepare(self, model_output): 34 | return self.scene_engine.compute_scene_prob(**model_output) 35 | 36 | 37 | class CenterSingle(GeneralEnv): 38 | def __init__(self, device, **kwargs): 39 | super(CenterSingle, self).__init__(1, device, **kwargs) 40 | 41 | 42 | class DistributeFour(GeneralEnv): 43 | def __init__(self, device, **kwargs): 44 | super(DistributeFour, self).__init__(4, device, **kwargs) 45 | 46 | 47 | class DistributeNine(GeneralEnv): 48 | def __init__(self, device, **kwargs): 49 | super(DistributeNine, self).__init__(9, device, **kwargs) 50 | 51 | 52 | class OutCenterSingle(GeneralEnv): 53 | def __init__(self, device, **kwargs): 54 | super(OutCenterSingle, self).__init__(1, device, **kwargs) 55 | 56 | 57 | class InCenterSingleOutCenterSingle(object): 58 | def __init__(self, device, **kwargs): 59 | self.in_center_single = CenterSingle(device, **kwargs) 60 | self.out_center_single = OutCenterSingle(device, **kwargs) 61 | 62 | def prepare(self, model_output): 63 | in_component = [] 64 | out_component = [] 65 | for element in model_output: 66 | in_component.append(element[:, :, 1:, :]) 67 | out_component.append(element[:, :, :1, :]) 68 | in_scene_prob, in_scene_logprob = self.in_center_single.prepare(in_component) 69 | out_scene_prob, out_scene_logprob = self.out_center_single.prepare( 70 | out_component 71 | ) 72 | return (in_scene_prob, out_scene_prob), (in_scene_logprob, out_scene_logprob) 73 | 74 | 75 | class InDistributeFourOutCenterSingle(object): 76 | def __init__(self, device, **kwargs): 77 | self.in_distribute_four = DistributeFour(device, **kwargs) 78 | self.out_center_single = OutCenterSingle(device, **kwargs) 79 | 80 | def prepare(self, model_output): 81 | in_component = [] 82 | out_component = [] 83 | for element in model_output: 84 | in_component.append(element[:, :, 1:5, :]) 85 | out_component.append(element[:, :, :1, :]) 86 | in_scene_prob, in_scene_logprob = self.in_distribute_four.prepare(in_component) 87 | out_scene_prob, out_scene_logprob = self.out_center_single.prepare( 88 | out_component 89 | ) 90 | return (in_scene_prob, out_scene_prob), (in_scene_logprob, out_scene_logprob) 91 | 92 | 93 | class LeftCenterSingleRightCenterSingle(object): 94 | def __init__(self, device, **kwargs): 95 | self.left_center_single = CenterSingle(device, **kwargs) 96 | self.right_center_single = CenterSingle(device, **kwargs) 97 | 98 | def prepare(self, model_output): 99 | left_component = [] 100 | right_component = [] 101 | for element in model_output: 102 | left_component.append(element[:, :, :1, :]) 103 | right_component.append(element[:, :, 1:, :]) 104 | left_scene_prob, left_scene_logprob = self.left_center_single.prepare( 105 | left_component 106 | ) 107 | right_scene_prob, right_scene_logprob = self.right_center_single.prepare( 108 | right_component 109 | ) 110 | return (left_scene_prob, right_scene_prob), ( 111 | left_scene_logprob, 112 | right_scene_logprob, 113 | ) 114 | 115 | 116 | class UpCenterSingleDownCenterSingle(object): 117 | def __init__(self, device, **kwargs): 118 | self.up_center_single = CenterSingle(device, **kwargs) 119 | self.down_center_single = CenterSingle(device, **kwargs) 120 | 121 | def prepare(self, model_output): 122 | up_component = [] 123 | down_component = [] 124 | for element in model_output: 125 | up_component.append(element[:, :, :1, :]) 126 | down_component.append(element[:, :, 1:, :]) 127 | up_scene_prob, up_scene_logprob = self.up_center_single.prepare(up_component) 128 | down_scene_prob, down_scene_logprob = self.down_center_single.prepare( 129 | down_component 130 | ) 131 | return (up_scene_prob, down_scene_prob), (up_scene_logprob, down_scene_logprob) 132 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Abductive Rule Learner with Context Awareness (ARLC) 2 | 3 |
4 | 5 |
6 | 7 | This repo contains the code for the Abductive Rule Learner with Context Awareness (ARLC), a probabilistic abductive reasoner for solving Raven's progressive matrices (RPM). 8 | The repo features the code used to run experiments in three publications: 9 | - Giacomo Camposampiero, Michael Hersche, Aleksandar Terzić, Roger Wattenhofer, Abu Sebastian and Abbas Rahimi. *Towards Learning Abductive Reasoning using VSA Distributed Representations*. 18th International Conference on Neural-Symbolic Learning and Reasoning (NeSy) **[Spotlight]**, 2024. [[Paper]](http://arxiv.org/abs/2406.19121) 10 | - Michael Hersche, Giacomo Camposampiero, Roger Wattenhofer, Abu Sebastian and Abbas Rahimi. *Towards Learning to Reason: Comparing LLMs with Neuro-Symbolic on Arithmetic Relations in Abstract Reasoning*. Neural Reasoning and Mathematical Discovery (NEURMAD) @ AAAI, 2025. [[Paper]](https://arxiv.org/pdf/2412.05586) 11 | - Giacomo Camposampiero‡, Michael Hersche‡, Roger Wattenhofer, Abu Sebastian and Abbas Rahimi. *Can Large Reasoning Models do Analogical Reasoning under Perceptual Uncertainty?*. arXiv, 2025. [[Paper]]() 12 | 13 | ‡ these authors contributed equally. 14 | 15 | ## Build the Environment 🛠️ 16 | 17 | #### Hardware 18 | You will need a machine with a CUDA-enabled GPU and the Nvidia SDK installed to compile the CUDA kernels. We tested our methods on an NVIDA Tesla V100 GPU with CUDA Version 11.3.1. 19 | 20 | #### Installing Dependencies 21 | 22 | The `mamba` software is required for running the code. You can create a new mamba environment using 23 | 24 | ```bash 25 | mamba create --name arlc python=3.7 26 | mamba activate arlc 27 | ``` 28 | 29 | To install PyTorch 1.11 and CUDA, use 30 | ```bash 31 | mamba install pytorch==1.11.0 torchvision==0.12.0 cudatoolkit=11.3 -c pytorch -c conda-forge 32 | ``` 33 | 34 | Clone and install the [neuro-vsa repo](https://github.com/IBM/neuro-vector-symbolic-architectures) (some of their utils are re-used in this project) 35 | ```bash 36 | git clone https://github.com/IBM/neuro-vector-symbolic-architectures.git 37 | cd neuro-vector-symbolic-architectures 38 | pip install -e . --no-dependencies 39 | ``` 40 | 41 | Finally, clone and install this repo 42 | ```bash 43 | git clone https://github.com/IBM/abductive-rule-learner-with-context-awareness.git 44 | cd abductive-rule-learner-with-context-awareness 45 | pip install -r requirements.txt 46 | pip install -e . 47 | pre-commit install 48 | ``` 49 | 50 | We suggest to format the code of the entire repository to improve its readability. 51 | To do so, please install and run `black` 52 | ```bash 53 | pip install black 54 | black abductive-rule-learner-with-context-awareness/ 55 | ``` 56 | 57 | 58 | #### I-RAVEN Dataset 59 | You can find the instructions to download and pre-process the data in the `data` folder. 60 | 61 | 62 | ## Run our Experiments 🔬 63 | You can replicate the main experiments shown in the paper with the following scripts 64 | ```bash 65 | # ARLC learned from data 66 | ./experiments/arlc_learn.sh 67 | # ARLC initialized with programming, then learned 68 | ./experiments/arlc_progr_to_learn.sh 69 | # ARLC programmed and evaluated 70 | ./experiments/arlc_progr.sh 71 | ``` 72 | 73 | To replicate our ablations on the introduced contributions, run 74 | ```bash 75 | # line 1 ablation table 76 | # obtained with the code from https://github.com/IBM/learn-vector-symbolic-architectures-rule-formulations, modified to run with multiple random seeds 77 | 78 | # line 2 ablation table 79 | ./experiments/ablations/learnvrf_nopn_2x2.sh 80 | 81 | # line 3 ablation table 82 | ./experiments/ablations/context.sh 83 | 84 | # line 4 ablation table 85 | # same as ./experiments/arlc_learn.sh 86 | ``` 87 | 88 | To replicate our OOD experiments, run 89 | ```bash 90 | ./experiments/arlc_ood.sh 91 | ``` 92 | 93 | ## I-RAVEN-X dataset evaluation 94 | To replicate the results reported in _Towards Learning Abductive Reasoning using VSA Distributed Representations_ (Hersche et al., 2024) on the novel I-RAVEN-X dataset, 95 | run the experiments in the `experiments/iravenx` folder. 96 | - `arlc_learn_iravenx_50.sh` allows to train an ARLC model from scratch, and evaluate it on both the full I-RAVEN-X dataset and the subset of Arithmetic rules. 97 | - `arlc_program_eval.sh` allows to evaluate the programmed ARLC on both the full I-RAVEN-X dataset and the subset of Arithmetic rules. 98 | 99 | ## I-RAVEN-X with perceptual uncertainty dataset evaluation 100 | To replicate the results reported in *Can Large Reasoning Models do Analogical Reasoning under Perceptual Uncertainty?* (Camposampiero, Hersche et al., 2025) on the novel I-RAVEN-X dataset with perceptual uncertainty, run the experiments in the `experiments/uncertainty` folder. 101 | 102 | ## Citation 📚 103 | If you use the work released here for your research, please consider citing our paper: 104 | ``` 105 | @inproceedings{camposampiero2024towards, 106 | title={Towards Learning Abductive Reasoning using VSA Distributed Representations}, 107 | author={Camposampiero, Giacomo and Hersche, Michael and Terzi{\'c}, Aleksandar and Wattenhofer, Roger and Sebastian, Abu and Rahimi, Abbas}, 108 | booktitle={18th International Conference on Neural-Symbolic Learning and Reasoning (NeSy)}, 109 | year={2024}, 110 | month={sep} 111 | } 112 | ``` 113 | 114 | 115 | ## License 🔏 116 | Please refer to the LICENSE file for the licensing of our code. Our implementation relies on [PrAE](https://github.com/WellyZhang/PrAE) released under GPL v3.0 and [Learn-VRF](https://github.com/IBM/learn-vector-symbolic-architectures-rule-formulations) released under GPL v3.0, as well as on [In-Context Analgoical Reasoning with Pre-Trained Language Models](https://github.com/hxiaoyang/lm-raven) distributed under the MIT license. 117 | -------------------------------------------------------------------------------- /arlc/utils/parsing.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | import argparse 7 | 8 | 9 | def eval_parse_args(): 10 | arg_parser = argparse.ArgumentParser() 11 | arg_parser.add_argument("--path", type=str) 12 | arg_parser.add_argument("--seeds", type=int) 13 | args = arg_parser.parse_args() 14 | return args 15 | 16 | 17 | def parse_args(): 18 | arg_parser = argparse.ArgumentParser( 19 | description="NVSA lernable backend training and evaluation on RAVEN" 20 | ) 21 | arg_parser.add_argument("--n", type=int) 22 | arg_parser.add_argument("--run_name", type=str) 23 | arg_parser.add_argument("--mode", type=str, default="train", help="Train/test") 24 | arg_parser.add_argument("--exp_dir", type=str, default="results/") 25 | arg_parser.add_argument("--dataset", type=str, default="iraven") 26 | arg_parser.add_argument("--data_dir", type=str, default="dataset/") 27 | arg_parser.add_argument("--dyn_range", type=int, default=-1) 28 | arg_parser.add_argument("--rule_type", type=str, default="arlc") 29 | arg_parser.add_argument("--num_terms", type=int, default=12) 30 | arg_parser.add_argument( 31 | "--resume", type=str, default="", help="Resume from a initialized model" 32 | ) 33 | arg_parser.add_argument("--seed", type=int, default=1234, help="Random number seed") 34 | arg_parser.add_argument("--run", type=int, default=0, help="Run id") 35 | 36 | # Dataset 37 | arg_parser.add_argument( 38 | "--partition", 39 | type=str, 40 | default="", 41 | ) 42 | arg_parser.add_argument( 43 | "--config", 44 | type=str, 45 | default="center_single", 46 | help="The configuration used for training", 47 | ) 48 | arg_parser.add_argument( 49 | "--gen_attribute", 50 | type=str, 51 | default="", 52 | help="Generalization experiment [Type, Size, Color]", 53 | ) 54 | arg_parser.add_argument( 55 | "--gen_rule", 56 | type=str, 57 | default="", 58 | help="Generalization experiment [Arithmetic, Constant, Progression, Distribute_Three]", 59 | ) 60 | arg_parser.add_argument("--n-train", type=int, default=None) 61 | 62 | # Training hyperparameters 63 | arg_parser.add_argument( 64 | "--model", 65 | type=str, 66 | default="LearnableFormula", 67 | help="Model used in the reasoner (LearnableFormula, MLP)", 68 | ) 69 | arg_parser.add_argument( 70 | "--epochs", type=int, default=50, help="The number of training epochs" 71 | ) 72 | arg_parser.add_argument("--batch_size", type=int, default=4, help="Size of batch") 73 | arg_parser.add_argument("--lr", type=float, default=0.01, help="Learning rate") 74 | arg_parser.add_argument( 75 | "--weight-decay", 76 | type=float, 77 | default=0, 78 | help="Weight decay of optimizer, same as l2 reg", 79 | ) 80 | arg_parser.add_argument( 81 | "--num_workers", type=int, default=8, help="Number of workers for data loader" 82 | ) 83 | arg_parser.add_argument( 84 | "--clip", 85 | type=float, 86 | default=10, 87 | help="Max value/norm in gradient clipping (now l2 norm)", 88 | ) 89 | arg_parser.add_argument( 90 | "--vsa_conversion", 91 | action="store_true", 92 | default=False, 93 | help="Use or not the VSA converter", 94 | ) 95 | arg_parser.add_argument( 96 | "--vsa_selection", 97 | action="store_true", 98 | default=False, 99 | help="Use or not the VSA selector", 100 | ) 101 | arg_parser.add_argument( 102 | "--context_superposition", 103 | action="store_true", 104 | default=False, 105 | help="Use or not the VSA selector", 106 | ) 107 | arg_parser.add_argument( 108 | "--program", 109 | action="store_true", 110 | default=False, 111 | help="Program the model with golden weights", 112 | ) 113 | arg_parser.add_argument("--evaluate-rule", action="store_true") 114 | arg_parser.add_argument( 115 | "--loss_fn", type=str, default="CosineLoss", help="Loss to use in the training" 116 | ) 117 | arg_parser.add_argument( 118 | "--num_rules", type=int, default=5, help="Number of rules per each attribute" 119 | ) 120 | arg_parser.add_argument("--annealing", type=int, default=-1) 121 | arg_parser.add_argument( 122 | "--rule_selector_temperature", 123 | type=float, 124 | default=0.01, 125 | help="Temperature used in the rule selector's softmax", 126 | ) 127 | arg_parser.add_argument( 128 | "--rule_selector", type=str, default="weight", help="Can be sample or weight" 129 | ) 130 | arg_parser.add_argument( 131 | "--shared_rules", 132 | action="store_true", 133 | default=False, 134 | help="Share the same rules across different attributes", 135 | ) 136 | arg_parser.add_argument( 137 | "--hidden_layers", 138 | type=int, 139 | default=3, 140 | help="Number of hidden MLP layers to use in the neural model", 141 | ) 142 | 143 | # NVSA backend settings 144 | arg_parser.add_argument( 145 | "--nvsa-backend-d", type=int, default=1024, help="VSA dimension in backend" 146 | ) 147 | arg_parser.add_argument( 148 | "--nvsa-backend-k", type=int, default=4, help="Number of blocks in VSA vectors" 149 | ) 150 | arg_parser.add_argument( 151 | "--orientation-confounder", 152 | type=int, 153 | default=0, 154 | ) 155 | arg_parser.add_argument( 156 | "--entropy", 157 | action="store_true", 158 | default=False, 159 | ) 160 | arg_parser.add_argument( 161 | "--sigma", 162 | type=float, 163 | default=0.1, 164 | ) 165 | args = arg_parser.parse_args() 166 | return args 167 | -------------------------------------------------------------------------------- /arlc/execution.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | import torch as t 7 | import torch.nn as nn 8 | from arlc.rule_templates import ( 9 | ExtendedGeneralLearnableFormula, 10 | IravenxGeneralLearnableFormula, 11 | IravenVGeneralLearnableFormula, 12 | ) 13 | from arlc.utils.vsa import VSAConverter 14 | 15 | 16 | class RuleLevelReasoner(nn.Module): 17 | def __init__( 18 | self, 19 | device, 20 | constellation, 21 | model, 22 | hidden_layers, 23 | dictionary, 24 | vsa_conversion=False, 25 | vsa_selection=False, 26 | context_superposition=False, 27 | num_rules=5, 28 | shared_rules=False, 29 | program=False, 30 | rule_type="arlc", 31 | num_terms=12, 32 | n=3, 33 | ): 34 | super(RuleLevelReasoner, self).__init__() 35 | self.device = device 36 | self.constellation = constellation 37 | self.model = model 38 | self.program = program 39 | self.rule_type = rule_type 40 | self.num_terms = num_terms 41 | self.num_panels = n 42 | self.d = dictionary.shape[1] * dictionary.shape[2] 43 | self.k = dictionary.shape[1] 44 | self.vsa_conversion = vsa_conversion 45 | self.vsa_selection = vsa_selection 46 | self.context_superposition = context_superposition 47 | self.vsa_converter = VSAConverter( 48 | device, self.constellation, dictionary, dictionary_type="Continuous" 49 | ) 50 | self.num_rules = num_rules 51 | self.rules_set = RulesSet( 52 | model=self.model, 53 | hidden_layers=hidden_layers, 54 | num_rules=self.num_rules, 55 | d_in=2 * self.d, 56 | d_out=-1, 57 | d_vsa=self.d, 58 | k=self.k, 59 | context_superpostion=self.context_superposition, 60 | context_keys=None, 61 | program=self.program, 62 | rule_type=self.rule_type, 63 | num_terms=self.num_terms, 64 | num_panels=self.num_panels, 65 | ) 66 | 67 | def forward(self, scene_prob, targets=None, distribute=False): 68 | # convert logprob to VSAs 69 | scene_vsa = self.vsa_converter(scene_prob) 70 | # flatten scene 71 | scene = {} 72 | for attr in scene_vsa._fields: 73 | if attr in ["position", "number"] and not distribute: 74 | scene[attr] = None 75 | else: 76 | scene[attr] = t.flatten( 77 | getattr(scene_vsa, attr), 78 | start_dim=len(getattr(scene_vsa, attr).shape) - 2, 79 | ) 80 | scene = type(scene_vsa)(**scene) 81 | # set indices for test panels 82 | if self.num_panels == 10: 83 | test_indeces = [9, 19] 84 | elif self.num_panels == 3: 85 | test_indeces = [2, 5] 86 | elif self.num_panels == 5: 87 | test_indeces = [4, 9] 88 | # compute output vectors 89 | output = dict() 90 | tests = dict() 91 | candidates = dict() 92 | for attr in scene._fields: 93 | if attr in ["position", "number"] and not distribute: 94 | tests[attr] = output[attr] = candidates[attr] = None 95 | else: 96 | tests[attr] = getattr(scene, attr)[:, test_indeces] 97 | output[attr] = self.rules_set(getattr(scene, attr)) 98 | candidates[attr] = getattr(scene, attr)[:, -8:] 99 | # compile them in named tuples and return 100 | output = type(scene_vsa)(**output) 101 | tests = type(scene_vsa)(**tests) 102 | candidates = type(scene_vsa)(**candidates) 103 | return output, candidates, tests 104 | 105 | def anneal_softmax(self): 106 | for rule in self.rules_set.rules: 107 | rule.rule.anneal_softmax() 108 | 109 | 110 | class RulesSet(nn.Module): 111 | def __init__( 112 | self, 113 | model, 114 | hidden_layers, 115 | num_rules, 116 | d_in, 117 | d_out, 118 | d_vsa, 119 | k, 120 | context_superpostion=False, 121 | context_keys=None, 122 | program=None, 123 | rule_type="arlc", 124 | num_terms=12, 125 | num_panels=3, 126 | ): 127 | super(RulesSet, self).__init__() 128 | rule_class = GeneralRule 129 | if program: 130 | rules = ["add", "sub", "dist3", "progr"] 131 | else: 132 | rules = [None] * num_rules 133 | self.rules = nn.ModuleList( 134 | [ 135 | rule_class( 136 | model, 137 | hidden_layers, 138 | d_in, 139 | d_out, 140 | d_vsa, 141 | k, 142 | context_superpostion, 143 | context_keys, 144 | program_rule, 145 | num_terms=num_terms, 146 | num_panels=num_panels, 147 | ) 148 | for program_rule in rules 149 | ] 150 | ) 151 | 152 | def forward(self, attribute): 153 | output_list = [ 154 | rule(attribute).reshape((attribute.shape[0], 3, -1)) for rule in self.rules 155 | ] 156 | outputs = t.stack(output_list, dim=1) 157 | return outputs 158 | 159 | 160 | class GeneralRule(nn.Module): 161 | def __init__( 162 | self, 163 | model, 164 | hidden_layers, 165 | d_in, 166 | d_out, 167 | d_vsa, 168 | k, 169 | context_superposition=False, 170 | context_keys=None, 171 | program_rule=None, 172 | num_terms=12, 173 | num_panels=3, 174 | ): 175 | super(GeneralRule, self).__init__() 176 | self.d_in = d_in 177 | self.d_out = d_out 178 | self.d = d_vsa 179 | self.k = k 180 | self.context_superposition = context_superposition 181 | self.context_keys = context_keys 182 | if num_panels == 10: 183 | # I-RAVEN-X 184 | self.a3_indeces = range(0, 9) 185 | self.a6_indeces = range(10, 19) 186 | self.a9_indeces = range(20, 29) 187 | self.a3_context_indeces = [range(10, 20), range(20, 29)] 188 | self.a6_context_indeces = [range(0, 10), range(20, 29)] 189 | self.a9_context_indeces = [range(0, 10), range(10, 19)] 190 | self.rule = IravenxGeneralLearnableFormula( 191 | examples_len=9, 192 | context_len=19, 193 | k=self.k, 194 | num_terms=num_terms, 195 | program_rule=program_rule, 196 | ) 197 | elif num_panels == 5: 198 | # I-RAVEN-V 199 | self.a3_indeces = range(0, 4) 200 | self.a6_indeces = range(5, 9) 201 | self.a9_indeces = range(10, 14) 202 | self.a3_context_indeces = [range(5, 10), range(10, 14)] 203 | self.a6_context_indeces = [range(0, 5), range(10, 14)] 204 | self.a9_context_indeces = [range(0, 5), range(5, 9)] 205 | self.rule = IravenVGeneralLearnableFormula( 206 | examples_len=4, 207 | context_len=9, 208 | k=self.k, 209 | num_terms=num_terms, 210 | program_rule=program_rule, 211 | ) 212 | elif num_panels == 3 and num_terms == 12: 213 | # I-RAVEN 214 | self.a3_indeces = [0, 1] 215 | self.a6_indeces = [3, 4] 216 | self.a9_indeces = [6, 7] 217 | self.a3_context_indeces = [[3, 4, 5], [6, 7]] 218 | self.a6_context_indeces = [[0, 1, 2], [6, 7]] 219 | self.a9_context_indeces = [[0, 1, 2], [3, 4]] 220 | self.rule = ExtendedGeneralLearnableFormula( 221 | examples_len=2, context_len=5, k=self.k, program_rule=program_rule 222 | ) 223 | 224 | def forward(self, x): 225 | a3 = self.rule( 226 | x=x[:, self.a3_indeces], 227 | ctx=t.cat([x[:, idx] for idx in self.a3_context_indeces], dim=1), 228 | ) 229 | a6 = self.rule( 230 | x=x[:, self.a6_indeces], 231 | ctx=t.cat([x[:, idx] for idx in self.a6_context_indeces], dim=1), 232 | ) 233 | a9 = self.rule( 234 | x=x[:, self.a9_indeces], 235 | ctx=t.cat([x[:, idx] for idx in self.a9_context_indeces], dim=1), 236 | ) 237 | return t.cat((a3, a6, a9), dim=1) 238 | -------------------------------------------------------------------------------- /arlc/utils/raven/raven_one_hot.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | import torch 7 | from arlc.utils.const import DIM_ONEHOT 8 | 9 | 10 | def smooth_dist(q, sigma=0.1): 11 | if sigma < 0: 12 | return bin_dist(q, -sigma) 13 | elif sigma == 0: 14 | return q 15 | l = torch.argmax(q) 16 | temp = torch.arange(0, q.shape[0]) 17 | temp = torch.div(torch.abs(temp - l), -2 * sigma) 18 | temp = torch.exp(temp) 19 | temp = temp / temp.sum() 20 | return temp 21 | 22 | 23 | def bin_dist(q, threshold): 24 | """Smooth a probability distribution by binning. 25 | In practice, the method takes a tensor representing a PMF and returns a tensor where 26 | the probability of the most likely value PT is sampled in [q, 1], and the probability of 27 | its two neighbours PN1 and PN2 are sampled in [0, 1-q] and 1-PT-PN1, respectively. 28 | 29 | Args: 30 | q (torch.Tensor): input PMF tensor 31 | """ 32 | l = torch.argmax(q) 33 | q[l] = threshold + (1 - threshold) * torch.rand(1) 34 | right_index = (l + 1) % q.shape[0] 35 | left_index = (l - 1) % q.shape[0] 36 | q[right_index] = (1 - q[l]) * torch.rand(1) 37 | q[left_index] = 1 - q[l] - q[right_index] 38 | return q / q.sum() 39 | 40 | 41 | def create_one_hot(puzzle, panel_constellation, sigma=0.1): 42 | eps = 10 ** (-10) 43 | batch_size, num_panels, _, num_att = puzzle.shape 44 | 45 | if panel_constellation == "center_single": 46 | exist_prob = torch.ones((batch_size, num_panels, 1, 2)) * eps 47 | type_prob = torch.ones((batch_size, num_panels, 1, DIM_ONEHOT)) * eps 48 | size_prob = torch.ones((batch_size, num_panels, 1, DIM_ONEHOT)) * eps 49 | color_prob = torch.ones((batch_size, num_panels, 1, DIM_ONEHOT)) * eps 50 | angle_prob = torch.ones((batch_size, num_panels, 1, DIM_ONEHOT)) * eps 51 | confounders_prob = [ 52 | torch.ones((batch_size, num_panels, 1, DIM_ONEHOT)) * eps 53 | for _ in range(num_att - 5) 54 | ] 55 | exist_prob[:, :, 0, 1] = 1 56 | for bs in range(batch_size): 57 | for i in range(num_panels): 58 | exist_prob[bs, i, 0] = smooth_dist(exist_prob[bs, i, 0], sigma) 59 | type_prob[bs, i, 0, int(puzzle[bs, i, 0, 4])] = 1 60 | type_prob[bs, i, 0] = smooth_dist(type_prob[bs, i, 0], sigma) 61 | size_prob[bs, i, 0, int(puzzle[bs, i, 0, 3])] = 1 62 | size_prob[bs, i, 0] = smooth_dist(size_prob[bs, i, 0], sigma) 63 | color_prob[bs, i, 0, int(puzzle[bs, i, 0, 2])] = 1 64 | color_prob[bs, i, 0] = smooth_dist(color_prob[bs, i, 0], sigma) 65 | angle_prob[bs, i, 0, int(puzzle[bs, i, 0, 1])] = 1 66 | angle_prob[bs, i, 0] = smooth_dist(angle_prob[bs, i, 0], sigma) 67 | for j in range(len(confounders_prob)): 68 | confounders_prob[j][bs, i, 0, int(puzzle[bs, i, 0, 5 + j])] = 1 69 | confounders_prob[j][bs, i, 0] = smooth_dist( 70 | confounders_prob[j][bs, i, 0], sigma 71 | ) 72 | att_prob = { 73 | "exist": torch.log(exist_prob), 74 | "type": torch.log(type_prob), 75 | "size": torch.log(size_prob), 76 | "color": torch.log(color_prob), 77 | "angle": torch.log(angle_prob), 78 | } 79 | conf_prob = { 80 | f"confounder{i}": torch.log(confounders_prob[i]) 81 | for i in range(len(confounders_prob)) 82 | } 83 | return {**att_prob, **conf_prob} 84 | 85 | if panel_constellation == "distribute_four": 86 | exist_prob = torch.ones((batch_size, num_panels, 4, 2)) * eps 87 | type_prob = torch.ones((batch_size, num_panels, 4, DIM_ONEHOT)) * eps 88 | size_prob = torch.ones((batch_size, num_panels, 4, DIM_ONEHOT)) * eps 89 | color_prob = torch.ones((batch_size, num_panels, 4, DIM_ONEHOT)) * eps 90 | angle_prob = torch.ones((batch_size, num_panels, 4, DIM_ONEHOT)) * eps 91 | 92 | for bs in range(batch_size): 93 | for i in range(num_panels): 94 | temp = [0, 1, 2, 3] 95 | for j in range(4): 96 | if puzzle[bs, i, j, 0] == -1: 97 | k = temp[0] 98 | exist_prob[bs, i, k, 0] = 1 99 | exist_prob[bs, i, k] = smooth_dist(exist_prob[bs, i, k]) 100 | temp.remove(k) 101 | else: 102 | k = int(puzzle[bs, i, j, 0]) 103 | exist_prob[bs, i, k, 1] = 1 104 | exist_prob[bs, i, k] = smooth_dist(exist_prob[bs, i, k]) 105 | type_prob[bs, i, k, int(puzzle[bs, i, j, 4])] = 1 106 | type_prob[bs, i, k] = smooth_dist(type_prob[bs, i, k]) 107 | size_prob[bs, i, k, int(puzzle[bs, i, j, 3])] = 1 108 | size_prob[bs, i, k] = smooth_dist(size_prob[bs, i, k]) 109 | color_prob[bs, i, k, int(puzzle[bs, i, j, 2])] = 1 110 | color_prob[bs, i, k] = smooth_dist(color_prob[bs, i, k]) 111 | angle_prob[bs, i, k, int(puzzle[bs, i, j, 1])] = 1 112 | angle_prob[bs, i, k] = smooth_dist(angle_prob[bs, i, k]) 113 | temp.remove(k) 114 | return { 115 | "exist": torch.log(exist_prob), 116 | "type": torch.log(type_prob), 117 | "size": torch.log(size_prob), 118 | "color": torch.log(color_prob), 119 | "angle": torch.log(angle_prob), 120 | } 121 | 122 | if panel_constellation == "distribute_nine": 123 | exist_prob = torch.ones((batch_size, num_panels, 9, 2)) * eps 124 | type_prob = torch.ones((batch_size, num_panels, 9, DIM_ONEHOT)) * eps 125 | size_prob = torch.ones((batch_size, num_panels, 9, DIM_ONEHOT)) * eps 126 | color_prob = torch.ones((batch_size, num_panels, 9, DIM_ONEHOT)) * eps 127 | angle_prob = torch.ones((batch_size, num_panels, 9, DIM_ONEHOT)) * eps 128 | 129 | for bs in range(batch_size): 130 | for i in range(num_panels): 131 | temp = [0, 1, 2, 3, 4, 5, 6, 7, 8] 132 | for j in range(9): 133 | if puzzle[bs, i, j, 0] == -1: 134 | k = temp[0] 135 | exist_prob[bs, i, k, 0] = 1 136 | exist_prob[bs, i, k] = smooth_dist(exist_prob[bs, i, k]) 137 | temp.remove(k) 138 | else: 139 | k = int(puzzle[bs, i, j, 0]) 140 | exist_prob[bs, i, k, 1] = 1 141 | exist_prob[bs, i, k] = smooth_dist(exist_prob[bs, i, k]) 142 | type_prob[bs, i, k, int(puzzle[bs, i, j, 4])] = 1 143 | type_prob[bs, i, k] = smooth_dist(type_prob[bs, i, k]) 144 | size_prob[bs, i, k, int(puzzle[bs, i, j, 3])] = 1 145 | size_prob[bs, i, k] = smooth_dist(size_prob[bs, i, k]) 146 | color_prob[bs, i, k, int(puzzle[bs, i, j, 2])] = 1 147 | color_prob[bs, i, k] = smooth_dist(color_prob[bs, i, k]) 148 | angle_prob[bs, i, k, int(puzzle[bs, i, j, 1])] = 1 149 | angle_prob[bs, i, k] = smooth_dist(angle_prob[bs, i, k]) 150 | temp.remove(k) 151 | return { 152 | "exist": torch.log(exist_prob), 153 | "type": torch.log(type_prob), 154 | "size": torch.log(size_prob), 155 | "color": torch.log(color_prob), 156 | "angle": torch.log(angle_prob), 157 | } 158 | 159 | if ( 160 | panel_constellation == "left_right" 161 | or panel_constellation == "up_down" 162 | or panel_constellation == "in_out_single" 163 | ): 164 | exist_prob = torch.ones((batch_size, num_panels, 2, 2)) * eps 165 | type_prob = torch.ones((batch_size, num_panels, 2, DIM_ONEHOT)) * eps 166 | size_prob = torch.ones((batch_size, num_panels, 2, DIM_ONEHOT)) * eps 167 | color_prob = torch.ones((batch_size, num_panels, 2, DIM_ONEHOT)) * eps 168 | angle_prob = torch.ones((batch_size, num_panels, 2, DIM_ONEHOT)) * eps 169 | exist_prob[:, :, 0, 1] = 1 170 | exist_prob[:, :, 1, 1] = 1 171 | 172 | for bs in range(batch_size): 173 | for i in range(num_panels): 174 | for j in range(2): 175 | k = int(puzzle[bs, i, j, 0]) 176 | exist_prob[bs, i, k] = smooth_dist(exist_prob[bs, i, k]) 177 | type_prob[bs, i, k, int(puzzle[bs, i, j, 4])] = 1 178 | size_prob[bs, i, k, int(puzzle[bs, i, j, 3])] = 1 179 | color_prob[bs, i, k, int(puzzle[bs, i, j, 2])] = 1 180 | angle_prob[bs, i, k, int(puzzle[bs, i, j, 1])] = 1 181 | return ( 182 | torch.log(exist_prob), 183 | torch.log(type_prob), 184 | torch.log(size_prob), 185 | torch.log(color_prob), 186 | torch.log(angle_prob), 187 | ) 188 | 189 | if panel_constellation == "in_out_four": 190 | exist_prob = torch.ones((batch_size, num_panels, 5, 2)) * eps 191 | type_prob = torch.ones((batch_size, num_panels, 5, DIM_ONEHOT)) * eps 192 | size_prob = torch.ones((batch_size, num_panels, 5, DIM_ONEHOT)) * eps 193 | color_prob = torch.ones((batch_size, num_panels, 5, DIM_ONEHOT)) * eps 194 | angle_prob = torch.ones((batch_size, num_panels, 5, DIM_ONEHOT)) * eps 195 | 196 | for bs in range(batch_size): 197 | for i in range(num_panels): 198 | temp = [0, 1, 2, 3, 4] 199 | for j in range(5): 200 | if puzzle[bs, i, j, 0] == -1: 201 | k = temp[0] 202 | exist_prob[bs, i, k, 0] = 1 203 | exist_prob[bs, i, k] = smooth_dist(exist_prob[bs, i, k]) 204 | temp.remove(k) 205 | else: 206 | k = int(puzzle[bs, i, j, 0]) 207 | exist_prob[bs, i, k, 1] = 1 208 | exist_prob[bs, i, k] = smooth_dist(exist_prob[bs, i, k]) 209 | type_prob[bs, i, k, int(puzzle[bs, i, j, 4])] = 1 210 | type_prob[bs, i, k] = smooth_dist(type_prob[bs, i, k]) 211 | size_prob[bs, i, k, int(puzzle[bs, i, j, 3])] = 1 212 | size_prob[bs, i, k] = smooth_dist(size_prob[bs, i, k]) 213 | color_prob[bs, i, k, int(puzzle[bs, i, j, 2])] = 1 214 | color_prob[bs, i, k] = smooth_dist(color_prob[bs, i, k]) 215 | angle_prob[bs, i, k, int(puzzle[bs, i, j, 1])] = 1 216 | angle_prob[bs, i, k] = smooth_dist(angle_prob[bs, i, k]) 217 | temp.remove(k) 218 | return ( 219 | torch.log(exist_prob), 220 | torch.log(type_prob), 221 | torch.log(size_prob), 222 | torch.log(color_prob), 223 | torch.log(angle_prob), 224 | ) 225 | -------------------------------------------------------------------------------- /arlc/rule_templates.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | import torch as t 7 | import torch.nn as nn 8 | from nvsa.reasoning.vsa_block_utils import ( 9 | block_binding2, 10 | block_unbinding2, 11 | ) 12 | 13 | 14 | def oh(tensor, idx, val=10001.0): 15 | tensor[idx] = val 16 | return tensor 17 | 18 | 19 | class ExtendedGeneralLearnableFormula(nn.Module): 20 | def __init__(self, examples_len, context_len, k, program_rule=None): 21 | super(ExtendedGeneralLearnableFormula, self).__init__() 22 | self.k = k 23 | self.context_len = context_len 24 | self.examples_len = examples_len 25 | self.program_rule = program_rule 26 | if program_rule: 27 | self.program_weights(rule=program_rule) 28 | else: 29 | self.init_terms(12, examples_len + context_len + 1) 30 | self.softmax = nn.Softmax(dim=-1) 31 | self.T = 1 32 | 33 | def init_terms(self, num_terms, num_panels): 34 | terms = list() 35 | for _ in range(num_terms): 36 | terms.append(nn.Parameter(t.randn(num_panels))) 37 | self.terms = nn.ParameterList(terms) 38 | 39 | def program_weights(self, rule, device="cuda"): 40 | print(f"Programming {rule}") 41 | # init every term with the identity 42 | self.terms = [] 43 | for i in range(12): 44 | self.terms.append(oh(t.zeros(8, device=device), -1)) 45 | if rule == "add": 46 | self.terms[0] = oh(t.zeros(8, device=device), 0) # +x1 47 | self.terms[1] = oh(t.zeros(8, device=device), 1) # +x2 48 | self.terms[2] = oh(t.zeros(8, device=device), 4) # +c3 49 | self.terms[7] = oh(t.zeros(8, device=device), 2) # -c1 50 | self.terms[8] = oh(t.zeros(8, device=device), 3) # -c2 51 | elif rule == "sub": 52 | self.terms[0] = oh(t.zeros(8, device=device), 0) # +x1 53 | self.terms[6] = oh(t.zeros(8, device=device), 1) # -x2 54 | self.terms[2] = oh(t.zeros(8, device=device), 2) # +c1 55 | self.terms[7] = oh(t.zeros(8, device=device), 3) # -c2 56 | self.terms[8] = oh(t.zeros(8, device=device), 4) # -c3 57 | elif rule == "dist3": 58 | self.terms[0] = oh(t.zeros(8, device=device), 2) # +c1 59 | self.terms[1] = oh(t.zeros(8, device=device), 3) # +c2 60 | self.terms[2] = oh(t.zeros(8, device=device), 4) # +c3 61 | self.terms[6] = oh(t.zeros(8, device=device), 0) # -x1 62 | self.terms[7] = oh(t.zeros(8, device=device), 1) # -x2 63 | elif rule == "progr": 64 | self.terms[0] = oh(t.zeros(8, device=device), 1) # +x2 65 | self.terms[1] = oh(t.zeros(8, device=device), 1) # +x2 66 | self.terms[6] = oh(t.zeros(8, device=device), 0) # -x1 67 | 68 | def add_identity(self, x): 69 | identity = t.zeros_like(x[:, 0]).unsqueeze(1) 70 | identity[:, :, :, 0] = 1 71 | x_with_identity = t.cat((x, identity), dim=1) 72 | return x_with_identity 73 | 74 | def forward(self, x, ctx): 75 | x = t.cat([x, ctx], dim=1) 76 | x = x.reshape(x.shape[0], x.shape[1], self.k, -1) 77 | x = self.add_identity(x) 78 | x = x.view(x.shape[0], x.shape[1], -1) 79 | 80 | def wcomb(weights, input): 81 | attn_score = self.softmax(weights.unsqueeze(0).unsqueeze(0) / self.T) 82 | term = ( 83 | t.matmul(attn_score.repeat(input.shape[0], 1, 1), input) 84 | .squeeze(1) 85 | .view(input.shape[0], self.k, -1) 86 | ) 87 | return term 88 | 89 | def bind_seq(seq): 90 | seq_len = len(seq) 91 | res = seq[0] 92 | for i in range(1, seq_len): 93 | res = block_binding2(res, seq[i]) 94 | return res 95 | 96 | n = bind_seq([wcomb(t, x) for t in self.terms[: len(self.terms) // 2]]) 97 | d = bind_seq([wcomb(t, x) for t in self.terms[len(self.terms) // 2 :]]) 98 | output = block_unbinding2(n, d) 99 | output = output.view(output.shape[0], -1) 100 | return output 101 | 102 | def __str__(self): 103 | tl = self.context_len + self.examples_len + 1 104 | cfp = nn.functional.one_hot(self.softmax(self.terms[0]).argmax(), tl) 105 | # add + terms 106 | for i in range(1, len(self.terms) // 2): 107 | cfp += nn.functional.one_hot(self.softmax(self.terms[i]).argmax(), tl) 108 | 109 | cfm = -nn.functional.one_hot( 110 | self.softmax(self.terms[len(self.terms) // 2]).argmax(), tl 111 | ) 112 | # add - terms 113 | for i in range(len(self.terms) // 2 + 1, len(self.terms)): 114 | cfm -= nn.functional.one_hot(self.softmax(self.terms[i]).argmax(), tl) 115 | 116 | cf = cfp + cfm 117 | 118 | terms = [f"x{i+1}" for i in range(self.examples_len)] + [ 119 | f"c{i+1}" for i in range(self.context_len) 120 | ] 121 | hr_rule = [ 122 | f"{'+' if cf[i]>0 else ''}{cf[i]}{x} " for i, x in enumerate(terms) if cf[i] 123 | ] 124 | return "".join(hr_rule) 125 | 126 | def anneal_softmax(self): 127 | self.T = 0.01 128 | 129 | 130 | class IravenxGeneralLearnableFormula(nn.Module): 131 | def __init__(self, examples_len, context_len, k, num_terms=12, program_rule=None): 132 | super(IravenxGeneralLearnableFormula, self).__init__() 133 | self.k = k 134 | self.context_len = context_len 135 | self.examples_len = examples_len 136 | self.program_rule = program_rule 137 | self.num_terms = num_terms 138 | if program_rule: 139 | self.program_weights(rule=program_rule, num_terms=num_terms) 140 | else: 141 | self.init_terms(num_terms, examples_len + context_len + 1) 142 | self.softmax = nn.Softmax(dim=-1) 143 | 144 | def init_terms(self, num_terms, num_panels): 145 | terms = list() 146 | for _ in range(num_terms): 147 | terms.append(nn.Parameter(t.randn(num_panels))) 148 | # terms.append(nn.Parameter(oh(t.zeros(num_panels), -1, 1))) 149 | # for i in range(12, 13): 150 | # terms[i].data = oh(t.zeros(num_panels), i-12, 1) # -x1 -x2 ... -x9 151 | self.terms = nn.ParameterList(terms) 152 | 153 | def add_identity(self, x): 154 | identity = t.zeros_like(x[:, 0]).unsqueeze(1) 155 | identity[:, :, :, 0] = 1 156 | x_with_identity = t.cat((x, identity), dim=1) 157 | return x_with_identity 158 | 159 | def forward(self, x, ctx): 160 | x = t.cat([x, ctx], dim=1) 161 | x = x.reshape(x.shape[0], x.shape[1], self.k, -1) 162 | x = self.add_identity(x) 163 | x = x.view(x.shape[0], x.shape[1], -1) 164 | 165 | def wcomb(weights, input): 166 | attn_score = self.softmax(weights.unsqueeze(0).unsqueeze(0)) 167 | term = ( 168 | t.matmul(attn_score.repeat(input.shape[0], 1, 1), input) 169 | .squeeze(1) 170 | .view(input.shape[0], self.k, -1) 171 | ) 172 | return term 173 | 174 | def bind_seq(seq): 175 | seq_len = len(seq) 176 | res = seq[0] 177 | for i in range(1, seq_len): 178 | res = block_binding2(res, seq[i]) 179 | return res 180 | 181 | n = bind_seq([wcomb(t, x) for t in self.terms[: len(self.terms) // 2]]) 182 | d = bind_seq([wcomb(t, x) for t in self.terms[len(self.terms) // 2 :]]) 183 | 184 | output = block_unbinding2(n, d) 185 | output = output.view(output.shape[0], -1) 186 | return output 187 | 188 | def program_weights(self, rule, num_terms, device="cuda"): 189 | print(f"Programming {rule}") 190 | 191 | # init every term with the identity 192 | self.terms = [] 193 | for i in range(num_terms): 194 | self.terms.append(oh(t.zeros(29, device=device), -1)) 195 | 196 | if rule == "constant": 197 | self.terms[0] = oh(self.terms[0], 0) # +x1 198 | self.terms[1] = oh(self.terms[0], 0) # +x1 199 | self.terms[12] = oh(self.terms[0], 0) # -x1 200 | 201 | elif rule == "add": 202 | for i in range(9): 203 | self.terms[i] = oh(t.zeros(29, device=device), i) # +x1 +x2 ... +x9 204 | self.terms[9] = oh(t.zeros(29, device=device), 0) # +x1 205 | self.terms[12] = oh(t.zeros(29, device=device), 0) # -x1 206 | 207 | elif rule == "sub": 208 | for i in range(12, 21): 209 | self.terms[i] = oh( 210 | t.zeros(29, device=device), i - 12 211 | ) # -x1 -x2 ... -x9 212 | self.terms[0] = oh(t.zeros(29, device=device), 0) # +x1 213 | self.terms[1] = oh(t.zeros(29, device=device), 0) # +x1 214 | 215 | elif rule == "dist3": 216 | for i in range(10): 217 | self.terms[i] = oh( 218 | t.zeros(29, device=device), 9 + i 219 | ) # +c1 +c2 ... +c10 220 | for i in range(13, 22): 221 | self.terms[i] = oh( 222 | t.zeros(29, device=device), i - 13 223 | ) # -x1 -x2 ... -x9 224 | 225 | elif rule == "progr": 226 | self.terms[0] = oh(t.zeros(29, device=device), 1) # +x2 227 | self.terms[1] = oh(t.zeros(29, device=device), 8) # +x9 228 | self.terms[12] = oh(t.zeros(29, device=device), 0) # -x1 229 | 230 | def __str__(self): 231 | tl = self.context_len + self.examples_len + 1 232 | cfp = nn.functional.one_hot(self.softmax(self.terms[0]).argmax(), tl) 233 | # add + terms 234 | for i in range(1, len(self.terms) // 2): 235 | cfp += nn.functional.one_hot(self.softmax(self.terms[i]).argmax(), tl) 236 | 237 | cfm = -nn.functional.one_hot( 238 | self.softmax(self.terms[len(self.terms) // 2]).argmax(), tl 239 | ) 240 | # add - terms 241 | for i in range(len(self.terms) // 2 + 1, len(self.terms)): 242 | cfm -= nn.functional.one_hot(self.softmax(self.terms[i]).argmax(), tl) 243 | 244 | cf = cfp + cfm 245 | 246 | terms = ( 247 | [f"x{i+1}" for i in range(self.examples_len)] 248 | + [f"c{i+1}" for i in range(self.context_len)] 249 | + ["e"] 250 | ) 251 | hr_rule = [ 252 | f"{'+' if cf[i]>0 else ''}{cf[i]}{x} " for i, x in enumerate(terms) if cf[i] 253 | ] 254 | return "".join( 255 | hr_rule 256 | ) 257 | 258 | 259 | class IravenVGeneralLearnableFormula(nn.Module): 260 | def __init__(self, examples_len, context_len, k, num_terms=12, program_rule=None): 261 | super(IravenVGeneralLearnableFormula, self).__init__() 262 | self.k = k 263 | self.context_len = context_len 264 | self.examples_len = examples_len 265 | self.program_rule = program_rule 266 | self.num_terms = num_terms 267 | if program_rule: 268 | self.program_weights(rule=program_rule, num_terms=num_terms) 269 | else: 270 | self.init_terms(num_terms, examples_len + context_len + 1) 271 | self.softmax = nn.Softmax(dim=-1) 272 | 273 | def init_terms(self, num_terms, num_panels): 274 | terms = list() 275 | for _ in range(num_terms): 276 | terms.append(nn.Parameter(t.randn(num_panels))) 277 | # terms.append(nn.Parameter(oh(t.zeros(num_panels), -1, 1))) 278 | # for i in range(12, 13): 279 | # terms[i].data = oh(t.zeros(num_panels), i-12, 1) # -x1 -x2 ... -x9 280 | self.terms = nn.ParameterList(terms) 281 | 282 | def add_identity(self, x): 283 | identity = t.zeros_like(x[:, 0]).unsqueeze(1) 284 | identity[:, :, :, 0] = 1 285 | x_with_identity = t.cat((x, identity), dim=1) 286 | return x_with_identity 287 | 288 | def forward(self, x, ctx): 289 | x = t.cat([x, ctx], dim=1) 290 | x = x.reshape(x.shape[0], x.shape[1], self.k, -1) 291 | x = self.add_identity(x) 292 | x = x.view(x.shape[0], x.shape[1], -1) 293 | 294 | def wcomb(weights, input): 295 | attn_score = self.softmax(weights.unsqueeze(0).unsqueeze(0)) 296 | term = ( 297 | t.matmul(attn_score.repeat(input.shape[0], 1, 1), input) 298 | .squeeze(1) 299 | .view(input.shape[0], self.k, -1) 300 | ) 301 | return term 302 | 303 | def bind_seq(seq): 304 | seq_len = len(seq) 305 | res = seq[0] 306 | for i in range(1, seq_len): 307 | res = block_binding2(res, seq[i]) 308 | return res 309 | 310 | n = bind_seq([wcomb(t, x) for t in self.terms[: len(self.terms) // 2]]) 311 | d = bind_seq([wcomb(t, x) for t in self.terms[len(self.terms) // 2 :]]) 312 | 313 | output = block_unbinding2(n, d) 314 | output = output.view(output.shape[0], -1) 315 | return output 316 | 317 | def program_weights(self, rule, num_terms, device="cuda"): 318 | print(f"Programming {rule}") 319 | 320 | # init every term with the identity 321 | self.terms = [] 322 | for i in range(num_terms): 323 | self.terms.append(oh(t.zeros(14, device=device), -1)) 324 | 325 | if rule == "constant": 326 | self.terms[0] = oh(self.terms[0], 0) # +x1 327 | self.terms[1] = oh(self.terms[0], 0) # +x1 328 | self.terms[11] = oh(self.terms[0], 0) # -x1 329 | 330 | elif rule == "add": 331 | for i in range(4): 332 | self.terms[i] = oh(t.zeros(14, device=device), i) # +x1 +x2 ... +x9 333 | self.terms[4] = oh(t.zeros(14, device=device), 0) # +x1 334 | self.terms[11] = oh(t.zeros(14, device=device), 0) # -x1 335 | 336 | elif rule == "sub": 337 | for i in range(10, 14): 338 | self.terms[i] = oh( 339 | t.zeros(14, device=device), i - 10 340 | ) # -x1 -x2 ... -x9 341 | self.terms[0] = oh(t.zeros(14, device=device), 0) # +x1 342 | self.terms[1] = oh(t.zeros(14, device=device), 0) # +x1 343 | 344 | elif rule == "dist3": 345 | for i in range(5): 346 | self.terms[i] = oh( 347 | t.zeros(14, device=device), 4 + i 348 | ) # +c1 +c2 ... +c10 349 | for i in range(10, 14): 350 | self.terms[i] = oh( 351 | t.zeros(14, device=device), i - 10 352 | ) # -x1 -x2 ... -x9 353 | 354 | elif rule == "progr": 355 | self.terms[0] = oh(t.zeros(14, device=device), 1) # +x2 356 | self.terms[1] = oh(t.zeros(14, device=device), 3) # +x3 357 | self.terms[10] = oh(t.zeros(14, device=device), 0) # -x1 358 | 359 | def __str__(self): 360 | tl = self.context_len + self.examples_len + 1 361 | cfp = nn.functional.one_hot(self.softmax(self.terms[0]).argmax(), tl) 362 | # add + terms 363 | for i in range(1, len(self.terms) // 2): 364 | cfp += nn.functional.one_hot(self.softmax(self.terms[i]).argmax(), tl) 365 | cfm = -nn.functional.one_hot( 366 | self.softmax(self.terms[len(self.terms) // 2]).argmax(), tl 367 | ) 368 | # add - terms 369 | for i in range(len(self.terms) // 2 + 1, len(self.terms)): 370 | cfm -= nn.functional.one_hot(self.softmax(self.terms[i]).argmax(), tl) 371 | cf = cfp + cfm 372 | terms = [f"x{i+1}" for i in range(self.examples_len)] + [ 373 | f"c{i+1}" for i in range(self.context_len) 374 | ] 375 | hr_rule = [ 376 | f"{'+' if cf[i]>0 else ''}{cf[i]}{x} " for i, x in enumerate(terms) if cf[i] 377 | ] 378 | return "".join( 379 | hr_rule 380 | ) 381 | -------------------------------------------------------------------------------- /arlc/utils/raven/extraction.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | import numpy as np 7 | import os 8 | import xml.etree.ElementTree as ET 9 | import argparse 10 | import tqdm 11 | 12 | parser = argparse.ArgumentParser(description="NVSA") 13 | parser.add_argument("--data_path", type=str, default="/dccstor/saentis/data/I-RAVEN") 14 | 15 | pos_num_rule_idx_map_four = { 16 | "Constant": 0, 17 | "Progression_One_Pos": 1, 18 | "Progression_Mone_Pos": 2, 19 | "Arithmetic_Plus_Pos": 3, 20 | "Arithmetic_Minus_Pos": 4, 21 | "Distribute_Three_Left_Pos": 5, 22 | "Distribute_Three_Right_Pos": 6, 23 | "Progression_One_Num": 7, 24 | "Progression_Mone_Num": 8, 25 | "Arithmetic_Plus_Num": 9, 26 | "Arithmetic_Minus_Num": 10, 27 | "Distribute_Three_Left_Num": 11, 28 | "Distribute_Three_Right_Num": 12, 29 | } 30 | 31 | pos_num_rule_idx_map_nine = { 32 | "Constant": 0, 33 | "Progression_One_Pos": 1, 34 | "Progression_Mone_Pos": 2, 35 | "Progression_Two_Pos": 3, 36 | "Progression_Mtwo_Pos": 4, 37 | "Arithmetic_Plus_Pos": 5, 38 | "Arithmetic_Minus_Pos": 6, 39 | "Distribute_Three_Left_Pos": 7, 40 | "Distribute_Three_Right_Pos": 8, 41 | "Progression_One_Num": 9, 42 | "Progression_Mone_Num": 10, 43 | "Progression_Two_Num": 11, 44 | "Progression_Mtwo_Num": 12, 45 | "Arithmetic_Plus_Num": 13, 46 | "Arithmetic_Minus_Num": 14, 47 | "Distribute_Three_Left_Num": 15, 48 | "Distribute_Three_Right_Num": 16, 49 | } 50 | 51 | type_rule_idx_map = { 52 | "Constant": 0, 53 | "Progression_One": 1, 54 | "Progression_Mone": 2, 55 | "Progression_Two": 3, 56 | "Progression_Mtwo": 4, 57 | "Distribute_Three_Left": 5, 58 | "Distribute_Three_Right": 6, 59 | } 60 | 61 | size_rule_idx_map = { 62 | "Constant": 0, 63 | "Progression_One": 1, 64 | "Progression_Mone": 2, 65 | "Progression_Two": 3, 66 | "Progression_Mtwo": 4, 67 | "Arithmetic_Plus": 5, 68 | "Arithmetic_Minus": 6, 69 | "Distribute_Three_Left": 7, 70 | "Distribute_Three_Right": 8, 71 | } 72 | 73 | color_rule_idx_map = { 74 | "Constant": 0, 75 | "Progression_One": 1, 76 | "Progression_Mone": 2, 77 | "Progression_Two": 3, 78 | "Progression_Mtwo": 4, 79 | "Arithmetic_Plus": 5, 80 | "Arithmetic_Minus": 6, 81 | "Distribute_Three_Left": 7, 82 | "Distribute_Three_Right": 8, 83 | } 84 | 85 | type_idx_rule_map = { 86 | 0: "Constant", 87 | 1: "Progression_One", 88 | 2: "Progression_Mone", 89 | 3: "Progression_Two", 90 | 4: "Progression_Mtwo", 91 | 5: "Distribute_Three_Left", 92 | 6: "Distribute_Three_Right", 93 | } 94 | size_idx_rule_map = { 95 | 0: "Constant", 96 | 1: "Progression_One", 97 | 2: "Progression_Mone", 98 | 3: "Progression_Two", 99 | 4: "Progression_Mtwo", 100 | 5: "Arithmetic_Plus", 101 | 6: "Arithmetic_Minus", 102 | 7: "Distribute_Three_Left", 103 | 8: "Distribute_Three_Right", 104 | } 105 | color_idx_rule_map = { 106 | 0: "Constant", 107 | 1: "Progression_One", 108 | 2: "Progression_Mone", 109 | 3: "Progression_Two", 110 | 4: "Progression_Mtwo", 111 | 5: "Arithmetic_Plus", 112 | 6: "Arithmetic_Minus", 113 | 7: "Distribute_Three_Left", 114 | 8: "Distribute_Three_Right", 115 | } 116 | 117 | pos_num_rule_four_idx_map = { 118 | 0: "Constant", 119 | 1: "Progression_One_Pos", 120 | 2: "Progression_Mone_Pos", 121 | 3: "Arithmetic_Plus_Pos", 122 | 4: "Arithmetic_Minus_Pos", 123 | 5: "Distribute_Three_Left_Pos", 124 | 6: "Distribute_Three_Right_Pos", 125 | 7: "Progression_One_Num", 126 | 8: "Progression_Mone_Num", 127 | 9: "Arithmetic_Plus_Num", 128 | 10: "Arithmetic_Minus_Num", 129 | 11: "Distribute_Three_Left_Num", 130 | 12: "Distribute_Three_Right_Num", 131 | } 132 | 133 | pos_num_rule_nine_idx_map = { 134 | 0: "Constant", 135 | 1: "Progression_One_Pos", 136 | 2: "Progression_Mone_Pos", 137 | 3: "Progression_Two_Pos", 138 | 4: "Progression_Mtwo_Pos", 139 | 5: "Arithmetic_Plus_Pos", 140 | 6: "Arithmetic_Minus_Pos", 141 | 7: "Distribute_Three_Left_Pos", 142 | 8: "Distribute_Three_Right_Pos", 143 | 9: "Progression_One_Num", 144 | 10: "Progression_Mone_Num", 145 | 11: "Progression_Two_Num", 146 | 12: "Progression_Mtwo_Num", 147 | 13: "Arithmetic_Plus_Num", 148 | 14: "Arithmetic_Minus_Num", 149 | 15: "Distribute_Three_Left_Num", 150 | 16: "Distribute_Three_Right_Num", 151 | } 152 | 153 | 154 | def get_pos_num_rule( 155 | rule_idx, comp_idx, num_elements, pos_num_rule_idx_map, xml_panels, xml_rules 156 | ): 157 | index_name = xml_rules[rule_idx][0].attrib["name"] 158 | attrib_name = xml_rules[rule_idx][0].attrib["attr"][:3] 159 | if index_name == "Progression": 160 | if attrib_name == "Num": 161 | first = int(xml_panels[0][0][comp_idx][0].attrib["Number"]) 162 | second = int(xml_panels[1][0][comp_idx][0].attrib["Number"]) 163 | if second == first + 1: 164 | index_name += "_One_Num" 165 | if second == first - 1: 166 | index_name += "_Mone_Num" 167 | if second == first + 2: 168 | index_name += "_Two_Num" 169 | if second == first - 2: 170 | index_name += "_Mtwo_Num" 171 | if attrib_name == "Pos": 172 | all_position = eval(xml_panels[0][0][comp_idx][0].attrib["Position"]) 173 | first = [] 174 | for entity in xml_panels[0][0][comp_idx][0]: 175 | first.append(all_position.index(eval(entity.attrib["bbox"]))) 176 | second = [] 177 | for entity in xml_panels[1][0][comp_idx][0]: 178 | second.append(all_position.index(eval(entity.attrib["bbox"]))) 179 | third = [] 180 | for entity in xml_panels[2][0][comp_idx][0]: 181 | third.append(all_position.index(eval(entity.attrib["bbox"]))) 182 | fourth = [] 183 | for entity in xml_panels[3][0][comp_idx][0]: 184 | fourth.append(all_position.index(eval(entity.attrib["bbox"]))) 185 | fifth = [] 186 | for entity in xml_panels[4][0][comp_idx][0]: 187 | fifth.append(all_position.index(eval(entity.attrib["bbox"]))) 188 | sixth = [] 189 | for entity in xml_panels[5][0][comp_idx][0]: 190 | sixth.append(all_position.index(eval(entity.attrib["bbox"]))) 191 | seventh = [] 192 | for entity in xml_panels[6][0][comp_idx][0]: 193 | seventh.append(all_position.index(eval(entity.attrib["bbox"]))) 194 | eighth = [] 195 | for entity in xml_panels[7][0][comp_idx][0]: 196 | eighth.append(all_position.index(eval(entity.attrib["bbox"]))) 197 | if ( 198 | len( 199 | set(map(lambda index: (index + 1) % num_elements, first)) 200 | - set(second) 201 | ) 202 | == 0 203 | and len( 204 | set(map(lambda index: (index + 1) % num_elements, second)) 205 | - set(third) 206 | ) 207 | == 0 208 | and len( 209 | set(map(lambda index: (index + 1) % num_elements, fourth)) 210 | - set(fifth) 211 | ) 212 | == 0 213 | and len( 214 | set(map(lambda index: (index + 1) % num_elements, fifth)) 215 | - set(sixth) 216 | ) 217 | == 0 218 | and len( 219 | set(map(lambda index: (index + 1) % num_elements, seventh)) 220 | - set(eighth) 221 | ) 222 | == 0 223 | ): 224 | index_name += "_One_Pos" 225 | if ( 226 | len( 227 | set(map(lambda index: (index - 1) % num_elements, first)) 228 | - set(second) 229 | ) 230 | == 0 231 | and len( 232 | set(map(lambda index: (index - 1) % num_elements, second)) 233 | - set(third) 234 | ) 235 | == 0 236 | and len( 237 | set(map(lambda index: (index - 1) % num_elements, fourth)) 238 | - set(fifth) 239 | ) 240 | == 0 241 | and len( 242 | set(map(lambda index: (index - 1) % num_elements, fifth)) 243 | - set(sixth) 244 | ) 245 | == 0 246 | and len( 247 | set(map(lambda index: (index - 1) % num_elements, seventh)) 248 | - set(eighth) 249 | ) 250 | == 0 251 | ): 252 | index_name += "_Mone_Pos" 253 | if ( 254 | len( 255 | set(map(lambda index: (index + 2) % num_elements, first)) 256 | - set(second) 257 | ) 258 | == 0 259 | and len( 260 | set(map(lambda index: (index + 2) % num_elements, second)) 261 | - set(third) 262 | ) 263 | == 0 264 | and len( 265 | set(map(lambda index: (index + 2) % num_elements, fourth)) 266 | - set(fifth) 267 | ) 268 | == 0 269 | and len( 270 | set(map(lambda index: (index + 2) % num_elements, fifth)) 271 | - set(sixth) 272 | ) 273 | == 0 274 | and len( 275 | set(map(lambda index: (index + 2) % num_elements, seventh)) 276 | - set(eighth) 277 | ) 278 | == 0 279 | ): 280 | index_name += "_Two_Pos" 281 | if ( 282 | len( 283 | set(map(lambda index: (index - 2) % num_elements, first)) 284 | - set(second) 285 | ) 286 | == 0 287 | and len( 288 | set(map(lambda index: (index - 2) % num_elements, second)) 289 | - set(third) 290 | ) 291 | == 0 292 | and len( 293 | set(map(lambda index: (index - 2) % num_elements, fourth)) 294 | - set(fifth) 295 | ) 296 | == 0 297 | and len( 298 | set(map(lambda index: (index - 2) % num_elements, fifth)) 299 | - set(sixth) 300 | ) 301 | == 0 302 | and len( 303 | set(map(lambda index: (index - 2) % num_elements, seventh)) 304 | - set(eighth) 305 | ) 306 | == 0 307 | ): 308 | index_name += "_Mtwo_Pos" 309 | if index_name.endswith("_One_Pos_Mone_Pos"): 310 | if np.random.uniform() >= 0.5: 311 | index_name = "Progression_One_Pos" 312 | else: 313 | index_name = "Progression_Mone_Pos" 314 | if index_name == "Arithmetic": 315 | if attrib_name == "Num": 316 | first = int(xml_panels[0][0][comp_idx][0].attrib["Number"]) 317 | second = int(xml_panels[1][0][comp_idx][0].attrib["Number"]) 318 | third = int(xml_panels[2][0][comp_idx][0].attrib["Number"]) 319 | if third == first + second + 1: 320 | index_name += "_Plus_Num" 321 | if third == first - second - 1: 322 | index_name += "_Minus_Num" 323 | if attrib_name == "Pos": 324 | all_position = eval(xml_panels[0][0][comp_idx][0].attrib["Position"]) 325 | first = [] 326 | for entity in xml_panels[0][0][comp_idx][0]: 327 | first.append(all_position.index(eval(entity.attrib["bbox"]))) 328 | second = [] 329 | for entity in xml_panels[1][0][comp_idx][0]: 330 | second.append(all_position.index(eval(entity.attrib["bbox"]))) 331 | third = [] 332 | for entity in xml_panels[2][0][comp_idx][0]: 333 | third.append(all_position.index(eval(entity.attrib["bbox"]))) 334 | if set(third) == set(first).union(set(second)): 335 | index_name += "_Plus_Pos" 336 | if set(third) == set(first) - set(second): 337 | index_name += "_Minus_Pos" 338 | if index_name == "Distribute_Three": 339 | if attrib_name == "Num": 340 | first = int(xml_panels[0][0][comp_idx][0].attrib["Number"]) 341 | second_left = int(xml_panels[5][0][comp_idx][0].attrib["Number"]) 342 | second_right = int(xml_panels[4][0][comp_idx][0].attrib["Number"]) 343 | if second_left == first: 344 | index_name += "_Left_Num" 345 | if second_right == first: 346 | index_name += "_Right_Num" 347 | if attrib_name == "Pos": 348 | all_position = eval(xml_panels[0][0][comp_idx][0].attrib["Position"]) 349 | first = [] 350 | for entity in xml_panels[0][0][comp_idx][0]: 351 | first.append(all_position.index(eval(entity.attrib["bbox"]))) 352 | second_left = [] 353 | for entity in xml_panels[5][0][comp_idx][0]: 354 | second_left.append(all_position.index(eval(entity.attrib["bbox"]))) 355 | second_right = [] 356 | for entity in xml_panels[4][0][comp_idx][0]: 357 | second_right.append(all_position.index(eval(entity.attrib["bbox"]))) 358 | if set(second_left) == set(first): 359 | index_name += "_Left_Pos" 360 | if set(second_right) == set(first): 361 | index_name += "_Right_Pos" 362 | return pos_num_rule_idx_map[index_name] 363 | 364 | 365 | def get_type_rule( 366 | rule_idx, comp_idx, num_elements, pos_num_rule_idx_map, xml_panels, xml_rules 367 | ): 368 | index_name = xml_rules[rule_idx][1].attrib["name"] 369 | if index_name == "Progression": 370 | first = int(xml_panels[0][0][comp_idx][0][0].attrib["Type"]) 371 | second = int(xml_panels[1][0][comp_idx][0][0].attrib["Type"]) 372 | if second == first + 1: 373 | index_name += "_One" 374 | if second == first - 1: 375 | index_name += "_Mone" 376 | if second == first + 2: 377 | index_name += "_Two" 378 | if second == first - 2: 379 | index_name += "_Mtwo" 380 | if index_name == "Distribute_Three": 381 | first = int(xml_panels[0][0][comp_idx][0][0].attrib["Type"]) 382 | second_left = int(xml_panels[5][0][comp_idx][0][0].attrib["Type"]) 383 | second_right = int(xml_panels[4][0][comp_idx][0][0].attrib["Type"]) 384 | if second_left == first: 385 | index_name += "_Left" 386 | if second_right == first: 387 | index_name += "_Right" 388 | return type_rule_idx_map[index_name] 389 | 390 | 391 | def get_size_rule( 392 | rule_idx, comp_idx, num_elements, pos_num_rule_idx_map, xml_panels, xml_rules 393 | ): 394 | index_name = xml_rules[rule_idx][2].attrib["name"] 395 | if index_name == "Progression": 396 | first = int(xml_panels[0][0][comp_idx][0][0].attrib["Size"]) 397 | second = int(xml_panels[1][0][comp_idx][0][0].attrib["Size"]) 398 | if second == first + 1: 399 | index_name += "_One" 400 | if second == first - 1: 401 | index_name += "_Mone" 402 | if second == first + 2: 403 | index_name += "_Two" 404 | if second == first - 2: 405 | index_name += "_Mtwo" 406 | if index_name == "Arithmetic": 407 | first = int(xml_panels[0][0][comp_idx][0][0].attrib["Size"]) 408 | second = int(xml_panels[1][0][comp_idx][0][0].attrib["Size"]) 409 | third = int(xml_panels[2][0][comp_idx][0][0].attrib["Size"]) 410 | if third == first + second + 1: 411 | index_name += "_Plus" 412 | if third == first - second - 1: 413 | index_name += "_Minus" 414 | if index_name == "Distribute_Three": 415 | first = int(xml_panels[0][0][comp_idx][0][0].attrib["Size"]) 416 | second_left = int(xml_panels[5][0][comp_idx][0][0].attrib["Size"]) 417 | second_right = int(xml_panels[4][0][comp_idx][0][0].attrib["Size"]) 418 | if second_left == first: 419 | index_name += "_Left" 420 | if second_right == first: 421 | index_name += "_Right" 422 | return size_rule_idx_map[index_name] 423 | 424 | 425 | def get_color_rule( 426 | rule_idx, comp_idx, num_elements, pos_num_rule_idx_map, xml_panels, xml_rules 427 | ): 428 | index_name = xml_rules[rule_idx][3].attrib["name"] 429 | if index_name == "Progression": 430 | first = int(xml_panels[0][0][comp_idx][0][0].attrib["Color"]) 431 | second = int(xml_panels[1][0][comp_idx][0][0].attrib["Color"]) 432 | if second == first + 1: 433 | index_name += "_One" 434 | if second == first - 1: 435 | index_name += "_Mone" 436 | if second == first + 2: 437 | index_name += "_Two" 438 | if second == first - 2: 439 | index_name += "_Mtwo" 440 | if index_name == "Arithmetic": 441 | first = int(xml_panels[0][0][comp_idx][0][0].attrib["Color"]) 442 | second = int(xml_panels[1][0][comp_idx][0][0].attrib["Color"]) 443 | third = int(xml_panels[2][0][comp_idx][0][0].attrib["Color"]) 444 | fourth = int(xml_panels[3][0][comp_idx][0][0].attrib["Color"]) 445 | fifth = int(xml_panels[4][0][comp_idx][0][0].attrib["Color"]) 446 | sixth = int(xml_panels[5][0][comp_idx][0][0].attrib["Color"]) 447 | if (third == first + second) and (sixth == fourth + fifth): 448 | index_name += "_Plus" 449 | if (third == first - second) and (sixth == fourth - fifth): 450 | index_name += "_Minus" 451 | if index_name == "Distribute_Three": 452 | first = int(xml_panels[0][0][comp_idx][0][0].attrib["Color"]) 453 | second_left = int(xml_panels[5][0][comp_idx][0][0].attrib["Color"]) 454 | second_right = int(xml_panels[4][0][comp_idx][0][0].attrib["Color"]) 455 | if second_left == first: 456 | index_name += "_Left" 457 | if second_right == first: 458 | index_name += "_Right" 459 | return color_rule_idx_map[index_name] 460 | 461 | 462 | def main(): 463 | args = parser.parse_args() 464 | DATA_PATH = args.data_path 465 | constellation_name_list = [ 466 | "center_single", 467 | "distribute_four", 468 | "distribute_nine", 469 | "left_center_single_right_center_single", 470 | "up_center_single_down_center_single", 471 | "in_center_single_out_center_single", 472 | "in_distribute_four_out_center_single", 473 | ] 474 | save_name_list = [ 475 | "center_single_extracted_with_rules", 476 | "distribute_four_extracted_with_rules", 477 | "distribute_nine_extracted_with_rules", 478 | "left_right_extracted_with_rules", 479 | "up_down_extracted_with_rules", 480 | "in_out_single_extracted_with_rules", 481 | "in_out_four_extracted_with_rules", 482 | ] 483 | 484 | obj_name_list = ["train", "val", "test"] 485 | my_bbox = { 486 | "in_out_four_extracted_with_rules": { 487 | "[0.5, 0.5, 1, 1]": 0, 488 | "[0.42, 0.42, 0.15, 0.15]": 1, 489 | "[0.42, 0.58, 0.15, 0.15]": 2, 490 | "[0.58, 0.42, 0.15, 0.15]": 3, 491 | "[0.58, 0.58, 0.15, 0.15]": 4, 492 | }, 493 | "in_out_single_extracted_with_rules": { 494 | "[0.5, 0.5, 1, 1]": 0, 495 | "[0.5, 0.5, 0.33, 0.33]": 1, 496 | }, 497 | "up_down_extracted_with_rules": { 498 | "[0.25, 0.5, 0.5, 0.5]": 0, 499 | "[0.75, 0.5, 0.5, 0.5]": 1, 500 | }, 501 | "left_right_extracted_with_rules": { 502 | "[0.5, 0.25, 0.5, 0.5]": 0, 503 | "[0.5, 0.75, 0.5, 0.5]": 1, 504 | }, 505 | "distribute_nine_extracted_with_rules": { 506 | "[0.16, 0.16, 0.33, 0.33]": 0, 507 | "[0.16, 0.5, 0.33, 0.33]": 1, 508 | "[0.16, 0.83, 0.33, 0.33]": 2, 509 | "[0.5, 0.16, 0.33, 0.33]": 3, 510 | "[0.5, 0.5, 0.33, 0.33]": 4, 511 | "[0.5, 0.83, 0.33, 0.33]": 5, 512 | "[0.83, 0.16, 0.33, 0.33]": 6, 513 | "[0.83, 0.5, 0.33, 0.33]": 7, 514 | "[0.83, 0.83, 0.33, 0.33]": 8, 515 | }, 516 | "distribute_four_extracted_with_rules": { 517 | "[0.25, 0.25, 0.5, 0.5]": 0, 518 | "[0.25, 0.75, 0.5, 0.5]": 1, 519 | "[0.75, 0.25, 0.5, 0.5]": 2, 520 | "[0.75, 0.75, 0.5, 0.5]": 3, 521 | }, 522 | } 523 | for w in range(len(constellation_name_list)): 524 | file_type = constellation_name_list[w] 525 | save_name = save_name_list[w] 526 | path = os.path.join(DATA_PATH, save_name) 527 | path_train, path_val, path_test = ( 528 | os.path.join(path, "train"), 529 | os.path.join(path, "val"), 530 | os.path.join(path, "test"), 531 | ) 532 | os.makedirs(path, exist_ok=True) 533 | os.makedirs(path_train, exist_ok=True) 534 | os.makedirs(path_val, exist_ok=True) 535 | os.makedirs(path_test, exist_ok=True) 536 | for n in range(len(obj_name_list)): 537 | count = 0 538 | obj_name = obj_name_list[n] 539 | for j in tqdm.tqdm(range(10001)): 540 | try: 541 | tree = ET.parse( 542 | "{0}/{1}/RAVEN_{2}_{3}.xml".format( 543 | DATA_PATH, file_type, j, obj_name 544 | ) 545 | ) 546 | except: 547 | continue 548 | root = tree.getroot() 549 | xml_panels = root[0] 550 | xml_rules = root[1] 551 | rule_idx = 0 552 | comp_idx = 0 553 | num_elements = 9 554 | pos_num_rule_idx_map = pos_num_rule_idx_map_four 555 | if file_type == "distribute_four": 556 | num_elements = 4 557 | pos_num_rule_idx_map = pos_num_rule_idx_map_four 558 | elif file_type == "distribute_nine": 559 | num_elements = 9 560 | pos_num_rule_idx_map = pos_num_rule_idx_map_nine 561 | args = [ 562 | rule_idx, 563 | comp_idx, 564 | num_elements, 565 | pos_num_rule_idx_map, 566 | xml_panels, 567 | xml_rules, 568 | ] 569 | pos_num_rule = np.array(get_pos_num_rule(*args)) 570 | color_rule = np.array(get_color_rule(*args)) 571 | size_rule = np.array(get_size_rule(*args)) 572 | type_rule = np.array(get_type_rule(*args)) 573 | rules = np.array([pos_num_rule, color_rule, size_rule, type_rule]) 574 | idx_panel = 0 575 | for panel in root[0]: 576 | idx = 0 577 | for component in panel[0]: 578 | for entity in component[0]: 579 | a = entity.attrib 580 | angle, color, size, typ, bbox = ( 581 | int(a.get("Angle")), 582 | int(a.get("Color")), 583 | int(a.get("Size")), 584 | int(a.get("Type")) - 1, 585 | a.get("bbox"), 586 | ) 587 | pos = ( 588 | my_bbox[save_name][bbox] 589 | if save_name != "center_single_extracted_with_rules" 590 | else 0 591 | ) 592 | ext_comp = [pos, angle, color, size, typ] 593 | ext_comp = np.expand_dims(ext_comp, axis=0) 594 | ext_panel = ( 595 | ext_comp 596 | if (idx == 0) 597 | else np.concatenate((ext_panel, ext_comp), axis=0) 598 | ) 599 | idx = idx + 1 600 | c = 9 - idx 601 | if c > 0: 602 | filler = np.ones((c, 5)) * (-1) 603 | ext_panel = np.concatenate((ext_panel, filler), axis=0) 604 | ext_panel = np.expand_dims(ext_panel, axis=0) 605 | ext_sample = ( 606 | ext_panel 607 | if (idx_panel == 0) 608 | else np.concatenate((ext_sample, ext_panel), axis=0) 609 | ) 610 | idx_panel = idx_panel + 1 611 | file = np.load( 612 | "{0}/{1}/RAVEN_{2}_{3}.npz".format( 613 | DATA_PATH, file_type, j, obj_name 614 | ) 615 | ) 616 | filename = "{0}/{1}/{2}/RAVEN_{3}_{2}.npz".format( 617 | DATA_PATH, save_name, obj_name, count 618 | ) 619 | np.savez( 620 | filename, 621 | target=file["target"], 622 | predict=file["predict"], 623 | image=file["image"], 624 | meta_matrix=file["meta_matrix"], 625 | meta_structure=file["meta_structure"], 626 | meta_target=file["meta_target"], 627 | structure=file["structure"], 628 | extracted_meta=ext_sample, 629 | rules=rules, 630 | ) 631 | count = count + 1 632 | print("finished with: ", save_name) 633 | 634 | 635 | if __name__ == "__main__": 636 | main() 637 | -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | # *----------------------------------------------------------------------------* 2 | # * Copyright (C) 2024 IBM Inc. All rights reserved * 3 | # * SPDX-License-Identifier: GPL-3.0-only * 4 | # *----------------------------------------------------------------------------* 5 | 6 | import json 7 | import os 8 | import random 9 | from tqdm import tqdm 10 | import numpy as np 11 | import torch 12 | import torch.optim as optim 13 | from torch.utils.data import DataLoader 14 | from torch.utils.tensorboard import SummaryWriter 15 | from collections import defaultdict as dd 16 | 17 | import arlc.utils.raven.env as reasoning_env 18 | from arlc.utils.averagemeter import AverageMeter 19 | from arlc.utils.checkpath import check_paths, save_checkpoint 20 | from arlc.datasets import GeneralIRAVENDataset 21 | from arlc.execution import RuleLevelReasoner 22 | from arlc.selection import RuleSelector 23 | from arlc.utils.vsa import generate_nvsa_codebooks 24 | import arlc.losses as losses 25 | from arlc.utils.raven.raven_one_hot import create_one_hot 26 | from arlc.utils.parsing import parse_args 27 | from arlc.utils.general import iravenx_rule_map, iravenx_index_map 28 | 29 | 30 | def compute_loss_and_scores( 31 | outputs, 32 | tests, 33 | candidates, 34 | targets, 35 | loss_fn, 36 | distribute, 37 | params=None, 38 | confounders=False, 39 | attr_entropy=None, 40 | use_entropy=False, 41 | ): 42 | loss = 0 43 | scores = 0 44 | att_scores = dd(lambda: 0) 45 | for attr in outputs._fields: 46 | # in constellations without position, do not compute a loss on it 47 | if not distribute and (attr == "position" or attr == "number"): 48 | continue 49 | # if confounders is turned off, disregard confounding attributes 50 | if (attr == "angle" or "confounder" in attr) and not confounders: 51 | pass 52 | # add to the loss the contribution of attr 53 | else: 54 | # attribute entropy regularization 55 | entropy_reg = np.clip(attr_entropy[attr], 0.1, 10) if use_entropy else 1 56 | loss += ( 57 | loss_fn( 58 | getattr(outputs, attr), 59 | torch.cat( 60 | ( 61 | getattr(tests, attr), 62 | getattr(candidates, attr)[ 63 | torch.arange(getattr(candidates, attr).shape[0]), 64 | targets, 65 | ].unsqueeze(1), 66 | ), 67 | dim=1, 68 | ), 69 | ).mean(dim=-1) 70 | / entropy_reg 71 | ) 72 | # compute attribute scores 73 | att_scores[attr] = loss_fn.score( 74 | getattr(outputs, attr)[:, -1].unsqueeze(1).repeat(1, 8, 1), 75 | getattr(candidates, attr), 76 | ) 77 | scores += att_scores[attr] / entropy_reg 78 | 79 | return loss, scores, [att_scores[att] for att in ["type", "color", "size"]] 80 | 81 | 82 | def train(args, env, device, confounders=False): 83 | """ 84 | Training and validation of learnable NVSA backend 85 | """ 86 | 87 | def inference_epoch(epoch, loader, train=True): 88 | if train: 89 | model.train() 90 | if args.config == "in_out_four": 91 | model2.train() 92 | rule_selector.train() 93 | else: 94 | model.eval() 95 | if args.config == "in_out_four": 96 | model2.eval() 97 | rule_selector.eval() 98 | 99 | # Define tracking meters 100 | loss_avg = AverageMeter("Loss", ":.3f") 101 | acc_avg = AverageMeter("Accuracy", ":.3f") 102 | 103 | for counter, (extracted, targets, all_action_rule) in enumerate(tqdm(loader)): 104 | extracted, targets, all_action_rule = ( 105 | extracted.to(device), 106 | targets.to(device), 107 | all_action_rule.to(device), 108 | ) 109 | att_logprob = create_one_hot(extracted, args.config, args.sigma) 110 | model_output = {k: v.to(device) for k, v in att_logprob.items()} 111 | scene_prob, _ = env.prepare(model_output) 112 | if args.config in ["center_single", "distribute_four", "distribute_nine"]: 113 | outputs, candidates, tests = model( 114 | scene_prob, targets, distribute=distribute 115 | ) 116 | outputs, attr_entropy = rule_selector( 117 | outputs, tests, candidates, targets 118 | ) 119 | loss, scores, _ = compute_loss_and_scores( 120 | outputs, 121 | tests, 122 | candidates, 123 | targets, 124 | loss_fn, 125 | "distribute" in args.config, 126 | [p.data for p in model.parameters() if p.requires_grad], 127 | confounders=confounders, 128 | attr_entropy=attr_entropy, 129 | use_entropy=args.entropy, 130 | ) 131 | else: 132 | outputs1, candidates1, tests1 = model( 133 | scene_prob[0], distribute=distribute 134 | ) 135 | outputs1, attr_entropy1 = rule_selector( 136 | outputs1, tests1, candidates1, targets 137 | ) 138 | if args.config == "in_out_four": 139 | outputs2, candidates2, tests2 = model2(scene_prob[1]) 140 | else: 141 | outputs2, candidates2, tests2 = model(scene_prob[1]) 142 | 143 | outputs2, attr_entropy2 = rule_selector( 144 | outputs2, tests2, candidates2, targets 145 | ) 146 | loss1, scores1, _ = compute_loss_and_scores( 147 | outputs1, 148 | tests1, 149 | candidates1, 150 | targets, 151 | loss_fn, 152 | distribute=args.config == "in_out_four", 153 | confounders=confounders, 154 | attr_entropy=attr_entropy1, 155 | use_entropy=args.entropy, 156 | ) 157 | loss2, scores2, _ = compute_loss_and_scores( 158 | outputs2, 159 | tests2, 160 | candidates2, 161 | targets, 162 | loss_fn, 163 | distribute=False, 164 | confounders=confounders, 165 | attr_entropy=attr_entropy2, 166 | use_entropy=args.entropy, 167 | ) 168 | loss = 0.8 * loss1 + 0.2 * loss2 169 | scores = 0.8 * scores1 + 0.2 * scores2 170 | 171 | predictions = torch.argmax(scores, dim=-1) 172 | accuracy = ((predictions == targets).sum() / len(targets)) * 100 173 | loss_avg.update(loss.item(), extracted.size(0)) 174 | acc_avg.update(accuracy.item(), extracted.size(0)) 175 | acc_avg.update(accuracy.item(), extracted.size(0)) 176 | 177 | if train: 178 | optimizer.zero_grad() 179 | loss.backward() 180 | if args.clip: 181 | torch.nn.utils.clip_grad_norm_( 182 | parameters=train_param, max_norm=args.clip, norm_type=2.0 183 | ) 184 | optimizer.step() 185 | 186 | if train: 187 | print( 188 | "Epoch {}, Total Iter: {}, Train Avg Loss: {:.6f}, Train Avg Accuracy: {:.6f}".format( 189 | epoch, counter, loss_avg.avg, acc_avg.avg 190 | ) 191 | ) 192 | writer.add_scalar("loss/training", loss_avg.avg, epoch) 193 | writer.add_scalar("accuracy/training", acc_avg.avg, epoch) 194 | else: 195 | print( 196 | "Epoch {}, Valid Avg Loss: {:.6f}, Valid Avg Acc: {:.4f}".format( 197 | epoch, loss_avg.avg, acc_avg.avg 198 | ) 199 | ) 200 | writer.add_scalar("loss/validation", loss_avg.avg, epoch) 201 | writer.add_scalar("accuracy/validation", acc_avg.avg, epoch) 202 | for r in model.rules_set.rules: 203 | print(str(r.rule)) 204 | return acc_avg.avg 205 | 206 | # Set random seed 207 | np.random.seed(args.seed) 208 | torch.manual_seed(args.seed) 209 | if args.cuda: 210 | torch.cuda.manual_seed(args.seed) 211 | torch.backends.cudnn.benchmark = False 212 | 213 | writer = SummaryWriter(args.log_dir) 214 | 215 | # Init model 216 | model = RuleLevelReasoner( 217 | args.device, 218 | args.config, 219 | model=args.model, 220 | hidden_layers=args.hidden_layers, 221 | dictionary=args.backend_cb, 222 | vsa_conversion=args.vsa_conversion, 223 | vsa_selection=args.vsa_selection, 224 | context_superposition=args.context_superposition, 225 | num_rules=args.num_rules, 226 | shared_rules=args.shared_rules, 227 | program=args.program, 228 | num_terms=args.num_terms, 229 | n=args.n, 230 | ) 231 | model.to(args.device) 232 | if args.config == "in_out_four": 233 | model2 = RuleLevelReasoner( 234 | args.device, 235 | "center_single", 236 | model=args.model, 237 | hidden_layers=args.hidden_layers, 238 | dictionary=args.backend_cb, 239 | vsa_conversion=args.vsa_conversion, 240 | vsa_selection=args.vsa_selection, 241 | context_superposition=args.context_superposition, 242 | num_rules=args.num_rules, 243 | shared_rules=args.shared_rules, 244 | program=args.program, 245 | num_terms=args.num_terms, 246 | n=args.n, 247 | ) 248 | model2.to(args.device) 249 | 250 | distribute = "distribute" in args.config or "in_out_four" == args.config 251 | # Init loss 252 | loss_fn = getattr(losses, args.loss_fn)() 253 | 254 | rule_selector = RuleSelector( 255 | loss_fn, args.rule_selector_temperature, rule_selector=args.rule_selector 256 | ) 257 | 258 | # Init optimizers 259 | train_param = list(model.parameters()) 260 | if args.config == "in_out_four": 261 | train_param += list(model2.parameters()) 262 | optimizer = optim.AdamW(train_param, args.lr, weight_decay=args.weight_decay) 263 | 264 | # Load all checkpoints 265 | rule_path = os.path.join(args.resume, "checkpoint.pth.tar") 266 | if os.path.isfile(rule_path): 267 | checkpoint = torch.load(rule_path) 268 | model.load_state_dict(checkpoint["state_dict_model"]) 269 | if args.config == "in_out_four": 270 | model2.load_state_dict(checkpoint["state_dict_model2"]) 271 | best_accuracy = checkpoint["best_accuracy"] 272 | start_epoch = checkpoint["epoch"] 273 | optimizer.load_state_dict(checkpoint["optimizer"]) 274 | print( 275 | "=> loaded checkpoint '{}' at Epoch {:.3f}".format( 276 | rule_path, checkpoint["epoch"] 277 | ) 278 | ) 279 | else: 280 | best_accuracy = 0 281 | start_epoch = 0 282 | 283 | # Dataset loader 284 | train_set = GeneralIRAVENDataset( 285 | "train", 286 | args.data_dir, 287 | constellation_filter=args.config, 288 | rule_filter=args.gen_rule, 289 | attribute_filter=args.gen_attribute, 290 | n_train=args.n_train, 291 | maxval=args.dyn_range, 292 | partition=args.partition, 293 | n=args.n, 294 | n_confounders=args.orientation_confounder, 295 | ) 296 | train_loader = DataLoader( 297 | train_set, 298 | batch_size=args.batch_size, 299 | shuffle=True, 300 | num_workers=args.num_workers, 301 | ) 302 | val_set = GeneralIRAVENDataset( 303 | "val", 304 | args.data_dir, 305 | constellation_filter=args.config, 306 | rule_filter=args.gen_rule, 307 | attribute_filter=args.gen_attribute, 308 | n_train=args.n_train, 309 | maxval=args.dyn_range, 310 | partition=args.partition, 311 | n=args.n, 312 | ) 313 | val_loader = DataLoader( 314 | val_set, batch_size=args.batch_size * 15, num_workers=args.num_workers 315 | ) 316 | 317 | # training loop starts 318 | for epoch in range(start_epoch, args.epochs): 319 | inference_epoch(epoch, loader=train_loader, train=True) 320 | with torch.no_grad(): 321 | accuracy = inference_epoch(epoch, loader=val_loader, train=False) 322 | 323 | # store model(s) 324 | is_best = accuracy > best_accuracy 325 | best_accuracy = max(accuracy, best_accuracy) 326 | if args.config == "in_out_four": 327 | save_checkpoint( 328 | { 329 | "epoch": epoch + 1, 330 | "state_dict_model": model.state_dict(), 331 | "state_dict_model2": model2.state_dict(), 332 | "best_accuracy": accuracy, 333 | "optimizer": optimizer.state_dict(), 334 | }, 335 | is_best, 336 | savedir=args.checkpoint_dir, 337 | ) 338 | else: 339 | save_checkpoint( 340 | { 341 | "epoch": epoch + 1, 342 | "state_dict_model": model.state_dict(), 343 | "best_accuracy": best_accuracy, 344 | "accuracy": accuracy, 345 | "optimizer": optimizer.state_dict(), 346 | }, 347 | is_best, 348 | savedir=args.checkpoint_dir, 349 | ) 350 | return writer 351 | 352 | 353 | def test(args, env, device, writer=None, dset="RAVEN", confounders=False): 354 | """ 355 | Testing of NVSA backend 356 | """ 357 | 358 | def test_epoch(): 359 | model.eval() 360 | if args.config == "in_out_four": 361 | model2.eval() 362 | rule_selector.eval() 363 | 364 | loss_avg = AverageMeter("Loss", ":.3f") 365 | acc_avg = AverageMeter("Accuracy", ":.3f") 366 | rule_acc_avg = { 367 | rule: AverageMeter("Accuracy", ":.3f") for rule in iravenx_rule_map.keys() 368 | } 369 | 370 | for extracted, targets, all_action_rule in tqdm(test_loader): 371 | extracted, targets, all_action_rule = ( 372 | extracted.to(device), 373 | targets.to(device), 374 | all_action_rule.to(device), 375 | ) 376 | att_logprob = create_one_hot(extracted, args.config, args.sigma) 377 | model_output = {k: v.to(device) for k, v in att_logprob.items()} 378 | scene_prob, _ = env.prepare(model_output) 379 | if args.config in ["center_single", "distribute_four", "distribute_nine"]: 380 | outputs, candidates, tests = model(scene_prob, distribute=distribute) 381 | outputs, attr_entropy = rule_selector(outputs, tests) 382 | loss, scores, attscores = compute_loss_and_scores( 383 | outputs, 384 | tests, 385 | candidates, 386 | targets, 387 | loss_fn, 388 | "distribute" in args.config, 389 | [p.data for p in model.parameters() if p.requires_grad], 390 | confounders=confounders, 391 | attr_entropy=attr_entropy, 392 | use_entropy=args.entropy, 393 | ) 394 | 395 | else: 396 | outputs1, candidates1, tests1 = model( 397 | scene_prob[0], distribute=distribute 398 | ) 399 | outputs1, attr_entropy1 = rule_selector(outputs1, tests1) 400 | if args.config == "in_out_four": 401 | outputs2, candidates2, tests2 = model2( 402 | scene_prob[1], distribute=False 403 | ) 404 | else: 405 | outputs2, candidates2, tests2 = model( 406 | scene_prob[1], distribute=False 407 | ) 408 | outputs2, attr_entropy2 = rule_selector(outputs2, tests2) 409 | loss1, scores1, _ = compute_loss_and_scores( 410 | outputs1, 411 | tests1, 412 | candidates1, 413 | targets, 414 | loss_fn, 415 | distribute=args.config == "in_out_four", 416 | confounders=confounders, 417 | attr_entropy=attr_entropy1, 418 | use_entropy=args.entropy, 419 | ) 420 | loss2, scores2, _ = compute_loss_and_scores( 421 | outputs2, 422 | tests2, 423 | candidates2, 424 | targets, 425 | loss_fn, 426 | distribute=False, 427 | confounders=confounders, 428 | attr_entropy=attr_entropy2, 429 | use_entropy=args.entropy, 430 | ) 431 | loss = (loss1 + loss2) / 2 432 | scores = (scores1 + scores2) / 2 433 | 434 | # accuracy and loss computation 435 | predictions = torch.argmax(scores, dim=-1) 436 | accuracy = ((predictions == targets).sum() / len(targets)) * 100 437 | loss_avg.update(loss.item(), extracted.size(0)) 438 | acc_avg.update(accuracy.item(), extracted.size(0)) 439 | ##### <- rule accuracy computation 440 | if args.evaluate_rule: 441 | expanded_rules = all_action_rule[:, 1:] 442 | batch_size = all_action_rule.shape[0] 443 | for rule in iravenx_index_map.keys(): 444 | total = 0 445 | correct = 0 446 | rule_mask = expanded_rules == rule 447 | if rule_mask.any(): 448 | if args.sigma >= 0: 449 | for attribute_idx in range(0, 3): 450 | max_scores = attscores[attribute_idx].max(dim=-1).values 451 | not_max_entropy = ~torch.all( 452 | attscores[attribute_idx] 453 | == attscores[attribute_idx][:, 0:1], 454 | dim=1, 455 | ) 456 | correct += ( 457 | ( 458 | torch.logical_and( 459 | attscores[attribute_idx][ 460 | torch.arange(batch_size), targets 461 | ] 462 | == max_scores, 463 | torch.logical_and( 464 | rule_mask[:, attribute_idx], 465 | not_max_entropy, 466 | ), 467 | ) 468 | ) 469 | .sum() 470 | .item() 471 | ) 472 | total += rule_mask[:, attribute_idx].sum().item() 473 | else: 474 | correct += ( 475 | (rule_mask.sum(dim=-1) * (predictions == targets)) 476 | .sum() 477 | .item() 478 | ) 479 | total += rule_mask.sum().item() 480 | 481 | rule_acc = correct / total if total else 0 482 | rule_acc_avg[iravenx_index_map[rule]].update( 483 | rule_acc, extracted.size(0) 484 | ) 485 | ##### -> rule accuracy computation 486 | 487 | # Save final result as npz (and potentially in Tensorboard) 488 | if args.resume == "": 489 | if writer is not None: 490 | writer.add_scalar("accuracy/testing-{}".format(dset), acc_avg.avg, 0) 491 | np.savez( 492 | args.save_dir + "result_{:}.npz".format(dset), loss=acc_avg.avg 493 | ) 494 | else: 495 | args.save_dir = args.resume.replace("ckpt/", "save/") 496 | np.savez( 497 | args.save_dir + "result_{:}.npz".format(dset), loss=acc_avg.avg 498 | ) 499 | 500 | print("Test Avg Accuracy: {:.4f}".format(acc_avg.avg)) 501 | if args.evaluate_rule: 502 | for rule in iravenx_rule_map.keys(): 503 | print(f"Rule {rule} Avg Accuracy: {rule_acc_avg[rule].avg * 100:.2f}") 504 | for r in model.rules_set.rules: 505 | print(str(r.rule)) 506 | return { 507 | **{"acc": acc_avg.avg}, 508 | **{rule: rule_acc_avg[rule].avg * 100 for rule in iravenx_rule_map.keys()}, 509 | } 510 | 511 | # Load all checkpoint 512 | model_path = os.path.join(args.resume, "model_best.pth.tar") 513 | # model_path = os.path.join(args.resume, "checkpoint.pth.tar") 514 | print(model_path) 515 | if os.path.isfile(model_path): 516 | checkpoint = torch.load(model_path) 517 | print( 518 | "=> loaded checkpoint '{}', epoch {}, with accuracy {:.3f}".format( 519 | model_path, checkpoint["epoch"], checkpoint["best_accuracy"] 520 | ) 521 | ) 522 | else: 523 | print( 524 | f"Careful! The model is not loaded from checkpoint. Program is: {args.program}" 525 | ) 526 | # raise ValueError("No checkpoint found at {:}".format(model_path)) 527 | test_acc = dict() 528 | configs = [ 529 | "center_single", 530 | # "distribute_four", 531 | # "distribute_nine", 532 | # "left_right", 533 | # "up_down", 534 | # "in_out_single", 535 | # "in_out_four", 536 | ] 537 | for config in configs: 538 | args.config = config 539 | env = reasoning_env.get_env(args.configs_map[args.config], device) 540 | # Init the model 541 | model = RuleLevelReasoner( 542 | args.device, 543 | config, 544 | model=args.model, 545 | hidden_layers=args.hidden_layers, 546 | dictionary=args.backend_cb, 547 | vsa_conversion=args.vsa_conversion, 548 | vsa_selection=args.vsa_selection, 549 | context_superposition=args.context_superposition, 550 | num_rules=args.num_rules, 551 | shared_rules=args.shared_rules, 552 | program=args.program, 553 | num_terms=args.num_terms, 554 | n=args.n, 555 | ) 556 | model.to(device) 557 | if not args.program: 558 | model.load_state_dict(checkpoint["state_dict_model"]) 559 | if args.data_dir == "/dccstor/saentis/data/I-RAVEN": 560 | model.anneal_softmax() 561 | if config == "in_out_four": 562 | model2 = RuleLevelReasoner( 563 | args.device, 564 | "center_single", 565 | model=args.model, 566 | hidden_layers=args.hidden_layers, 567 | dictionary=args.backend_cb, 568 | vsa_conversion=args.vsa_conversion, 569 | vsa_selection=args.vsa_selection, 570 | context_superposition=args.context_superposition, 571 | num_rules=args.num_rules, 572 | shared_rules=args.shared_rules, 573 | program=args.program, 574 | num_terms=args.num_terms, 575 | n=args.n, 576 | ) 577 | model2.to(device) 578 | if not args.program: 579 | model2.load_state_dict(checkpoint["state_dict_model"]) 580 | distribute = "distribute" in config or "in_out_four" == config 581 | # Init loss 582 | loss_fn = getattr(losses, args.loss_fn)() 583 | 584 | rule_selector = RuleSelector( 585 | loss_fn, args.rule_selector_temperature, rule_selector=args.rule_selector 586 | ) 587 | 588 | # Dataset loader 589 | test_set = GeneralIRAVENDataset( 590 | "test", 591 | args.data_dir, 592 | constellation_filter=config, 593 | rule_filter=args.gen_rule, 594 | attribute_filter=args.gen_attribute, 595 | maxval=args.dyn_range, 596 | partition=args.partition, 597 | n=args.n, 598 | n_confounders=args.orientation_confounder, 599 | ) 600 | test_loader = DataLoader( 601 | test_set, batch_size=args.batch_size, num_workers=args.num_workers 602 | ) 603 | print("Evaluating on {}".format(config)) 604 | with torch.no_grad(): 605 | acc = test_epoch() 606 | test_acc[config] = acc 607 | 608 | with open(os.path.join(args.resume, f"eval.json"), "w") as fp: 609 | json.dump(test_acc, fp) 610 | return writer 611 | 612 | 613 | def main(): 614 | args = parse_args() 615 | 616 | args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 617 | args.cuda = torch.cuda.is_available() 618 | 619 | # Use a rng for reproducible results 620 | rng = np.random.default_rng(seed=args.seed) 621 | 622 | torch.manual_seed(args.seed) 623 | np.random.seed(args.seed) 624 | random.seed(args.seed) 625 | os.environ["PYTHONHASHSEED"] = str(args.seed) 626 | 627 | # Load or define new codebooks 628 | backend_cb_cont, backend_cb_discrete = generate_nvsa_codebooks(args, rng) 629 | 630 | args.backend_cb_discrete = backend_cb_discrete 631 | args.backend_cb_cont = backend_cb_cont 632 | 633 | print(f"Sigma: {args.sigma}") 634 | 635 | if args.model == "LearnableFormula": 636 | args.backend_cb = backend_cb_cont 637 | else: 638 | args.backend_cb = backend_cb_discrete 639 | 640 | # backend for training/testing 641 | input_configs = [ 642 | "center_single", 643 | "left_right", 644 | "up_down", 645 | "in_out_single", 646 | "distribute_four", 647 | "in_out_four", 648 | "distribute_nine", 649 | ] 650 | output_configs = [ 651 | "center_single", 652 | "left_center_single_right_center_single", 653 | "up_center_single_down_center_single", 654 | "in_center_single_out_center_single", 655 | "distribute_four", 656 | "in_distribute_four_out_center_single", 657 | "distribute_nine", 658 | ] 659 | args.configs_map = dict(zip(input_configs, output_configs)) 660 | 661 | env = reasoning_env.get_env(args.configs_map[args.config], args.device) 662 | 663 | if args.mode == "train": 664 | args.exp_dir = os.path.join(args.exp_dir, args.run_name, str(args.seed)) 665 | args.checkpoint_dir = os.path.join(args.exp_dir, "ckpt") 666 | args.save_dir = os.path.join(args.exp_dir, "save") 667 | args.log_dir = os.path.join(args.exp_dir, "log") 668 | check_paths(args) 669 | 670 | # Run the actual training 671 | writer = train(args, env, args.device, confounders=args.orientation_confounder) 672 | 673 | # Do final testing 674 | args.resume = args.checkpoint_dir 675 | writer = test( 676 | args, env, args.device, writer, confounders=args.orientation_confounder 677 | ) 678 | 679 | writer.close() 680 | 681 | elif args.mode == "test": 682 | test(args, env, args.device, confounders=args.orientation_confounder) 683 | 684 | 685 | if __name__ == "__main__": 686 | main() 687 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . --------------------------------------------------------------------------------