What-Can-Neural-Networks-Reason-About: A Python repository from NNReasoning

Pytorch implementation for abstract reasoning tasks - [What Can Neural Networks Reason About?]

File Descriptions

All files ending in *generator.py generate a type of dataset for a given task (e.g. subset-sum, Monster trainer, etc):

The tasks described in our experiment section corresponds to the following subtype IDs (IDs are authors' favorite random numbers):

Furthest pair: subtype 0

Maximum value difference: subtype 2

Subset sum: subtype 12

Monster trainer: subtype 15

Requirements

Python 2.7
numpy
pytorch 0.4.0

Usage

n_iter: number of message passing layers in GNN
n_objects: number of graph nodes in the input dataset

See appendix G for the set of hyperparameters we use for each task

For tasks Furthest pair (subtype 0) and Maximum value difference (subtype 2):

K: number of animal types (6 by default)
n_objects: number of animals (25 by default)

Use $ python treasure_generator.py --K=6 --n_objects=25 --coord_size=8 --age_range=100 --coord_range=20 to generate the animal world dataset

and $ python main.py --model=$MODEL_NAME --lr=$LEARNING_RATE --batch_size=$BATCH_SIZE --hidden_dim=$HIDDEN_DIMENSION --data=$DATA_FILE.pickle --epochs=$EPOCHS --fc_output_layer=$NUMBER_OF_LAYERS_IN_OUTPUT_FC --mlp_layer=$NUMBER_OF_LAYERS_IN_GNN_AGGREGATION --K=6 --n_iter=3 --n_objects=25 --coord_size=8 --age_range=100 --coord_range=20 --subtype=0 (or --subtype=2 for task Maximum value difference) to train.

For task subset sum:

Use $ python subset_sum_generator.py to generate the dataset

For task Monster trainer:

max_level: number of other objects (except self) in the monster trainer task

Use $ python monster_trainer_generator.py --max_level=10 to generate the monster trainer dataset