Architopes: An Architecture Modification for Composite Pattern Learning, Increased Expressiveness, and Reduced Training Time

Requirements

To install requirements:

Install Anaconda version 4.8.2.
Create Conda Environment

# cd into the same directory as this README file.

conda create python=3.7.6 --name architopes \
conda activate architopes \
pip install -r requirements.txt

Training and Evaluation

Preprocess

In this step, latitude and longitude are mapped to Euclidean coordinates about the projected extrinsic mean to remove the effect of earth’s curvature and then the data is split into distinct set using the proximity to ocean. After each set is split into train and test using 30% for test. Skip this step to use the same exact train and test sets used to produce teh results in the paper.

python3 ./code/preprocessing.py \
--source_file './data/raw/housing.csv' \
--sink_path './data/data'

Specify the parameters related to each set and the space of hyper parameters in ./Training_Evaluation/Grid.py. Default parameters used to obtain the results in the paper can be found in the appendix.
Train and save the results for the ffNN and ffNN-tope using the following command:

python3 ./code/train_eval.py \
--is_test 'F'  \
--is_manual 'T' \
--n_iter 200 \
--n_jobs 30 \
--result_path './results/train_eval' \
--data_path './data/data'

Compile Results

python ./code/postprocessing.py \
--result_path './results/train_eval' \
--compile_path './results/compiled' \

Pre-trained Models

Pre-trained models are found in ./pretrained_models. pretrained_models.ipynb is provided for easier use.

Results

Our model achieves the following performance on :

California Housing Price Dataset

The house prices were multiplied by 10^(-5) to avoid exploding gradient issues.

For Train:

Model name	MAE	MAPE	MSE
ffNN-tope	0.285	15.01	0.210
ffNN	0.297	16.41	0.211
ffNN-dp	0.411	19.94	0.395

For Test: