mark_twain

CS7150 Deep Learning Final Project

Applying Deep Learning to the Numerai data science competition.

Code authors

Jackson Neal
Rohit Barve
Quay Dragon

GPU Access

NOTE: GPU configuration is optional but recommended for compute intensive jobs.

1. SSH to login.discovery.neu.edu
2. Request GPU partition

make req-gpu

3. Load anaconda and cuda modules

module load anaconda3/2022.01

module load cuda/11.1

4. Activate pytorch_env_training environment

source activate pytorch_env_training

Installation

make init

Execution

Demo Predictions

Load saved BASE model, load Numerai data, execute predictions. Predictions formatted for submission to Numerai and output to ./predictions.csv in project root.

# no GPU
make demo-base

# use GPU
make demo-base-gpu

Training walk through with EDA available in code_walkthrough.ipynb

Single Run Train Configuration

1. Modify single run hyperparameters in single.yaml
2. Run trainer with one of:

# no GPU
make single

# use GPU
make single-gpu

3. View Tensorboard logs

make tb-logs

Weights and Biases Sweep Run Train Configuration

NOTE: Executing sweep runs requires an api key

1. Modify sweep run hyperparameters in sweep.yaml
2. Run trainer with on of:

# no GPU
make sweep

# use GPU
make sweep-gpu

3. View sweep results at https://wandb.ai/cs7150-jn

Predictions From Model Checkpoint

Load saved model configs and weights to execute predictions. File paths are relative to project root. Run predictions with one of:

# no GPU
make predict ckpt=path/to/.ckpt hparams=path/to/.yaml

# use GPU
make predict-gpu ckpt=path/to/.ckpt hparams=path/to/.yaml

Submit Predictions to Numerai

NOTE: Submission requires api keys in local .env file

make submit model=(BASE|AEMLP|TMLP|CAE)