This repository contains code to train, tune, and make/submit predictions using the Numerox API to the Numerai Data Science Tournament.
https://docs.numer.ai/tournament/learn
To train the models or make predictions I recommend that you first create a virtual environment and install the requirements file.
Use Python 3.7 to avoid some import issues I encountered:
virtualenv --python=/usr/bin/python3.7 <path/to/new/virtualenv/>- On Posix systems:
source /path/to/ENV/bin/activate - On Windows (where this repo was developed):
\path\to\env\Scripts\activate
conda create -n yourenvname python=x.x anacondasource activate yourenvname
Once you have your virtual environment set up, install the requirements:
pip install -r requirements.txt
The API has a few basic components that come together in the predict.py module. There are Models and Trainers that compose a train_and_predict_model() function within the predict module. A Model is built on top of the Numerox API to make submission a bit simpler; a Trainer contains functionality to train, load, save models and also submit predictions - locally or from/to an s3 bucket.
- Create a parameter dictionary:
EXAMPLE_PARAMS = { 'depth': 7, 'learning_rate': 0.1, 'l2': 0.01, 'iterations': 100 }- Use parameter dictionary as an argument to
params=EXAMPLE_PARAMSto thetrain_and_predict_<model-name>_model()function.
- To then train, make predictions, and submit predictions run the following command. After training is complete the model weights will be saved to disk with the model name attached to it like this:
<model-name>_model_trained_<competition-name>. You also have the option of saving the model to an s3 bucket (more on that below):
python predict.py --model <model-name> --load-model <bool> --save-model <bool> --submit <bool>
- Setup ennvironment variables for AWS s3 bucket:
export BUCKET=<bucket-name>export AWS_ACCESS_KEY_ID=<access-key-id>export AWS_SECRET_KEY=<secret-key>
- Default setup is for models to be loaded to and from an s3 bucket, so run
predict.pymodule as is. If you want to also save the models locally change the code by callingtrainer.save/load_model_locally()methods.
- Setup ennvironment variables for NumerAPI:
export NUMERAI_PUBLIC_ID=<public-id>export NUMERAI_SECREY_KEY=<secret-key>
- Set the
--submitparameter to True when runningpython predict.py --submit
With numerai_train as your working directory, run the following from the command line:
python -m pytest tests/tests_unit.py -v
WIP
polyaxon login --username=root --password=rootpasswordpolyaxon project create --name=numerai_training --description='Train models on polyaxon'polyaxon init numerai_training- CPU:
polyaxon run -f configs/polyaxon_cpu.yaml - GPU:
polyaxon run -f configs/polyaxon_gpu.yaml
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent