pytorch
training script implementing a feedforward network on the Rossman Store Sales competition.
This is a tabular time-series dataset, traditionally the domain of gradient boosted tree libraries like xgboost
. However, recent advances in generalized embeddings (mostly stemming from the NLP world) have put feedforward neural networks on par with GBTs in terms of performance.
This pytorch
implementation adapts the fastai
model presented in Lesson 3 of the FastAI course.
To run code and notebooks in a Spell workspace:
spell jupyter --lab \
--github-url https://github.com/spellml/feedforward-rossman.git \
--pip kaggle \
--env KAGGLE_USERNAME=YOUR_USERNAME \
--env KAGGLE_KEY=YOUR_KEY \
feedforward-rossman
spell run \
--machine-type V100 \
--github-url https://github.com/spellml/feedforward-rossman.git \
--pip kaggle \
--env KAGGLE_USERNAME=YOUR_USERNAME \
--env KAGGLE_KEY=YOUR_KEY \
"chmod +x /spell/scripts/download_data.sh /spell/scripts/upgrade_env.sh; /spell/scripts/download_data.sh; /spell/scripts/upgrade_env.sh; python /spell/models/model_4.py"
spell run \
--machine-type V100 \
--github-url https://github.com/spellml/feedforward-rossman.git \
--pip kaggle \
--env KAGGLE_USERNAME=YOUR_USERNAME \
--env KAGGLE_KEY=YOUR_KEY \
"chmod +x /spell/scripts/download_data.sh /spell/scripts/upgrade_env.sh; /spell/scripts/download_data.sh; /spell/scripts/upgrade_env.sh; python /spell/models/model_4.py"