This repo contains code of the 5th place solution in Quora Insincere Question Classification. The solution is written in Python 3, based on PyTorch. The experiments are organized by sacred. For details of the solution, check my kernel on Kaggle.
A modified version of temporal convolutional network gotten from here is provided in branch tcn. The f1 is about 0.02 lower than my best rnn model.
Firstly, you need to download data here
and put the data into ./input folder. The parameters can be modified in config.yaml.
(comments after the challenge: Better result can be achieved with batch size 768 in the given time limit)
To start an experiment, run
python expr.py
To generate a submission, run
python script.py
To run a hypothesis test, run
python stats.py
The experimental results are assumed to be tracked in mongodb.