Starter code: Kaggle Toxic Comment Classification Challenge

Here, at Neptune we enjoy participating in the Kaggle competitions. Toxic Comment Classification Challenge is especially interesting because it touches important issue of online harassment.

Ensemble our predictions in the cloud!

You need to be registered to neptune.ml to be able to use our predictions for your ensemble models.

click start notebook
choose browse button
select the neptune_ensembling.ipynb file from this repository.
choose worker type: gcp-large takes over an hour gcp-gpu-medium less 20min
run first few cells to load our predictions on the held out validation set along with the labels
train your second level, ensemble model
load our predictions on the test set
feed our test set predictions to your ensemble model and get final predictions
save your submission file
click on browse files and find your submission file to download it.

Running the notebook as is got 0.9849 on the LB.

The idea

We are contributing starter code that is easy to use and extend. We did it before with Cdiscount’s Image Classification Challenge and we believe that it is correct way to open data science to the wider community and encourage more people to participate in Challenges. This starter is ready-to-use end-to-end solution. Since all computations are organized in separate steps, it is also easy to extend. Check devbook.ipynb for more information about different pipelines.

Now we want to go one step further and invite you to participate in the development of this analysis pipeline. At the later stage of the competition (early February) we will invite top contributors to join our team on Kaggle.

Contributing

You are welcome to extend this pipeline and contribute your own models or procedures. Please refer to the CONTRIBUTING for more details.

Installation

option 1: Neptune cloud

on the neptune site

register to receive $100 in GPU time
log in
create new project named toxic: Follow the link Projects (top bar, left side), then click New project button. This action will generate project-key TOX, which is already listed in the neptune_config.yaml.

run setup commands

$ git clone https://github.com/neptune-ml/kaggle-toxic-starter.git
$ pip3 install neptune-cli
$ neptune login

start experiment

$ neptune send experiment_manager.py --environment keras-2.0-gpu-py3 --worker gcp-gpu-medium --config neptune_config.yaml -- train_evaluate_predict_pipeline --pipeline_name glove_lstm

Happy Training :)

Refer to Neptune documentation and Getting started: Neptune Cloud for more.

option 2: local install

Please refer to the Getting started: local instance for installation procedure.

Solution visualization

Below end-to-end pipeline is visualized. You can run exactly this one!

We have also prepared something simpler to just get you started:

User support

There are several ways to seek help:

Read project's Wiki, where we publish descriptions about the code, pipelines and neptune.
Kaggle discussion is our primary way of communication.
You can submit an issue directly in this repo.

epxjk82/kaggle-toxic-starter