This is a clone from radek and scales the training to 100K iamges per class with different data loading strategy.

The code in this repo is all you need to make a first submission to the Quick, Draw! Kaggle Competition. It uses the FastAi library.

For additional information please refer to the discussion thread on Kaggle forums.

I provide instructions on how to run the code below.

This code is based on code from a fast.ai MOOC that will be publicly available in Jan 2019

You can find an earlier version of this starter pack here. This iteration eliminates some of the rough edges that resulted in a relatively low score and also introduces the data block API. A major change is that here I am generating drawings on the fly - this should help with experimentation.

1. You need to have the FastAi library up and running. You can find installation instructions here.
2. You will also need to download the competition data. The competition can be found under this url. If you do not have a Kaggle account you will need to register. There are many ways to download the data - I use the Kaggle CLI. To set it up you will need to generate an API key - all this information is available in the Kaggle CLI repository. If you do not want to set this up at this point, you can download the data from the competition's data tab.
3. We will need test_simplified.csv and train_simplified.zip. Please put them in the data directory. If you were to download one of the files using the Kaggle CLI, the command (executed from the root of the repository) would be kaggle competitions download -c quickdraw-doodle-recognition -f test_simplified.csv -p data.
4. We now need to unzip the downloaded files. cd into the data folder and create a new directory train. Extract the downloaded files by executing unzip train_simplified.zip -d train.
5. Open first_submission.ipynb. If you have kaggle CLI installed and configured, you can uncomment the last line. Hit run all. See you on the leaderboards :)