This repository contains the codes used for the study on the quickdraw dataset conducted for the CentraleSupélec's Deep Learning class.
The goal of the project was to train models to classify elements from the quickdraw dataset.
Quickdraw is a game release by Google that prompts the participant with the name of an item (for instance drums), the participant then has 20 seconds to draw it and make a neural network guess the correct item name. Thanks to this game, the Quickdraw team collected a lot of annotated images and made them available publicly, there are 345 different items and each has at least
The notebooks are located in the notebooks/ directory:
- lstm_quickdraw.ipynb is used for the training of LSTMs on the strokes data
- strokes2image_quickdraw.ipynb is used for the training of CNNs on images generated from strokes
- cnn_28x28bitmaps_training.ipynb is used for the training of CNNs on grayscale 28x28 bitmaps
To be able to run the notebooks you will need the following dependencies:
- pytorch
- torchvision
- opencv
- numpy
For a detailed list of dependencies, you can have a look at this file
Results can be found in the results/ directory in markdown files, there is also a helper script to generate graphs of results. For CNNs on bitmaps, results are in two notebooks: cnn_bitmaps_plot_train_val_results.ipynb for training and validation results and cnn_bitmaps_test_results.ipynb for test results.
In the mesocentre/ are the configuration for running the training jobs on a PBS queue, as well as a conda environment configuration.
Finally the report source code can be found here