1. Install PyTorch from http://pytorch.org

2. Create a new virtual environment

pyenv virtualenv 3.11.9 a3
pyenv activate a3

3. Run the following command to install additional dependencies

pip install poetry
poetry install

Download the training/validation/test images from here. The test image labels are not provided.

Run the script main.py to train your model.

As the model trains, model checkpoints are saved to files such as model_x.pth to the current working directory. You can take one of the checkpoints and run:

python evaluate.py --data [data_dir] --model [model_file] --model_name [model_name]

That generates a file kaggle.csv that you can upload to the private kaggle competition website.

We recommend you use an online logger like Weights and Biases to track your experiments. This allows to visualise and compare every experiment you run. In particular, it could come in handy if you use google colab as you might easily loose track of your experiments when your sessions ends.

Note that currently, the code does not support such a logger. It should be pretty straightforward to set it up.

Adapted from Rob Fergus and Soumith Chintala https://github.com/soumith/traffic-sign-detection-homework.
Origial adaptation done by Gul Varol: https://github.com/gulvarol
New Sketch dataset and code adaptation done by Ricardo Garcia and Charles Raude: https://github.com/rjgpinel, http://imagine.enpc.fr/~raudec/