Object recognition and computer vision 2023/2024

Assignment 3: Sketch image classification

Requirements

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

2. Run the following command to install additional dependencies

pip install -r requirements.txt

Dataset

We will be using a dataset containing 250 different classes of sketches adapted from the classifysketch dataset. Download the training/validation/test images from here. The test image labels are not provided.

Training and validating your model

Run the script main.py to train your model.

Modify main.py, model.py and data.py for your assignment, with an aim to make the validation score better.

• By default the images are loaded and resized to 64x64 pixels and normalized to zero-mean and standard deviation of 1. See data.py for the data_transforms.

• When changing models, you should also add support for your model in the ModelFactory class in model_factory.py. This allows to not having to modify the evaluation script after the model has finished training.

Evaluating your model on the test set

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.

Logger

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.

Acknowledgments

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/