alexredplanet / mushroomclassifier Goto Github PK

This is an old project from 2020 on predicting 1000 mushroom species from image data.

The data was kindly provided by mushroomobserver.org.

I am revisiting this as the dataset is highly interesting as a playground to test ideas. This is because the data contains:

• highly noisy labels (labels are from a user voting system)
• feature noise (images are uploaded by individual users)
• highly imbalanced classes

The current version of the project uses simple deep learning models (ResNet variants) and the FastAI library. One interesting idea I toyed with here was a hierarchical model, where a genus level classifier was first trained, and then for each genus, a species level classifier was trained. I tried this as it would allow for new mushroom species to be gradually added to the model over time, with much faster retraining than a single species level model.

I have not yet implemented solutions to the noise or imbalance problems, although I will gradually add them in future.

I have uploaded a small report which I submitted as part of the course Principles of Data Science at the University of Sydney in 2021, which explains some of the reasoning behind the hierarchical model.

I haven't uploaded the model files as they are very large, nor have I uploaded the image files. The notebooks will download all relevant data and generate all models.