Course project for CMU 17-737 AI for Social Good. Data is collected from GeoLifeCLEF 2020
Set up a local virtual environment to install and manage package versions.
virtualenv -p python3 venvsource venv/bin/activatepip3 install -r requirements.txt
Given an input of annotations (one of annotations_train.json or annotations_val.json), this script will read and re-format the JSON object accordingly.
There are two outputs: (1) a reformatted JSON which includes the category_id (label) with each record, and (2) a list of all included filenames, sorted.
Optionally, this script can also filter out all category_id's above a certain value.
i.e. It's useful to restrict the eventual ML model to the top 20 classes for easier training and eval.
Example usage: python3 json_parser.py annotations_train.json (see the file for how the filtering works)
This is a bash script which, given a list of filenames (the output of json_parser, for example, annotations_train_filtered_fr.txt), downloads all the az patches, and deletes all the subdirectories not found in the filelist.
Given a set of input .npy files, (if you used the bash script above, the remaining directories in \patches) this script will load each out and separate out the data format into n-by-d formatting matrices, ready to be used for machine learning.
Mostly still an example for now
Given a set of input processed .npy files (the output of process_patches.py), this script will perform a very basic CNN training, and output the train/test accuracy, as well as the loss curve.