Multi-class, single target classification of birds sold in wildlife marketplaces

• as a general thing - if you are altering or moving any data it is better to make a copy first, so you are making sure that you are not altering the original data and file/folder structure

• Structure of data and labels
  • images of bird species from different sensors are stored in folders under data_root/high, there is one bird per image (single target)
  • these are named species_name_index_source.fileextension
  • In the overall annotations_new.json file is an entry for every image, its annotation (which is 1-1, i.e. there is one annotation per image) and the overall class, or category that it belongs to

Folder structure

• preprocessing folder
  • annotations_backup: copies of the annotations files that are generated by the scripts in the preprocessing folder - when running the code, these are stored in the ./data folder, as they are read in along with the imagePaths
  • train.json is a random split (though using the same random seed for reproducibility) which contains 80% of the original data, across all initial 29 classes (training_18_08.json is an older version of the split, which contained some bad data and one class, the coal tit, which has since been completely removed since the crops were corrupted)
  • val.json is 20% of the original data, and generated at the same time as train.json (val_18_08.json)
  • upsampling json - is an attempt to balance the classes. In this file, it contains balanced data across all 29 classes

delete_backup_folders.py

• I copied the data initially into folders, where folder_name/random_images was the subset for each species at the beginning that I knew were contributed by different labellers and thus using different sensors (or these were images collected online)

• delete_files.py - this was used to delete files with _random in the file name when these were incorrectly copied to the wrong folder at the beginning - this code could be used to remove images from sources that might not be relevant (i.e. from a certain sensor if hte image is labelled with species_name_sensor_source.fileextension

• visualisation_training.py and visualisation_validation.py is what prints the distribution of data across classes in both the training and validation sets, by loading the training and validation .jsons for visualisation

• configs folder contains the default config .yaml which has most of the hyperparameters and CometML logging details

• in the 'all_model_states' = these are currently in the .gitignore file, but there is now a folder which is renamed with whatever experiment name is in the default config file, so you now have individual folders per experiment, with a config file also named with the same experiment name

• scripts folder not really being used by me

• ct_classifier

• contains most if not all of the training code and dataset structure and information

• the class_mapping.pickle is a list of all the class_names, it is generated in the json_generate_80_20.py file from the preprocesing folder

• dataset.py file: contains the FixedHeightResize class which does the padding on the model

• also performs the other transformations on the images like the augmentations

• testing metrics

• evaluation metrics on the val set

• histogram scores for upsampling, plotting the average precision on the val data

• • plot the confusion matrix
  • plot the species that were high confusion, so any overall score of less than 0.7, we showed a brief sample to see what species were getting confused with each other

• the train.py file contains the definitions for the dataloaders, loading the model (whether starting from model epoch 0 or retraining from an already saved model). You can either load the training data in the .json or upsampling data (in the dl_train dataloader)

  • the training loop is defined as a while loop
  • it defines the statistics you want to measure
  • it also initialises the CometML experiments