This repo contains a condensed version of the code used in the paper Confidence-Driven Hierarchical Classification of Cultivated Plant Stresses by Logan Frank, Christopher Wiegman, Dr. Jim Davis, and Dr. Scott Shearer.

The contents in this repo are organized as follows:

• base_classifiers contains all code pertaining to the base classifiers used in this paper, with network architectures defined in the architectures subdirectory.
• davis contains all code used to reimplement the paper Hierarchical Semantic Labeling with Adaptive Confidence.
• hedging contains all code used to reimplement the paper Hedging Your Bets: Optimizing Accuracy-Specificity Trade-offs in Large Scale Visual Recognition
• trees contains example hierarchy txt files
• utils contains some utility functions and classes

To use our code begin with installing the necessary requirements, assuming you already have some environment with Python 3.7 and pip: pip install -r requirements.txt

You may adapt and use the code provided for training your own base classifier (run everything from the top-level of this repo):

python base_classifiers/train.py \
    --image_dir <path-to-your-dataset-according-to-pytorch-image-folder-class> \
    --network_dir <wherever-you-want-to-save-your-network-related-data> \
    --results_dir <wherever-you-want-to-save-your-results-from-training> \
    --dataset <name-of-your-dataset> \
    --name <whatever-you-want-to-name-your-experiment> \
    --network <whichever-network-you-want-to-use> \
    --batch_size <some-int> \
    --learning_rate <some-float> \
    --num_epochs <some-int> \
    --balance_dataset <whether-or-not-you-want-to-balance-the-training-dataset>

Currently, our code only supports the ResNet-18 and small CNN networks used in our paper. More can easily be added. If you choose to balance the training dataset (using instance replication), it is strongly recommended you choose a batch size that is a multiple of your number of training classes so there is the same number of examples from each class in every batch.

Once a base classifier has been trained, the approach we applied to our datasets can be ran with:

python davis/pipeline.py \
    --image_dir <path-to-your-dataset-according-to-pytorch-image-folder-class> \
    --network_path <path-to-your-base-classifier-pth-file> \
    --tree_path <path-to-your-hierarchy-txt-file> \
    --results_dir <wherever-you-want-to-save-your-results-from-training> \
    --dataset <name-of-your-dataset> \
    --name <whatever-you-want-to-name-your-experiment> \
    --network <whichever-network-you-want-to-use-as-your-base-classifier> \
    --nbins <number-of-desired-histogram-bins-for-calibration> \
    --priors <what-type-of-priors> \
    --confidence <some-float-for-confidence-threshold>

By default, we run with nbins == 10, priors == 'equal' ('data' or 'manual' can also be used), and confidence == 0.5 (confidence should be [0, 1]).