The present work is implementation of DROC on FathomNet Dataset for one-class classification of Marine objects. It is based on the work of Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister. Please refer below paper for more information. Learning and Evaluating Representations for Deep One-class Classification published at ICLR 2021 as a conference paper by Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister.
This directory has deep one-class classification which includes self-supervised deep representation learning from one-class data, and a classifier using discriminative model OC-SVM. It also contains Baseline model based on state of the art OC-SVM.
The requirements.txt includes all the dependencies for this project.
Dataset/ includes an instruction how to download and prepare data for
FathomNet dataset. The dataset needs to be downloaded using the scripts prepare_data001, prepare_data002,... given in Dataset folder.
Please download dataset prior to model training.
The options for the experiments are specified through the command line arguments.
The detailed explanation can be found in train_and_eval_loop.py. Scripts for
running experiments can be found
-
Contrastive learning with distribution augmentation in home directory:
run_contrastive_da.sh(paths should be set prior in the code) -
To run baseline:
run_Baseline.sh(paths should be set prior in the code) -
Other details for code are in source folder (Network architecture, util, etc.)
After running train_and_eval_loop.py using run_contrastive_da.sh, the evaluation results can be found in
$Output/.../stats/summary.json, where ... is specified as classwise folder.
-
For model prediction:
Model_presiction.sh(paths should be updated in the file prior) -
Model can be found in saved_model folder in home directory