README.md

MMDetection scripts

For MMDetection instructions see the MMDetection README.

Because MMDetection requires importing the pipeline class defined numpy_loader.py to load inputs, all MMDetection scripts are within a single directory. Similarly, common_vars.py defines common config parameters.

All other scripts are adapted from the tools/ scripts provided in the MMDetection repository.

Structure

├── README.md                       <- This file
├── analyze_results.py              <- Plots loss/mAP curves
├── browse_dataset.py               <- Visualise annotations and dataset
├── combine_evaluation_scores.py    <- Calculate mean/std of evaluation scores
├── common_vars.py                  <- Common configurations
├── numpy_loader.py                 <- Pipeline class to load NumPy inputs
├── slurm_bulk_test.sh              <- Submit multiple evaluation jobs
├── slurm_submit.sh                 <- Submit a single training job
├── submit_all_seeds.sh             <- Run multiple trainings with different seeds
├── test.py                         <- Evaluate a trained model
└── train.py                        <- Train a model

Visualisation

For plotting the training dataset, I've set up a copy of the config used for MaskRCNN R50 training but with augmentations removed. You can plot it with something like

python browse_dataset.py --output-dir /path/to.output/ --channel RGB --not-show ~/Wahn/configs/mmdet/common/plotting_config.py

Training

TBD

Testing

TBD

Analyzing results

You should set TOPDIR to the top-level work directory of the training, then run

python analyze_results.py $TOPDIR/mask_*.py $TOPDIR/evaluation/eval_results.pkl $TOPDIR/evaluation/images_0.3/ --show-score-thr 0.3 --topk 50