Repository files navigation

Usesnetharn (https://gitlab.kitware.com/computer-vision/netharn) to writethe boilerplate for training loops.

Scripts takekwcoco datasets as inputs. Seehttps://gitlab.kitware.com/computer-vision/kwcoco for how to format in theextended-COCO format (regular MS-COCO files will also work).

To train a detection model seebioharn/detect_fit.py.

To train a classification model seebioharn/clf_fit.py.

To predict with a pretrained detection model seebioharn/detect_predict.py.

To predict with a pretrained classification model seebioharn/clf_predict.py.

To evaluate ROC / PR-curves with a pretrained detection model and truth seebioharn/detect_eval.py.

To evaluate ROC / PR-curves with a pretrained classification model and truth seebioharn/clf_eval.py.

Current supported detection models include

• YOLO-v2
• EfficientDet
• MaskRCNN - Requires mmdet
• CascadeRCNN - Requires mmdet
• RetinaNet - Requires mmdet

Older versions of bioharn were previously targeting mmdet 1.0 revision4c94f10d0ebb566701fb5319f5da6808df0ebf6a but we are now targeting v2.0.

This repo is a component of the VIAME project:https://github.com/VIAME/VIAME

some of the data for this project can be found here

https://data.kitware.com/#collection/58b747ec8d777f0aef5d0f6a

Notes for mmcv install on cuda 10.1 with torch 1.5:

See:https://github.com/open-mmlab/mmcv

pipinstallmmcv-full==latest+torch1.5.0+cu101-fhttps://openmmlab.oss-accelerate.aliyuncs.com/mmcv/dist/index.html

To train a simple detector let use the kwcoco toy data to make sure we can fitto a small dummy dataset. Lets use the kwcoco CLI to generate toy training. Forthis test we will forgo a validation set.

kwcoco toydata --key=shapes1024 --dst=toydata.kwcoco.json

For our notable hyperparameters we are going to use:

• --optim=adam - we will use the ADAM optimizer for faster convergence (may also want to trysgd).

• --lr=1e-4 - we will start with a small learning rate

• --decay=1e-4 - we will use weight decay regularization of1e-4 to encourage smaller network weights.

• --window_dims=full - which means that each batch item will sample a full image.

• --input_dims=512,512 - which means we are going to resize each image to H=512, W=512 (using letterboxing to preserve aspect ratio) before inputting the item to the network.

• --schedule=step-16-22 - will divide the learning rate by 10 at epoch 16 and 22.

• --augment=medium - will do random flips, crops, and color jitter for augmentation (--augment=complex will do much more and--augment=simple will only do flips and crops).

• --num_batches=auto - determines the number of batches per epoch. If auto it will use the entire dataset. If you set it to a number if will use that many batches per epoch with random sampling with replacement. This is useful if you are going to use over/undersampling via the--balance CLI arg.

--batch_size=8 will use 8 items (sampled windows) per batch.

--bstep=8 will run 8 batches before backpropagating (approximates a larger batch size)

Seepython -m bioharn.detect_fit --help for help on all available options.

python -m bioharn.detect_fit \    --name=det_baseline_toydata \    --workdir=$HOME/work/bioharn \    --train_dataset=./toydata.kwcoco.json \    --arch=retinanet \    --optim=adam \    --lr=1e-4 \    --schedule=step-16-22 \    --augment=medium \    --input_dims=512,512 \    --window_dims=full \    --window_overlap=0.0 \    --normalize_inputs=imagenet \    --num_batches=auto \    --workers=4 --xpu=auto --batch_size=8 --bstep=8 \    --sampler_backend=cog

This should start producing reasonable training-set bounding boxes after a fewminutes of training.

Because we are using netharn, training this detection model will write a"training directory" in your work directory. This directory will be a functionof your "name" and the hash of the learning-relevant hyperparameters.

In this case the training directory will be in:$HOME/work/bioharn/fit/runs/det_baseline_toydata/qxvodtak and for convenience there willbe a symlink to this directory in$HOME/work/bioharn/fit/name/det_baseline_toydata. Example training (andvalidation if specified) images will be written to themonitor directory. Iftensorboard and matplotlib are installed themonitor directory will alsocontain atensorboard subdirectory with loss curves as they are produced.