training.rst

Training

To train a new FCN, use the command-line interface (CLI) application bob binseg train, available on your prompt. To use this CLI, you must define the input dataset that will be used to train the FCN, as well as the type of model that will be trained. You may issue bob binseg train --help for a help message containing more detailed instructions.

To replicate our results, use our main application bob binseg train followed by the model configuration, and dataset configuration files, and/or command-line options. Use bob binseg train --help for more information.

Tip

We strongly advice training with a GPU (using --device="cuda:0"). Depending on the available GPU memory you might have to adjust your batch size (--batch).

Baseline Benchmarks

The following table describes recommended batch sizes for 24Gb of RAM GPU card, for supervised training of baselines. Use it like this:

# change <model> and <dataset> by one of items bellow
$ bob binseg train -vv <model> <dataset> --batch-size=<see-table> --device="cuda:0"
# check results in the "results" folder

Models / Datasets	:py:mod:`drive <bob.ip.binseg.configs.datasets.drive>`	:py:mod:`stare <bob.ip.binseg.configs.datasets.stare>`	:py:mod:`chasedb1 <bob.ip.binseg.configs.datasets.chasedb1>`	:py:mod:`iostar-vessel <bob.ip.binseg.configs.datasets.iostarvessel>`	:py:mod:`hrf <bob.ip.binseg.configs.datasets.hrf1168>`
:py:mod:`unet <bob.ip.binseg.configs.models.unet>`	4	2	2	2	1
:py:mod:`hed <bob.ip.binseg.configs.models.hed>`	8	4	4	4	1
:py:mod:`driu <bob.ip.binseg.configs.models.driu>` / :py:mod:`driu-bn <bob.ip.binseg.configs.models.driubn>`	8	5	4	4	1
:py:mod:`m2unet <bob.ip.binseg.configs.models.m2unet>`	16	6	6	6	1

Tip

Instead of the default configurations, you can pass the full path of your customized dataset and model files. You may :ref:`copy any of the existing configuration resources <bob.ip.binseg.cli.config.copy>` and change them locally. Once you're happy, you may use the newly created files directly on your training command line. For example, suppose you wanted to slightly change the drive pre-processing pipeline. You could do the following:

$ bob binseg config copy drive my_drive_remix.py
# edit my_drive_remix.py to your needs
$ bob binseg train -vv <model> ./my_drive_remix.py --batch-size=<see-table> --device="cuda:0"

Tip

If you are at Idiap, you may install the package gridtk (conda install gridtk) on your environment, and submit the job like this:

$ jman submit --queue=gpu --memory=24G --name=m2unet-drive -- bob binseg train --device='cuda:0' ... #paste the rest of the command-line

Combined Vessel Dataset (COVD)

The following table describes recommended batch sizes for 24Gb of RAM GPU card, for supervised training of COVD- systems. Use it like this:

# change <model> and <dataset> by one of items bellow
$ bob binseg train -vv <model> <dataset> --batch-size=<see-table> --device="cuda:0"

Models / Datasets	:py:mod:`covd-drive <bob.ip.binseg.configs.datasets.starechasedb1iostarhrf544>`	:py:mod:`covd-stare <bob.ip.binseg.configs.datasets.drivechasedb1iostarhrf608>`	:py:mod:`covd-chasedb1 <bob.ip.binseg.configs.datasets.drivestareiostarhrf960>`	:py:mod:`covd-iostar-vessel <bob.ip.binseg.configs.datasets.drivestarechasedb1hrf1024>`	:py:mod:`covd-hrf <bob.ip.binseg.configs.datasets.drivestarechasedb1iostar1168>`
:py:mod:`driu <bob.ip.binseg.configs.models.driu>` / :py:mod:`driu-bn <bob.ip.binseg.configs.models.driubn>`	4	4	2	2	2
:py:mod:`m2unet <bob.ip.binseg.configs.models.m2unet>`	8	4	4	4	4

Combined Vessel Dataset (COVD) and Semi-Supervised Learning (SSL)

The following table describes recommended batch sizes for 24Gb of RAM GPU card, for semi-supervised learning of COVD- systems. Use it like this:

# change <model> and <dataset> by one of items bellow
$ bob binseg train -vv --ssl <model> <dataset> --batch-size=<see-table> --device="cuda:0"

Models / Datasets	:py:mod:`covd-drive-ssl <bob.ip.binseg.configs.datasets.starechasedb1iostarhrf544ssldrive>`	:py:mod:`covd-stare-ssl <bob.ip.binseg.configs.datasets.drivechasedb1iostarhrf608sslstare>`	:py:mod:`covd-chasedb1-ssl <bob.ip.binseg.configs.datasets.drivestareiostarhrf960sslchase>`	:py:mod:`covd-iostar-vessel-ssl <bob.ip.binseg.configs.datasets.drivestarechasedb1hrf1024ssliostar>`	:py:mod:`covd-hrf-ssl <bob.ip.binseg.configs.datasets.drivestarechasedb1iostar1168sslhrf>`
:py:mod:`driu-ssl <bob.ip.binseg.configs.models.driussl>` / :py:mod:`driu-bn-ssl <bob.ip.binseg.configs.models.driubnssl>`	4	4	2	1	1
:py:mod:`m2unet-ssl <bob.ip.binseg.configs.models.m2unetssl>`	4	4	2	2	2

Using your own dataset

To use your own dataset, we recommend you read our instructions at :py:mod:`bob.ip.binseg.configs.datasets.csv`, and setup a CSV file describing input data and ground-truth (segmentation maps). Then, prepare a configuration file by copying our configuration example and edit it to apply the required transforms to your input data. Once you are happy with the result, use it in place of one of our datasets:

$ bob binseg config copy csv-dataset-example mydataset.py
# edit mydataset following instructions
$ bob binseg train ... mydataset.py ...