autocontext

Contains a python script to apply the autocontext method to an ilastik project. Currently, this only works on hdf5, tiff and bmp data.

Example usage (training)

Before using the script, you must create an ilastik project file:

Create an ilastik pixel classification project and add one or more datasets.
Select some features.
Add some labels.
Save the project and exit ilastik.

Now you can use the autocontext script. Open a terminal and run a command like the following:

python autocontext.py --help
python autocontext.py --train myproject.ilp --ilastik /usr/local/ilastik/run_ilastik.sh
python autocontext.py --train myproject.ilp --ilastik /usr/local/ilastik/run_ilastik.sh --cache training/cache
python autocontext.py --train infile.ilp -o outfile.ilp --ilastik /usr/local/ilastik/run_ilastik.sh

Example usage (batch prediction)

You can use the autocontext in combination with the ilastik batch prediction. Lets say you want to use batch prediction on the files to_predict0.h5/raw and to_predict1.h5/raw. First, you have to train the autocontext (see above). The trained autocontext is saved in the cache folder, say this is the folder training/cache. Now you can call the batch prediction:

python autocontext.py --batch_predict training/cache --ilastik /usr/local/ilastik/run_ilastik.sh --cache prediction/cache --files to_predict0.h5/raw to_predict1.h5/raw

Please keep in mind, that you need a cache folder for the batch prediction, too. It may be a good idea to use different cache folders for training and batch prediction.

Forwarding arguments to ilastik

All command line arguments that are not used by autocontext are forwarded to ilastik. See [http://ilastik.org/documentation/pixelclassification/headless.html] (http://ilastik.org/documentation/pixelclassification/headless.html) for a full list of ilastik options. Example:

python autocontext.py --batch_predict training/cache --ilastik /usr/local/ilastik/run_ilastik.sh --cache prediction/cache --files to_predict0.h5/raw to_predict1.h5/raw --output_filename_format {nickname}_Probabilities.h5 --output_internal_path my_personal_export_key

There are a few exceptions:

The options --headless and --project are ignored, since they are predefined by the autocontext.
Since you only need the ilastik results from the last autocontext iteration, the options --output_format, --output_filename_format, --output_internal_path are only taken into account in the last iteration.

Prevent OSError in autocontext iteration

If possible, replace your ilastik.py by autocontxt/ilastik_mods/ilastik-1.1.X/ilastik.py and start autocontext with the --predict_file flag. This prevents the OSError "Argument list too long" in the prediction step of the autocontext iteration.

Dependencies

Python packages (all included in the python installation from ilastik):

vigra
numpy
colorama
h5py

Other:

ilastik

TODO

Support ilastik projects with datasets other than hdf5, tiff or bmp.