Headless Mode Operation - Using ilastik as a command line tool


ilastik has an interactive graphical user interface for training a classifier and refining your results. Once you’re happy with your classifier, you may wish to apply it to several other images using the Batch Processing Applet. Still, if you wish to run ilastik from inside your automation scripts or in environments with no graphical capabilities (e.g.: HPC clusters), then you can use the “headless” mode, which will allow you to run ilastik as a command line tool.

Except for the Carving Workflow, all workflows are able to run in headless mode.

In order to run ilastik in headless mode, you will need to use the graphical user interface to create a project and train a classifier by manually drawing annotations as usual (see, e.g. Pixel Classification Workflow or Object Classification Workflow for instructions on how to train your classifier). Once you’re done with the training, save your project and quit ilastik.

When running ilastik in headless mode, you must specify at least:

  • The --headless flag, to trigger headless operation;
  • Your pre-trained ilastik .ilp project file, via the --project argument (e.g. --project=MyPixelClassProj.ilp);
  • One or more files to be processed as a batch. Those should always be given as the last arguments to the command line.

For example, this command will run your classifier over 2 additional images (here we assume your classifier was trained on 2D images):


$ cd ilastik-1.3.2-Linux/
$ ./run_ilastik.sh --headless --project=MyProject.ilp my_next_image1.png my_next_image2.png


$ ./ilastik-1.3.2-OSX.app/Contents/ilastik-release/run_ilastik.sh  --headless --project=MyProject.ilp my_next_image1.png my_next_image2.png


$ cd "\Program Files\ilastik-1.3.2"
$ .\ilastik.exe --headless --project=MyProject.ilp my_next_image1.png my_next_image2.png

Note: the following examples use linux shell syntax, but the options are the same for all platforms.

Important Notes:

  • Except for options passed with the --export_source argument (see below), don’t specify any values with space-characters in them, even if inside quotes (e.g. "my file name.h5", or "(10, 20)"). Due to a bug in cpython those might be misinterpreted.
  • For paths to hdf5 datasests (either input or output), ilastik uses the same conventions as the generic h5ls utility. That is, the hdf5 dataset name should be appended to the file path: /path/to/my_file.h5/internal/path/to/dataset.
  • In the current “headless” implementation of object classification, the entire image is loaded into RAM in one go, and then object classification is run on it. Therefore, there is a limit to how large your input image can be.
  • ilastik’s command line options use underlines rather than dashes to separate words (e.g.: --cutout_subregion and not --cutout-subregion). If you mix them up, you might get strange errors. See Known Issues for more info.

Controlling RAM and CPU resources

By default, ilastik will use all available CPU cores (as detected by Python’s “multiprocessing” module), including “virtual” cores if your CPU supports hyperthreading (like most modern Intel processors) and all RAM available on your machine. CPU and RAM resources can be controlled with environment variables or a config file as described here.

Input options

Required settings:

  • --headless Invokes headless classification mode.
  • --project The path to your project file, which you have already used to train a classifier.

Optional input settings:

  • --readonly Open the project in read-only mode; necessary for using a single project from multiple processes.
  • --input_axes specify meaning of your axes. In some cases ilastik’s default guess for the axes is not correct. In such cases, supply the axes explicitly (e.g. --input_axes=zcyx for a 3D dataset with channels and the axes in the “zcyx” order). Note that the order should be given in “C-order”, so the fastest varying axes comes last. It should be the same order that you would give in the dataset properties. Furthermore, if stack input is used, the input_axes define axes per slice.

Using hdf5 for memory efficiency

We recommend to use hdf5 for input and output over, for example, TIF stacks. Indeed, TIF stacks cannot be written/read block-wise, hence the whole data has to be loaded and kept in memory. This can lead to excessive memory usage and slowness.

In order to convert your TIF slices into hdf5 datasets, you can check the dedicated Fiji Plugin.

Using stack input

If you are dealing with 3D data in the form of an image sequence (e.g. a tiff stack), then use globstring syntax to tell ilastik which images to combine for each volume. Furthermore, the axis along which should be stacked must be given with the --stack_along command line parameter. You can stack either over the channel, time, or the z-axis. So valid values for this option are c, t, and z, respectively.

$ ls ~/mydata/
my_stack_1.png        my_stack_2.png        my_stack_3.png        my_stack_4.png
my_other_stack_1.png  my_other_stack_2.png  my_other_stack_3.png  my_other_stack_4.png

$ ./run_ilastik.sh --headless \
                   --project=MyProject.ilp \
                   --stack_along="c" \
                   "~/mydata/my_stack_*.png" "~/mydata/my_other_stack_*.png"

Note: The use of quotation marks around file names in the above example are critical. The * in each input argument must be provided to ilastik, NOT auto-expanded by the shell before ilastik sees the command!

Output Options

By default, ilastik will export the results in hdf5 format, stored to the same directory as the input image. However, you can customize the output location and format with extra parameters. For example:

$ ./run_ilastik.sh --headless \
                   --project=MyProject.ilp \
                   --output_format=tiff \
                   --output_filename_format=/tmp/results/{nickname}_results.tiff \
                   my_next_image1.png my_next_image2.png

Here’s a quick summary of each command-line option provided by the headless interface. For the most part, these map directly to the corresponding controls in the Data Export Settings Window. No matter what settings you use, the list of input files to process must come after all other items in the command (as shown in the example above).

Optional output settings:

  • --export_source The data to export, which in general should be an option from the ‘Source’ dropdown in the Data Export Applet for your workflow. See the individual section of this page for the workflow you’re interested in.
  • --output_format The file format to store your results in. Some formats are less flexible than others and therefore cannot be combined with every option here. Choices are: bmp, gif, hdr, jpeg, jpg, pbm, pgm, png, pnm, ppm, ras, tif, tiff, xv, bmp sequence, gif sequence, hdr sequence, jpeg sequence, jpg sequence, pbm sequence, pgm sequence, png sequence, pnm sequence, ppm sequence, ras sequence, tif sequence, tiff sequence, xv sequence, multipage tiff, multipage tiff sequence, hdf5, compressed hdf5, numpy, dvid.
  • --output_filename_format The path to the output file to write. A few “magic” placeholders can be used in these settings. The placeholders are automatically expanded to values specific for the dataset. These are useful when you are exporting multiple datasets:
    • {dataset_dir} - the directory containing the original raw dataset corresponding these export results
    • {nickname} - expands to the raw input file basename. E.g.: if your input file is called myImage.png, {nickname} will expand to myImage
    • {roi} - The region-of-interest as specified in the --cutout_subregion setting.
    • {x_start}, {x_stop}, {y_start}, {y_stop}, etc - Specific axis start/stop boundaries for the region-of-interest
    • {slice_index} - The index of each slice in an exported image sequence (required for all image sequence formats, not allowed with any other format).
    • {result_type} - the selected workflow-specific ilastik result source, as in the drop-down of the Data Export Applet, or as specified by the --export_source command line flag.
  • --output_internal_path (HDF5 only) Specifies the name of the HDF5 dataset to write to. (Default: /exported_data)
  • --cutout_subregion Subregion from the original to imput to operate on. The expected format is [(axis_1_start,axis_2_start,...,axis_n_start),(axis_1_stop,axis_2_stop,...,axis_n_stop)]. The order of the axis is the same order as in your input image. The start values are inclusive and the stop values are exclusive. You can replace any value with None to have it be interpreted as 0 in the start values or the maximum value in that axis for the stop values. If you specify a region that exceeds the shape of your input image, ilastik will crop your selection so that it fits the input shape. E.g.: If your image is a png file with axis x y c, then you could run a headless prediction on a 50x50 square starting from 10x10 and containing all channels by specifying --cutout_subregion="[(10,10,None),(60,60,None)]".
  • --export_dtype The pixel type to convert your results to. Choices are: uint8, uint16, uint32, int8, int16, int32, float32, float64. Note that some formats don’t support every pixel type.
  • --output_axis_order Transpose the storage order of the results. For example, this affects the sliced dimension for stack outputs.
  • --pipeline_result_drange Pipeline result data range (min,max) BEFORE normalization, e.g. "(0.0,1.0)"
  • --export_drange Exported data range (min,max) AFTER normalization, e.g. "(0,255)"

Headless Mode for Pixel Classification

When running the Pixel Classification Workflow in headless mode, the available options for the --export_source flag are:

  • "Probabilities", which exports a multi-channel image where pixel values represent the probability that that pixel belongs to the class represented by that channel;
  • "Simple Segmentation", which produces a single-channel image where the (integer) pixel values indicate the class to which a pixel belongs. For this image, every pixel with the same value should belong to the same class of pixels;
  • Uncertainty, which produces an image where pixel intensity is proportional to the uncertainty found when trying to classify that pixel;
  • Features, which outputs a multi-channel image where each channel represents one of the computed pixel features;
  • Labels, which outputs an image representing the users’ manually created annotations.

Headless Mode for Autocontext Workflow

The Autocontext workflow is essentially two runs of pixel classification, with the latter one using the output of the first one as an additional input. When running the Autocontext Workflow in headless mode, the available options for the --export_source are analogous to those of the simple Pixel Classification Workflow, but you can export the results of any of the two pixel classification stages involved in this workflow. Here’s the full list of options:

  • "Probabilities Stage 1";
  • "Probabilities Stage 2";
  • "Propabilities All Stages": exports both "Probabilities Stage 1" and "Probabilities Stage 2" into a single multi-channel image;
  • "Simple Segmentation Stage 1";
  • "Simple Segmentation Stage 2";
  • "Uncertainty Stage 1";
  • "Uncertainty Stage 2";
  • "Features Stage 1";
  • "Features Stage 2";
  • "Labels Stage 1";
  • "Labels Stage 2";
  • "Input Stage 1": exports your raw input image that was fed into the first stage of the workflow;
  • "Input Stage 2": exports the input received by the second Pixel Classification stage in the workflow.

Headless Mode for Object Classification

When running the Object Classification Workflow in headless mode, the available options for the --export_source flag are:

  • "Object Predictions" (default when nothing is specified), which exports a label image of the object class predictions;
  • "Object Probabilities", which exports a multi-channel image volume of object prediction probabilities instead of a label image (one channel for each prediction class);
  • "Blockwise Object Predictions" or "Blockwise Object Probabilities", which is analogous to "Object Predictions" and "Object Probabilities" respectively, but the image will be processed in independent blocks, which is useful when your image won’t fit in RAM. Note that the values for the block size and halo cannot be configured via command line parameters and must be set in yout .ilp project file via the graphical user interface. See Blockwise Object Classification Applet.
  • "Pixel Probabilities", which exports the pixel prediction images of the pixel classification part of that workflow. Only valid if you specified an .ilp --project that was created with the “Pixel Classification + Object Classification” workflow.

Depending on which variant of the Object Classification Workflow you used to create your .ilp project file, you may need to provide more than one input image for each volume of data you want to process (e.g. “Raw Data” and “Segmentation Image” for , or “Raw Data” and “Prediction Maps”). These input images correspond to the tabs in the Input Data applet of the Object Classification Workflow and must be provided on the command-line. They serve to provide the information necessary to segment the input image into objects which can then be classified by the trained classifier that was saved into your .ilp project file.

To specify which is which, prefix the list of input files with either --raw_data, --segmentation_image, or --prediction_maps accordingly. Here’s an example of invoking ilastik with an Object Classification Workflow that takes a Segmentation Image as input:

$ ./run_ilastik.sh --headless \
                   --project=MyObjClassFromPredictMap.ilp \
                   --table_filename=/tmp/exported_object_features.csv \
                   --export_source="Object Predictions" \
                   --raw_data "my_grayscale_stack_1/*.png" \
                   --segmentation_image my_unclassified_objects_1.h5/binary_segmentation_volume

If you provide a path for the --table_filename output, ilastik will export a .csv file of the computed object features that were used during classification, indexed by object id.

So, the example command above produces 2 files:

  • a prediction image (a label image),
  • a .csv file containing a row for each object and columns for each of the features your project uses. This file also contains the probability value for each object.

If you are processing more than one volume in a single command, provide all inputs of a given type in sequence:

$ ./run_ilastik.sh --headless \
--project=MyObjClassFromSegmentation.ilp \
--raw_data "my_grayscale_stack_1/*.png" "my_grayscale_stack_2/*.png" "my_grayscale_stack_3/*.png" \
--segmentation_image my_unclassified_objects_1.h5/binary_segmentation_volume my_unclassified_objects_2.h5/binary_segmentation_volume my_unclassified_objects_3.h5/binary_segmentation_volume

Headless Mode for Boundary-based Segmentation with Multicut

When running the headless mode for the Multicut Workflow, the following flags define the input data to be used:

  • --raw_data: path to the raw data that should be processed.
  • --probabilities: path to boundary probability map. Note that it should be generated in the same way as the probability map that was used to train the project (e.g. with the [Pixel Classification Workflow]. I.e. same number (and meaning) of channels.

For the --export_source the only available option is "Multicut Segmentation".

An example invocation is given below:

$ ./run_ilastik.sh
--headless \
--project="/path/to/project/MyProject.ilp" \
--raw_data="/path/to/raw/data.h5" \
--probabilities="/path/to/boundary/probability/data.h5" \
--export_source="Multicut Segmentation" \

Headless Mode for Counting Workflow

When running the Object Classification Workflow in headless mode, the only available option for the --export_source is Probabilities.

Additional arguments:

  • --csv-export-file (optional): File path to which a .csv with total object counts should be exported.

Headless Mode for Tracking

Note: There is no headless mode for Manual Tracking.

Depending on the type of [Tracking Workflow], the following flags are available to specify input data:

  • --raw_data: path to the raw data that should be processed.
  • --prediction_maps: path to probability map for the objects you want to track. Note that it should be generated in the same way as the probability map that was used to train the project (e.g. with the [Pixel Classification Workflow]. I.e. same number (and meaning) of channels.
  • --segmentation_image: segmentation image (single channel, integer image) where 0 is interpreted as background and values > 0 are interpreted as objects.

For the --export_source the following values are available:

  • Object-Identities: Integer image with object-IDs of every object as image file
  • Tracking-Result: Tracking results as an integer image where each object is assigned a gray value corresponding to its lineage ID
  • Merger-Result: export only the detections where the optimization decided that it contains more than one object
  • Plugin: Specify Plugin export. Need to further specify --export_plugin which can be any of:
    • CSV-Table: Plugin to export the ilastik tracking results to a CSV table
    • Fiji-MaMuT: Plugin to export the ilastik tracking results to Fiji’s MaMuT
    • Multi-Worm-Tracker: Plugin to export the ilastik tracking results in the Multi-Worm Tracker format (.blobs and .summary files)
    • JSON: JSON export for use of some tracking analysis tools developed by the ilastik tracking guys
    • CellTrackingChallenge: Tracking format used in the ISBI Cell Tracking Challenges
    • H5-Event-Sequence: H5 event sequence export for use of some tracking analysis tools developed by the ilastik tracking guys
    • Contours-With-Head: Plugin to export the ilastik tracking results as contours with the head location (head index)
    • Contours: Plugin to export the ilastik tracking results as contours (used mainly to export larvae contours for Zlatics Lab)

See the following example invocation that produces a csv-table via the plugin export:

    $ ./run_ilastik.sh
    --headless \
    --project="/path/to/project/MyProject.ilp" \
    --raw_data="/path/to/raw/data.h5" \
    --prediction_maps="/path/to/boundary/probability/data.h5" \
    --export_source="Plugin" \
    --export_plugin="CSV-Table" \

Running distributed ilastik via MPI (potentially trough SLURM)

You can run some ilastik headless workflows as a distributed MPI application. This is functionally equivalent to (though far more efficient than) invoking ilastik multiple times, each time with a different --cutout_subregion, and saving all those executions as tiles of a single .n5 dataset.


Not all workflows can be sensibly run in parallel like this; Pixel Classification is a perfect candidate, because each tile can be processed independent of its neighbors. On the other hand, a workflow like Tracking, in which objects of interest often migrate between tiles, would not work at all in this implementation of the distributed operation.

At the moment, only .n5 files can be output from a --distributed invocation; Setting --output_format to anything different than n5 will be ignored.


In order to run ilastik distributed, you will need:

  • Either the mpiexec executable in your PATH or acess to a SLURM installation that is backed by MPI (which is the case for most HPC clusters);
  • the mpi4py python library; Note that if you’re running ilastik in an HPC cluster, you should NOT install mpi4py via conda, since that installation will come with its own precompiled MPI binaries, which will probably not work optimally (if at all) with the MPI installation of your HPC. Instead, install mpi4py via pip, and allow the binaries to be compiled using the MPI headers and C compilers made available in your particular HPC cluster.


When running distributed, you can make use of the following command line options:

  • --distributed Required. This directs ilastik to distribute its workload among its workers. Failing to set this flag will launch independent instances of ilastik in each of your workers, each of which processing the entirety of the input file;
  • --distributed-block-roi Optional. Determines the dimensions of the blocks used to split the input data in distributed mode. Values can be either:
    • An integer, which will be interpreted as if the following dict was passed in: {'x': value, 'y': value, 'z': value, 't': 1, 'c': None}
    • or a literal python Dict[str, Optional[int]], with keys in 'xyztc'. Missing keys will default like so: {'x': 256, 'y': 256, 'z': 256, 't': 1, 'c': None}. Use None anywhere in the dict to mean “the whole dimension”.

    Though optional, it is recommended to set --distributed-block-roi to a sensible value - ideally one that matches the natural tiling of your --raw-data and that can fit in your worker memory.

To run ilastik distributed, you must invoke it either through mpiexec or srun (or sbatch), and pass it the --distributed flag. Following is an exemple of running ilastik via mpiexec using 4 workers and processing blocks of 150px in length, 300px in height and 3 channels. Note the quoting when specifying --distributed-block-roi; to prevent the quotes around the axis names to be removed by the shell, the entirety of the dict is surrounded by single quotes:

$ mpiexec -n 4 ./run_ilastik.sh --headless \
                                --distributed \
                                --distributed-block-roi '{"x": 150, "y": 300, "z": 1, "c": 3}' \
                                --output_filename_format=/tmp/results.n5 \
                                --output_format=n5 \
                                --project=MyPixelClassificationProject.ilp \

Running your own Python scripts

For developers and power-users, you can run your own ilastik-dependent python scripts using the interpreter shipped within the ilastik install tree. The interpreter is located in the bin directory:

# Linux
$ ./ilastik-1.3.2-Linux/bin/python -c "import ilastik; print ilastik.__version__"

# Mac
$ ./ilastik-1.3.2-OSX.app/Contents/ilastik-release/bin/python -c "import ilastik; print ilastik.__version__"

Known Issues

ilastik’s headless mode will sometimes throw exceptions and output a stack trace instead of letting you known why your command line arguments are wrong. Though those issues are being worked on, here are some hints and workarounds you can use to get by:

  • ‘RuntimeError: Could not find one or more input files. See logged errors.’

    If you’re sure that all files passed as arguments to ilastik actually exist, then ilastik might be misinterpreting some arguments as files to be processed rather than command line flags. For example, if you use --output-format png (with a dash between ‘output’ and ‘format’) instead of --output_format png (with an underline), then the words --output-format and png might be interpreted as files to be processed by the workflow, and those probably won’t be found in your machine, which will trigger the error.