GraXpert Interface
Truco
The first GraXpert interface was written in C and simply called the GraXpert executable. However for various reasons this approach proved unreliable for a significant minority of users and with Siril 1.4.0-beta2 an updated interface has been released as a Python script. This directly interfaces with the GraXpert model files and provides much better performance and integration, and since it builds on the way the Python interface uses the native PyPI modules library to install the correct versions of dependecies for each user's machine it is expected to be more reliable and better able to make use of different types of GPU.
Owing to the way GraXpert distributes its AI models from an Amazon S3 instance, a copy of the GraXpert program needs to be available in order to download model files. The model files you can download are dependent on the version of GraXpert installed therefore it is recommended to install 3.1.0-RC2 which provides access to the stellar and background object deconvolution models as well as the denoising and background extraction models that have been available since version 3.0.0.
The old GraXpert interface was marked as deprecated in 1.4.0-beta2 and has since been removed from the code base. This documentation covers the new Python-based interface.
Siril provides an interface to the AI models provided by the third party GraXpert tool. This is a third party piece of Free and Open Source software that provides sophisticated background removal and denoising routines. There is a lengthy history of collaboration between Siril and GraXpert, and the RBF algorithm used in Siril's core background removal function was contributed by the GraXpert developers several years ago.
In order to be used, the path to the working GraXpert installation must be set in .
Installation and Configuration
The GraXpert-AI.py script can be installed in the usual way, by going to the burger menu -> Get Scripts and selecting the check box next to the script in the list.
To configure the script you need to tell Siril where your GraXpert installation is. To do this, go to Preferences -> Miscellaneous and select the GraXpert executable using the appropriate file chooser widget.
General Aspects of the GUI
The GUI is shown below:
GraXpert script interface, showing denoising options.
The title bar changes to show the currently selected operation, and at the top of the window are the script title and version number. A dropdown allows you to select the operation you wish to perform. How this dropdown is populated depends on the version of GraXpert you have available. It will always feature Denoising and Background Extraction, even if no locla models are available (see the section on Model Manager for information on how to download models). If you have locally available deconvolution models, or if you have set the location of a GraXpert version that supports deconvolution, you will also see Deconvolution (Stars) and Deconvolution (Objects), however if these models are not available locally and you don't have a version of GraXpert that is capable of downloading them they will not be shown.
Next follows an area for operation-specific parameters. If you select Denoising you will find a widget to control strength, if you select Background Extraction you will find widgets for smoothing and the type of background extraction to perform, and so on. These are described in more detail below.
The Advanced setting provides a means of controlling the inferencing batch size and whether or not to use GPU acceleration. It is generally a good idea to leave the batch size at the default: it sets how many chunks of the image can be processed in parallel, and increasing it may result in out-of-memory errors.
Truco
The GraXpert models use ONNX to perform inferencing. This is a hardware- agnostic Machine Learning framework that provides various runtimes targeting different types of acceleration hardware. There is a specific NVidia runtime, a DirectML runtime that provides generic Windows GPU acceleration, a MacOS runtime and Linux runtimes for NVidia, AMD and Intel GPUs. The script will automatically try to install the most suitable runtime for your hardware, however in the event that it does not work there is always a CPU runtime to revert to - simply disable the GPU acceleration checkbox and please submit a bug report as this is new functionality and we are especially keen to hear of any issues relating to the ONNX runtime selection.
Finally, the UI provides Apply and Close buttons that perform the expected functions and a Model Manager button that opens a tool to discover and download versions of the AI models.
When opening the GUI, it will detect whether a single image or a sequence is loaded. If a single image is loaded it will process it and update the result in Siril, whereas if a sequence is loaded it will process all the selected frames of the sequence and create a new sequence with a prefix dependent on the operation being performed.
Advertencia
In Siril 1.4, Python scripting cannot subscribe to Siril event notifications so it is not possible for the script to detect if the image has been closed or if a change between a single image and a sequence occurs after opening the tool. Therefore if you do this, you should close and reopen the script GUI.
Background Extraction
The GraXpert background extraction AI model offers two parameters:
Smoothing - this determines the smoothness of the background model that is generated.
Correction type - this determines the type of correction that is applied: you can choose from subtraction or division.
Note that the GraXpert script does not provide an interface to the classical background removal algorithms provided by GraXpert. RBF is already natively available in Siril, and Spline and Kriging do not generally provide better performance than RBF.
Deconvolution
From version 3.1 onwards, GraXpert supports deconvolution. With the models downloaded, Siril supports both modes provided by GraXpert:
Object deconvolution
Stellar deconvolution
These modes can be selected using the Operation dropdown.
Both deconvolution modes have two parameters, Strength, which takes values between 0.0 (no deconvolution) and 1.0 (maximum deconvolution) and Image FWHM (in pixels), which also takes values between 0.0 (prioritises fine structure retention) and 10.0 (prioritise deconvolution ability). If you set it to zero you can see a very (too) small effect on processed image, if so set it higher.
Denoising
To use the GraXpert AI denoising model, select Denoising in the Operations dropdown. This operation has only one parameter, Strength, which takes values between 0.0 (no denoising) and 1.0 (maximum denoising).
Denoising can take a little time, especially for large images and / or if using a CPU for inferencing. However once the denoising is done the denoised image is cached, so experimenting with different strengths is extremly quick.
Model Manager
The script provides a model manager to discover and download model versions. Whereas inferencing is done directly in the script, model download requires use of the actual GraXpert program because the models are stored in an Amazon S3 instance and GraXpert stores its credentials internally. The model manager will show radio buttons for each model that the installed version of GraXpert is capable of downloading. For v3.0.x this means denoising and background extraction; for v3.1.0-rcx this also adds stellar and object deconvolution.
The list of remotely available models is shown. If you want to discover models for a different AI operation, change the radio button to the operation you want a model for and click Refresh. A model can be downloaded by selecting it in the list and clicking Download selected model. Note that the download may take some time in the case of large models or slow network connections and GraXpert provides little feedback, so be patient. The models are stored in the GraXpert user data directory.
Advertencia
Do not move the GraXpert models from the GraXpert data directory structure where they are saved by the model manager. This is where the script expects to find them and it will ignore models put in any other location.
Once a model is downloaded it will automatically be populated in the model version dropdown in the main UI when the relevant AI operation is selected.
GraXpert script interface, showing the model manager.
Scripting
The GraXpert-AI.py script provides a command-line interface for use in scripting. This uses the pyscript command as shown below.
Siril command line
pyscript scriptname.py [script_argv]
Operation Parameters
The following parameters are used to specify the operation to perform:
-denoise specifies denoising
-deconv_obj specifies object deconvolution
-deconv_stellar specifies stellar deconvolution
-bge specifies background extraction
General Parameters
The following parameters are recognised in all modes:
-h / --help prints help to the Siril console and exits
-model specifies the model version to use (e.g. -model="3.0.2") If this argument is omitted it defaults to the latest available version.
-batch specifies the AI batch size
-gpu specifies to use the GPU (this is the default)
-nogpu speficies to use CPU inferencing only
-listmodels causes the script to list the available model versions for the specified operation and exit
Truco
Note that this script does not provide a -seq parameter as it expects to work on a sequence already loaded in Siril. When scripting, this can be achieved using the load_seq command first to load the desired sequence.
Background Extraction Parameters
-smoothing= This option sets the smoothing parameter, between 0.0-1.0
-correction= This option sets the correction type and may be either "subtraction" or "division"
-keep-bg This option specifies that the modelled background should be saved as well. For a filename such as myfile.fit, the background will be saved as myfile_bg.fit
Denoising Parameter
-strength= This option sets the strength parameter, between 0.0-1.0
Deconvolution Parameters
The same parameters are available for both stellar and object deconvolution:
-strength This option sets the deconvolution strength, between 0.0-1.0
-psfsize This option sets the PSF size, as described above
Examples
pyscript GraXpert_AI.py -denoise -model='3.0.1'
Or, if calling it from another Python script:
siril.cmd("pyscript", "GraXpert-AI.py", "-denoise", "-model='3.0.1'")