Embedding niftynet into a custom python code (a.k.a. module API)

[somsol]

I want to have a python script in which I apply some preprocessing steps to my data and then feed it to a neural network of choice. How can I do this without the need to write the processed data into disk prior to calling niftynet.main()? Is there a way to input the data that is already in memory into a neural network? Likewise, is there a way to get the output of the network without saving it to disk (for postprocessing purposes)?

Any help would be most appreciated.

[wyli]

duplicate of #99?

[tvercaut]

Also of relevance: https://cmiclab.cs.ucl.ac.uk/CMIC/NiftyNet/issues/240

[tvercaut]

For ease of reference, I am copy-pasting the content of https://cmiclab.cs.ucl.ac.uk/CMIC/NiftyNet/issues/240 here.

Using NiftyNet as a module

It is currently not trivial to run NiftyNet as part of a separate python script. It would be nice to have an alternative to run_application() that let you do the following:

nn = ApplicationDriver()
nn.initialise_application(system_param, input_data_param)
my_output_data_dict = nn.process_image(my_input_data_dict)

This would enable NN to be used more easily in complex pipelines, and integrated more easily into interactive systems.

This would require a few modifications:

1. a process_image function that takes in a dictionary of data (matching what you would get from a reader?), processes it as if it had been read from a file, and returns the results as would be written to a file.
2. to facilitate this, the reading/writing to files would need to be decoupled from the preprocessing and the sampler systems.

[somsol]

Does it mean that you're still working on this? Please let me know when you get there. Thanks

[ericspod]

Work is still ongoing with the pull request associated with this ticket.

[bor2093]

Hi guys,
I think this feature is really great. I wonder if there is any progress in merging to dev branch...

[srivathsapv]

Any updates on this feature?