Deepinterpolation revived: training, transfer, and inference with more flexibility

[alejoe91]

Refactored and extended the deepinterpolation submodule.

The existing code was only compatible with a pre-trained model from Neuropixels1,0 with interleaved zeros (so that the final shape was 384,2).

Training

With this PR, one can train a DI model from ANY SpikeInterface object using the train_deepinterpolation function. The desired_shape parameter controls how the input should be reshaped to make the input image.

For example, a NP1 probe could be treated as follows:

• desired_shape=(192,2): the 4 staggered columns are "aligned" into two columns of 192 electrodes
• desired_shape=(96, 4): each "real" column is a different image column (in this shape the channels would need to be reordered first).

To use the previously built model (384, 2) , one can use the zero-channel pad to "manually" interleave zeros in the original recording.

Transfer

The existing_model_path allows one to do transfer learning from an existing model

fine_tuned_model = train_deepinterpolation(recording, ..., existing_model_path="path-to-existing")

Inference

Same as before, but the derived shape is inferred from the network:

recording_di = spre.deepinterpolate(recording, model_path="path-to-model", **kwargs)

NOTE: it is advised to zscore the recording prior to using DeepInterpolation, but this is not currently enforced.

[alejoe91]

@khl02007 FYI ;)

[alejoe91]

@jeromelecoq

I updated the train_deepinteprolation function as wollows:

• You can now pass a single recording or a list of recordings as input. In case of a recording with multiple segments (e.g., several play/pause) or multiple recordings, I modified the _calculate_list_samples so that intervals at borders are exluded.
• You can optionally pass test_recordings (a single recording or a recording list) which can be a different recording used for the set
• Unified the correct usage of total_samples (and removed overridden functions)
• Set mean_squared_error as default loss
• Extended the tests

Would you have time to review it? It's ready to merge on my side :)

[jeromelecoq]

I think the SequentialGenerator might have some of this logic already?

[alejoe91]

to exclude borders? I couldn't find it!

[jeromelecoq]

I wonder if this is needed given that you are inheriting from the SequentialGenerator

[alejoe91]

in this case it's needed because of the borders exclusion

[jeromelecoq]

To what extent using seconds is helpful here? I am always worried about edge conditions created by wrapping code.

[alejoe91]

I think that using second here is much more convenient. E.g. I'd rather use 600 to indicate 10 minutes, than 18000000 (easy to forget, add a 0 here)

[jeromelecoq]

I would consider only exposing the parameters you care about, not all of them.

[jeromelecoq]

For example, which parameters will in practice never change?

[alejoe91]

you tell me :) In my opinion there's no harm in exposing everything with defaults!

[alejoe91]

@samuelgarcia we can merge this

[samuelgarcia]

Enormous and amazing work camarade.
I hope this will be usefull and usable.
As a general comments I think I prefer a unique file for this kind preprocessing even if the file is enormous.
I understand that it was more easy while develpping to have multiple and sub folder file.

[alejoe91]

Enormous and amazing work camarade. I hope this will be usefull and usable. As a general comments I think I prefer a unique file for this kind preprocessing even if the file is enormous. I understand that it was more easy while develpping to have multiple and sub folder file.

Thanks! The multiple files are useful because I can do lazy imports without conditional definitions ;)