.save_to_folder() fails for the InjectDriftingTemplatesRecording

[cwindolf]

When trying to save a hybrid recording, I get the error TypeError: Object of type DriftingTemplates is not JSON serializable. This makes sense, because DriftingTemplates inherits from Templates which is a dataclass and not a BaseExtractor, so it will not be .to_dict()ed by https://github.com/SpikeInterface/spikeinterface/blob/main/src/spikeinterface/core/base.py#L423 even though it has a .to_dict() method.

I'm not sure how to resolve this? One could maybe make Templates subclass BaseExtractor while still being a dataclass? Although it does feel a bit strange to json-serialize a large array of templates. Anyway, wanted to see what you all think, thanks!

[alejoe91]

Thanks! One simple option is to make the InjectDriftingTemplatesRecording not JSON serializable, so it will be saved to pkl at all instances

[cwindolf]

Oh! That sounds good. Happy to make a PR for that -- is it done by implementing check_if_json_serializable(self)? Or would Templates need to inherit from BaseExtractor?

[alejoe91]

@cwindolf you can just add this at the end of the init:

self._serializability["json"] = False

[h-mayorquin]

Mmm the Templates should be json serializable, it has a dict:

spikeinterface/src/spikeinterface/core/template.py

Lines 250 to 261 in ed9d1b5

	def to_dict(self):
	return {
	"templates_array": self.templates_array,
	"sparsity_mask": None if self.sparsity_mask is None else self.sparsity_mask,
	"channel_ids": self.channel_ids,
	"unit_ids": self.unit_ids,
	"sampling_frequency": self.sampling_frequency,
	"nbefore": self.nbefore,
	"is_scaled": self.is_scaled,
	"probe": self.probe.to_dict() if self.probe is not None else None,
	}

Could you share a minimal example where this fails?

[cwindolf]

@h-mayorquin it does have a to_dict() method, but it doesn't inherit from BaseExtractor, so it won't go through the usual code path (see the line I link to in the top post). Sorry, my example is a bit of a pain to share, but I think any example should work. Note that one needs to make sure the sorting used is a dumpable one, else provenance.json just has an error message and you won't get to the point of trying to serialize the templates.

I think for the same reason @alejoe91 just setting self._serializability["json"] is not enough, because Templates is not a BaseExtractor so it won't know how to reconstruct itself -- the code path that's hit when I then later try to load up the full hybrid recording object via sc.load_extractor(my_hybrid_recording_dir / "provenance.json") then goes throughBaseExtractor.load() stuff. I think Templates needs to subclass BaseExtractor, but I don't know enough about how these things work yet...

[alejoe91]

@h-mayorquin maybe we just need to extend the SIJsonEncoder to support Templates?

[alejoe91]

@cwindolf no we don't need to inherit from BaseExtractor, because that's for a Recording or Sorting object.

@h-mayorquin let's discuss about this tomorrow!

[h-mayorquin]

OK. Let's discuss it tomorrow.

[alejoe91]

@cwindolf can you test this one #3130? With this PR the provenance is saved to pkl if the recording is not JSON serializable (and this is the case for then InjectDriftingTemplatesRecording)

[cwindolf]

Thanks @alejoe91 ! This is great. rec.check_serializability(type='json') fails now and there is a provenance.pkl saved. I can access the .drifting_templates.templates_array_moved property and everything. I did this on a .frame_slice() for testing like so:

rec.frame_slice(0, 1000).save_to_folder(scratch_dir)
rec1 = si.load_extractor(scratch_dir / "provenance.pkl")
rec1._kwargs["parent_recording"].drifting_templates.templates_array_moved.shape

[alejoe91]

@cwindolf you can also do this! (better API)

rec1.get_parent().drifting_templates.templates_array_moved.shape

[cwindolf]

Thanks, good to know!