Adapter.reshape()

[jerrymhuang]

Related to #508 .

Example usage:

sim = {"x": np.arange(6)}

adapter = (
    bf.adapters.Adapter()
    # ...
    .reshape("x", to=(3, 2))
    # ...
)

For this to be stateless, the original shape of data should not be stored. Therefore, the forward and inverse operations are identical.

[codecov]

Codecov Report

❌ Patch coverage is 48.00000% with 13 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
bayesflow/adapters/transforms/reshape.py	58.82%	7 Missing ⚠️
bayesflow/adapters/adapter.py	14.28%	6 Missing ⚠️

Files with missing lines	Coverage Δ
bayesflow/adapters/transforms/__init__.py	100.00% <100.00%> (ø)
bayesflow/adapters/adapter.py	83.87% <14.28%> (-2.03%)	⬇️
bayesflow/adapters/transforms/reshape.py	58.82% <58.82%> (ø)

[stefanradev93]

Thanks Jerry! The current operation can potentially break an inference pipeline. One alternative compatible with varying shape sizes) is to store and update the shape of the most recent array after each forward transform. This renders the adapter "semi-stateful", since it needs to be called at least once in a forward manner to be able to go in reverse, but this is a given for both the training and the inference phases.

[vpratz]

I think this uv.lock file slipped into the PR by accident.

[vpratz]

I agree with Stefan that storing a state is fine here.
In my opinion, the functionality should also be able to handle batches of data of varying size. I'm not sure what the best interface for this would be. Maybe we can in/exclude axis and only reshape over the specified dimensions, keeping the others fixed? So for example to reshape over the last dimension, we could do .reshape("key", (2,4), include_axis=[-1]) to reshape a batch of shape (batch_size, set_size, 8) into a shape of (batch_size, set_size, 2, 4).