Extend Python API with a new `model.reshape()` overload

[p-wysocki]

Context

Currently there are multiple ways to reshape model inputs in Python API. They are defined as bindings to C++ code.

A new way to reshape has been suggested - implicit input reshaping with a list of input shapes. An example of new, desired overload/Python API function:

1. model has three inputs: A, B and C
2. model is being reshaped with a list of shapes:
   model.reshape([[2,2], [1, 3, 224, 244], [10]])
3. the given shapes are assigned to respective inputs:
   A.shape = [2, 2]
B.shape = [1, 3, 244, 244]
C.shape = [10]

Python API model binding needs to be extended with a new overload which will add this functionality.

What needs to be done?

1. Add a new overload for model.reshape in model C++ bindings
2. Add tests for the new functionality in the correct test file - test_model.py
3. Submit a PR
4. Wait for review

Resources

• Contribution guide - start here!
• How to extend OpenVINO Python API
• How to test OpenVINO Python API
• Blog post on contributing to OpenVINO

Contact points

@p-wysocki
@jiwaszki
@akuporos

Don't hesitate to reach out, we're here to help!

Ticket: 110444

[ilya-lavrenov]

Do we need any placeholder if I want to change only specific shapes?
E.g. model has 3 inputs and old last one needs to be changed.

[p-wysocki]

I guess in that case the user should use dict overload, where a specific input can be reshaped. Either that or we do the placeholder you suggested, but I think that whatever the placeholder would be, it wouldn't be clear. We can't do -1 since it means a dynamic shape, None I think would look a bit weird, maybe some string? I guess it's a thing to discuss.

cc @jiwaszki @akuporos

[dev-seek]

hey @p-wysocki @ilya-lavrenov I want to work on this can you please assign it to me

[dev-seek]

thanks a lot for assigning me , i will surely come up with doubts or updates if any

[p-wysocki]

@dev-seek Thanks for taking a look, we're here to answer any questions. :)

[jiwaszki]

I guess in that case the user should use dict overload, where a specific input can be reshaped. Either that or we do the placeholder you suggested, but I think that whatever the placeholder would be, it wouldn't be clear. We can't do -1 since it means a dynamic shape, None I think would look a bit weird, maybe some string? I guess it's a thing to discuss.

cc @jiwaszki @akuporos

As for placeholder values, there is a problem with adding them. What should be a concept instead of integers? Enums like ov.PartialShape.KEEP and consequence of it ov.Dimension.KEEP? Or just a string value KEEP?

I think that placeholder values are not in the scope of this task and @p-wysocki solution with dict is suitable for now (e.x. {1: [1, 21, 37]}).

In general case user could perform something like model.reshape([model.inputs[0].shape, [1, 22, 48], model.inputs[0].partial_shape) where shape/partial_shape can be used interchangeably depending on dynamism of given model inputs. It's worth adding a test-case(s) for such usage.

[dev-seek]

hey @p-wysocki @jiwaszki Just to clearify can i build the whole on Debian as here -> https://github.com/openvinotoolkit/openvino/blob/master/docs/dev/build_linux.md its written for Ubuntu can i do the same for Debian as Ubuntu is based on Debian

[p-wysocki]

Hi @dev-seek, as far as I know this instruction works for Debian. If you encounter any issues please let us know.

[Jay-sanjay]

hey is the issue resolve I want to contribute . I am new to open-source contribution

[dev-seek]

Hey @Jay-sanjay i am already contributing to it