[PyAPI] Tensor creation from Pillow Image

[akuporos]

Details:

• Tensor creation from Pillow Image

Tickets:

• CVS-156053

[Wovchena]

Genai also reshapes and casts dtype: https://github.com/openvinotoolkit/openvino.genai/blob/b395127b973ac1276c8c028c9889447ebb68ed5b/samples/python/visual_language_chat/visual_language_chat.py#L39

That can be a problem because ov::Tensor doesn't allow casting. Does numpy preserve Image's layout and dtype? I see the test verifies that, but the used shape and dtype are kind of mainstream, so that doesn't explain what happens to uncommon shapes and types.

@popovaan, do you remember why reshape() and astype() are there?

[popovaan]

reshape() is needed here, because pic.getdata() returns flattened data. For example for 100x200x3 image pic.getdata() will have shape 20000x3.
astype(np.uint8) is needed because pic.getdata() contains int64 values and if we remove the cast, we get error from genai:
RuntimeError: Exception from src/core/src/runtime/tensor.cpp:96:
Exception from src/inference/src/dev/make_tensor.cpp:66:
Tensor data with element type i64, is not representable as pointer to u8

[popovaan]

I just checked np.array(pic) works fine and preserves shapes and type. I don't remember why I used pic.getdata(), looks like we can replace it.

[Wovchena]

@ilya-lavrenov, if image reading can be reduced to ov.Tensor(np.array(Image.open(path).convert("RGB"))), is Tensor creation from Pillow Image is still needed? I'd not add it.

[Wovchena]

openvinotoolkit/openvino.genai#1763

[ilya-lavrenov]

@ilya-lavrenov, if image reading can be reduced to ov.Tensor(np.array(Image.open(path).convert("RGB"))), is Tensor creation from Pillow Image is still needed? I'd not add it.

up to PR authors.

ov.Tensor(np.array(Image.open(path).convert("RGB")))

People need to know about this np.array..

[p-wysocki]

In my opinion for pure user convenience we could have this binding. Constructing a Tensor from PIL image out of the box is more user friendly than a oneliner, in which the image has to be opened, converted to RGB and then cast to np.array - the user has to know/remember about all of these.

[Wovchena]

This binding doesn't remove the need to know about opening and conversion to RGB.

[p-wysocki]

Still, the way I see it is this PR allows users to omit the np.array casting and makes Python API more convenient and working out of the box, which I think aligns with Python values as a language. I agree it's rather a minor change and we could do without it, but out of the two options, I think it's a change for the better. IMO this PR should be merged.

[t-jankowski]

Ok for core part (looks like only code style fixes)