Dynamic input shape support?

[VicHofs]

This is similar to something mentioned in #16, but the codebase seems to have changed since then, so I thought I'd create a new issue.

I'd like to try using this library with one of my models, which has a dynamic input size ([?, ?, ?, 1] / [-1, -1, -1, 1] on the first layer, which is later resized in ResizeBilinear). However, when I called model.run, I got the following error, which I assume is due to the library not handling this first layer size as dynamic:

TFLite: Input Buffer size (16384) does not match the Input Tensor's expected size (4)! Make sure to resize the input values accordingly.

I am passing in an image as a grayscale tfCore.Tensor, img (64 x 64 x 1), which I am converting to a Float32Array with a new Float32Array(await img.data()). Is there any way currently to use models with dynamic size? I'm also willing to look into implementing this myself; any tips in that regard would also be appreciated.

[maintenance-hans]

Guten Tag, Hans here! 🍻

Thank you for your detailed description of ze issue! It looks like you're running into a problem with dynamic input sizes in your model. That's a valid concern. However, I would recommend checking ze documentation for hints about supported input shapes and making sure ze model is correctly set up for dynamic inputs.

If you can provide ze logs, such as from adb logcat or Xcode, showing any runtime errors when you call model.run, that would help mrousavy diagnose ze issue better. Please check and post those logs if you can.

In any case, danke for your insights, and let’s keep zis discussion open!

Note: If you think I made a mistake, please ping @mrousavy to take a look.

[VicHofs]

I couldn't find any native-level logs about this issue. The message I get is not that mysterious, though, and it appears as a console warning.

@mrousavy I think this is all set to take a look at; thanks!