Edgetpu: Compiler stuck

[prashant45]

Hi,

The edgetpu_compiler is stuck for a long time when I am compiling my model.

model-Conv2D-quant.tflite.tar.gz

[Namburger]

Isn't this the same model from https://github.com/google-coral/edgetpu/issues/212#issue-689291359?
I'm getting that same error instead of compiler being stuck

[prashant45]

@Namburger

Not exactly.
https://github.com/google-coral/edgetpu/issues/212#issue-689291359 uses Conv1D layer in tensorflow. If you look closely there are Reshape layers before the Conv layers. I think tensorflow-lite projects the ops to 4 dims using Reshape layer. Check the original input dimension.

To avoid that, I changed this model input to 4 dims and used Conv2D layers. You can see that there are no Reshape layer before Conv2D layer.

model-details.txt

I am using,

tensorflow 2.3
Edge TPU Compiler version 14.1.317412892

[Namburger]

maybe you uploaded the wrong model?
the md5sum for both model from 212 and this one is giving me:

9917860e13b6f75b5cf6bd71e3b7558d

:)

[prashant45]

I renamed the file, I hope that helps. The confusion might be because of the same filename for both models. Sorry for the confusion.

[Namburger]

I see, thanks, I just downloaded the new one and trying to compile now, it seems to be stuck as reported
Let me take a look and will update shortly

[Namburger]

@prashant45 I've reach out to another team regarding this, will keep you updated

[Edit] We opened an internal bug for this!

[Namburger]

Update: This is fixed internally for us, I don't have a release date for this fix yet, unfortunately.
Here is the compiled model, thanks for submitting the issue!

[alexanderfrey]

@Namburger Very much waiting towards the new version ! Is there any eta ?

[Namburger]

@alexanderfrey I have no ETA, so sorry :(
we're having tons of roadblock, expect some library changes very soon though