Didn't find op for builtin opcode 'RESIZE_BILINEAR' version '3' about edgetpu HOT 9 CLOSED

Hello, I have the same problem, I have installed the latest compiler and switch between different versions of tf(1.14.0 , 1.15.0, 2.2.0, 2.4.0) still can't get this work, anyone know how to solve the problem? Thanks.

from edgetpu.

@hgaiser Unfortunately, it is not possible to link to the tf version with the compiler. As per our compatibility overview, we still suggest that tf1.15 is use for all training and tflite quantization conversions. The list of supported operations will always be hosted here
The good new is that we are expecting a new compiler version to be coming out very soon, with this we'll see many bug fixes and tf2.0 compatibility. Stay tune in our /news page for updates.

from edgetpu.

@Namburger thank you for your quick response! Just to verify, when the new compiler comes out, should it support the new RESIZE_BILINEAR op? Or will it support ops from tensorflow 2.0 / 2.1 only?

Also, is the compiler the only way to run models on a TPU? What I mean to ask is if you know a workaround while I'm waiting for the new compiler.

from edgetpu.

@hgaiser new compiler is out since I last posted this, actually!
The new version is 2.0.29* can you try to compiler your model first? I believe an apt upgrade will get you the new version. Unfortunately, despite much effort from our team, the new compiler is still new fully compatible with tf2.0 yet, but most new ops will pass through.

from edgetpu.

Awesome, thank you for the update :) It seems to work, I haven't tested it on the TPU yet, but the compiler succeeded without errors, so I'll close this issue.

🎉

from edgetpu.

@hgaiser Is this a new MLIR based compiler?

from edgetpu.

@bhack it's a compiler to map model operations to the tpu:
https://coral.ai/docs/edgetpu/compiler/

from edgetpu.

I meant if the new compiler was based on MLIR
https://mlir.llvm.org/
https://github.com/tensorflow/tensorflow/tree/master/tensorflow/compiler/mlir/lite

from edgetpu.

@bhack no this is the "standard" compiler found here.

from edgetpu.