QDQ node for weight tensor of Con2D undergoes Constant folding (enabled for node using tf type=FakeQuantWithMinMaxVarsPerChannel)

[rado82]

I am doing some experiment on using QAT for a sample model. Looks like QDQ node for the weight tensor of Conv operation is always folded during onnx generation.

Version of various packages are as follows:
tensorflow version is 2.8.2
tf2onnx version is 1.11.1
tf model optimization toolkit version is 0.7.2

I am using tf model optimization to apply fake quantization nodes and using tf2onnx to convert the frozen graph from pb to onnx representation. I always get the weight tensor for the conv2d undergo constant folding during tf2onnx conversion. I can clearly see from the visualization of the frozen graph, there is a fake node introduced for weights.

To reproduce:
Colab pynb Link: https://colab.research.google.com/drive/1Y_LhhWtJejv5teHgQslMPQdwebyHY1GD?usp=sharing

Netron vis of Pb file (fed as input to tf2onnx)

Netron vis of Generated onnx

Checking the previous issues here , I found this. Though tf.quantize_and_dequantize_v2 is used in earlier issue. Here I am using tf model optimization which uses other tf quantization API's

[hwangdeyu]

Hi @rado82, Thanks for the issue.
It seems you want the Conv weight do not be constanted when the type is type=FakeQuantWithMinMaxVarsPerChannel. And it would be very helpful if you can provide the tf code that produced the pb model file.

[mbrookhart]

The TF code is attached in the collab notebook in the original post?

I just found this issue, I'm seeing similar behavior with a saved graphdef QAT model. I get a lot of INFO - folding node using tf type=FakeQuantWithMinMaxVars for my model with the latest tf2onnx. If I edit the script to disable tf constant node folding here and here I get this error instead:

ValueError: make_sure failure: Unable to convert node FakeQuantWithMinMaxArgs with narrow_range=1

I'll try to see if I can reduce this to a unit test. Thanks!