WARNING: The shape inference of ::LearnablePerTensorAffine type is missing

[bobby20180331]

WARNING: The shape inference of ::LearnablePerTensorAffine type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
when I use:

input_shape={'data': [1, 3, 224, 224]}
convert_deploy(model, BackendType.ONNX_QNN, input_shape)

can we ignore this warning ?

[bobby20180331]

And Then I got an error.

Traceback (most recent call last):
File "inference_fx.py", line 280, in
main(args)
File "inference_fx.py", line 80, in main
run_evaluation(model, data_loader_eval, device, args)
File "inference_fx.py", line 141, in run_evaluation
c_result=convert_deploy(model, BackendType.ONNX_QNN, input_shape)
File "/root/anaconda3/envs/c2ffxt18/lib/python3.7/site-packages/MQBench-0.0.6-py3.7.egg/mqbench/convert_deploy.py", line 184, in convert_deploy
convert_function(deploy_model, **kwargs)
File "/root/anaconda3/envs/c2ffxt18/lib/python3.7/site-packages/MQBench-0.0.6-py3.7.egg/mqbench/convert_deploy.py", line 138, in deploy_qparams_tvm
ONNXQNNPass(onnx_model_path).run(model_name)
File "/root/anaconda3/envs/c2ffxt18/lib/python3.7/site-packages/MQBench-0.0.6-py3.7.egg/mqbench/deploy/deploy_onnx_qnn.py", line 277, in run
self.replace_op_pass()
File "/root/anaconda3/envs/c2ffxt18/lib/python3.7/site-packages/MQBench-0.0.6-py3.7.egg/mqbench/deploy/deploy_onnx_qnn.py", line 187, in replace_op_pass
self.replace_add_to_qlinearadd(node, idx)
File "/root/anaconda3/envs/c2ffxt18/lib/python3.7/site-packages/MQBench-0.0.6-py3.7.egg/mqbench/deploy/deploy_onnx_qnn.py", line 119, in replace_add_to_qlinearadd
assert first_input_node.op_type in FAKE_QUANTIZE_OP
AssertionError

I have traced the vars value as followed:
first_input_node.op_type:Mul
FAKE_QUANTIZE_OP: ['FakeQuantizeLearnablePerchannelAffine', 'FixedPerChannelAffine', 'FakeQuantizeDSQPerchannel', 'LearnablePerTensorAffine', 'FixedPerTensorAffine', 'FakeQuantizeDSQPertensor']

[zmzds]

请问你解决了吗？我也遇到这个问题了

[YeRen123455]

@zmzds @bobby20180331 请问你们解决上面的问题了吗，我一直复现不出来example里面的ImageNet的例子，怀疑是转ONNX的时候，这个Error导致的问题。

[github-actions]

This issue has not received any updates in 120 days. Please reply to this issue if this still unresolved!

[jmkho]

Hi, I also encountered this problem except mine returns several types of error

WARNING: The shape inference of ::FixedPerTensorAffine type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding 
it in symbolic function.
WARNING: The shape inference of ::FixedPerChannelAffine type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding
 it in symbolic function.

Any idea as to why this happened and how to solve them? Thanks!

[PannenetsF]

Hi, I also encountered this problem except mine returns several types of error

WARNING: The shape inference of ::FixedPerTensorAffine type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding 
it in symbolic function.
WARNING: The shape inference of ::FixedPerChannelAffine type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding
 it in symbolic function.

Any idea as to why this happened and how to solve them? Thanks!

You should disable the fake quantize in the export stage, as there are no implementations in the onnx. The MQBench processes the weight to onnx and activation to some jsons.