[OSX] [Resnet18] Parameter replacement usage

[QianqianF]

Issue Type

Documentation Feature Request, Others

onnx2tf version number

1.13.12

onnx version number

1.13.1

onnxruntime version number

1.15.0

onnxsim (onnx_simplifier) version number

0.4.17

tensorflow version number

2.13.0rc0

Download URL for ONNX

https://github.com/PINTO0309/onnx2tf/releases/download/0.0.2/resnet18-v1-7.onnx

Parameter Replacement JSON

{
    "format_version": 1,
    "operations": [
      {
        "op_name": "resnetv15_pool1_fwd/Mean",
        "param_target": "attributes", 
        "param_name": "axes",
        "values": [1, 2] # also tried [0, 1]
      }
    ]
}

Description

1. Purpose: Research development
2. What:
   I tried runningonnx2tf -i resnet18-v1-7.onnx -cotof or onnx2tf -i resnet18-v1-7.onnx -ois data:1,3,224,224 or onnx2tf -i resnet18-v1-7.onnx -ois data:1,3,224,224 -prf replace.json. Attached is the original onnx graph for resnet18 and the converted tflite-float32 graph respectively

It looks like the problem is caused by the conversion from GlobalAveragePool to ReduceMean being applied to the wrong axis.

The output of running onnx2tf -i resnet18-v1-7.onnx -cotof is attached here

3. How: So I tried resolving it with using replace.json. However, it's not very successful and I suspect that it's due to me supplying the wrong op_name. I have tried Mean__34 and resnetv15_pool1_fwd/Mean. This might be a really stupid question: where should I look to find the correct op_name especially in this case the onnx operation is converted to a few different tf operations?

4. Why: I would really like to figure out how to use the package properly so I can convert my other pytorch models to tflite for deploying on edge devices.

5. Resources: the example parameter replacement json in documentation

[PINTO0309]

It was not clear to me why parameter substitution was necessary. It just fixes the batch size.

onnx2tf -i resnet18-v1-7.onnx -cotof -b 1

INFO: onnx_output_name: resnetv15_stage2_relu1_fwd tf_output_name: tf.nn.relu_7/Relu:0 shape: (1, 128, 28, 28) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage2_conv4_fwd tf_output_name: tf.math.add_15/Add:0 shape: (1, 128, 28, 28) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage2__plus1 tf_output_name: tf.math.add_17/Add:0 shape: (1, 128, 28, 28) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage2_activation1 tf_output_name: tf.nn.relu_8/Relu:0 shape: (1, 128, 28, 28) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage3_conv2_fwd tf_output_name: tf.math.add_18/Add:0 shape: (1, 256, 14, 14) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage3_conv0_fwd tf_output_name: tf.math.add_19/Add:0 shape: (1, 256, 14, 14) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage3_relu0_fwd tf_output_name: tf.nn.relu_9/Relu:0 shape: (1, 256, 14, 14) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage3_conv1_fwd tf_output_name: tf.math.add_20/Add:0 shape: (1, 256, 14, 14) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage3__plus0 tf_output_name: tf.math.add_22/Add:0 shape: (1, 256, 14, 14) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage3_activation0 tf_output_name: tf.nn.relu_10/Relu:0 shape: (1, 256, 14, 14) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage3_conv3_fwd tf_output_name: tf.math.add_23/Add:0 shape: (1, 256, 14, 14) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage3_relu1_fwd tf_output_name: tf.nn.relu_11/Relu:0 shape: (1, 256, 14, 14) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage3_conv4_fwd tf_output_name: tf.math.add_24/Add:0 shape: (1, 256, 14, 14) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage3__plus1 tf_output_name: tf.math.add_26/Add:0 shape: (1, 256, 14, 14) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage3_activation1 tf_output_name: tf.nn.relu_12/Relu:0 shape: (1, 256, 14, 14) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage4_conv2_fwd tf_output_name: tf.math.add_27/Add:0 shape: (1, 512, 7, 7) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage4_conv0_fwd tf_output_name: tf.math.add_28/Add:0 shape: (1, 512, 7, 7) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage4_relu0_fwd tf_output_name: tf.nn.relu_13/Relu:0 shape: (1, 512, 7, 7) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage4_conv1_fwd tf_output_name: tf.math.add_29/Add:0 shape: (1, 512, 7, 7) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage4__plus0 tf_output_name: tf.math.add_31/Add:0 shape: (1, 512, 7, 7) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage4_activation0 tf_output_name: tf.nn.relu_14/Relu:0 shape: (1, 512, 7, 7) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage4_conv3_fwd tf_output_name: tf.math.add_32/Add:0 shape: (1, 512, 7, 7) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage4_relu1_fwd tf_output_name: tf.nn.relu_15/Relu:0 shape: (1, 512, 7, 7) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage4_conv4_fwd tf_output_name: tf.math.add_33/Add:0 shape: (1, 512, 7, 7) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage4__plus1 tf_output_name: tf.math.add_35/Add:0 shape: (1, 512, 7, 7) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_stage4_activation1 tf_output_name: tf.nn.relu_16/Relu:0 shape: (1, 512, 7, 7) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_pool1_fwd tf_output_name: tf.math.reduce_mean/Mean:0 shape: (1, 512, 1, 1) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: flatten_170 tf_output_name: tf.reshape/Reshape:0 shape: (1, 512) dtype: float32 validate_result:  Matches 
INFO: onnx_output_name: resnetv15_dense0_fwd tf_output_name: Placeholder:0 shape: (1, 1000) dtype: float32 validate_result:  Matches 

[QianqianF]

Thanks for the fast reply. Fixing the batch size did not fix it for me somehow.

INFO: onnx_op_type: GlobalAveragePool onnx_op_name: resnetv15_pool1_fwd
INFO:  input_name.1: resnetv15_stage4_activation1 shape: [1, 512, 7, 7] dtype: float32
INFO:  output_name.1: resnetv15_pool1_fwd shape: [1, 512, 1, 1] dtype: float32
INFO: tf_op_type: reduce_mean
INFO:  input.1.input_tensor: name: tf.nn.relu_16/Relu:0 shape: (1, 7, 7, 512) dtype: <dtype: 'float32'>
INFO:  input.2.axis: val: [1, 2]
INFO:  input.3.keepdims: val: True
INFO:  output.1.output: name: tf.math.reduce_mean/Mean:0 shape: (1, 7, 1, 1) dtype: <dtype: 'float32'>

INFO: onnx_op_type: Flatten onnx_op_name: flatten_170
INFO:  input_name.1: resnetv15_pool1_fwd shape: [1, 512, 1, 1] dtype: float32
INFO:  output_name.1: flatten_170 shape: [1, 512] dtype: float32
ERROR: The trace log is below.

...


ValueError: Exception encountered when calling layer "tf.reshape" (type TFOpLambda).

Cannot reshape a tensor with 7 elements to shape [1,512] (512 elements) for '{{node tf.reshape/Reshape}} = Reshape[T=DT_FLOAT, Tshape=DT_INT32](Placeholder, tf.reshape/Reshape/shape)' with input shapes: [1,1,7,1], [2] and with input tensors computed as partial shapes: input[1] = [1,512].

Call arguments received by layer "tf.reshape" (type TFOpLambda):
  • tensor=tf.Tensor(shape=(1, 1, 7, 1), dtype=float32)
  • shape=['1', '512']
  • name='flatten_170'

[PINTO0309]

It is a problem specific to your environment.

pip show onnx onnxruntime onnxruntime-gpu onnxsim tensorflow onnx2tf

$ docker login ghcr.io
Username (xxxx): {Enter}
Password: {Personal Access Token}
Login Succeeded

$ docker run --rm -it \
-v `pwd`:/workdir \
-w /workdir \
ghcr.io/pinto0309/onnx2tf:1.13.12

To begin with, your log shows that the conversion was successful. https://n.ethz.ch/~qifeng/download/log.txt There is an error in the version of the tools you reported at the beginning of this issue. Perhaps there is a problem with onnxruntime.

I will not investigate environment specific issues. If you cannot solve the problem on your own, you should use Docker.

[QianqianF]

Thanks a lot! I'll look into it.

[QianqianF]

Update: problem resolved after moving from osx to linux system (in case anyone is having the same issue).

[PINTO0309]

Seems to be a bug in onnxruntime for OSX or TensorFlow.

[QianqianF]

Ah I figured out what the problem was. I changed ~/.keras/keras.json to be "image_data_format": "channels_first" when trying out a different converter. After changing it back to channels_last, everything is working again on OSX.