ONNXRuntimeError after applying fix https://github.com/fastmachinelearning/qonnx/issues/150

[RicDen]

Hello, I seem to have a similar issue to this one #150, when working with qonnx in FINN. I am trying to bring a custom model through the workflow. The model is the LRASPP model from the torchvision library https://docs.pytorch.org/vision/main/models/lraspp.html

The layers causing problems are the following:

Node with empty input:
input: "Mul_8_out0"
input: ""
input: ""
input: "Resize_0_param0"
output: "Resize_0_out0"
name: "Resize_0"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "half_pixel"
  type: STRING
}
attribute {
  name: "cubic_coeff_a"
  f: -0.75
  type: FLOAT
}
attribute {
  name: "mode"
  s: "linear"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}

----------------------------------------
Node with empty input:
input: "Mul_8_out0"
input: ""
input: ""
input: "Resize_0_param0"
output: "Resize_0_out0"
name: "Resize_0"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "half_pixel"
  type: STRING
}
attribute {
  name: "cubic_coeff_a"
  f: -0.75
  type: FLOAT
}
attribute {
  name: "mode"
  s: "linear"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}

----------------------------------------
Node with empty input:
input: "Add_10_out0"
input: ""
input: ""
input: "Resize_1_param0"
output: "global_out"
name: "Resize_1"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "half_pixel"
  type: STRING
}
attribute {
  name: "cubic_coeff_a"
  f: -0.75
  type: FLOAT
}
attribute {
  name: "mode"
  s: "linear"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}

----------------------------------------
Node with empty input:
input: "Add_10_out0"
input: ""
input: ""
input: "Resize_1_param0"
output: "global_out"
name: "Resize_1"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "half_pixel"
  type: STRING
}
attribute {
  name: "cubic_coeff_a"
  f: -0.75
  type: FLOAT
}
attribute {
  name: "mode"
  s: "linear"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}

----------------------------------------

The current FINN main branch does not yet apply the fix, so I went ahead and checked out the latest qonnx release.

I indeed don't fail at the check_all_tensor_shapes_specified anymore.

However, I now fail during onnx runtime with:

Fail                                      Traceback (most recent call last)
Cell In[13], [line 2](vscode-notebook-cell:?execution_count=13&line=2)
----> [2](vscode-notebook-cell:?execution_count=13&line=2) output_dict = oxe.execute_onnx(model, input_dict)
      3 produced_finn = output_dict[list(output_dict.keys())[0]]

File /home/user/finn/deps/qonnx/src/qonnx/core/onnx_exec.py:179, in execute_onnx(model, input_dict, return_full_exec_context, start_node, end_node)
    176 if get_sanitize_quant_tensors() != 0:
    177     # round input values to match quantization annotation
    178     execution_context = sanitize_quant_values(model, node.input, execution_context)
--> [179](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a224c696e2d72696364656e67656c227d.vscode-resource.vscode-cdn.net/home/user/finn/deps/qonnx/src/qonnx/core/onnx_exec.py:179) execute_node(node, execution_context, graph, return_full_exec_context, opset_version)
    180 if get_sanitize_quant_tensors() != 0:
    181     # round output values to quantization annotation
    182     execution_context = sanitize_quant_values(model, node.output, execution_context)

File /home/user/finn/deps/qonnx/src/qonnx/core/onnx_exec.py:88, in execute_node(node, context, graph, return_full_exec_context, opset_version)
     85     input_dict[inp] = context[inp]
     87 sess = rt.InferenceSession(node_model.SerializeToString())
---> [88](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a224c696e2d72696364656e67656c227d.vscode-resource.vscode-cdn.net/home/user/finn/deps/qonnx/src/qonnx/core/onnx_exec.py:88) output_list = sess.run(None, input_dict)
     90 for output_ind in range(len(node.output)):
     91     # get the name of the target buffer from node.output
     92     outp = node.output[output_ind]

File /usr/local/lib/python3.10/dist-packages/onnxruntime/capi/onnxruntime_inference_collection.py:220, in Session.run(self, output_names, input_feed, run_options)
    218     output_names = [output.name for output in self._outputs_meta]
    219 try:
--> [220](https://vscode-remote+ssh-002dremote-002b7b22686f73744e616d65223a224c696e2d72696364656e67656c227d.vscode-resource.vscode-cdn.net/usr/local/lib/python3.10/dist-packages/onnxruntime/capi/onnxruntime_inference_collection.py:220)     return self._sess.run(output_names, input_feed, run_options)
    221 except C.EPFail as err:
    222     if self._enable_fallback:

Fail: [ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running Resize node. Name:'Resize_1' Status Message: upsamplebase.h:422 ScalesValidation Scale value should be greater than 0.

Do you have any suggestions for this scenario? Should I report this as a new bug?

[klassen9]

Hey RicDen, I could reproduce your error only when setting an invalid sizes input (Resize_X_param0 in your case). The sizes input is translated into the scales input somewhere along the way. These two inputs convey the same information and are just different representations that can be translated into each other. A negative sizes input translates into a negative scales input, hence the error.

Maybe take a look at the model export? It does not seem like a qonnx problem.

[RicDen]

Hey klassen9,
Thank you very much for the input. Indeed it seemed like sizes and scales somehow got mixed up with the onnx conversion.
However. now everything seems fine and the issue has resolved itself.