Why are Some ONNX Model Layers Ignored by Profiler ?

[markreeves225]

I've been using the profiling tool from the repository and observed that it seems to ignore certain layers of ONNX models during profiling.

The following opType are being ignored:

• Reshape
• Unsqueeze
• Constant
• Cast
• Gather

I understand that every layer consumes memory when loaded into a system, so I'm curious about the reason behind this behavior.
Please shed some light on why certain layers might be ignored by the profiling tool?

# Code snippet to run profiler
# Download model from ONNX zoo
wget https://media.githubusercontent.com/media/onnx/models/main/Natural_Language_Processing/roberta_Opset18_transformers/roberta_Opset18.onnx

# Run profiler from CLI
python -m onnx_tool -m profile -i 'roberta_Opset18.onnx' -f outputFromProfile.profilerOutput

After running the command, manually checked the output and found that some layers were not part of the output.

[markreeves225]

To print all layers 's profile details

1 Clone repo

git clone https://github.com/ThanatosShinji/onnx-tool.git

2 Make changes in the init.py.

# These ops have no computation
# NoMacsOps = (
#     'Identity', 'Constant', 'Shape', 'Squeeze', 'Unsqueeze', 'Reshape', 'ConstantOfShape', 'Cast', 'Pad', 'Concat',
#     'Slice', 'Gather'
# )

#--------------------------------Change to empty
NoMacsOps = ()

3 Install

pip install <local_dir>/onnx-tool/