Search and modify layer/module outputs by name

[Izan_C_G]

Hi, I wish to extract/modify several module outputs from Resnet50 pre-trained model. I desire to do so before the dequantized layers. I decided to see the model graph with Netron to know which ones I want. Because of that, I transformed the torchvision model to a onnx model with the following code:

weights = models.ResNet50_QuantizedWeights.DEFAULT
model = models.resnet50(weights=weights, quantize=True)

dummy_input = torch.randn(125, 3, 224, 224)
input_names = [ "actual_input" ]
output_names = [ "output" ]

torch.onnx.export(model, 
                  dummy_input,
                  "resnet50torch2-int8prova.onnx",
                  input_names=input_names,
                  output_names=output_names
                  )

The resulting graph is the following:

The output I want is named “onnx::DequantizeLinear_148”, which is also the input of a dequantize layer. However, my problem is that I don’t know how to identify this output in the pytorch model. When I print the model, there are no names and the model’s graph changes completely:

How can I search the outputs by name and then modify or obtain for example, with some hooks? Any information or background or example code will help.

Thank you.

[HDCharles]

there is no dequantize op in the pytorch model.

you converted to onnx, which took a bunch of pytorch ops and converted them to onnx ops. This conversion isn’t 1 to 1 but often 1 to many, so it doesn’t exist. You may want to ask the onnx folks how you retrieve this conversion mapping.