How to read the output of jit.trace

1

Here is the output from torch.jit.trace on one of my model.

Screenshot from 2017-10-28 13-16-26
Screenshot from 2017-10-28 13-16-26865×224 220 KB

Most of these are intuitive but I couldn’t understand the meaning of certain keywords. Do we have a good documentation for explaining what is in there? If not, could somebody help me understand what does “Handle” and “uses” in this line mean?

%9 : Float(900, 100), %10 : Handle = ^Addmm(1, 1, False)(%3, %1, %7), uses = [[%11.i0], []];

2

You can do:

import torch.onnx
import torch.onnx.utils
...
trace, out = torch.jit.get_trace_graph(model, args=(input,))
trace = torch.onnx.utils._optimize_graph(trace.graph(), torch.onnx.OperatorExportTypes.ONNX)

This will give you more human-readable trace with onnx ops. Though, _optimize_graph is internal function.

1 Like