Add shape functions and ignore un-consumed outputs

[ashay]

This PR contains two commits:

[graph] ignore nodes whose outputs are not consumed

If the input graph contains a node (e.g. an initializer) whose output is
not consumed by any node, then the graph initialization fails.  More
specifically, in looking up the list of consumer nodes, the code indexes
into the `consumedby` dictionary, but the key (i.e. the output
name) does not exist in the dictionary since the output is unused.

This patch makes a quick fix by first checking whether the output name
is present in the dictionary before looking up the value in it.

and

[shape] add shape functions for GridSample and DepthToSpace nodes

This patch registers shape functions for the GridSample and DepthToSpace
nodes based on the ONNX operator spec.

[ashay]

To test the shape functions for GridSample and DepthToSpace nodes, use the following scripts to generate the ONNX files:

import onnx

from onnxscript import opset19, script, FLOAT

@script()
def this_script(X: FLOAT[1, 1, 4, 4], G: FLOAT[1, 6, 6, 2]) -> FLOAT[1, 1, 6, 6]:
    return opset19.GridSample(X, G)

onnx_model = this_script.to_model_proto()
onnx.save(onnx_model, "/tmp/a.onnx")

import onnx

from onnxscript import opset19, script, FLOAT

@script()
def this_script(X: FLOAT[1, 8, 2, 3]) -> FLOAT[1, 2, 4, 6]:
    return opset19.DepthToSpace(X, blocksize=2)

onnx_model = this_script.to_model_proto()
onnx.save(onnx_model, "/tmp/b.onnx")

And then reproduce the error using the following commands:

$ python3 -m onnx_tool --in /tmp/a.onnx --out /dev/null
$ python3 -m onnx_tool --in /tmp/b.onnx --out /dev/null

[ThanatosShinji]

Thanks for your contribution! For personal reseason , your code review will be delayed for 3 days. Thanks again!

Best Regards

---

From: Ashay Rane @.> Sent: Wednesday, October 4, 2023 5:59:36 AM To: ThanatosShinji/onnx-tool @.> Cc: Subscribed @.***> Subject: Re: [ThanatosShinji/onnx-tool] Add shape functions and ignore un-consumed outputs (PR #53)

To test the shape functions for GridSample and DepthToSpace nodes, use the following scripts to generate the ONNX files:

import onnx

from onnxscript import opset19, script, FLOAT

@script() def this_script(X: FLOAT[1, 1, 4, 4], G: FLOAT[1, 6, 6, 2]) -> FLOAT[1, 1, 6, 6]: return opset19.GridSample(X, G)

onnx_model = this_script.to_model_proto() onnx.save(onnx_model, "/tmp/a.onnx")

import onnx

from onnxscript import opset19, script, FLOAT

@script() def this_script(X: FLOAT[1, 8, 2, 3]) -> FLOAT[1, 2, 4, 6]: return opset19.DepthToSpace(X, blocksize=2)

onnx_model = this_script.to_model_proto() onnx.save(onnx_model, "/tmp/b.onnx")

And then reproduce the error using the following commands:

$ python3 -m onnx_tool --in /tmp/a.onnx --out /dev/null $ python3 -m onnx_tool --in /tmp/b.onnx --out /dev/null

— Reply to this email directly, view it on GitHubhttps://github.com/ThanatosShinji/onnx-tool/pull/53#issuecomment-1745792396, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AZZIQRUK25WV5BS7IAVL5RLX5SDERAVCNFSM6AAAAAA5RUP27OVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONBVG44TEMZZGY. You are receiving this because you are subscribed to this thread.Message ID: @.***>

[ashay]

No worries, thanks for agreeing to review the PR!

[ThanatosShinji]

Good work!