Add target for converting to channels last format #1581
Conversation
TomWildenhain-Microsoft
force-pushed
the
tom/channel_last
branch
from
143db68 to
b097bf2
Compare
|
This pull request introduces 1 alert when merging ffa34bb into d4be5af - view on LGTM.com new alerts:
|
TomWildenhain-Microsoft
force-pushed
the
tom/channel_last
branch
from
ffa34bb to
e7e9343
Compare
Signed-off-by: Tom Wildenhain <[email protected]>
TomWildenhain-Microsoft
force-pushed
the
tom/channel_last
branch
from
e3db072 to
8ba83c9
Compare
| if op.domain in [constants.ONNX_DOMAIN, constants.AI_ONNX_ML_DOMAIN]: | ||
| continue | ||
| # We still don't 100% trust the accuracy of all the shapes in graph.py, but for custom ops they are | ||
| # almost certainly accurate and onnx has no other way of knowing them. |
There was a problem hiding this comment.
Maybe one example would help here. Typical case: Reshape operator, Z = Reshape(X, Y) where X and Y are both inputs. There is no way to infer the shape of Z. Type should be always be known, if the shape is not known, the number of dimensions usually is. Sometimes, if a shape is not known, it can be guessed based on the forward nodes (Z = Gemm(Reshape(X, Y), W) where W has a known shape.
There was a problem hiding this comment.
This case wasn't actually what I was referring to. Having unknown dims added is fine, the bigger issue would be actually incorrect dims. While manipulating the graph object I have in the past seen shape data become inaccurate (after it has been used). We now have tests for this and ideally it should never happen but I'm not 100% sure all bugs have been fixed.
| args = get_args() | ||
|
|
||
| g, org_model_proto = load_graph(args.input) | ||
| g, org_model_proto = load_graph(args.input, args.target) |
There was a problem hiding this comment.
Maybe it would be worth adding an example in the documentation. Is it available through all APIs?
There was a problem hiding this comment.
I think we could but right now onnx does not support this so it would be a little iffy.
Signed-off-by: Tom Wildenhain [email protected]