[CoreML EP] Add support of BatchNormalization/Reshape/Global[Average/Max]Pool #6625
Conversation
|
/azp run Windows CPU CI Pipeline |
|
Azure Pipelines successfully started running 1 pipeline(s). |
onnxruntime/core/providers/coreml/builders/impl/batch_norm_op_builder.cc
Outdated
Show resolved
Hide resolved
| << input_size << "d shape"; | ||
| return false; | ||
| } | ||
|
|
There was a problem hiding this comment.
Can we infer the batch size to be 1 and treat it as 4d if input_size == 3? #Pending
There was a problem hiding this comment.
This will be handled in a future PR, will need to add 1 to the end of the shape and make this a 2d batch norm, such as {N,C,H} -> {N,C,H,1}
Will need to add a ExpandDimsLayer (unsqueeze) before bn and a SqueezeLayer after
This may also be required by other ops since CoreML has weird shape for some layers
Updated comments
In reply to: 573438207 [](ancestors = 573438207)
|
|
||
| // Reshape opset 4- uses attributes for new shape which we do not support for now | ||
| int GetMinSupportedOpSet(const Node& /* node */) const override { return 5; } | ||
| }; |
There was a problem hiding this comment.
FWIW opset 7 was the first opset ORT ever supported so anything prior to that is irrelevant
There was a problem hiding this comment.
There are some op is the same since opset 1, for simplicity, keep it this way
Description: [CoreML EP] Add support of BatchNormalization/Reshape/Global[Average/Max]Pool
Motivation and Context