Related to this issue: #1262
We discovered that aten::instance_norm is decomposed during export, emitting combination of reshape + reduce + batchnorm.
It is possible to skip decomposition in exporter, and try to convert to onnx::InstanceNormalization instead. This needs torchlib support for this operator.
cc @xiaowuhu, @fatcat-z, @justinchuby, @titaiwangms
Related to this issue: #1262
We discovered that
aten::instance_normis decomposed during export, emitting combination of reshape + reduce + batchnorm.It is possible to skip decomposition in exporter, and try to convert to onnx::InstanceNormalization instead. This needs torchlib support for this operator.
cc @xiaowuhu, @fatcat-z, @justinchuby, @titaiwangms