onnx-tensorflow
onnx-tensorflow copied to clipboard
Support NHWC compute format in concat and conv's pad op.
This MR add NHWC support to Concat
operator, and Pad
operator created by Conv
.
Original behaviour
Concat
Concat didn't support NHWC format, so tensors must be transposed back to NCHW before concat, and may be converted to NHWC to do other compuation.
Pad of Conv
When explicit padding is used in Conv, the pad was done before input tensor was transposed, which also leads to a pair of unresolvable Transpose
operators.
Changes
This MR added NHWC support to concat
operator, and pad was done on transposed input of conv.