Created by: iclementine
- add
conv2d/conv2d_transpose/conv3d/conv3d_transpose
inpaddle.nn.functional
; - minor fix for ConvTransposeOp (InferShape's behavior in compile time and run time. Now it only infers shape at run time or input's shape is not -1.).
- update error message for COnvTransposeOp's InferShape.
- fix the bug that
fluid.layers.conv3d_transopse
does not passoutput_size
to OP. - fix
fluid.layers.conv3d_transpose
'soutput_size
's size, (3 instead of 2). - support tuple of int as
output_size
(in addition to int and list[int]). - unify the way to handle all kinds of padding( in all the 4 apis.)
- use a different order of keyword arguments in conv2d/conv3d,
padding
now precedesstride
. This is the same asconv2d_transpose
andconv3d_transpose
, but different than that influid.layers.conv2d
andfluid.layers.conv3d
.
请注册或登录再回复