You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
f"Interpolation works differently is align_corners is False for {resize_mode} mode in PyTorch and TensorRT."
)
else:
raiseRuntimeError(
f"Interpolation mode is {resize_mode} which is not supported by TensorRT."
)
The LINEAR resize mode supports linear (1D), bilinear (2D), trilinear (3D) interpolations, but only bilinear is accepted in dynamo.
An error is thrown saying Interpolation works differently if align_corners is False for bilinear mode in PyTorch and TensorRT. when resize_mode == "bilinear" and align_corners is False. That is quite wrong because the coordinate_transformation is incorrectly set to ASYMMETRIC when align_corners=False.
NEAREST, LINEAR, CUBIC
resize modes, but onlyNEAREST
andLINEAR
are implemented.TensorRT/py/torch_tensorrt/dynamo/conversion/impl/upsample.py
Lines 36 to 47 in ad74a73
LINEAR
resize mode supportslinear (1D), bilinear (2D), trilinear (3D)
interpolations, but onlybilinear
is accepted in dynamo.Interpolation works differently if align_corners is False for bilinear mode in PyTorch and TensorRT.
whenresize_mode == "bilinear"
andalign_corners is False
. That is quite wrong because thecoordinate_transformation
is incorrectly set toASYMMETRIC
whenalign_corners=False
.TensorRT/py/torch_tensorrt/dynamo/conversion/impl/upsample.py
Lines 53 to 62 in ad74a73
The text was updated successfully, but these errors were encountered: