-
由 Pascal Lamblin 提交于
When using "valid" convolutions with subsampling, the gradient implementation was wrong (if dx or dy > 2) or slow. In that case, we now use the implementation in conv3D instead.
eaef458c
When using "valid" convolutions with subsampling, the gradient implementation was wrong (if dx or dy > 2) or slow. In that case, we now use the implementation in conv3D instead.