The SpectralConv2DTranspose layer is a new Lipschitz layer equivalent to the tf.keras Conv2DTranspose.
Note that some arguments must be set to a specific value to be supported:
padding must be same
dilation_rate must be (1, 1)
output_padding must be None
Like the SpectralConv2D layer, the kernel is normalized using the spectral normalization, then orthogonalized with Björck algorithm.
thib-s
commented
2 years ago
The
SpectralConv2DTransposelayer is a new Lipschitz layer equivalent to the tf.kerasConv2DTranspose. Note that some arguments must be set to a specific value to be supported:paddingmust besamedilation_ratemust be(1, 1)output_paddingmust beNoneLike the SpectralConv2D layer, the kernel is normalized using the spectral normalization, then orthogonalized with Björck algorithm.