Strided Transposed convolution issue

[EngineerKhan]

I am trying to implement the strided transposed convolution.

This is my code:

x = jnp.ones((50,50))
y = jnp.ones((3,3))
ans = lax.conv_general_dilated(x[None, None], y[None, None], (1,1),[(0,0),(0,0)],(2,2))

But I cannot verify the ans size with the formula in section4.5 from convolution arithmetic.
Am I doing anything wrong in using the formula? I have been unable to relate the conv_general_dilated()'s formula for transposed convolution properly I think.

[hawkinsp]

Have you seen the lax.conv_transpose wrapper? https://jax.readthedocs.io/en/latest/_autosummary/jax.lax.conv_transpose.html

(You can look at its implementation to see how it does its padding calculation, if you like.)

[EngineerKhan]

I have checked it but it's not allowing me to use the padding (even the default [(0,0),(0,0)] I am using here). Can you recommend the equivalent conv_transpose formulation for that?

Also, can you just explain why does it give 97x97 in the above case, rather than 101 by s(w-1)+f?