xnn.cnn.models.spherical.SphericalResBlock#

class xnn.cnn.models.spherical.SphericalResBlock(n_in, n_out, b_in, b_out, s2_input, n_alpha=None, n_beta=2, n_gamma=2, max_beta=0.39269908169872414, normalization='batch', activation='relu')[source]#

Bases: Module

ResNet block of the QM7 network (Sec. 5.4): correlation, batch norm, activation, SO(3) correlation, batch norm, plus a shortcut of the input.

Parameters:

  • n_in (int) – Channels in and out.

  • n_out (int) – Channels in and out.

  • b_in (int) – Bandwidths in and out.

  • b_out (int) – Bandwidths in and out.

  • s2_input (bool) – Whether the input is a spherical signal (first block) or an SO(3) signal.

  • n_alpha (Optional[int]) – Filter support (near-identity grids of n_beta rings at polar angles up to max_beta with n_alpha azimuths, and n_gamma values of gamma for SO(3) filters); n_alpha=None uses 2 b_in.

  • n_beta (int) – Filter support (near-identity grids of n_beta rings at polar angles up to max_beta with n_alpha azimuths, and n_gamma values of gamma for SO(3) filters); n_alpha=None uses 2 b_in.

  • n_gamma (int) – Filter support (near-identity grids of n_beta rings at polar angles up to max_beta with n_alpha azimuths, and n_gamma values of gamma for SO(3) filters); n_alpha=None uses 2 b_in.

  • max_beta (float) – Filter support (near-identity grids of n_beta rings at polar angles up to max_beta with n_alpha azimuths, and n_gamma values of gamma for SO(3) filters); n_alpha=None uses 2 b_in.

  • normalization (str or None) – "batch" (the paper) or None.

  • activation (str) – Activation between the two correlations.

forward(x)[source]#

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Parameters:

x (Tensor)

Return type:

Tensor