xnn.cnn.models.cnn3d#
Conventional 3D CNN over voxelized atomic environments.
The baseline the 3D steerable CNN of Weiler et al. (NeurIPS 2018) is
compared with: the same architecture “simply without the constraint of
being equivariant for rotation” (supplement Sec. 3). Each block is a
Conv3d with an unconstrained kernel, the Gaussian low-pass filter before
any strided downsampling (Sec. 4.4.2, the smooth_stride of the paper’s
experiments), optional batch normalization and an element-wise activation;
the last feature map is pooled by a global average into the per-atom
features of the VoxelPotential readout.
Because its kernels are unconstrained, the energy this model predicts is not invariant under rotations of the structure (only under its translations and the permutations of its atoms); it is the control that shows what the steerable kernels buy.
Classes
|
Conventional 3D convolutional potential over voxelized environments. |
|
One block of the conventional CNN: convolution, low-pass stride, normalization, activation. |