4cbe889c27f5a0c272c54b4e1065ef3525c2fc09,s2cnn/ops/gpu/so3_mm.py,SO3_mm,backward,#SO3_mm#Any#,15

Before Change

        if self.needs_input_grad[0]:
            gradx_cuda_kernel = _setup_so3mm_cuda_kernel(nl=nl, ni=nbatch, nj=nfeature_in, nk=nfeature_out, trans_y_feature=True)
            gradx = torch.cuda.FloatTensor(nspec, nbatch, nfeature_in, 2)
            gradx_cuda_kernel(gradz, y, gradx)

        if self.needs_input_grad[1]:
            grady_cuda_kernel = _setup_so3mm_cuda_kernel(nl=nl, ni=nfeature_out, nj=nfeature_in, nk=nbatch, trans_out_feature=True, conj_x=True, trans_x_spec=True, trans_x_feature=True)

After Change

        if self.needs_input_grad[1]:
            grady_cuda_kernel = _setup_so3mm_cuda_kernel(nl=nl, ni=nfeature_out, nj=nfeature_in, nk=nbatch, trans_out_feature=True, conj_x=True, trans_x_spec=True, trans_x_feature=True)
            gradx = gradz.new_empty((nspec, nfeature_in, nfeature_out, 2))
            grady_cuda_kernel(gradz, x, grady)

        return gradx, grady

Instances