Class _grid_encode

gridencoder/grid.py:24–89

}

class _grid_encode(Function):
    @staticmethod
    @custom_fwd
    def forward(ctx, inputs, embeddings, offsets, per_level_scale, base_resolution, calc_grad_inputs=False, gridtype=0, align_corners=False, interpolation=0):
        # inputs: [B, D], float in [0, 1]
        # embeddings: [sO, C], float
        # offsets: [L + 1], int
        # RETURN: [B, F], float

        inputs = inputs.contiguous()

        B, D = inputs.shape # batch size, coord dim
        L = offsets.shape[0] - 1 # level
        C = embeddings.shape[1] # embedding dim for each level
        S = np.log2(per_level_scale) # resolution multiplier at each level, apply log2 for later CUDA exp2f
        H = base_resolution # base resolution

        # manually handle autocast (only use half precision embeddings, inputs must be float for enough precision)
        # if C % 2 != 0, force float, since half for atomicAdd is very slow.
        if torch.is_autocast_enabled() and C % 2 == 0:
            embeddings = embeddings.to(torch.half)

        # L first, optimize cache for cuda kernel, but needs an extra permute later
        outputs = torch.empty(L, B, C, device=inputs.device, dtype=embeddings.dtype)

        if calc_grad_inputs:
            dy_dx = torch.empty(B, L * D * C, device=inputs.device, dtype=embeddings.dtype)
        else:
            dy_dx = None

        _backend.grid_encode_forward(inputs, embeddings, offsets, outputs, B, D, C, L, S, H, dy_dx, gridtype, align_corners, interpolation)

        # permute back to [B, L * C]
        outputs = outputs.permute(1, 0, 2).reshape(B, L * C)

        ctx.save_for_backward(inputs, embeddings, offsets, dy_dx)
        ctx.dims = [B, D, C, L, S, H, gridtype, interpolation]
        ctx.align_corners = align_corners

        return outputs
    
    @staticmethod
    #@once_differentiable
    @custom_bwd
    def backward(ctx, grad):

        inputs, embeddings, offsets, dy_dx = ctx.saved_tensors
        B, D, C, L, S, H, gridtype, interpolation = ctx.dims
        align_corners = ctx.align_corners

        # grad: [B, L * C] --> [L, B, C]
        grad = grad.view(B, L, C).permute(1, 0, 2).contiguous()

        grad_embeddings = torch.zeros_like(embeddings)

        if dy_dx is not None:
            grad_inputs = torch.zeros_like(inputs, dtype=embeddings.dtype)
        else: