Function sample_one

core/utils/utils.py:35–74

Input: -img (N, C, H, W) (flow) -shiftx, shifty (N, c, H, W)

(img, shiftx, shifty, weight)

        return w11, w12, w21, w22

    def sample_one(img, shiftx, shifty, weight):
        """
		Input:
			-img (N, C, H, W) (flow)
			-shiftx, shifty (N, c, H, W)
		"""

        N, C, H, W = img.size()

        # flatten all (all restored as Tensors)
        flat_shiftx = shiftx.view(-1)
        flat_shifty = shifty.view(-1)
        flat_basex = torch.arange(0, H, requires_grad=False).view(-1, 1)[None, None].cuda().long().repeat(N, C, 1, W).view(-1)
        flat_basey = torch.arange(0, W, requires_grad=False).view(1, -1)[None, None].cuda().long().repeat(N, C, H, 1).view(-1)
        flat_weight = weight.view(-1)
        flat_img = img.view(-1)

        # The corresponding positions in I1
        idxn = torch.arange(0, N, requires_grad=False).view(N, 1, 1, 1).long().cuda().repeat(1, C, H, W).view(-1)
        idxc = torch.arange(0, C, requires_grad=False).view(1, C, 1, 1).long().cuda().repeat(N, 1, H, W).view(-1)
        # ttype = flat_basex.type()
        idxx = flat_shiftx.long() + flat_basex
        idxy = flat_shifty.long() + flat_basey


        # recording the inside part the shifted
        mask = idxx.ge(0) & idxx.lt(H) & idxy.ge(0) & idxy.lt(W)

        # Mask off points out of boundaries
        ids = (idxn*C*H*W + idxc*H*W + idxx*W + idxy)
        ids_mask = torch.masked_select(ids, mask).clone().cuda()

        #(zero part - gt) -> difference
        # difference back propagate -> No influence! Whether we do need mask? mask?
                # put (add) them together
        # Note here! accmulate fla must be true for proper bp
        img_warp = torch.zeros([N*C*H*W, ]).cuda() 
        img_warp.put_(ids_mask, torch.masked_select(flat_img*flat_weight, mask), accumulate=True)

        return img_warp.view(N, C, H, W)

    """
        flow should be a pytorch tensor