Method __call__

monai/transforms/spatial/array.py:3584–3628

Args: img: shape must be (num_channels, H, W[, D]), randomize: whether to execute `randomize()` function first, defaults to True.

(self, img: torch.Tensor, randomize: bool = True)

            return None

    def __call__(self, img: torch.Tensor, randomize: bool = True) -> torch.Tensor:
        """
        Args:
            img: shape must be (num_channels, H, W[, D]),
            randomize: whether to execute `randomize()` function first, defaults to True.
        """
        if randomize:
            self.randomize()

        if self._do_transform:
            input_shape = img.shape[1:]
            # Clamp each axis to at least 1 so F.interpolate never sees a zero-sized dimension.
            target_shape = tuple(max(1, int(np.round(s * self.zoom_factor))) for s in input_shape)

            # Use F.interpolate directly on a plain tensor to avoid mutating the global
            # set_track_meta flag, which is not thread-safe (see GitHub issue #8409).
            img_t = convert_to_tensor(img, track_meta=False)
            # F.interpolate requires float input and a batch dimension; cast matches
            # the default dtype=float32 that Resize uses internally.
            img_float = img_t.unsqueeze(0).to(dtype=torch.float32)

            downsample_mode = str(self.downsample_mode)
            upsample_mode = str(self.upsample_mode)
            # align_corners is only valid for linear/bilinear/bicubic/trilinear modes
            _align_corners_modes = {"linear", "bilinear", "bicubic", "trilinear"}
            downsample_align_corners = self.align_corners if downsample_mode in _align_corners_modes else None
            upsample_align_corners = self.align_corners if upsample_mode in _align_corners_modes else None

            img_downsampled = torch.nn.functional.interpolate(
                img_float, size=target_shape, mode=downsample_mode, align_corners=downsample_align_corners
            )
            img_upsampled_t = torch.nn.functional.interpolate(
                img_downsampled, size=input_shape, mode=upsample_mode, align_corners=upsample_align_corners
            ).squeeze(0)

            # copy metadata from original image to down-and-upsampled image,
            # respecting the caller's get_track_meta() setting.
            img_upsampled = cast(torch.Tensor, convert_to_tensor(img_upsampled_t, track_meta=get_track_meta()))
            if isinstance(img_upsampled, MetaTensor):
                img_upsampled.copy_meta_from(img)

            return img_upsampled

        else:
            return img


class ConvertBoxToPoints(Transform):