Refactor the AnchorGenerator implementation #3045
Conversation
…ssary vars and simplifying code.
| # (scales, aspect_ratios) are usually an element of zip(self.scales, self.aspect_ratios) | ||
| # This method assumes aspect ratio = height / width for an anchor. | ||
| def generate_anchors(self, scales, aspect_ratios, dtype=torch.float32, device="cpu"): | ||
| # type: (List[int], List[float], int, Device) -> Tensor # noqa: F821 |
There was a problem hiding this comment.
Any reason why the dtype was defined as int earlier?
There was a problem hiding this comment.
because torchscript didn't support torch.dtype before. Maybe it's supported now so we can switch to using it
| return base_anchors.round() | ||
|
|
||
| def set_cell_anchors(self, dtype, device): | ||
| # type: (int, Device) -> None # noqa: F821 |
There was a problem hiding this comment.
Unclear why dtype is declared as int and device as Device instead of torch.device.
There was a problem hiding this comment.
see above, those were limitations of torchscript that might have since then been fixed
| self.set_cell_anchors(dtype, device) | ||
| anchors_over_all_feature_maps = self.cached_grid_anchors(grid_sizes, strides) | ||
| anchors = torch.jit.annotate(List[List[torch.Tensor]], []) | ||
| for i, (image_height, image_width) in enumerate(image_list.image_sizes): |
There was a problem hiding this comment.
Thanks! This is due to some last minute refactoring before the release that kept those variables still around.
| self.set_cell_anchors(dtype, device) | ||
| anchors_over_all_feature_maps = self.cached_grid_anchors(grid_sizes, strides) | ||
| anchors = torch.jit.annotate(List[List[torch.Tensor]], []) | ||
| for i, (image_height, image_width) in enumerate(image_list.image_sizes): |
There was a problem hiding this comment.
Thanks! This is due to some last minute refactoring before the release that kept those variables still around.
…ssary vars and simplifying code. (pytorch#3045)
Adding Python type hints, correcting incorrect types, removing unnecessary vars and simplifying code.