Adjusting anchor box stride for Faster - RCNN

[ddarolfi]

I attempted to register my own custom anchor generator class in anchor_generator.py and copied the DefaultAnchorGenerator making one modification, changing strides to equal a list of integers, rather than the [x.stride for x in input_shape] it is by default.
Using this custom anchor generator results in a error "AttributeError: 'torch.device' object has no attribute 'device'".
So I have two questions:

1. Is this the correct way to adjust anchor box strides?
2. What are the stride values by default? As in what are the values in input_shape by default? This article (https://medium.com/@hirotoschwert/digging-into-detectron-2-part-4-3d1436f91266) suggests the stride is 64 for the p6 grid. If so I would like to change this as I am trying to utilize much smaller anchor boxes, and 64 would not be suitable.

Here is a section of my custom anchor generator, my modification is in the from_config definition, everything else is the same as the DefaultAnchorGenerator

@ANCHOR_GENERATOR_REGISTRY.register()
class CustomAnchorGenerator(nn.Module):
"""
Compute anchors in the standard ways described in
"Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks".
"""

box_dim: torch.jit.Final[int] = 4
"""
the dimension of each anchor box.
"""

@configurable
def __init__(self, *, sizes, aspect_ratios, strides, offset=0.5):
    """
    This interface is experimental.

    Args:
        sizes (list[list[float]] or list[float]):
            If ``sizes`` is list[list[float]], ``sizes[i]`` is the list of anchor sizes
            (i.e. sqrt of anchor area) to use for the i-th feature map.
            If ``sizes`` is list[float], ``sizes`` is used for all feature maps.
            Anchor sizes are given in absolute lengths in units of
            the input image; they do not dynamically scale if the input image size changes.
        aspect_ratios (list[list[float]] or list[float]): list of aspect ratios
            (i.e. height / width) to use for anchors. Same "broadcast" rule for `sizes` applies.
        strides (list[int]): stride of each input feature.
        offset (float): Relative offset between the center of the first anchor and the top-left
            corner of the image. Value has to be in [0, 1).
            Recommend to use 0.5, which means half stride.
    """
    super().__init__()

    self.strides = strides
    self.num_features = len(self.strides)
    sizes = _broadcast_params(sizes, self.num_features, "sizes")
    aspect_ratios = _broadcast_params(aspect_ratios, self.num_features, "aspect_ratios")
    self.cell_anchors = self._calculate_anchors(sizes, aspect_ratios)

    self.offset = offset
    assert 0.0 <= self.offset < 1.0, self.offset

@classmethod
def from_config(cls, cfg, input_shape: List[ShapeSpec]):
    return {
        "sizes": cfg.MODEL.ANCHOR_GENERATOR.SIZES,
        "aspect_ratios": cfg.MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS,
        "strides": [2, 2, 4, 8, 16],
        "offset": cfg.MODEL.ANCHOR_GENERATOR.OFFSET,
    } ......

`

[ddarolfi]

I have figured out what the default strides are by adding a print statement in the anchor generator. The default strides are [4, 8, 16, 32, 64]. My first question still remains though. Is there a way to change these strides? My modification returns a error, so is there somewhere else I should be changing it?