Edit Anchor Boxes sizes in model config

[FraPochetti]

Hi all,
I see here that I can change the anchors aspect ratios in the model config.
What about the anchors sizes?
E.g. this is what I would do in torchvision

AnchorGenerator(sizes=(32, 64, 128, 256, 512,), aspect_ratios=(0.25, 0.5, 1.0, 2.0, 4.0))

Is this supported in EffDet?
Or maybe the sizes are being automatically/dynamically inferred based on image size (and/or other params)?

Thanks!

[rwightman]

The box sizes are computed relative to the feature map sizes w/ the aspect ratios and anchor scale config. I based that on code from the TF impl, which shares most of the impl with Tensorflow Retinanet models, although some variations of that code do alow explicitly setting the sizes...

It's a non-trivial amount of code to wrap ones head around, and have a TODO to simplify and maybe switch to the torchvision approach but it's well off of current priorities and projects....

I think you can tune it fairly easily for datasets where the scales of the objects differ from a typical coco by tuning the anchor scale (which can be specified as one scalar or a sequence per feature level.

efficientdet-pytorch/effdet/anchors.py

Lines 235 to 285 in b3dcb20

	def _generate_configs(self):
	"""Generate configurations of anchor boxes."""
	anchor_configs = {}
	feat_sizes = self.feat_sizes
	for level in range(self.min_level, self.max_level + 1):
	anchor_configs[level] = []
	for scale_octave in range(self.num_scales):
	for aspect in self.aspect_ratios:
	anchor_configs[level].append(
	((feat_sizes[0][0] // feat_sizes[level][0],
	feat_sizes[0][1] // feat_sizes[level][1]),
	scale_octave / float(self.num_scales), aspect,
	self.anchor_scales[level - self.min_level]))
	return anchor_configs
	def _generate_boxes(self):
	"""Generates multiscale anchor boxes."""
	boxes_all = []
	for _, configs in self.config.items():
	boxes_level = []
	for config in configs:
	stride, octave_scale, aspect, anchor_scale = config
	base_anchor_size_x = anchor_scale * stride[1] * 2 ** octave_scale
	base_anchor_size_y = anchor_scale * stride[0] * 2 ** octave_scale
	if isinstance(aspect, Sequence):
	aspect_x = aspect[0]
	aspect_y = aspect[1]
	else:
	aspect_x = np.sqrt(aspect)
	aspect_y = 1.0 / aspect_x
	anchor_size_x_2 = base_anchor_size_x * aspect_x / 2.0
	anchor_size_y_2 = base_anchor_size_y * aspect_y / 2.0
	x = np.arange(stride[1] / 2, self.image_size[1], stride[1])
	y = np.arange(stride[0] / 2, self.image_size[0], stride[0])
	xv, yv = np.meshgrid(x, y)
	xv = xv.reshape(-1)
	yv = yv.reshape(-1)
	boxes = np.vstack((yv - anchor_size_y_2, xv - anchor_size_x_2,
	yv + anchor_size_y_2, xv + anchor_size_x_2))
	boxes = np.swapaxes(boxes, 0, 1)
	boxes_level.append(np.expand_dims(boxes, axis=1))
	# concat anchors on the same level to the reshape NxAx4
	boxes_level = np.concatenate(boxes_level, axis=1)
	boxes_all.append(boxes_level.reshape([-1, 4]))
	anchor_boxes = np.vstack(boxes_all)
	anchor_boxes = torch.from_numpy(anchor_boxes).float()
	return anchor_boxes

[FraPochetti]

Thanks Ross!