How to get our own anchors?

[BrianKenobi]

I have the same problem as issue #2394. Without autoanchor, I can't get my recall more than 0.663. So I am thinking about how to get my own data's anchor so I can have a better result. Thanks.

well that's an improvement at least. Though I see your BPR is only 0.65, which means that your recall and mAP will never be higher than 0.65. You should try to supply your own manual anchors in your yolov5s.yaml file if possible to better match your data:

yolov5/models/yolov5s.yaml

Lines 6 to 11 in 38ff499

	# anchors
	anchors:
	- [10,13, 16,30, 33,23] # P3/8
	- [30,61, 62,45, 59,119] # P4/16
	- [116,90, 156,198, 373,326] # P5/32

Originally posted by @glenn-jocher in #2394 (comment)

[glenn-jocher]

@BrianKenobi YOLOv5 anchors are saved as Detect() layer attributes on model creation, and updated as necessary by autoanchor before training starts. Their exact location is here:

yolov5/models/yolo.py

Lines 34 to 35 in fab5085

	self.register_buffer('anchors', a) # shape(nl,na,2)
	self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2)

You can examine the anchors of any trained YOLOv5 model like this:

Input

import torch

# Model
model = torch.hub.load('ultralytics/yolov5', 'yolov5s', autoshape=False)  # hub model
# -- or --
model = torch.load('yolov5s.pt')['model']  # local model

# Anchors
m = model.model[-1]  # Detect() layer
print(m.anchor_grid.squeeze())  # print anchors

Output

#           x     y
tensor([[[ 10.,  13.],
         [ 16.,  30.],
         [ 33.,  23.]],  # P3/8-small

        [[ 30.,  61.],
         [ 62.,  45.],
         [ 59., 119.]],  # P4/16-medium

        [[116.,  90.],
         [156., 198.],
         [373., 326.]]], dtype=torch.float16)  # P5/32-large

[BrianKenobi]

@glenn-jocher thank you!
in the line :

model = torch.load('yolov5s.pt')['model'] # local model

I'm not sure what should I put in ['model'].
So I only change 'yolov5s.pt' to my trained YoloV5x model 'best.pt'.
And the result is "ModuleNotFoundError: No module named 'models'"

[BrianKenobi]

I solved the problem by put the code in the YoloV5 folder. Thank you @glenn-jocher again!

But as the result, I still get the default anchor.
I am wondering how can I calculate my own data's anchor because using default anchors I won't get my recall higher than 0.663.

[glenn-jocher]

@BrianKenobi yes this is true. YOLOv5 models must either be loaded within the yolov5/ directory or via PyTorch Hub to enabled access to required module imports. To load from anywhere via PyTorch Hub:

model = torch.hub.load('ultralytics/yolov5', 'custom', path_or_model='best.pt')  # custom model

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.