Darknet: Anchors for own dataset

Created on 26 Jun 2018 · 16Comments · Source: pjreddie/darknet

Hi all,

I'm training on my own dataset and assume the anchors need to be modified to fit my data. Does the code take care of this automatically or is there a way to generate anchors?

I've seen this for the Keras implementation of YOLO for anchor generation: https://github.com/experiencor/keras-yolo2/blob/master/gen_anchors.py

Thanks!

Source

RJVisee44

👍1

Most helpful comment

@RyanCodes44 Yes.

for Yolo v2: width=704 height=576 in cfg-file
./darknet detector calc_anchors data/hand.data -num_of_clusters 5 -width 22 -height 18 -show
for Yolo v3: width=704 height=576 in cfg-file
./darknet detector calc_anchors data/hand.data -num_of_clusters 9 -width 704 -height 576 -show

And you can use any images with any sizes.

AlexeyAB on 3 Jul 2018

👍10

All 16 comments

I also tried this:
./darknet detector calc_anchors data/hand.data -num_of_clusters 9 -width 720 -height 405 -show

as I am using the Linux implementation. but nothing happened. I believe the link I attached above is the same or similar to this: https://github.com/AlexeyAB/darknet/blob/master/scripts/gen_anchors.py

RJVisee44 on 26 Jun 2018

👍1

@RyanCodes44 Hi,

You should use this repository to get anchors: https://github.com/AlexeyAB/darknet
By using this command for Yolo v3 (or v2):
./darknet detector calc_anchors data/hand.data -num_of_clusters 9 -width 720 -height 405 -show

You can use this script only for Yolo v2 (not v3):
https://github.com/AlexeyAB/darknet/blob/master/scripts/gen_anchors.py

AlexeyAB on 26 Jun 2018

👍7

@AlexeyAB thank you. I am using Yolov2 so I generated anchors using the script. If i set -num_clusters=9 in that script I get 18 anchors (and 1 average IoU I think). Do I use all of the anchors?

RJVisee44 on 26 Jun 2018

For Yolo v2 use:
./darknet detector calc_anchors data/hand.data -num_of_clusters 5 -width 22 -height 18 -show

720/32=22
405/32=18

Also: https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

Note: If you changed width= or height= in your cfg-file, then new width and height must be divisible by 32.

AlexeyAB on 26 Jun 2018

👍3

Okay awesome. Thank you for your help.

RJVisee44 on 27 Jun 2018

@AlexeyAB does the height/width have to be the same height/width of the images in the dataset? I ask because 720 and 405 are not divisible by 32.

RJVisee44 on 3 Jul 2018

@RyanCodes44

height and width in cfg file and for calc_anchors should be divisible by 32.
height and width of your images can be any for each image: one image 100x50, another image 500x700, 1920x1080, ...

AlexeyAB on 3 Jul 2018

@AlexeyAB so for calc_anchors:

./darknet detector calc_anchors data/hand.data -num_of_clusters 5 -width 22 -height 18 -show

The width and height does not have to be the width and height of the images in my dataset (720x405). They only have to be divisible by 32?

RJVisee44 on 3 Jul 2018

@RyanCodes44 Yes.

for Yolo v2: width=704 height=576 in cfg-file
./darknet detector calc_anchors data/hand.data -num_of_clusters 5 -width 22 -height 18 -show
for Yolo v3: width=704 height=576 in cfg-file
./darknet detector calc_anchors data/hand.data -num_of_clusters 9 -width 704 -height 576 -show

And you can use any images with any sizes.

AlexeyAB on 3 Jul 2018

👍10

Thanks!

RJVisee44 on 3 Jul 2018

@AlexeyAB Hi @AlexeyAB thank you for your answers .
After train with 3000 iterations and I used yolov2 when i test the result of detect is not good such as my goal is to detect the iris of eyes but my result
marw

it can be caused by the anchros.
Remark: the size of my all images is :320*280
Thank you very much @AlexeyAB to help me

such as my file configuration is :
[net]
batch=64
subdivisions=64
width=520
height=520
channels=3
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1

learning_rate=0.001
max_batches = 120000
policy=steps
steps=-1,100,80000,100000
scales=.1,10,.1,.1

[convolutional]
batch_normalize=1
filters=16
size=3
stride=1
pad=1
activation=leaky

[maxpool]
size=2
stride=2

[convolutional]
batch_normalize=1
filters=32
size=3
stride=1
pad=1
activation=leaky

[maxpool]
size=2
stride=2

[convolutional]
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=leaky

[maxpool]
size=2
stride=2

[convolutional]
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=leaky

[maxpool]
size=2
stride=2

[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky

[maxpool]
size=2
stride=2

[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky

[maxpool]
size=2
stride=1

[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky

#

[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky

[convolutional]
size=1
stride=1
pad=1
filters=30
activation=linear

[region]
anchors = 0.738768,0.874946, 2.42204,2.65704, 4.30971,7.04493, 10.246,4.59428, 12.6868,11.8741
bias_match=1
classes=1
coords=4
num=5
softmax=1
jitter=.2
rescore=1

object_scale=5
noobject_scale=1
class_scale=1
coord_scale=1

absolute=1
thresh = .6
random=1

sarratouil on 18 Mar 2019

there are three fields (anchors) in the cfg file.
Which ones do i change and replace with my new calculated anchors??

[net]

Testing

batch=1

subdivisions=1

Training

batch=64
subdivisions=64
width=416
height=416
channels=3 #changed from 3 to 1
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1

learning_rate=0.001
burn_in=1000
max_batches = 4000
policy=steps
steps=3200,3600
scales=.1,.1

[convolutional]
batch_normalize=1
filters=32
size=3
stride=1
pad=1
activation=leaky

Downsample

[convolutional]
batch_normalize=1
filters=64
size=3
stride=2
pad=1
activation=leaky

[convolutional]
batch_normalize=1
filters=32
size=1
stride=1
pad=1
activation=leaky

[convolutional]
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=leaky