YOLOv5 Focus() Layer

[glenn-jocher]

I've received lots of interest in the YOLOv5 🚀 Focus layer, so I've written a short documentation here on it. I created the Focus layer myself when evolving the YOLOv3 architecture into YOLOv5, and did not adopt it from other sources. The main purpose of the Focus layer is to reduce layers, reduce parameters, reduce FLOPS, reduce CUDA memory, increase forward and backward speed while minimally impacting mAP.

yolov5/models/common.py

Lines 163 to 171 in 17b0f71

	class Focus(nn.Module):
	# Focus wh information into c-space
	def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups
	super(Focus, self).__init__()
	self.conv = Conv(c1 * 4, c2, k, s, p, g, act)
	# self.contract = Contract(gain=2)
	def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2)
	return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))

The YOLOv5 Focus layer replaces the first 3 YOLOv3 layers with a single layer:

I settled on the current Focus layer design after a significant effort profiling alternative designs to the YOLOv3 input layers, both for immediate forward/backward/memory profiling results, and also comparing full 300 epoch COCO trainings to determine the effect on mAP. You can profile the Focus layer very easily against the YOLOv3 original layers replaced using the YOLOv5 profile() function:

# Profile
import torch.nn as nn
from models.common import Focus, Conv, Bottleneck
from utils.torch_utils import profile 

m1 = Focus(3, 64, 3)  # YOLOv5 Focus layer
m2 = nn.Sequential(Conv(3, 32, 3, 1), Conv(32, 64, 3, 2), Bottleneck(64, 64))  # YOLOv3 first 3 layers

results = profile(input=torch.randn(16, 3, 640, 640), ops=[m1, m2], n=10, device=0)  # profile both 10 times at batch-size 16

In our YOLOv5 Google Colab notebook I get the following result:

GPU device=0

YOLOv5 🚀 v5.0-405-gfad57c2 torch 1.9.0+cu102 CUDA:0 (Tesla T4, 15109.75MB)

      Params      GFLOPs  GPU_mem (GB)  forward (ms) backward (ms)                   input                  output
        7040       23.07         2.259         16.65          54.1       (16, 3, 640, 640)      (16, 64, 320, 320)  # Focus
       40160       140.7         7.522         77.86         331.9       (16, 3, 640, 640)      (16, 64, 320, 320)

CPU device='cpu'

      Params      GFLOPs  GPU_mem (GB)  forward (ms) backward (ms)                   input                  output
        7040       23.07         0.000         882.1          2029       (16, 3, 640, 640)      (16, 64, 320, 320)  # Focus
       40160       140.7         0.000          4513          8565       (16, 3, 640, 640)      (16, 64, 320, 320)

[developer0hye]

@glenn-jocher

Is there comparative results between the YOLOV5 with YOLOv3 original layers and Focus layer in terms of mAP?

[developer0hye]

@glenn-jocher Thanks a lot!

[developer0hye]

@glenn-jocher
Is there any progressive on this issue?

[glenn-jocher]

@developer0hye no, sorry, I started one training which crashed for some reason and I didn't have time to investigate.

[developer0hye]

@glenn-jocher It's okay! Thanks!

[luohao123]

@glenn-jocher Does it have a result? what's the gap between w/o Focus

[hussein-MA]

@glenn-jocher Thank you for your great work
I have question, why I do not see the calling for focus in yolov5s.yaml?
backbone:
[from, number, module, args]

[[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2
[-1, 1, Conv, [128, 3, 2]], # 1-P2/4
[-1, 3, C3, [128]],
[-1, 1, Conv, [256, 3, 2]], # 3-P3/8
[-1, 6, C3, [256]],
[-1, 1, Conv, [512, 3, 2]], # 5-P4/16
[-1, 9, C3, [512]],
[-1, 1, Conv, [1024, 3, 2]], # 7-P5/32
[-1, 3, C3, [1024]],
[-1, 1, SPPF, [1024, 5]], # 9
]

Is the first line [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 represents the Focus?
May I know where the focus is called?

[glenn-jocher]

@hussein-MA Focus() was replaced by Conv() for better exportability in release v6.0. See release notes for details:
https://github.com/ultralytics/yolov5/releases/tag/v6.0

[hussein-MA]

@glenn-jocher Thank you

[soumya997]

What is the purpose of using a focus structure? slicing the image into 4 parts and then passing them through a conv layer, how does it help, in yolov5? thanks.

cc: @glenn-jocher