Darknet: SAM Block in Yolov4

[duongphilong1989]

I'm researching the network architecture of Yolov4.
In the paper of Yolov4, modified SAM block is one of BoS for detector. But I can't find any SAM block in cfg file.
Can you help me to find it in architecture of Yolov4 ? Thank so much.

[WongKinYiu]

@AlexeyAB Maybe we can create a model zoo?

• yolov4.cfg (weights):
  
  
  
  • width=608 height=608 in cfg: 43.5% [email protected]:0.95
  
  
• yolov4-mish.cfg (weights):
  
  
  
  • width=608 height=608 in cfg: 43.8% [email protected]:0.95
  
  
• yolov4-sam-mish.cfg (weights):
  
  
  
  • width=608 height=608 in cfg: 44.0% [email protected]:0.95
  
  

[AlexeyAB]

@WongKinYiu

I think - Yes.

1. Should we make another repository for Model Zoo, or just add list in the Readme / Wiki? https://github.com/AlexeyAB/darknet/wiki

2. What file storage should we use?

• Google Disk (limit 15 GB) / Baidu / something else...?

• Git Large File Storage (LFS) https://git-lfs.github.com/

• GitHub - Assets in realeases? https://github.com/AlexeyAB/darknet/releases
  It seems there is recommended limit 5 GB: https://help.github.com/en/github/managing-large-files/what-is-my-disk-quota

[WongKinYiu]

@AlexeyAB Hello,

My google drive has 10TB space.
I will use markdown format to list the models with cfg and weights.
And also add some training suggestion for different gpus (4gb, 8gb, 12gb, 16gb, ...).
Will share you the document at last 17 June (after cvprw presentation).

[duongphilong1989]

@AlexeyAB , @WongKinYiu Many thanks for your responses.
Sorry, I have one more question. In SAM block, you used concatenate or add?
In figure 5 (modified SAM), I think it's concatenate, but this symbol (x) is different from the symbol (*) in modified PAN.

[WongKinYiu]

Oh, it is multiplication.

[duongphilong1989]

@AlexeyAB , @WongKinYiu
I try to draw the model of the network architecture of Yolov4
Can you help me to check and validate them?
Many thanks for your help.

[MrPhu]

Hello, Thanks for your beautiful images.
I just found out that your SAM block seems to be wrong. (I am sorry if my answer is incorrect).
I take an example of the bottom SAM block (128):

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

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

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

[sam]
from=-2

The Conv3x3/256 should be moved to the main branch and there is one more Conv1x1/256 in the sub-branch (before the sigmoid).
They are similar in other 2 SAM blocks.
Sincerely.

[scamianbas]

Hi ! Congrats for your beautiful images.
Which parts of this blocks can you customize with the cfg file ?
In YoloV5's yaml file it looks like Neck is not configurable ...
Thanks !