Pretrained yolov3-spp-ultralytics.pt is not compatible with cfg/yolov3-asff.cfg

[anhnktp]

KeyError: "weights/yolov3-spp-ultralytics.pt is not compatible with cfg/yolov3-asff.cfg. Specify --weights '' or specify a --cfg compatible with weights/yolov3-spp-ultralytics.pt. See #657"

Could you provide yolov3-asff pretrained ?
thank you!

[github-actions]

Hello @anhnktp, thank you for your interest in our work! Please visit our Custom Training Tutorial to get started, and see our Google Colab Notebook, Docker Image, and GCP Quickstart Guide for example environments.

If this is a bug report, please provide screenshots and minimum viable code to reproduce your issue, otherwise we can not help you.

[glenn-jocher]

@anhnktp your code is out of date. git pull or reclone.

[anhnktp]

@glenn-jocher tks. I will have a try !

[anhnktp]

@glenn-jocher Thanks for quick reply. I have a question, in yolov3-asff.cfg, I see with 80 class -> 258 filter. May be formular (6 + num_class) * 3 = filter ?

[glenn-jocher]

@anhnktp yes, this is 85 * 3 plus 3 output channels for the asff operations.

Honestly though we could not get performance improvements with ASFF, I would recommend training with yolov3-spp with all default settings if you are training a custom dataset. If you are doing research, then go ahead and experiment!

[anhnktp]

@glenn-jocher After pull latest code, the bug still remain. I only change num_class in yolov3-asff.cfg from 80 -> 2 & filter from 258 -> 24. It seems pretrained model doesn't compatible with asff config. Do you know why ?

[glenn-jocher]

@anhnktp asff is a special cfg, it needs 3 extra output channels at each output layer, so you'd need 27 channels in each place you have 24 now.

[glenn-jocher]

@anhnktp ah nevermind, it seems you already did this. Then as the error message says, your pretrained weights are not compatible with the cfg. You need to train from scratch:

python train.py --weights ''

[ChrisLiiiii]

@glenn-jocher I train my custom dataset which has 4 classes, and I set the classes=4 in [yolo] and change all the filters=258 to filters=30 in the yolov3-asff.cfg as you said. I train it:
python train.py --weights '' --cfg 'cfg/yolov3-asff.cfg' --cache-images
But I got this:

Do you know where I am wrong？

[glenn-jocher]

@ChrisLiiiii yes you are correct, they are not compatible. If you want to use this cfg you should train from scratch python train.py --weights ''

[glenn-jocher]

@ChrisLiiiii ah, I see the problem. asff needs 3 extra filters per output layer, which is why there are 258 and not 255. So add 3 filters: filters=33.

[ChrisLiiiii]

@glenn-jocher I have tried your suggestion to change 30 to 33, but I still get this similar results:

In my opinion, the dataset has 4 classes so the filters need to be changed to (4+5)*3+3=30 instead of 33, is that right?
And another thing I'm not sure about is whether I should change all the filters=258 to filters=30? Other than that, I only change the [yolo] classes to 4.
I'm sorry I bothered you when you were so busy, I really appreciate your help!

[glenn-jocher]

@ChrisLiiiii ah, yes if the default would have been 27 filters with yolov3-spp then yes you should simply add 3 to that value. If its not working then it may be an error somewhere else in the cfg. I'd start with yolov3-spp.cfg anyway, we could not gete asff to show improvement in our own tests.

[github-actions]

This issue is stale because it has been open 30 days with no activity. Remove Stale label or comment or this will be closed in 5 days.

[ZZY-666]

@ChrisLiiiii ah, yes if the default would have been 27 filters with yolov3-spp then yes you should simply add 3 to that value. If its not working then it may be an error somewhere else in the cfg. I'd start with yolov3-spp.cfg anyway, we could not gete asff to show improvement in our own tests.

I have the same problem as the friend above. When I + 3, I reported the same error. Please tell me how to solve it. Thank you.

[ChrisLiiiii]

@ChrisLiiiii ah, yes if the default would have been 27 filters with yolov3-spp then yes you should simply add 3 to that value. If its not working then it may be an error somewhere else in the cfg. I'd start with yolov3-spp.cfg anyway, we could not gete asff to show improvement in our own tests.

I have the same problem as the friend above. When I + 3, I reported the same error. Please tell me how to solve it. Thank you.

I had this problem because I didn't set ASFF = True in the models.py

yolov3/models.py

Line 169 in bf34ae0

ASFF = False # https://arxiv.org/abs/1911.09516

You can also check to see if there is a problem here

[ZZY-666]

@ChrisLiiiii ah, yes if the default would have been 27 filters with yolov3-spp then yes you should simply add 3 to that value. If its not working then it may be an error somewhere else in the cfg. I'd start with yolov3-spp.cfg anyway, we could not gete asff to show improvement in our own tests.

I have the same problem as the friend above. When I + 3, I reported the same error. Please tell me how to solve it. Thank you.

I had this problem because I didn't set ASFF = True in the models.py

  yolov3/models.py


     Line 169
  in
  bf34ae0





    
      
       ASFF = False  # https://arxiv.org/abs/1911.09516 

You can also check to see if there is a problem here

Thank you very much for your suggestion. I can do the experiment now！Could you tell me your precision when you run yolov3-asff? Does it have a high precision?Thank you again!

[ChrisLiiiii]

@ChrisLiiiii ah, yes if the default would have been 27 filters with yolov3-spp then yes you should simply add 3 to that value. If its not working then it may be an error somewhere else in the cfg. I'd start with yolov3-spp.cfg anyway, we could not gete asff to show improvement in our own tests.
I have the same problem as the friend above. When I + 3, I reported the same error. Please tell me how to solve it. Thank you.
I had this problem because I didn't set ASFF = True in the models.py

  yolov3/models.py


     Line 169
  in
  bf34ae0





    
      
       ASFF = False  # https://arxiv.org/abs/1911.09516 

You can also check to see if there is a problem here

Thank you very much for your suggestion. I can do the experiment now！Could you tell me your precision when you run yolov3-asff? Does it have a high precision?Thank you again!

I did the experiment on my custom dataset.
To be honest, ASFF did not improve the accuracy of my experiment, probably because of the lack of pre-trained weights.

[ZZY-666]

@ChrisLiiiii ah, yes if the default would have been 27 filters with yolov3-spp then yes you should simply add 3 to that value. If its not working then it may be an error somewhere else in the cfg. I'd start with yolov3-spp.cfg anyway, we could not gete asff to show improvement in our own tests.
I have the same problem as the friend above. When I + 3, I reported the same error. Please tell me how to solve it. Thank you.
I had this problem because I didn't set ASFF = True in the models.py
yolov3/models.py

 Line 169

in
bf34ae0

   ASFF = False  # https://arxiv.org/abs/1911.09516 

You can also check to see if there is a problem here

Thank you very much for your suggestion. I can do the experiment now！Could you tell me your precision when you run yolov3-asff? Does it have a high precision?Thank you again!

I did the experiment on my custom dataset.
To be honest, ASFF did not improve the accuracy of my experiment, probably because of the lack of pre-trained weights.

If there is no pre training weight, the results of more training rounds should be the same. Thank you for your answer. I will communicate with you after the experiment. Before + 3, the accuracy is about 0.06, and I hope to have a good experimental result after the change.

[ChrisLiiiii]

@ZZY-666 Do you mean the accuracy of your experiment is only 6%? That's quite bad. Maybe your targets are very difficult to be detected?

[ZZY-666]

@ZZY-666 Do you mean the accuracy of your experiment is only 6%? That's quite bad. Maybe your targets are very difficult to be detected?

Haha，It is before the number of filters is increased by 3, the dataset can reach 97% in v3. After the change(+3) is in training now, hope to have a good result.

[ChrisLiiiii]

@ZZY-666 Do you mean the accuracy of your experiment is only 6%? That's quite bad. Maybe your targets are very difficult to be detected?

Haha，It is before the number of filters is increased by 3, the dataset can reach 97% in v3. After the change(+3) is in training now, hope to have a good result.

Come on. If your accuracy has improved, I hope I can do the same.

[ZZY-666]

@ZZY-666 Do you mean the accuracy of your experiment is only 6%? That's quite bad. Maybe your targets are very difficult to be detected?

Haha，It is before the number of filters is increased by 3, the dataset can reach 97% in v3. After the change(+3) is in training now, hope to have a good result.

Come on. If your accuracy has improved, I hope I can do the same.

OK, I'll let you know when the results come out. Keep in touch. Thank you!

[ZZY-666]

@ChrisLiiiii ah, yes if the default would have been 27 filters with yolov3-spp then yes you should simply add 3 to that value. If its not working then it may be an error somewhere else in the cfg. I'd start with yolov3-spp.cfg anyway, we could not gete asff to show improvement in our own tests.

I have the same problem as the friend above. When I + 3, I reported the same error. Please tell me how to solve it. Thank you.

I had this problem because I didn't set ASFF = True in the models.py

  yolov3/models.py


     Line 169
  in
  bf34ae0





    
      
       ASFF = False  # https://arxiv.org/abs/1911.09516 

You can also check to see if there is a problem here

兄弟我看你是武汉的 方便给个q吗 我yolov3-asff的训练结果还不错

[ZZY-666]

加了哈哈哈 我的昵称是🐷------------------&nbsp;原始邮件&nbsp;------------------ 发件人:&nbsp;&quot;Chris&amp;nbsp;Li&quot;<[email protected]&gt; 发送时间:&nbsp;2020年9月5日(星期六) 上午9:15 收件人:&nbsp;&quot;ultralytics/yolov3&quot;<[email protected]&gt;; 抄送:&nbsp;&quot;ZZY-666&quot;<[email protected]&gt;;&quot;Mention&quot;<[email protected]&gt;; 主题:&nbsp;Re: [ultralytics/yolov3] Pretrained yolov3-spp-ultralytics.pt is not compatible with cfg/yolov3-asff.cfg (#1097)

[JJQQJJQQ]

@ZZY-666 兄弟你好，请问你使用asff后模型是否有效果？Hello ，does the model work after you use AFSS module?