About the subdivision of batch size for training in the latest code !!

[J-LINC]

Search before asking

• [X] I have searched the YOLOv3 issues and discussions and found no similar questions.

Question

I would like to know if the latest version of YOLOv3/5 supports setting subdivision parameters to support general devices with higher batch_ Size, just like the previous version used YOLOv3. cfg

Additional

No response

[glenn-jocher]

@J-LINC hello,

Yes, you can set the subdivision parameter in the latest versions of YOLOv3/5 to support devices with higher batch sizes. In fact, increasing the subdivision value has been shown to improve training speed and stability for larger batch sizes.

You can find the subdivision parameter in the .cfg file under the [net] section. Simply adjust the value to your desired subdivision value.

Hope this helps! Let us know if you have any further questions or concerns.

Best regards, Glenn Jocher

[J-LINC]

@J-LINC hello,

Yes, you can set the subdivision parameter in the latest versions of YOLOv3/5 to support devices with higher batch sizes. In fact, increasing the subdivision value has been shown to improve training speed and stability for larger batch sizes.

You can find the subdivision parameter in the .cfg file under the [net] section. Simply adjust the value to your desired subdivision value.

Hope this helps! Let us know if you have any further questions or concerns.

Best regards, Glenn Jocher

wow, you mean the latest versions of YOLOV3/5 could use yolov3.cfg?just like: python train --cfg yolov3.cfg rather than python train --cfg yolov3.yaml? by the way, I want to know whether the model structure of the latest version of yolov3 is different from that of yolov5. The other settings, including training strategies and loss function, are the same as those of yolov5.

[glenn-jocher]

Hello @J-LINC,

Thank you for reaching out to us.

Regarding your first question, I apologize if my previous response was unclear. You cannot use the yolov3.cfg file with the latest versions of YOLOv3/5. What I meant to convey is that you can adjust the subdivision parameter in the .cfg file of YOLOv3/5 to support devices with higher batch sizes. Sorry for any confusion.

As for your second question, the model structures of the latest versions of YOLOv3 and YOLOv5 are different. While they share some similarities, such as the use of convolutional layers and anchor boxes, their architectural differences make them unique.

However, the training strategies and loss functions used in both YOLOv3 and YOLOv5 are similar, if not the same in some cases.

I hope this clears things up. Please let us know if you have any further questions or concerns.

Best regards, Glenn Jocher

[J-LINC]

Hello @J-LINC,

Thank you for reaching out to us.

Regarding your first question, I apologize if my previous response was unclear. You cannot use the yolov3.cfg file with the latest versions of YOLOv3/5. What I meant to convey is that you can adjust the subdivision parameter in the .cfg file of YOLOv3/5 to support devices with higher batch sizes. Sorry for any confusion.

As for your second question, the model structures of the latest versions of YOLOv3 and YOLOv5 are different. While they share some similarities, such as the use of convolutional layers and anchor boxes, their architectural differences make them unique.

However, the training strategies and loss functions used in both YOLOv3 and YOLOv5 are similar, if not the same in some cases.

I hope this clears things up. Please let us know if you have any further questions or concerns.

Best regards, Glenn Jocher

Well, I checked the codes of yolov3 and v5, loss.py. I found that in the latest version of the code, the codes of yolov3 and v5 on the loss function are exactly the same. I know that yolov5 uses better loss function, such as GIOU, and of course, some data enhancement strategies and learning rate decay strategies, which makes it impossible for me to use the original version of yolov3 for training. I only came up with one solution, which is to use the old version of the code repository, because I saw that the latest version is suitable for both YOLOv3 and v5, but all training methods and strategies are targeted at YOLOv5. Do you think what I said is correct?

[glenn-jocher]

Hello @J-LINC,

Thank you for taking a closer look at the code and raising your concerns.

Yes, you are correct that the latest version of the loss function code for both YOLOv3 and YOLOv5 is the same. However, as you mentioned, YOLOv5 uses additional loss functions such as GIOU and other training strategies and data augmentation techniques, which make it work differently from YOLOv3.

If you want to use the original YOLOv3 version for training, it is still possible as the code for the loss function is the same. However, keep in mind that you may not get the same performance as YOLOv5 as it has been optimized for the new loss functions and training strategies.

Using the old version of the code repository may be a possible solution but may not have the latest features and bug fixes. You can assess the tradeoffs and make a decision based on your specific needs.

I hope this clarifies your concerns. Let us know if you have any further questions or issues.

Best regards, Glenn Jocher

[J-LINC]

Hello @J-LINC,

Thank you for taking a closer look at the code and raising your concerns.

Yes, you are correct that the latest version of the loss function code for both YOLOv3 and YOLOv5 is the same. However, as you mentioned, YOLOv5 uses additional loss functions such as GIOU and other training strategies and data augmentation techniques, which make it work differently from YOLOv3.

If you want to use the original YOLOv3 version for training, it is still possible as the code for the loss function is the same. However, keep in mind that you may not get the same performance as YOLOv5 as it has been optimized for the new loss functions and training strategies.

Using the old version of the code repository may be a possible solution but may not have the latest features and bug fixes. You can assess the tradeoffs and make a decision based on your specific needs.

I hope this clarifies your concerns. Let us know if you have any further questions or issues.

Best regards, Glenn Jocher

This is too strange. v3 and v5 use the same loss function code, which means that if I train with the latest version of yolov3, I use the loss function of yolov5? That's not the original version. I'm curious whether the v3 and v5 repositories in your code repository are identical. I guess you separated v5 because you originally modified it on the basis of yolov3. That is to say, the latest version of yolov3 is the same as the first generation of yolov5. Some bug modifications in subsequent v5 will not be synchronized to the latest version of v3? Thank you for your patient answer. I really want to know, because I don't want to use the loss function of yolov5 incorrectly when training yolov3, but I can use the data enhancement and learning rate attenuation strategies used in v5

[J-LINC]

Hello @J-LINC,

Thank you for taking a closer look at the code and raising your concerns.

Yes, you are correct that the latest version of the loss function code for both YOLOv3 and YOLOv5 is the same. However, as you mentioned, YOLOv5 uses additional loss functions such as GIOU and other training strategies and data augmentation techniques, which make it work differently from YOLOv3.

If you want to use the original YOLOv3 version for training, it is still possible as the code for the loss function is the same. However, keep in mind that you may not get the same performance as YOLOv5 as it has been optimized for the new loss functions and training strategies.

Using the old version of the code repository may be a possible solution but may not have the latest features and bug fixes. You can assess the tradeoffs and make a decision based on your specific needs.

I hope this clarifies your concerns. Let us know if you have any further questions or issues.

Best regards, Glenn Jocher

Does it mean that if I want to use the loss function of the original v3, I have to comment out some code or add the old loss function? Or can we achieve quick switching by setting certain parameters? Unfortunately, I didn't find it

[glenn-jocher]

Hello @J-LINC,

Thank you for your question.

The loss function code for YOLOv3 and YOLOv5 is the same in the latest version of the repository. This means that if you train with YOLOv3, you will be using the same loss function code as YOLOv5. However, as you pointed out, YOLOv5 also uses additional loss functions and other training strategies that could improve performance.

If you want to use the original loss function code of YOLOv3, one way is to comment out the code for the additional loss functions used in YOLOv5. However, this may not be an ideal solution as it could have unintended consequences.

Another way to achieve quick switching between the two loss functions is by setting certain parameters. For example, you can set the loss function to use only the original loss function code of YOLOv3 by specifying the iou_loss parameter to be zero in the .cfg file. On the other hand, you can set the giou_loss parameter to be non-zero to include the additional loss functions used in YOLOv5.

I hope this helps answer your question. Please let us know if you have any other questions or concerns.

Best regards, Glenn Jocher

[J-LINC]

iou_loss

But the new version of YOLOv3 code no longer supports. cfg files, and I don't understand Where should iou_loss be set 0? can you give me some examples?Thanks

[J-LINC]

@glenn-jocher Sorry, I may have made a bit of a fuss about whether to use the original loss function. After all, v5 and v3 are both target detection, so it is acceptable for them to use the same and better loss function Do you think if I set the CIOU below to False, it means that I will be using the original version?

[glenn-jocher]

Hello @J-LINC,

Thank you for your question.

Setting the CIOU parameter to False in the loss function configuration means that the model will not use the Complete IoU (CIOU) loss function and instead use the regular IoU loss function. This does not necessarily mean that you are using the original version of the loss function, but it will use a simpler version of the loss function for bounding box regression.

Please note that the latest YOLOv3 code repository uses a unified loss function for both YOLOv3 and YOLOv5, which includes additional loss functions such as the GIoU and DIoU loss functions. If you want to use the original loss function of YOLOv3, you may need to use an older version of the code repository.

I hope this answers your question. Please let us know if you have any further questions or concerns.

Best regards, Glenn Jocher

[J-LINC]

Hello @J-LINC,

Thank you for your question.

Setting the CIOU parameter to False in the loss function configuration means that the model will not use the Complete IoU (CIOU) loss function and instead use the regular IoU loss function. This does not necessarily mean that you are using the original version of the loss function, but it will use a simpler version of the loss function for bounding box regression.

Please note that the latest YOLOv3 code repository uses a unified loss function for both YOLOv3 and YOLOv5, which includes additional loss functions such as the GIoU and DIoU loss functions. If you want to use the original loss function of YOLOv3, you may need to use an older version of the code repository.

I hope this answers your question. Please let us know if you have any further questions or concerns.

Best regards, Glenn Jocher

Thanks a lot!i got it!!!

[glenn-jocher]

Dear @J-LINC,

You're welcome! I'm glad to hear that my response helped. If you have any further questions or concerns, please don't hesitate to ask.

Best regards, Glenn Jocher