Train custom dataset

[blueclowd]

Hi @ruizengalways

Thanks for sharing the code. May I know if there is any tutorial showing how to train a custom dataset apart from COCO? Thanks

[ruizengalways]

You can use next() to iterate the dataset generator to know the details inside. I think it is not too hard to extend the generator to other datasets

[SuperDetected]

Excuse me!
Thanks for your codes.When I use your codes to train in the dataset COCO, the loss is always at 11.xx, and the MACE is always at 24.xx.They never go down.Could you tell me if there are something to adjust?

[ruizengalways]

Excuse me!
Thanks for your codes.When I use your codes to train in the dataset COCO, the loss is always at 11.xx, and the MACE is always at 24.xx.They never go down.Could you tell me if there are something to adjust?

How many epochs have you already trained? Could you please give me more details about your training?

[SuperDetected]

Thanks for your reply,I trained the dataset just with the epochs you have set in the codes: And the only thing I changed is the IMAGES_PER_GPU,the original value is 32 and I set it 2 like this:

…

------------------&nbsp;原始邮件&nbsp;------------------ 发件人:&nbsp;"Rui Zeng"<[email protected]&gt;; 发送时间:&nbsp;2020年5月27日(星期三) 上午10:03 收件人:&nbsp;"ruizengalways/PFNet"<[email protected]&gt;; 抄送:&nbsp;"知足~奔跑"<[email protected]&gt;;"Comment"<[email protected]&gt;; 主题:&nbsp;Re: [ruizengalways/PFNet] Train custom dataset (#1) Excuse me! Thanks for your codes.When I use your codes to train in the dataset COCO, the loss is always at 11.xx, and the MACE is always at 24.xx.They never go down.Could you tell me if there are something to adjust? How many epochs have you already trained? Could you please give me more details about your training? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[ruizengalways]

Increase the batch size and decrease the learning rate

Thanks for your reply,I trained the dataset just with the epochs you have set in the codes: And the only thing I changed is the IMAGES_PER_GPU,the original value is 32 and I set it 2 like this:
…
------------------ 原始邮件 ------------------ 发件人: "Rui Zeng"<[email protected]>; 发送时间: 2020年5月27日(星期三) 上午10:03 收件人: "ruizengalways/PFNet"<[email protected]>; 抄送: "知足~奔跑"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [ruizengalways/PFNet] Train custom dataset (#1) Excuse me! Thanks for your codes.When I use your codes to train in the dataset COCO, the loss is always at 11.xx, and the MACE is always at 24.xx.They never go down.Could you tell me if there are something to adjust? How many epochs have you already trained? Could you please give me more details about your training? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[SuperDetected]

OK, I will have a try,thanks.

…

------------------&nbsp;原始邮件&nbsp;------------------ 发件人:&nbsp;"Rui Zeng"<[email protected]&gt;; 发送时间:&nbsp;2020年5月27日(星期三) 下午3:07 收件人:&nbsp;"ruizengalways/PFNet"<[email protected]&gt;; 抄送:&nbsp;"知足~奔跑"<[email protected]&gt;;"Comment"<[email protected]&gt;; 主题:&nbsp;Re: [ruizengalways/PFNet] Train custom dataset (#1) Increase the batch size and decrease the learning rate Thanks for your reply,I trained the dataset just with the epochs you have set in the codes: And the only thing I changed is the IMAGES_PER_GPU,the original value is 32 and&nbsp;I set&nbsp;it 2&nbsp;like this: … ------------------&nbsp;原始邮件&nbsp;------------------ 发件人:&nbsp;"Rui Zeng"<[email protected]&gt;; 发送时间:&nbsp;2020年5月27日(星期三) 上午10:03 收件人:&nbsp;"ruizengalways/PFNet"<[email protected]&gt;; 抄送:&nbsp;"知足~奔跑"<[email protected]&gt;;"Comment"<[email protected]&gt;; 主题:&nbsp;Re: [ruizengalways/PFNet] Train custom dataset (#1) Excuse me! Thanks for your codes.When I use your codes to train in the dataset COCO, the loss is always at 11.xx, and the MACE is always at 24.xx.They never go down.Could you tell me if there are something to adjust? How many epochs have you already trained? Could you please give me more details about your training? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.