search

dd604 / refinedet.pytorch

A PyTorch inplementation of RefineDet
MIT License
35 stars 12 forks source link

train on own dataset #4

Open Shawn0Hsu opened 3 years ago

Shawn0Hsu commented 3 years ago

Hi, thank you for your work. I want to train my own data set on this network. There are only a few types and the data format is VOC format. May I ask what I need to modify, thank you!

dd604 commented 3 years ago

You can refer to the training on VOC, where data arranges as follow: VOCdevkit2007 |_ VOC2007 | JPEGImages | Annotations | ImageSets |_ SegmentationClass

Shawn0Hsu commented 3 years ago

You can refer to the training on VOC, where data arranges as follow: VOCdevkit2007 |_ VOC2007 | JPEGImages | Annotations | ImageSets |_ SegmentationClass

Okay, thank you. I have tried training on my own data set. There is only one category, but the training loss function will soon be nan. Do you know the reason? The training set is about 2000 images.

dd604 commented 3 years ago

You can refer to the training on VOC, where data arranges as follow: VOCdevkit2007 |_ VOC2007 | JPEGImages | Annotations | ImageSets |_ SegmentationClass

Okay, thank you. I have tried training on my own data set. There is only one category, but the training loss function will soon be nan. Do you know the reason? The training set is about 2000 images.

You can use a smaller learning rate.

Shawn0Hsu commented 3 years ago

You can refer to the training on VOC, where data arranges as follow: VOCdevkit2007 |_ VOC2007 | JPEGImages | Annotations | ImageSets |_ SegmentationClass

Okay, thank you. I have tried training on my own data set. There is only one category, but the training loss function will soon be nan. Do you know the reason? The training set is about 2000 images.

You can use a smaller learning rate.

Ok, thanks. I will try, thank you!

dd604 commented 3 years ago

You can refer to the training on VOC, where data arranges as follow: VOCdevkit2007 |_ VOC2007 | JPEGImages | Annotations | ImageSets |_ SegmentationClass

Okay, thank you. I have tried training on my own data set. There is only one category, but the training loss function will soon be nan. Do you know the reason? The training set is about 2000 images.

You can use a smaller learning rate.

Ok, thanks. I will try, thank you!

No problem. You can also consider using warmup.

shangshanghuiliyi commented 1 year ago

Where to modify the data to find the path, I don't want to follow the kind of data file storage form given by the author