search

zuokai / roialign

caffe
17 stars 4 forks source link

always get loss not a number, is there something need to do when trainning? #4

Closed botaichang closed 6 years ago

botaichang commented 6 years ago

when i trainning faster rcnn using roialign, i always get bbox_loss = nan, and cls_loss is very large, is there need other settings? but when i use same configurations and replace roialign with roipooling, the loss converge.

zuokai commented 6 years ago

can i see your prototxt?

zuokai commented 6 years ago

sometimes i found the loss diveraged，you can set the multiple intepolate times by setting is_multiple_interpolate = true, if this step still failed, you can train with roipooling first and finetune the model with roialign

botaichang commented 6 years ago

I just replace ROIppoling layer with ROIAlign in the prototxt. Yeah, I train with roipooling first and then finetune with roialign, but MAP is very low.

zuokai commented 6 years ago

It‘s my experience with roialign： 1）MAP 4 times cubic > 1 time cubic ≈ 4 times bilinear > 1 time bilinear 2) roialign is more effective with small objects and seems not very effective in large objects I didn’t see the situation that MAP is very low，so I don‘t know what’s happened with your models