soeaver / caffe-model

Caffe models (including classification, detection and segmentation) and deploy files for famouse networks
MIT License
1.28k stars 623 forks source link

About the multi-scale traing #62

Open suzhenghang opened 6 years ago

suzhenghang commented 6 years ago

Hi @soeaver , I try to add the multi-scale traing, but the convergence seems to be difficult; Without multi-scale traing, converge quickly. Do you meet this situation? Thanks in advance

suzhenghang commented 6 years ago

During training, some loss will be larger, such as 0.99, 0.73 ..., I try to imshow the preprocessed image and mask, I do not find something wrong

soeaver commented 6 years ago

Hi, you mean multi-scale training for semantic seg? Usually, ms training will lead to a slightly larger and unstable loss. I think you can look at the final result of the training.

suzhenghang commented 6 years ago

@soeaver , Thanks, ms training does lead to unstable loss.By the way, does ms training increase the IOU in your experiments?

soeaver commented 6 years ago

Actually, I didn't do a single-scale training experiment. But as many papers say, ms training and random flipping will improve 1-3% mIoU as PASCAL VOC dataset.

shiyuangogogo commented 6 years ago

Hi. @suzhenghang @soeaver . How do you implement multi-scale traing in Caffe?