OCR_detection_IC15 icon indicating copy to clipboard operation
OCR_detection_IC15 copied to clipboard

Published 20 hours ago • verified publisher icon Vipermdl

Too small loss on first epoch

Open mlcoop opened this issue 5 years ago • 7 comments

I have tried to train on icdar 2015 dataset. But loss is already 0.06 after firt epoch. Seems not all right @Vipermdl Could you please help me out?

mlcoop avatar Apr 15 '19 11:04 mlcoop

I have tried to train on icdar 2015 dataset. But loss is already 0.06 after firt epoch. Seems not all right @Vipermdl Could you please help me out?

not enough,i can reach 0.0028

novioleo avatar Apr 22 '19 13:04 novioleo

@novioleo could you please share your training config like? -resnet50 or resnet34? -batch size -how many epochs did you train? -didnt you overfit? -which dataset are you using to train?

mlcoop avatar Apr 23 '19 03:04 mlcoop

  1. not overfit at all.the val loss similar to train loss
  2. bs=16
  3. resnet50
  4. icdar2015 @AssylCooper

novioleo avatar Apr 23 '19 04:04 novioleo

@novioleo Thanx did it take you like weeks to get loss: 0.0028 ? because it trains really slowly

mlcoop avatar Apr 23 '19 04:04 mlcoop

@AssylCooper change your learning rate~i don't think it slow.i spend few days to train.

novioleo avatar Apr 23 '19 05:04 novioleo

@novioleo I'm trained it on icdar2017(7200 images) and now fine-tuning on icdar2015. Could you please share your weights with 0.0028 loss? Thanx

mlcoop avatar Apr 23 '19 05:04 mlcoop

sorry ,i can't @AssylCooper

novioleo avatar Apr 23 '19 05:04 novioleo