search

ZJULearning / pixel_link

Implementation of our paper 'PixelLink: Detecting Scene Text via Instance Segmentation' in AAAI2018
MIT License
766 stars 254 forks source link

Loss is not improving #44

Open kulkarnivishal opened 6 years ago

kulkarnivishal commented 6 years ago

Hi @dengdan @BowieHsu

I have been training the model with COCO dataset for a couple of days, and yet the loss is not improving. I used the default parameters and training with batch size set at 8. Looking forward to your suggestions. Please find the screenshot for reference:

Best, Vishal

image

kulkarnivishal commented 6 years ago

@GodOfSmallThings Any thoughts on this? What setting is ideal for COCO dataset

BryceXu90 commented 5 years ago

@kulkarnivishal 你好，这个问题是怎么解决的呢，我现在跑了20000多step，loss一直是2.3--2.4，一直没有改善，不知道怎么办