The loss does not decrease

WenmuZhou / PAN.pytorch

A unofficial pytorch implementation of PAN(PSENet2): Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network

Apache License 2.0

418 stars 112 forks source link

The loss does not decrease #3

Closed oysz2016 closed 5 years ago

oysz2016 commented 5 years ago

It's a great job.Before your update, I tried to train PAN, but the loss was still high until the end of the training.Does the current version support effective training and inference?

lgcy commented 5 years ago

i meet the problem too, the loss is non-convergence

ghost commented 5 years ago

@oysz2016 @lgcy

Which dataset you trained on?
How many epochs & hours did you train for?
What was your best your model_best loss?
did you modify config.py?
upload log?

lgcy commented 5 years ago

@deepseek i do not modify the config.py and trained on the icdar2015 for 600 epoch. Then, i got the model_best loss is 0.6

WenmuZhou commented 5 years ago

At present, the loss that can be achieved with the network in the pan paper is about 0.5. Change the get_model to get_model_pse1 can get better results, its still in the experiment

ghost commented 5 years ago

@WenmuZhou I see that you listed the backbone as resnet18, did the detection rate improve?

WenmuZhou commented 5 years ago

resnet 18 is used on paper, resnet50 is used in my experiment

ustczhouyu commented 5 years ago

@WenmuZhou 作者好，感谢分享代码，请问这个复现的代码有使用PA聚类重建的吗？还是只使用pse重建？