search

taoxugit / AttnGAN

MIT License
1.33k stars 415 forks source link

The losses of G were too large when training #61

Open imyzx2017 opened 4 years ago

imyzx2017 commented 4 years ago

image image

Here is my 3 stages generator's loss changing during training on CUB dataset, all the parameters were used the default setting.

During training, the D's loss was decrease, but the G's loss was too large(oscillate around 10), is that right? How to improve it? Or any property solution for parameters setting. Thanks!!!!

Piyushsharma567 commented 3 years ago

Hello I am a beginner so can you tell me how did you generated the graphs. Thank you

imyzx2017 commented 3 years ago

Hello I am a beginner so can you tell me how did you generated the graphs. Thank you

U can save your loss using tensorboard, and then visualize the graph by loading the log file in tensorboard.

ruipengZ commented 3 years ago

Hello! I wonder if you solved the problem. I also get quite large G loss. Is this normal? Do I need to train with extra epochs?

YibinLiu666 commented 1 year ago

Hello! I alse wonder if you solved the problem. I also get quite large G loss, so the R-presion is only 60.14%.

YibinLiu666 commented 1 year ago

image image

Here is my 3 stages generator's loss changing during training on CUB dataset, all the parameters were used the default setting.

During training, the D's loss was decrease, but the G's loss was too large(oscillate around 10), is that right? How to improve it? Or any property solution for parameters setting. Thanks!!!!

image image

Here is my 3 stages generator's loss changing during training on CUB dataset, all the parameters were used the default setting.

During training, the D's loss was decrease, but the G's loss was too large(oscillate around 10), is that right? How to improve it? Or any property solution for parameters setting. Thanks!!!!

Hello! I alse wonder if you solved the problem. I also get quite large G loss, so the R-presion is only 60.14%.