dali92002 / DE-GAN

Document Image Enhancement with GANs - TPAMI journal
GNU General Public License v3.0
181 stars 32 forks source link

predict_images are all black or white #2

Open minastinisofking opened 3 years ago

minastinisofking commented 3 years ago

Hello, DE-GAN author.I used watermark text dataset provided by you to train the model without changing any parameters in the model.But it seems doesn't work beacause the PSNR has always been below 1 and predict images are all black or white until I stop running in epoch 36.Can you please tell me where the problem is? Thank you!!! image image

dali92002 commented 3 years ago

Hello,

You can increase the weight of the binary_cross_entropy loss in line 209. Make it 500 instead of 100 as the paper stating.

Or, Try to add some batch normalization layers in the generator,

Or, do them both,

Regards,

On Thu, Dec 3, 2020 at 2:22 AM minastinisofking notifications@github.com wrote:

Hello, DE-GAN author.I used watermark text dataset provided by you to train the model without changing any parameters in the model.But it seems doesn't work beacause the PSNR has always been below 1 and predict images are all black or white until I stop running in epoch 36.Can you please tell me where the problem is? Thank you!!! [image: image] https://user-images.githubusercontent.com/54938769/100950892-61396180-3548-11eb-9053-a588d8487ec2.png [image: image] https://user-images.githubusercontent.com/54938769/100950947-79a97c00-3548-11eb-897f-9add02e3d526.png

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/dali92002/DE-GAN/issues/2, or unsubscribe https://github.com/notifications/unsubscribe-auth/AIQCYOT57MDLAWNHDGNKYOTSS3R43ANCNFSM4ULI5GSA .

gouravsinghbais commented 3 years ago

Hi, I am having the same issue and I have tried adding batch normalization layers and used binary_cross_entropy loss as 500 but it is still not improving. Anything else that we could do to get the results.

Regards

zzojin commented 2 years ago

I come to same problem

Ashwani-Dangwal commented 1 year ago

Hey there, was anyone able to train the model using the code provided here?