search

shepnerd / inpainting_gmcnn

Image Inpainting via Generative Multi-column Convolutional Neural Networks, NeurIPS2018
MIT License
428 stars 97 forks source link

Train on my own dataset #53

Open jinyu121 opened 4 years ago

jinyu121 commented 4 years ago

Hi~

I am training on my own dataset. In stage 2, I get this image(up to down: GT, image feed to the generator, result)

image

As you can see, a darker mask is still on the image.

Have you ever experienced that? Or, what should I do next? Wait for a longer time, or optimize some hyper-parameters? (Now the parameters are copied from your code and not changed)

shepnerd commented 4 years ago

The posted results did seem problematic. Did the first stage training converge? Could you share the training loss trend of both first and second stage?

jinyu121 commented 4 years ago

Stage 1:

image

image

image

At last, Stage 1 can get images like this(Up to down: gt, input, output):

image

Tensorboard record of Stage2 crashed😂

Seems I should wait for longer at Stage 1?

( 1st time to try GAN... May have tons of questions😂)

shepnerd commented 4 years ago

It seems you need to training the first stage longer since you conduct natural scenes inpainting. Besides, you may reverse the channel order of the shown images. It appears in BRG instead of RGB channels.