forty-lock/pepsi-fast_image_inpainting_with_parallel_decoding_network

pepsi-fast_image_inpainting_with_parallel_decoding_network's Issues

Dataset and trained networks

Could you please share the pretrained networks and the mask dataset? I would really appreciate this.

Many thanks, Jan

TypeError: decoder() missing 1 required positional argument: 'size2'

I_co = mm.decoder(vec_en, Height, reuse=False, name='G_de')

The decoder module takes in another parameter size2, we are supposed to send in width to it , right?

PEPSI++

I am very interested in your pepsi++ work, can you upload the code? Thanks.

How to calculate the time required to inpainting a single picture？

Hey, thank you for publishing the code and I have some questions.

Q1: How to calculate the time required to inpainting a single picture？

Q2: How to train it？

I am looking forward your reply!

Hinge loss for discriminator

Thanks for your brilliant work.

One question. In your papaer, you mentioned that you use hinge loss for discriminator. However, in provided code, only traditional SN-GAN losses are used. Is there exists any differences about the performance?

Thanks~

Hello, I would like to know the code of the specific Diet-PEPSI units

I am interested in your research，hope to get your response

Calculation of euclidian distance

Did I understand correctly, that you are using the following formula to calculate the euclidian distance in the CA-module?

d(a, b) = sqrt( ||a||^2 + ||b||^2 - 2ab )

PEPSI-Fast_image_inpainting_with_parallel_decoding_network/module.py

Line 223 in be4e487

DS1 = tf.expand_dims(tt, 0) - 2 * CS

Many thanks, Jan