Giters

terrybroad / Learned-Sim-Autoencoder-For-Video-Frames

An implementation of an autoencoder with a learned similarity metric for reconstructing video frames

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

pink noise //

kenkroft opened this issue · comments

commented

Hey (:

everything is working properly, I trained with a dataset of 1030 images, dividing by ten the minibatch number needed to get a model, and all I got is pink noise when I run the model after. An idea of what the problem is ?

ps : i had to go trought virtualenv and use tensorflow0.12 in order to make it work.

thank you !

commented

Hi I am really sorry but there were changes made to TensorFlow 0.8 that meant that training a new model with this code no longer works. I have tried extensively to debug the problem but have not been able too.

I have just editied to readme to make this more explicit to other people: e7056d4

An alternative to using my code would be to use the original chainer implementation by the paper authors: https://github.com/andersbll/autoencoding_beyond_pixels

Or to use the pix2pix: https://github.com/phillipi/pix2pix or pix2pixTensorflow: https://github.com/yenchenlin/pix2pix-tensorflow The codebase of the latter is very similiar to my this repo and you could just train image pairs to be the same images to get similar results to what this code did.

I'm sorry you wasted time on this I should have made this more explicit beforehand. I hope this other links are useful for you going forward.

commented

maybe If I try to train it using tensorflow 0.7 in virtualenv that could work ?

commented

If you can install it from binaries then absolutely! But I had a lot of trouble doing this and TensorFlow 0.7 is no longer supported by google :(

commented

hmhm, okay I'll try to find another way . I have a very hard time finding a real-time autoencoder for video stream. But thanks for your answers !