Thursday, November 22, 2018

Things Learned On the Road to GANs (Generative Adversarial Networks)

draft

Miscellaneous things learned working with Tensorflow/Keras and machine learning to generate images.

1. Use Containers to install Tensorflow

Containers are virtual machines - lite, and it is the least annoying way to install tensorflow on Linux/Centos 7.5.  

2. For CNNs, the GPU is many times faster.

Its true what they say, at least in part.  For convolution neural networks, the GPU speeds things up by at least an order of magnitude and possibly two.  With containers, you only need to install the Nvidia driver on the base machine / OS and use a different container.

3. It is not at all clear how the CNN relates to real convolutions as we know them from image processing, but it does seem as though using larger filters as we increase resolution avoids the problem of the picture subdividing into lots of little squares.




No comments:

Post a Comment