Integer-Hyper-prior-Network

Reproduction of "INTEGER NETWORKS FOR DATA COMPRESSION WITH LATENT-VARIABLE MODELS".

"Integer networks for data compression with latent-variable models" (ICLR 2019)
J. Ballé, N. Johnston, and D. Minnen
https://openreview.net/pdf?id=S1zz2i0cY7

In the file compare.py, a hyper synthesis transform with integer parameters is provided. This network is expected to work in a deterministic manner, that the output for the fixed input will remain unchanged no matter how the platform varies.

The details of this file are as follows:

• The function float_network() is the integer network, while integer_network() is an instance of the float network implemented with standard tfc library. The structure of these two networks should be the same;
• Both networks are evaluated on cpu and gpu, and here is a typical result:

  Integer Network: True . Error:0.0 
  Float Network: False . Error:-2.0726176330754242e-10 
  

  which shows that the integer network gives the same result across cpu and gpu, while the float network provides two different results (MSE=-2.0726176330754242e-10).
• The libraries used are tensorflow-gpu==1.15 and tensorflow-compression==1.3.