A utility library for zfp compression / decompression of Torch7 float tensors.
It (usually) gets better compression ratios than zlib (see https://github.com/jonathantompson/torchzlib for a Torch7 wrapper) for floating point types, particularly tensors that exhibit lots of spatial coherence. The caveat is that usually you need to set the accuracy value to lossy compression to see any gains.
You can read more about the method here: http://computation.llnl.gov/projects/floating-point-compression (based on the paper "Fixed-Rate Compressed Floating-Point Arrays" by Peter Lindstrom).
The main (and only) API entry point is a new class torch.ZFPTensor. This is a super simple class that creates a compressed ByteTensor of an input tensor and has a single decompress() method to return the original data.
The constructor signature is:
torch.ZFPTensor(tensor)Where tensor is the tensor to be compressed.
Usage:
require 'torchzfp'
require 'image'
data = image.lena():double() -- Can be double or float.
accuracy = 1e-4
dataCompressed = torch.ZFPTensor(data, accuracy) -- Compress data.
dataDecompressed = dataCompressed:decompress()