PAQ7 Released
Posted by Sachin Garg on 26th December 2005 | Permanent Link
Here is Matt Mahoney’s post to comp.compression in which he announced release of PAQ7. Some major improvements include models for image compression and significant speed improvements. He has also posted some benchmark results in the thread.
I just posted my newest compressor, PAQ7, to http://cs.fit.edu/~mmahoney/compression/#paq7
Compression is similar to PAQAR but about 3 times faster (still very slow). It includes models for color .bmp, .tiff, and .jpeg images, so gives better compression on these (but not as good as WinRK or Stuffit - how they do this well eludes me). It lacks a dictionary and a x86 model so compression is a bit worse than PAsQDa on English text and 32-bit WIndows .exe and .dll files.
This is a complete rewrite of PAQ6. It differs primarily in that it replaces the gradient descent model mixer with a neural network, which can be accelerated using MMX assembler (thus the better speed). For non x86-32 machines or if you don’t have NASM you can compile with -DNOASM (1/3 slower). I tested it under WIndows, Linux and Sparc Solaris for archive compatibility.
I will let Werner test on the maximumcompression.com corpus but in my own tests it takes first place on ohs.doc (due to a large embedded jpeg, which Stuffit missed), and english.dic, and second place on a couple other files.
I don’t know how Stuffit models jpeg (I haven’t seen their patent) but what I did was partially decode the image back to the DCT coefficients to provide context for the Huffman coded data.
I plan to add more models to PAQ8 but I wanted to get something released this year.
– Matt Mahoney