local interfaces to ipol algorithms
c76f627e — Enric Meinhardt-Llopis 1 year, 10 months ago
232cd603 — Enric Meinhardt-Llopis 2 years ago
23f747aa — Enric Meinhardt-Llopis 2 years ago
[p.tex] gtti doc


browse  log 



You can also use your local clone with git send-email.

CLIPOL: a command line interface to IPOL

All the implementation is contained on a single file "ipol", which is an
executable and importable python script.  This project has two separate goals:

1) provide a command line interface to *all* ipol articles
2) provide a unified IPOL interface for many languages (numpy, octave, lua, ...)

The idea is that each IPOL article is described by a single text file, written
in IDL (Ipol Description Language).  This file specifies the URL of the source
code of the article; the compilation instructions; the names, types and default
values of the input and output parameters; and the way to call the compiled
code of the article.

The IDL files of each IPOL algorithms are stored on the directory
"~/.config/ipol/idl/".   This can be created and filled-in manually, and the
files can be edited by hand using a text editor.  It can also be filled-in
automatically by running "ipol download_idls".

The downloaded and compiled sources, and the binaries, of an algorithm X are
stored on the directory "~/.cache/ipol/X" according to the following hierarchy:
	~/.cache/ipol/X/dl/src.tar.gz (original source code of the article)
	~/.cache/ipol/X/src/ (decompressed source code tree)
	~/.cache/ipol/X/bin/ (binary files associated to the article)
	~/.cache/ipol/X/tmp/ (temporary files created on each execution)

Each time that we want to run an algorithm "X", the file ~/.config/ipol/idl/X
is parsed and the tree under "~/.cache/ipol/X" is created if it didn't already
exist.  The original source code is downloaded and compiled locally.  Then the
algorithm is run to create the requested files.

Thus, for example, if we run

$ ipol ace lena.png out1.png
$ ipol ace lena.png out2.png

The files out1.png and out2.png will be identical (a color-corrected version of
the lena image) but the first run will be much slower because the whole source
code of the ACE algorithm will be downloaded and compiled.  The second run will
be very fast.  Thus, to try the ACE algorithm with several values of alpha we
can do

for i in {1..8}; do
	ipol ace lena.png lena_ace_$i.png alpha=$i

DESIGN PRINCIPLES (in decreasing order of importance)

1. adding a new ipol wrapper consists in writing a single text file
2. this text file must be language-agnostic
3. all functionalities exposed by the compiled code can be exposed
4. it must be easy to write wrappers for many languages using exactly the same
   wrapper text files
5. important targets (in decreasing order): numpy, matlab, octave, lua, C, C++
6. the interface must be as efficient as possible
7. it must be possible to call the original program in "raw" form
8. input parameters are raw strings, their verification is left to the backend


Each IPOL article exposes a single interface, in the form of a single function
with a well-defined signature.  The function receives an array of n objects and
returns an array of m objects, and the numbers n and m are constant and fixed.
There are only THREE possible types of objects:

	- a multidimensional array of floats
	- a float
	- a string

If you need integer values or booleans, you put them inside floats.  Strings
are to be avoided unless really necessary (e.g., if your program deals with

DESIDERATA (in python)

import ipol
import iio

x = iio.read("lena.png")   # x is an array of size 256x256
y = ipol.lsd(x)            # y is an array of size 351x7
z = ipol.nlm(x, sigma=3)   # z is like x, but denoised
f = ipol.phs(x, z)         # f is the horn-schunck flow from x to z
g = ipol.tvl1(x, z)        # f is the TVL1 flow from x to z