annotate test_exact.py @ 21:d395461b92ae tip

Lots and lots of modifications. Approximate recovery script working.

# -*- coding: utf-8 -*-
"""
Main script for exact reconstruction tests.
Author: Nicolae Cleju
"""
__author__ = "Nicolae Cleju"
__license__ = "GPL"
__email__ = "[e-mail]"


import numpy
import scipy.io
import math
import os
import time
import multiprocessing
import sys

# Try to do smart importing of matplotlib
try:
  import matplotlib
  if os.name == 'nt':
    print "Importing matplotlib with default (GUI) backend... "
  else:
    print "Importing matplotlib with \"Cairo\" backend... "
    matplotlib.use('Cairo')
  import matplotlib.pyplot as plt
  import matplotlib.cm as cm
  import matplotlib.colors as mcolors
except:
  print "FAIL"
  print "Importing matplotlib.pyplot failed. No figures at all"
  print "Try selecting a different backend"

currmodule = sys.modules[__name__]
printLock = None  # Lock for printing in a thread-safe way
 # Thread-safe variable to store number of finished tasks
currmodule.proccount = multiprocessing.Value('I', 0) # 'I' = unsigned int, see docs (multiprocessing, array)

# Contains pre-defined simulation parameters
import stdparams_exact

# Analysis operator and data generation functions
import AnalysisGenerate

# For exceptions
import pyCSalgos.BP.l1eq_pd
import pyCSalgos.NESTA.NESTA

# For plotting with right axes
import utils



def initProcess(share, ntasks, printLock):
    """
    Pool initializer function (multiprocessing)
    Needed to pass the shared variable to the worker processes
    The variables must be global in the module in order to be seen later in run_once_tuple()
    see http://stackoverflow.com/questions/1675766/how-to-combine-pool-map-with-array-shared-memory-in-python-multiprocessing
    """
    currmodule = sys.modules[__name__]
    currmodule.proccount = share
    currmodule.ntasks = ntasks
    currmodule._printLock = printLock


def generateTaskParams(globalparams):
  """
  Generate a list of task parameters (for parallel running)
  """
  taskparams = []
  SNRdb = globalparams['SNRdb'] 
  sigma = globalparams['sigma'] 
  d = globalparams['d'] 
  deltas = globalparams['deltas'] 
  rhos = globalparams['rhos'] 
  numvects = globalparams['numvects']
  algos = globalparams['algos']
  
  # Process parameters
  noiselevel = 1.0 / (10.0**(SNRdb/10.0));
  
  for delta in deltas:
      for rho in rhos:
          p = round(sigma*d);
          m = round(delta*d);
          l = round(d - rho*m);
  
          # Generate Omega and data based on parameters
          Omega = AnalysisGenerate.Generate_Analysis_Operator(d, p);
          # Optionally make Omega more coherent
          #U,S,Vt = numpy.linalg.svd(Omega);
          #Sdnew = S * (1+numpy.arange(S.size)) # Make D coherent, not Omega!
          #Snew = numpy.vstack((numpy.diag(Sdnew), numpy.zeros((Omega.shape[0] - Omega.shape[1], Omega.shape[1]))))
          #Omega = numpy.dot(U , numpy.dot(Snew,Vt))
        
          # Generate data
          x0,y,M,Lambda,realnoise = AnalysisGenerate.Generate_Data_Known_Omega(Omega, d,p,m,l,noiselevel, numvects,'l0')
          
          # Append task params
          taskparams.append((algos,Omega,y,M,x0))
  
  return taskparams

def processResults(params, taskresults):
  """
  Process the raw task results
  """
  deltas = params['deltas']
  rhos = params['rhos']
  algos = params['algos']
  
  # Init results
  meanmatrix = dict()
  elapsed = dict()
  for algo in algos:
    meanmatrix[algo[1]] = numpy.zeros((rhos.size, deltas.size))
    elapsed[algo[1]] = 0  

  # Process results  
  idx = 0
  for idelta,delta in zip(numpy.arange(deltas.size),deltas):
    for irho,rho in zip(numpy.arange(rhos.size),rhos):
      mrelerr,addelapsed = taskresults[idx]
      idx = idx+1
      for algotuple in algos: 
        meanmatrix[algotuple[1]][irho,idelta] = mrelerr[algotuple[1]]
        if meanmatrix[algotuple[1]][irho,idelta] < 0 or math.isnan(meanmatrix[algotuple[1]][irho,idelta]):
          meanmatrix[algotuple[1]][irho,idelta] = 0
        elapsed[algotuple[1]] = elapsed[algotuple[1]] + addelapsed[algotuple[1]]    
        
  procresults = dict()
  procresults['meanmatrix'] = meanmatrix
  procresults['elapsed'] = elapsed
  return procresults
         
def saveSim(params, procresults):
  """
  Save simulation to mat file
  """
  #tosaveparams = ['d','sigma','deltas','rhos','numvects','SNRdb']
  #tosaveprocresults = ['meanmatrix','elapsed']

  tosave = dict()
  tosave['meanmatrix'] = procresults['meanmatrix']
  tosave['elapsed'] = procresults['elapsed']
  tosave['d'] = params['d']
  tosave['sigma'] = params['sigma']
  tosave['deltas'] = params['deltas']
  tosave['rhos'] = params['rhos']
  tosave['numvects'] = params['numvects']
  tosave['SNRdb'] = params['SNRdb']
  tosave['saveplotbase'] = params['saveplotbase']
  tosave['saveplotexts'] = params['saveplotexts']
  # Save algo names as cell array
  obj_arr = numpy.zeros((len(params['algos']),), dtype=numpy.object)
  idx = 0
  for algotuple in params['algos']:
    obj_arr[idx] = algotuple[1]
    idx = idx+1
  tosave['algonames'] = obj_arr
  try:
    scipy.io.savemat(params['savedataname'], tosave)
  except:
    print "Save error"
    
def loadSim(savedataname):
  """
  Load simulation from mat file
  """
  mdict = scipy.io.loadmat(savedataname)
  
  params = dict()
  procresults = dict()
  
  procresults['meanmatrix'] = mdict['meanmatrix'][0,0]
  procresults['elapsed'] = mdict['elapsed']
  params['d'] = mdict['d']
  params['sigma'] = mdict['sigma']
  params['deltas'] = mdict['deltas']
  params['rhos'] = mdict['rhos']
  params['numvects'] = mdict['numvects']
  params['SNRdb'] = mdict['SNRdb']
  params['saveplotbase'] = mdict['saveplotbase'][0]
  params['saveplotexts'] = mdict['saveplotexts']
  
  algonames = mdict['algonames'][:,0]
  params['algonames'] = []
  for algoname in algonames:
    params['algonames'].append(algoname[0])
  
  return params, procresults

def plot(savedataname):
  """
  Plot results from a mat file.
  The files are saved in the current folder.
  """
  params, procresults = loadSim(savedataname)
  meanmatrix = procresults['meanmatrix']
  saveplotbase = params['saveplotbase']  
  saveplotexts = params['saveplotexts']
  algonames = params['algonames']
  
  for algoname in algonames:
    plt.figure()
    plt.imshow(meanmatrix[algoname], cmap=cm.gray, norm=mcolors.Normalize(0,1), interpolation='nearest',origin='lower')
    for ext in saveplotexts:
      plt.savefig(saveplotbase + algoname + '.' + ext, bbox_inches='tight')
    #plt.show()         
      
#==========================
# Main function
#==========================
def run(params):
  """
  Run simulation with given parameters
  """

  print "This is analysis recovery ABS exact script by Nic"
  print "Running simulation"

  algos = params['algos']
  d = params['d']
  sigma = params['sigma']
  deltas = params['deltas']
  rhos = params['rhos']
  numvects = params['numvects']
  SNRdb = params['SNRdb']
  if 'ncpus' in params:
    ncpus = params['ncpus']
  else:
    ncpus = None
  savedataname = params['savedataname']

  if ncpus is None:
      print "  Running in parallel with default",multiprocessing.cpu_count(),"threads using \"multiprocessing\" package"
      if multiprocessing.cpu_count() == 1:
          doparallel = False
      else:
          doparallel = True
  elif ncpus > 1:
      print "  Running in parallel with",ncpus,"threads using \"multiprocessing\" package"
      doparallel = True
  elif ncpus == 1:
      print "Running single thread"
      doparallel = False
  else:
      print "Wrong number of threads, exiting"
      return
      
  # Print summary of parameters
  print "Parameters:"
  print "  Running algorithms",[algotuple[1] for algotuple in algos]
  
  # Prepare parameters
  print "Generating task parameters..."
  taskparams = generateTaskParams(params)

  # Store global variables
  currmodule.ntasks = len(taskparams)
 
  # Run
  print "Running..."
  taskresults = []
  if doparallel:
    currmodule.printLock = multiprocessing.Lock()
    pool = multiprocessing.Pool(ncpus,initializer=initProcess,initargs=(currmodule.proccount,currmodule.ntasks,currmodule.printLock))
    taskresults = pool.map(run_once_tuple, taskparams)
  else:
    for taskparam in taskparams:
      taskresults.append(run_once_tuple(taskparam))

  # Process results
  procresults = processResults(params, taskresults)

  # Save
  saveSim(params, procresults)
    
  print "Finished."
  
def run_once_tuple(t):
  """
  Wrapper for run_once() that explodes the tuple argument t and shows
  the number of finished / remaining tasks
  """
  
  # Call run_once() here
  results = run_once(*t)
  
  if currmodule.printLock:
    currmodule.printLock.acquire()  
    
    currmodule.proccount.value = currmodule.proccount.value + 1
    print "================================"
    print "Finished task",currmodule.proccount.value,"/",currmodule.ntasks,"tasks remaining",currmodule.ntasks - currmodule.proccount.value,"/",currmodule.ntasks
    print "================================"

    currmodule.printLock.release()

  return results

def run_once(algos,Omega,y,M,x0):
  """
  Run single task (i.e. task function)
  """
  
  d = Omega.shape[1]  
  
  nalgos = len(algos)  
 
  xrec = dict()
  err = dict()
  relerr = dict()
  elapsed = dict()

  # Prepare storage variables for algorithms
  for i,algo in zip(numpy.arange(nalgos),algos):
    xrec[algo[1]]    = numpy.zeros((d, y.shape[1]))
    err[algo[1]]     = numpy.zeros(y.shape[1])
    relerr[algo[1]]  = numpy.zeros(y.shape[1])
    elapsed[algo[1]] = 0
  
  # Run algorithms
  for iy in numpy.arange(y.shape[1]):
    for algofunc,strname in algos:
      try:
        timestart = time.time()
        xrec[strname][:,iy] = algofunc(y[:,iy],M,Omega)
        elapsed[strname] = elapsed[strname] + (time.time() - timestart)
      except pyCSalgos.BP.l1eq_pd.l1eqNotImplementedError as e:
        if currmodule.printLock:
          currmodule.printLock.acquire()
          print "Caught exception when running algorithm",strname," :",e.message
          currmodule.printLock.release()
      except ValueError as e:
        if currmodule.printLock:
          currmodule.printLock.acquire()
          print "Caught ValueError exception when running algorithm",strname," :",e.message
          currmodule.printLock.release()
      err[strname][iy]    = numpy.linalg.norm(x0[:,iy] - xrec[strname][:,iy])
      relerr[strname][iy] = err[strname][iy] / numpy.linalg.norm(x0[:,iy])
  for algofunc,strname in algos:
    if currmodule.printLock:
      currmodule.printLock.acquire()
      print strname,' : avg relative error = ',numpy.mean(relerr[strname])
      currmodule.printLock.release()

  # Prepare results
  #mrelerr = dict()
  #for algotuple in algos:
  #  mrelerr[algotuple[1]] = numpy.mean(relerr[algotuple[1]])
  #return mrelerr,elapsed
  
  exactthr = 1e-6
  mrelerr = dict()
  for algotuple in algos:
    mrelerr[algotuple[1]] = float(numpy.count_nonzero(relerr[algotuple[1]] < exactthr)) / y.shape[1]
  return mrelerr,elapsed

  
def testMatlab():
    """
    For debugging only.
    Load parameters from a mat file saved by Matlab.
    """
    mdict = scipy.io.loadmat("E:\\CS\\Ale mele\\Analysis_ExactRec\\temp.mat")
    algos = stdparams_exact.std1()[0]
    res = run_once(algos, mdict['Omega'].byteswap().newbyteorder(),mdict['y'],mdict['M'],mdict['x0'])
  
def generateFig():
  """
  Generates figures from paper "Analysis-based sparse reconstruction with synthesis-based solvers".
  The figures are saved in the current folder.
  """
  run(stdparams_exact.params1)
  #plot(stdparams_exact.params1['savedataname'])
  utils.replot_exact(stdparams_exact.params1['savedataname'],
                     algonames = None, # will read them from mat file
                     doshow=False,
                     dosave=True,
                     saveplotbase=stdparams_exact.params1['saveplotbase'],
                     saveplotexts=stdparams_exact.params1['saveplotexts'])
  
# Script main
if __name__ == "__main__":
  
  # Set the number of cpus for paraller running (or comment to leave default = max)
  #stdparams_exact.paramstest['ncpus'] = 1

  # Run test parameters
  #stdparams_exact.paramstest['ncpus'] = 1
  #run(stdparams_exact.paramstest)
  #plot(stdparams_exact.paramstest['savedataname'])

  #stdparams_exact.params1['ncpus'] = 1
  generateFig()