diff scripts/ABSapprox.py @ 22:2dd78e37b23a

ABS approx script is working
Started working on parallel

--- a/scripts/ABSapprox.py	Tue Nov 08 14:45:35 2011 +0000
+++ b/scripts/ABSapprox.py	Wed Nov 09 00:11:14 2011 +0000
@@ -6,61 +6,133 @@
 """
 
 import numpy as np
+import scipy.io
+import math
+import matplotlib.pyplot as plt
+import matplotlib.cm as cm
 import pyCSalgos
 import pyCSalgos.GAP.GAP
 import pyCSalgos.SL0.SL0_approx
 
 # Define functions that prepare arguments for each algorithm call
-def gap_paramsetup(y,M,Omega,epsilon,lbd):
+def run_gap(y,M,Omega,epsilon):
   gapparams = {"num_iteration" : 1000,\
                    "greedy_level" : 0.9,\
                    "stopping_coefficient_size" : 1e-4,\
                    "l2solver" : 'pseudoinverse',\
                    "noise_level": epsilon}
-  return y,M,M.T,Omega,Omega.T,gapparams,np.zeros(Omega.shape[1])
-def sl0_paramsetup(y,M,Omega,epsilon,lbd):
+  return pyCSalgos.GAP.GAP.GAP(y,M,M.T,Omega,Omega.T,gapparams,np.zeros(Omega.shape[1]))[0]
+ 
+def run_sl0(y,M,Omega,D,U,S,Vt,epsilon,lbd):
   
   N,n = Omega.shape
-  D = np.linalg.pinv(Omega)
-  U,S,Vt = np.linalg.svd(D)
+  #D = np.linalg.pinv(Omega)
+  #U,S,Vt = np.linalg.svd(D)
   aggDupper = np.dot(M,D)
   aggDlower = Vt[-(N-n):,:]
   aggD = np.concatenate((aggDupper, lbd * aggDlower))
   aggy = np.concatenate((y, np.zeros(N-n)))
   
-  sigmamin = 0.01
-  sigma_decrease_factor = 0.8
+  sigmamin = 0.001
+  sigma_decrease_factor = 0.5
   mu_0 = 2
   L = 10
-  return aggD,aggy,epsilon,sigmamin,sigma_decrease_factor,mu_0,L
-
-def post_multiply_with_D(D,gamma):
-    return np.dot(D,gamma)
-def post_do_nothing(D,gamma):
-    return gamma
+  return pyCSalgos.SL0.SL0_approx.SL0_approx(aggD,aggy,epsilon,sigmamin,sigma_decrease_factor,mu_0,L)
 
 # Define tuples (algorithm setup function, algorithm function, name)
-gap = (gap_paramsetup, pyCSalgos.GAP.GAP.GAP, post_do_nothing, 'GAP')
-sl0 = (sl0_paramsetup, pyCSalgos.SL0.SL0_approx.SL0_approx, post_multiply_with_D, 'SL0_approx')
-#sl0 = (sl0_paramsetup, lambda x: np.dot(x[0],x[1]()), 'SL0_approx')
+gap = (run_gap, 'GAP')
+sl0 = (run_sl0, 'SL0_approx')
 
-# Main function
+# Define which algorithms to run
+#  1. Algorithms not depending on lambda
+algosN = gap,   # tuple
+#  2. Algorithms depending on lambda (our ABS approach)
+algosL = sl0,   # tuple
+
 def mainrun():
-
-  # Define which algorithms to run
-  algos = (gap, sl0)
-  numalgos = len(algos)
   
-  # Set up experiment parameters
-  sigma = 2.0;
-  delta = 0.8;
-  rho   = 0.15;
+  nalgosN = len(algosN)  
+  nalgosL = len(algosL)
+  
+  #Set up experiment parameters
+  d = 50;
+  sigma = 2.0
+  #deltas = np.arange(0.05,0.95,0.05)
+  #rhos = np.arange(0.05,0.95,0.05)
+  deltas = np.array([0.05,0.95])
+  rhos = np.array([0.05,0.95])
+  #deltas = np.array([0.05])
+  #rhos = np.array([0.05])
+  #delta = 0.8;
+  #rho   = 0.15;
   numvects = 10; # Number of vectors to generate
   SNRdb = 20.;    # This is norm(signal)/norm(noise), so power, not energy
+  # Values for lambda
+  #lambdas = [0 10.^linspace(-5, 4, 10)];
+  lambdas = np.concatenate((np.array([0]), 10**np.linspace(-5, 4, 10)))
+
+  meanmatrix = dict()
+  for i,algo in zip(np.arange(nalgosN),algosN):
+    meanmatrix[algo[1]]   = np.zeros((rhos.size, deltas.size))
+  for i,algo in zip(np.arange(nalgosL),algosL):
+    meanmatrix[algo[1]]   = np.zeros((lambdas.size, rhos.size, deltas.size))
+  
+  for idelta,delta in zip(np.arange(deltas.size),deltas):
+    for irho,rho in zip(np.arange(rhos.size),rhos):
+      
+      # Generate data and operator
+      Omega,x0,y,M,realnoise = genData(d,sigma,delta,rho,numvects,SNRdb)
+      
+      # Run algorithms
+      mrelerrN,mrelerrL = runonce(algosN,algosL,Omega,y,lambdas,realnoise,M,x0)
+      
+      for algotuple in algosN: 
+        meanmatrix[algotuple[1]][irho,idelta] = 1 - mrelerrN[algotuple[1]]
+        if meanmatrix[algotuple[1]][irho,idelta] < 0 or math.isnan(meanmatrix[algotuple[1]][irho,idelta]):
+          meanmatrix[algotuple[1]][irho,idelta] = 0
+      for algotuple in algosL:
+        for ilbd in np.arange(lambdas.size):
+          meanmatrix[algotuple[1]][ilbd,irho,idelta] = 1 - mrelerrL[algotuple[1]][ilbd]
+          if meanmatrix[algotuple[1]][ilbd,irho,idelta] < 0 or math.isnan(meanmatrix[algotuple[1]][ilbd,irho,idelta]):
+            meanmatrix[algotuple[1]][ilbd,irho,idelta] = 0
+   
+  #  # Prepare matrices to show
+  #  showmats = dict()
+  #  for i,algo in zip(np.arange(nalgosN),algosN):
+  #    showmats[algo[1]]   = np.zeros(rhos.size, deltas.size)
+  #  for i,algo in zip(np.arange(nalgosL),algosL):
+  #    showmats[algo[1]]   = np.zeros(lambdas.size, rhos.size, deltas.size)
+
+    # Save
+    tosave = dict()
+    tosave['meanmatrix'] = meanmatrix
+    tosave['d'] = d
+    tosave['sigma'] = sigma
+    tosave['deltas'] = deltas
+    tosave['rhos'] = rhos
+    tosave['numvects'] = numvects
+    tosave['SNRdb'] = SNRdb
+    tosave['lambdas'] = lambdas
+    try:
+      scipy.io.savemat('ABSapprox.mat',tosave)
+    except TypeError:
+      print "Oops, Type Error"
+      raise    
+  # Show
+  for algotuple in algosN:
+    plt.figure()
+    plt.imshow(meanmatrix[algotuple[1]], cmap=cm.gray, interpolation='nearest')
+  for algotuple in algosL:
+    for ilbd in np.arange(lambdas.size):
+      plt.figure()
+      plt.imshow(meanmatrix[algotuple[1]][ilbd], cmap=cm.gray, interpolation='nearest')
+  plt.show()
+  print "Finished."
+  
+def genData(d,sigma,delta,rho,numvects,SNRdb):
 
   # Process parameters
   noiselevel = 1.0 / (10.0**(SNRdb/10.0));
-  d = 50;
   p = round(sigma*d);
   m = round(delta*d);
   l = round(d - rho*m);
@@ -68,43 +140,73 @@
   # Generate Omega and data based on parameters
   Omega = pyCSalgos.GAP.GAP.Generate_Analysis_Operator(d, p);
   # Optionally make Omega more coherent
-  #[U, S, Vt] = np.linalg.svd(Omega);
-  #Sdnew = np.diag(S) * (1+np.arange(np.diag(S).size)); % Make D coherent, not Omega!
-  #Snew = [diag(Sdnew); zeros(size(S,1) - size(S,2), size(S,2))];
-  #Omega = U * Snew * V';
+  U,S,Vt = np.linalg.svd(Omega);
+  Sdnew = S * (1+np.arange(S.size)) # Make D coherent, not Omega!
+  Snew = np.vstack((np.diag(Sdnew), np.zeros((Omega.shape[0] - Omega.shape[1], Omega.shape[1]))))
+  Omega = np.dot(U , np.dot(Snew,Vt))
 
   # Generate data  
   x0,y,M,Lambda,realnoise = pyCSalgos.GAP.GAP.Generate_Data_Known_Omega(Omega, d,p,m,l,noiselevel, numvects,'l0');
+  
+  return Omega,x0,y,M,realnoise
 
-  # Values for lambda
-  #lambdas = [0 10.^linspace(-5, 4, 10)];
-  lambdas = np.concatenate((np.array([0]), 10**np.linspace(-5, 4, 10)))
+def runonce(algosN,algosL,Omega,y,lambdas,realnoise,M,x0):
+  
+  d = Omega.shape[1]  
+  
+  nalgosN = len(algosN)  
+  nalgosL = len(algosL)
   
   xrec = dict()
   err = dict()
   relerr = dict()
-  for i,algo in zip(np.arange(numalgos),algos):
-    xrec[algo[3]]   = np.zeros((lambdas.size, d, y.shape[1]))
-    err[algo[3]]    = np.zeros((lambdas.size, y.shape[1]))
-    relerr[algo[3]] = np.zeros((lambdas.size, y.shape[1]))
+
+  # Prepare storage variables for algorithms non-Lambda
+  for i,algo in zip(np.arange(nalgosN),algosN):
+    xrec[algo[1]]   = np.zeros((d, y.shape[1]))
+    err[algo[1]]    = np.zeros(y.shape[1])
+    relerr[algo[1]] = np.zeros(y.shape[1])
+  # Prepare storage variables for algorithms with Lambda    
+  for i,algo in zip(np.arange(nalgosL),algosL):
+    xrec[algo[1]]   = np.zeros((lambdas.size, d, y.shape[1]))
+    err[algo[1]]    = np.zeros((lambdas.size, y.shape[1]))
+    relerr[algo[1]] = np.zeros((lambdas.size, y.shape[1]))
   
+  # Run algorithms non-Lambda
+  for iy in np.arange(y.shape[1]):
+    for algofunc,strname in algosN:
+      epsilon = 1.1 * np.linalg.norm(realnoise[:,iy])
+      xrec[strname][:,iy] = algofunc(y[:,iy],M,Omega,epsilon)
+      err[strname][iy]    = np.linalg.norm(x0[:,iy] - xrec[strname][:,iy])
+      relerr[strname][iy] = err[strname][iy] / np.linalg.norm(x0[:,iy])
+  for algotuple in algosN:
+    print algotuple[1],' : avg relative error = ',np.mean(relerr[strname])  
+
+  # Run algorithms with Lambda
   for ilbd,lbd in zip(np.arange(lambdas.size),lambdas):
     for iy in np.arange(y.shape[1]):
-      for algosetupfunc,algofunc,algopostfunc,strname in algos:
+      D = np.linalg.pinv(Omega)
+      U,S,Vt = np.linalg.svd(D)
+      for algofunc,strname in algosL:
         epsilon = 1.1 * np.linalg.norm(realnoise[:,iy])
-        
-        inparams = algosetupfunc(y[:,iy],M,Omega,epsilon,lbd)
-        xrec[strname][ilbd,:,iy] = algopostfunc(algofunc(*inparams)[0])
-        
+        gamma = algofunc(y[:,iy],M,Omega,D,U,S,Vt,epsilon,lbd)
+        xrec[strname][ilbd,:,iy] = np.dot(D,gamma)
         err[strname][ilbd,iy]    = np.linalg.norm(x0[:,iy] - xrec[strname][ilbd,:,iy])
         relerr[strname][ilbd,iy] = err[strname][ilbd,iy] / np.linalg.norm(x0[:,iy])
-        
     print 'Lambda = ',lbd,' :'
-    for strname in relerr:
-      print '   ',strname,' : avg relative error = ',np.mean(relerr[strname][ilbd,:])
+    for algotuple in algosL:
+      print '   ',algotuple[1],' : avg relative error = ',np.mean(relerr[strname][ilbd,:])
 
-
-
+  # Prepare results
+  mrelerrN = dict()
+  for algotuple in algosN:
+    mrelerrN[algotuple[1]] = np.mean(relerr[algotuple[1]])
+  mrelerrL = dict()
+  for algotuple in algosL:
+    mrelerrL[algotuple[1]] = np.mean(relerr[algotuple[1]],1)
+  
+  return mrelerrN,mrelerrL
+  
 # Script main
 if __name__ == "__main__":
   mainrun()
\ No newline at end of file