Merge pull request #129 from daducci/fix_fista_update

[FIX] Fista update

commit/proximals.pyx

@@ -32,10 +32,16 @@ cpdef soft_thresholding(double [::1] x, double lam, int compartment_start, int c
     cdef:
         int i
     for i in xrange(compartment_start, compartment_start+compartment_size):
-        if x[i] <= lam:
-            x[i] = 0.0
-        else:
+        # if x[i] <= lam:
+        #     x[i] = 0.0
+        # else:
+        #     x[i] = x[i] - lam
+        if x[i] > lam:
             x[i] = x[i] - lam
+        elif x[i] < -lam:
+            x[i] = x[i] + lam
+        else:
+            x[i] = 0.0
     return np.asarray( x )
 
 
@@ -71,29 +77,18 @@ cpdef omega_group_sparsity(double [::1] x, int [::1] group_idx, int [::1] group_
         double omega = 0.0, gNorm, x_i
 
     if lam != 0:
-        if n == 2:
-            for k in xrange(nG):
-                N = group_size[k]
-                gNorm = 0.0
-                for i in xrange(j,j+N) :
-                    x_i = x[group_idx[i]]
-                    gNorm += x_i*x_i
-                omega += group_weight[k] * sqrt( gNorm )
-                j += N
-        elif n == np.inf:
-            for k in xrange(nG):
-                N = group_size[k]
-                gNorm = x[group_idx[j]]
-                for i in xrange(j+1,j+N) :
-                    x_i = x[group_idx[i]]
-                    if x_i > gNorm :
-                        gNorm = x_i
-                omega += group_weight[k] * gNorm
-                j += N
+        for k in xrange(nG):
+            N = group_size[k]
+            gNorm = 0.0
+            for i in xrange(j,j+N) :
+                x_i = x[group_idx[i]]
+                gNorm += x_i*x_i
+            omega += group_weight[k] * sqrt( gNorm )
+            j += N
     return lam*omega
 
 
-cpdef prox_group_sparsity( double [::1] x, int [::1] group_idx, int [::1] group_size, double [::1] group_weight, double lam, double n ) :
+cpdef prox_group_sparsity( double [::1] x, int [::1] group_idx, int [::1] group_size, double [::1] group_weight, double lam) :
     """
     References:
         [1] Jenatton et al. - `Proximal Methods for Hierarchical Sparse Coding`
@@ -109,33 +104,21 @@ cpdef prox_group_sparsity( double [::1] x, int [::1] group_idx, int [::1] group_
             x[i] = 0.0
 
     if lam != 0:
-        if n == 2 :
-            for k in xrange(nG) :
-                N = group_size[k]
-                gNorm = 0.0
-                for i in xrange(j,j+N) :
-                    x_i = x[group_idx[i]]
-                    gNorm += x_i*x_i
-                gNorm = sqrt( gNorm )
+        for k in xrange(nG) :
+            N = group_size[k]
+            gNorm = 0.0
+            for i in xrange(j,j+N) :
+                x_i = x[group_idx[i]]
+                gNorm += x_i*x_i
+            gNorm = sqrt( gNorm )
 
-                wl = group_weight[k] * lam
-                if gNorm <= wl :
-                    for i in xrange(j,j+N) :
-                        x[ group_idx[i] ] = 0.0
-                else :
-                    wl = (gNorm-wl)/gNorm
-                    for i in xrange(j,j+N) :
-                        x[ group_idx[i] ] *= wl
-                j += N
-        # elif n == np.inf :
-        # [TODO] TO be correctly implemented
-        #     for k in range(nG) :
-        #         idx = subtree[k]
-        #         # xn = max( v[idx] )
-        #         r = weight[k] * lam
-        #         for i in idx :
-        #             if v[i] <= r:
-        #                 v[i] = 0.0
-        #             else :
-        #                 v[i] -= r
+            wl = group_weight[k] * lam
+            if gNorm <= wl :
+                for i in xrange(j,j+N) :
+                    x[ group_idx[i] ] = 0.0
+            else :
+                wl = (gNorm-wl)/gNorm
+                for i in xrange(j,j+N) :
+                    x[ group_idx[i] ] *= wl
+            j += N
     return np.asarray( x )