def fastICAall(data,g,dg,resize = True):
nIC = data.shape[1]
data = center(data)
data = mix.whiten(data)
w = np.array([random_w(nIC) for i in range(nIC)])
epsilon = 0
w_temp = np.zeros_like(w)
while(not converge(epsilon,1e-5)):
print("Epsilon is currently ", epsilon)
for i in range(nIC):
w_temp[i] = newton_step(w[i],data,g,dg)
w_temp[i] /= np.linalg.norm(w_temp[i])
W = sym_decorrelation(w_temp)
epsilon = max(abs(np.diag(np.dot(W, w.T))))
w = np.copy(W)
# recover unmixed signals
dataest = np.array([np.dot(data,w[i]) for i in range(nIC)])
if resize:
dataest = dataest.T*(256- 127)/np.max(np.abs(dataest), axis = 1)
return (dataest + 127).T
else:
return dataest
def fastICA(data,g,dg):
data = center(data)
data = mix.whiten(data)
w = random_w(data.shape[1])
# epsilon is used to know if w converges
epsilon = 0
nbsteps = 0
# max nb of steps. Used to test different functons ( used because
# the algorithm might diverge)
MAX = 100
while(not converge(epsilon,1e-20) and nbsteps <= MAX):
w_temp = newton_step(w,data,g,dg)
w_temp /= np.linalg.norm(w_temp)
epsilon = np.dot(w,w_temp)
w = w_temp
nbsteps += 1
# if algorithm diverged , the nbsteps is negative
if nbsteps > MAX:
nbsteps = -1
return np.dot(data,w),np.dot(data,np.array([w[1],-w[0]])),nbsteps
