removed matlab files (50680bf4) · Commits · 钟慕尧 / tvgl

.gitignore

+1 −0

Original line number	Diff line number	Diff line
		*.pyc
		*~
		drafts/*
		results/*

DataHandler.py

+9 −1

Original line number	Diff line number	Diff line
		@@ -139,7 +139,15 @@ class DataHandler(object):
		""" Write networks """
		f.write("\n\n#Networks:\n\n")
		for k in range(alg.blocks):
		f.write("Block %s\n " % k)
		f.write("Block %s," % k)
		f.write(alg.blockdates[k] + "\n")
		if k > 0:
		f.write("Dev to prev,")
		f.write("{0:.3f},".format(alg.deviations[k-1]))
		if k < alg.blocks - 1:
		f.write("Dev to next,")
		f.write("{0:.3f}".format(alg.deviations[k]))
		f.write("\n")
		for feat in feats:
		f.write("," + feat)
		f.write("\n")

ParallelTVGL.py

+17 −8

Original line number	Diff line number	Diff line
		@@ -12,7 +12,7 @@ MAX_ITER = 10000
		def mp_parallel_tvgl((thetas, z0s, z1s, z2s, u0s, u1s, u2s,
		emp_cov_mat, lambd, beta, rho, nju,
		indexes, out_queue, prev_pipe, next_pipe,
		proc_index, stopping_criteria),
		proc_index, stopping_criteria, pen_func),
		last=False):
		try:
		iteration = 0
		@@ -38,7 +38,8 @@ def mp_parallel_tvgl((thetas, z0s, z1s, z2s, u0s, u1s, u2s,
		for j, i in zip(indexes[:end], range(nn)):
		a = (z0s[i] + z1s[i] + z2s[i] - u0s[i] - u1s[i] - u2s[i])/3
		at = a.transpose()
		m = nju*(a + at)/2 - emp_cov_mat[i]
		#m = nju*(a + at)/2 - emp_cov_mat[i]
		m = (a + at)/(2 * nju) - emp_cov_mat[i]
		d, q = np.linalg.eig(m)
		qt = q.transpose()
		sqrt_matrix = np.sqrt(d*2 + 4/njunp.ones(dimension))
		@@ -59,7 +60,7 @@ def mp_parallel_tvgl((thetas, z0s, z1s, z2s, u0s, u1s, u2s,
		""" Z1-Z2 Update """
		for i in range(1, n):
		a = thetas[i] - thetas[i-1] + u2s[i] - u1s[i-1]
		e = pf.group_lasso_penalty(a, beta, rho)
		e = getattr(pf, pen_func)(a, beta, rho)
		summ = thetas[i] + thetas[i-1] + u2s[i] + u1s[i-1]
		z1s[i-1] = 0.5*(summ - e)
		z2s[i] = 0.5*(summ + e)
		@@ -131,7 +132,8 @@ class ParallelTVGL(TVGL):
		None,
		self.pipes[i][0],
		i,
		stopping_criteria),))
		stopping_criteria,
		self.penalty_function),))
		elif i == self.processes - 1:
		p = multiprocessing.Process(target=mp_parallel_tvgl,
		args=((self.thetas[self.chunk * i:],
		@@ -151,7 +153,8 @@ class ParallelTVGL(TVGL):
		self.pipes[i-1][1],
		None,
		i,
		stopping_criteria), True))
		stopping_criteria,
		self.penalty_function), True))
		else:
		p = multiprocessing.Process(target=mp_parallel_tvgl,
		args=((self.thetas[self.chunk * i:self.chunk*(i+1)+1],
		@@ -171,7 +174,8 @@ class ParallelTVGL(TVGL):
		self.pipes[i-1][1],
		self.pipes[i][0],
		i,
		stopping_criteria),))
		stopping_criteria,
		self.penalty_function),))
		self.procs.append(p)

		def run_algorithm(self, max_iter=10000):
		@@ -190,6 +194,11 @@ class ParallelTVGL(TVGL):
		for p in self.procs:
		p.join()
		self.iteration = iteration
		print self.iteration
		#print self.iteration
		self.run_time = '{0:.3g}'.format(time.time() - start_time)
		self.thetas = [np.round(theta, self.roundup) for theta in self.thetas]
		self.final_tuning(True, MAX_ITER)
		#self.thetas = [np.round(theta, self.roundup) for theta in self.thetas]

		def terminate_pools(self):
		for p in self.procs:
		p.terminate()

TVGL.py

+32 −6

Original line number	Diff line number	Diff line
		@@ -9,13 +9,17 @@ class TVGL(object):
		np.set_printoptions(precision=3)

		def __init__(self, filename, blocks, lambd, beta,
		processes, penalty_function="Group Lasso"):
		processes, penalty_function="group_lasso",
		datecolumn=True):
		self.datecolumn = datecolumn
		self.processes = processes
		self.blocks = blocks
		self.penalty_function = penalty_function
		self.dimension = None
		self.emp_cov_mat = [0] * self.blocks
		self.real_thetas = [0] * self.blocks
		if self.datecolumn:
		self.blockdates = [0] * self.blocks
		self.read_data(filename)
		self.rho = self.get_rho()
		self.max_step = 0.1
		@@ -35,7 +39,7 @@ class TVGL(object):
		self.e = 1e-5
		self.roundup = 1

		def read_data(self, filename, comment="#", splitter=",", datecolumn=True):
		def read_data(self, filename, comment="#", splitter=","):
		with open(filename, "r") as f:
		comment_count = 0
		for i, line in enumerate(f):
		@@ -43,7 +47,7 @@ class TVGL(object):
		comment_count += 1
		else:
		if self.dimension is None:
		if datecolumn:
		if self.datecolumn:
		self.dimension = len(line.split(splitter)) - 1
		else:
		self.dimension = len(line.split(splitter))
		@@ -61,7 +65,9 @@ class TVGL(object):
		if i == 1:
		self.generate_real_thetas(line, splitter)
		continue
		if datecolumn:
		if count == 0 and self.datecolumn is True:
		start_date = line.strip().split(splitter)[0]
		if self.datecolumn:
		lst.append([float(x)
		for x in np.array(line.strip().
		split(splitter)[1:])])
		@@ -71,6 +77,9 @@ class TVGL(object):
		split(splitter))])
		count += 1
		if count == self.obs:
		if self.datecolumn:
		end_date = line.strip().split(splitter)[0]
		self.blockdates[block] = start_date + " - " + end_date
		datablck = np.array(lst)
		tp = datablck.transpose()
		self.emp_cov_mat[block] = np.real(
		@@ -176,20 +185,37 @@ class TVGL(object):

		def final_tuning(self, stopping_criteria, max_iter):
		self.thetas = [np.round(theta, self.roundup) for theta in self.thetas]
		#self.only_true_false_edges()
		self.terminate_pools()
		if stopping_criteria:
		print "\nIterations to complete: %s" % self.iteration
		else:
		print "\nMax iterations (%s) reached" % max_iter

		def only_true_false_edges(self):
		for k in range(self.blocks):
		for i in range(self.dimension - 1):
		for j in range(i + 1, self.dimension):
		if self.thetas[k][i, j] != 0:
		self.thetas[k][i, j] = 1
		self.thetas[k][j, i] = 1
		else:
		self.thetas[k][i, j] = 0
		self.thetas[k][j, i] = 0

		def temporal_deviations(self):
		deviations = np.zeros(self.blocks - 1)
		for i in range(0, self.blocks - 1):
		dif = self.thetas[i+1] - self.thetas[i]
		np.fill_diagonal(dif, 0)
		deviations[i] = np.linalg.norm(dif)
		print deviations
		try:
		self.deviations = deviations/max(deviations)
		self.dev_ratio = float(max(deviations))/float(np.mean(deviations))
		except ZeroDivisionError:
		self.deviations = deviations
		self.dev_ratio = 0
		print "Temp deviations ratio: {0:.3g}".format(self.dev_ratio)

		def correct_edges(self):

matlab/admm_algorithm.m

deleted100644 → 0

+0 −65

Original line number	Diff line number	Diff line
		function Theta = admm_algorithm(Theta,Z0,Z1,Z2,U0,U1,U2,...
		rho,lambda,beta,nju,empirical_covariance_matrices)

		max_iteration = 20000;
		index = 0;
		stopping_criteria = 0;

		total_stamps = size(Theta,1);
		dimension = size(Theta{1},1);

		while index < max_iteration && stopping_criteria == 0

		%% Theta-Update
		for i = 1:total_stamps
		A = (Z0{i}+Z1{i}+Z2{i}-U0{i}-U1{i}-U2{i})/3;
		M = nju*(A+A')/2 - empirical_covariance_matrices{i};
		[Q,D] = eig(M)

		Theta{i} = nju/2Q(D+sqrt(D^2 + 4/njudiag(ones(dimension,1))))Q';
		end

		%% Z0-Update
		for i = 1:total_stamps
		Z0{i} = soft_thres_od(Theta{i} + U0{i},lambda,rho); % soft threshold odd
		end

		%% Z1,Z2-Update
		for i = 2:total_stamps
		A = Theta{i} - Theta{i-1} + U2{i} - U1{i-1};
		E = group_lasso_penalty(A,2*beta/rho);
		Z_pre = 0.5*[Theta{i-1}+Theta{i}+U1{i}+U2{i};Theta{i-1}+Theta{i}+U1{i}+U2{i}] + ...
		0.5*[-E;E];
		Z1{i-1} = Z_pre(1:dimension,:);
		Z2{i} = Z_pre(dimension+1:end,:);
		end

		%% U0-Update
		for i = 1:total_stamps
		U0{i} = U0{i} + Theta{i} - Z0{i};
		end

		%% U1,U2-Update
		for i = 2:total_stamps
		U1{i-1} = U1{i-1} + Theta{i-1} - Z1{i-1};
		U2{i} = U2{i} + Theta{i} - Z2{i};
		end

		%% Stopping condition
		if index > 0
		fro_norm = 0;
		for i = 1:total_stamps
		dif = Theta{i} - Theta_pre{i};
		fro_norm = fro_norm + norm(dif,'fro');
		end
		if fro_norm < 1e-5
		stopping_criteria = 1;
		end
		end
		Theta_pre = Theta;
		index = index + 1;

		end
		index

		end
		No newline at end of file

Admin message