def get_gradient(self, data, params):
(samples_and_r_scales, actions) = unzip(data)
(samples, r_scales) = unzip(samples_and_r_scales)
gradient = self._get_gradient(samples, r_scales, actions)
self._compute_M_step(params, actions)
return gradient
def get_objective(self, data, params):
(samples, actions) = unzip(data)
(rewards, r_scales) = unzip(samples)
r_scales = np.array(r_scales)[:, np.newaxis]
actions = np.array(actions)[:, np.newaxis]
shifted_samples = np.array(rewards) + \
np.hstack([r_scales, -r_scales])
densities = self._get_densities(shifted_samples)
terms = np.dot(densities, self.ps)
return -np.mean(np.log(terms), axis=0)[0]
def _update_history(self, Xs, missing):
normed = unzip(self.basis_pairs_t)[1]
if self.filtering_history is None:
self.filtering_history = [
np.dot(X[-1, :], Phi) for (X, Phi) in zip(Xs, normed)
]
if self.keep_basis_history:
self.basis_history = [[
Phi[:, i][:, np.newaxis].T for i in xrange(self.k)
] for Phi in normed]
else:
for i in xrange(self.num_views):
current = self.filtering_history[i]
# TODO: put meaningful filler here; zeroes are not meaningful
new = np.zeros((1, self.k))
if not missing[i]:
new = np.dot(Xs[i][-1, :], normed[i])
self.filtering_history[i] = np.vstack([current, new])
# Update basis history
if self.keep_basis_history:
for j in xrange(self.k):
current = self.basis_history[i][j]
new = np.copy(current[-1, :])
if not missing[i]:
new = normed[i][:, j]
self.basis_history[i][j] = np.vstack([current, new])
def run_things_all_day_bb(data_dir, num_data, k, ds, pi_factors, lazy):
ds = [int(d) for d in ds.split()]
angles = [float(pf) * np.pi for pf in pi_factors.split()]
pre_As = [np.random.randn(2 * d, d) for d in ds]
As = [np.dot(pre_A.T, pre_A) for pre_A in pre_As]
(_, Qs) = unzip([np.linalg.eig(A) for A in As])
zipped = zip(ds, angles, Qs)
dynamics = [get_rotation(d, a, Q, P_inv=Q.T) for (d, a, Q) in zipped]
loaders = get_easy_DCCAPMLs(dynamics, num_data, k, ds, lazy=lazy)
names = ['N', 'k', 'ds', 'pi_factors']
vals = [
str(num_data),
str(k), '-'.join([str(d) for d in ds]),
'-'.join([str(pf) for pf in pi_factors])
]
dirname = get_ts('easy_cca' + '_' +
'_'.join([n + '-' + v for (n, v) in zip(names, vals)]))
dirpath = os.path.join(data_dir, dirname)
os.mkdir(dirpath)
for (i, l) in enumerate(loaders):
filename = 'view_' + str(i) + '.csv'
filepath = os.path.join(dirpath, filename)
np.savetxt(filepath, l.get_data(), delimiter=',')
def run(self):
printerval = self.num_data / 10
for t in range(self.num_data):
zipped = zip(self.agvs, self.servers)
for (agv, ds) in zipped:
agv.set_data(ds.get_data())
if self.delay is None or t % self.delay == 0:
XPhis = [agv.get_projected() for agv in self.agvs]
for agv in self.agvs:
for (n, XPhi) in enumerate(XPhis):
agv.update_neighbor_state(n, XPhi)
(self.Phis,
new_tccs) = unzip([agv.get_update() for agv in self.agvs])
self.tccs.append(new_tccs)
if False: #t % printerval == 0:
print('tccs', new_tccs)
zipped = zip(self.Phis, self.loaders)
XPhis = [np.dot(get_zm_uv(dl.get_data()), Phi) for (Phi, dl) in zipped]
tcc_mats = [
np.dot(XPhis[n].T, XPhis[m]) for n in range(self.num_views)
for m in range(n, self.num_views)
]
tccs = [np.trace(tccm) / self.k for tccm in tcc_mats]
print(tccs)
def update(self, Xs, Sxs, missing, etas):
if self.basis_pairs_t is None:
#print 'Initializing basis_pairs_t'
# Initialization of optimization variables
self.basis_pairs_t = agu.get_init_basis_pairs(Sxs, self.k)
self.num_rounds += 1
#print 'Updating history'
self._update_history(Xs, missing)
if self.verbose:
print "\tGetting updated basis estimates"
# Get updated canonical bases
#print 'Setting basis_pairs_t1'
self.basis_pairs_t1 = self._get_basis_updates(Xs, Sxs, missing, etas)
(unn, normed) = unzip(self.basis_pairs_t1)
#print 'Computing loss'
loss = gu.misc.get_objective(Xs, normed)
if self.verbose:
print "\tObjective:", loss
self.loss_history.append(loss)
# Check for convergence
unnormed_pairs = zip(unzip(self.basis_pairs_t)[0], unn)
#print 'Checking for convergence'
pre_converged = gu.misc.is_converged(unnormed_pairs, self.epsilons,
self.verbose)
# This is because bases are unchanged for missing data
self.converged = [
False if missing[i] else c for (i, c) in enumerate(pre_converged)
]
# Update iterates
self.basis_pairs_t = [(np.copy(unn_Phi), np.copy(Phi))
for unn_Phi, Phi in self.basis_pairs_t1]
return (self.basis_pairs_t, loss)
def get_coordinate_counts(self, data):
actions = unzip(data)[1]
total = len(actions)
acted = sum(actions)
coord_counts = np.zeros(self.get_parameter_shape())
coord_counts[0:2, :] = total
coord_counts[2:, :] = acted
return coord_counts
def _get_S_and_A_servers(self):
(njks, phi_jks,
self.nks) = unzip([n.get_fsvrg_params() for n in self.nodes])
self.n = sum(self.nks)
njs = sum(njks)
phi_js = njs / self.n
sjks = [(phi_js / phi_jk) for phi_jk in phi_jks]
S_servers = [SDS(sjk) for sjk in sjks]
omega_js = np.vstack((njk != 0).astype(float) for njk in njks)
ajs = get_sp(omega_js / self.num_nodes, -1)
A_server = SDS(ajs)
return (S_servers, A_server)
def get_gradients(Xs, basis_pairs):
m = len(Xs)
X_transforms = [(np.dot(X, unnormed), np.dot(X, normed))
for (X, (unnormed, normed)) in zip(Xs, basis_pairs)]
info = zip(Xs, basis_pairs, X_transforms)
for thang in info:
X = thang[0]
(unnormed_Phi, normed_Phi) = thang[1]
(unnormed_Phi_X, normed_Phi_X) = thang[2]
#print 'X', X
#print 'unnormed_Phi', unnormed_Phi
#print 'normed_Phi', normed_Phi
#print 'unnormed_Phi_X', unnormed_Phi_X
#print 'normed_Phi_X', normed_Phi_X
drdb.check_for_large_numbers(unnormed_Phi_X,
'appgrad.utils get_gradients',
'unnormed_Phi_X')
drdb.check_for_large_numbers(normed_Phi_X,
'appgrad.utils get_gradients',
'normed_Phi_X')
minus_term = sum(unzip(X_transforms)[1])
drdb.check_for_large_numbers(minus_term, 'appgrad.utils get_gradients',
'minus_term')
diffs = [(m-1)*unnormed - \
(minus_term - normed)
for (unnormed, normed) in X_transforms]
for diff in diffs:
#print 'Diff', diff
drdb.check_for_large_numbers(diff, 'appgrad.utils get_gradients',
'diff')
gradients = [
np.dot(X.T, diff) / X.shape[0] for (X, diff) in zip(Xs, diffs)
]
for g in gradients:
drdb.check_for_large_numbers(g, 'appgrad.utils get_gradients',
'gradient')
return gradients
def get_status(self):
return {
'k': self.k,
'converged': self.converged,
'bases': unzip(self.basis_pairs_t)[1],
'num_views': self.num_views,
'epsilons': self.epsilons,
'num_rounds': self.num_rounds,
'num_iters': self.num_iters,
'has_been_fit': self.has_been_fit,
'basis_pairs': self.basis_pairs_t,
'basis_history': self.basis_history,
'loss_history': self.loss_history,
'filtering_history': self.filtering_history
}
def _compute_S_and_A_servers(self):
(njks, phi_jks,
self.nks) = unzip([n.get_fsvrg_params() for n in self.nodes])
self.n = sum(self.nks)
njs = sum(njks)
phi_js = njs / self.n
sjks = [(phi_js / phi_jk)[:, np.newaxis] for phi_jk in phi_jks]
self.S_servers = [SDS(sjk) for sjk in sjks]
for (n, S_s) in zip(self.nodes, self.S_servers):
n.set_S_server(S_s)
omega_js = np.vstack(
(njk[:, np.newaxis] != 0).astype(float) for njk in njks)
ajs = get_sp(omega_js / self.num_nodes, -1)
self.A_server = SDS(ajs)
def compute_parameters(self):
w_t = np.copy(self.init_params)
for t in xrange(self.num_rounds):
local_grads = [n.get_gradient(np.copy(w_t))
for n in self.nodes]
grad_t = sum(local_grads) / self.num_nodes
updates = [n.get_update(
np.copy(w_t),
np.copy(grad_t))
for n in self.nodes]
(ws, objectives) = unzip(updates)
# TODO: Make sure this is a good place to project
w_t = sum(ws) / self.num_nodes
self.objectives.append(sum(objectives))
self.w = w_t
def compute_parameters(self):
w_t = np.copy(self.init_params)
for i in xrange(self.max_rounds):
local_grads = [
n.get_gradient(n.get_local(w_t)) for n in self.nodes
]
grad_t = np.vstack(local_grads) / self.num_nodes
updates = [
n.get_update(np.copy(w_t), np.copy(grad_t)) for n in self.nodes
]
(ws, objectives) = unzip(updates)
agg = self._get_aggregate_grad(ws, w_t)
w_t = w_t + agg
self.objectives.append([o[-1] for o in objectives])
self.w = w_t
def compute_parameters(self):
w_t = np.copy(self.init_params)
for i in xrange(self.max_rounds):
grad_t = np.zeros_like(w_t)
if i > 0:
local_grads = [
n.get_gradient(n.get_local(w_t)) for n in self.nodes
]
grad_t = np.vstack(local_grads) / self.num_nodes
(S_servers, A_server) = self._get_S_and_A_servers()
for (S_s, n) in zip(S_servers, self.nodes):
n.set_global_info(w_t, grad_t)
n.set_S_server(S_s)
for j in xrange(2**i):
self.num_rounds += 1
for n in self.nodes:
n.set_local_action()
for n in self.nodes:
n.compute_local_update()
updates = [n.get_update() for n in self.nodes]
(ws, objectives) = unzip(updates)
agg = self._get_aggregate_grad(ws, w_t, A_server)
w_t = w_t + agg
for (i, o) in enumerate(objectives):
self.objectives[i].append(o)
#for n in self.nodes:
# print n.objectives
self.w = w_t
def get_bases(self):
if not self.has_been_fit:
raise Exception('Model has not yet been fit.')
return unzip(self.basis_pairs_t)[1]
