def _build_model(self, data, feats, states):
self._state0 = np.mean(states, 0)
# Statistics for clamping
self._max_state_norm2 = np.max(np.sum(states * states, 0))
self._max_state_norm = np.sqrt(self._max_state_norm2)
self._max_state_coord = np.max(states, axis=1).reshape((-1, 1))
self._min_state_coord = np.min(states, axis=1).reshape((-1, 1))
# Horizon Prediction
s2_h_in = ula.khatri_rao_rowwise(states, feats.fut_act)
W_h = ridge(s2_h_in, feats.fut_obs, self._lambda['pred'])
W_rff2fo = ridge(feats.fut_obs, data.fut_obs, self._lambda['pred'])
self._W_h = np.dot(W_h, W_rff2fo)
# 1-Step Prediction
s2_1s_in = ula.khatri_rao_rowwise(states, feats.act)
W_1s = ridge(s2_1s_in, feats.obs, self._lambda['pred'])
W_rff2obs = ridge(feats.obs, data.obs, self._lambda['pred'])
self._W_1s = np.dot(W_1s, W_rff2obs)
def _estimate_condop(self, A, B, C, reg_lambda, div_lambda, importance_weights):
N,da = A.shape
db = B.shape[1]
dab = da*db
dbb = db*db
reg_out = np.empty((N,dab+dbb))
reg_out[:,:dab] = ula.khatri_rao_rowwise(A,B)
reg_out[:,dab:] = ula.khatri_rao_rowwise(B,B)
W = ridge(C, reg_out, reg_lambda, importance_weights)
est_reg_out = np.dot(C,W)
output = np.empty((N,dab))
for i in xrange(N):
C_ab = est_reg_out[i,:dab].reshape((da,db))
C_bb = est_reg_out[i,dab:].reshape((db,db))
C_a_b = ula.reg_rdivide_nopsd(C_ab, C_bb, div_lambda)
output[i,:] = C_a_b.reshape(-1)
return output,W
def _s1_regression_cond(self, data, feats, imp_weights):
print('Stage 1A Regression')
s1a_in = ula.khatri_rao_rowwise(feats.past, feats.fut_act)
s1a_out = feats.fut_obs
W_s1a = ridge(s1a_in, s1a_out, self._lambda['s1a'], imp_weights[0])
W_s1a = W_s1a.reshape((self._feat_dim.past, self._feat_dim.fut_act,
self._feat_dim.fut_obs))
W_s1a = W_s1a.transpose((0, 2, 1))
W_s1a = W_s1a.reshape((self._feat_dim.past, -1))
states = np.dot(feats.past, W_s1a)
print('Stage 1B Regression')
s1b_in = ula.khatri_rao_rowwise(feats.past, feats.exfut_act)
s1b_out = feats.exfut_obs
W_s1b = ridge(s1b_in, s1b_out, self._lambda['s1b'], imp_weights[1])
W_s1b = W_s1b.reshape((self._feat_dim.past, self._feat_dim.exfut_act,
self._feat_dim.exfut_obs))
W_s1b = W_s1b.transpose((0, 2, 1))
W_s1b = W_s1b.reshape((self._feat_dim.past, -1))
ex_states = np.dot(feats.past, W_s1b)
print('Stage 1C Regression')
s1c_in = ula.khatri_rao_rowwise(feats.past, feats.act)
s1c_out = feats.oo
W_s1c = ridge(s1c_in, s1c_out, self._lambda['s1c'], None)
W_s1c = W_s1c.reshape(
(self._feat_dim.past, self._feat_dim.act, self._feat_dim.oo))
W_s1c = W_s1c.transpose((0, 2, 1))
W_s1c = W_s1c.reshape((self._feat_dim.past, -1))
im_states = np.dot(feats.past, W_s1c)
self._dbg_W_s1a = W_s1a
self._dbg_W_s1b = W_s1b
self._dbg_W_s1c = W_s1c
return states, ex_states, im_states
def _extract_feats(self, data, U_old = None):
# Past
print('Past')
feats = structtype()
self._fext_past = self._feature_set['past']
feats.past = self._fext_past.build(data.past).process(data.past)
# Immediate
print('Immediate')
self._fext_obs = self._feature_set['obs']
feats.obs = self._fext_obs.build(data.obs).process(data.obs)
self._fext_act = self._feature_set['act']
feats.act = self._fext_act.build(data.act).process(data.act)
# Future
print('Future')
self._fext_fut_obs = self._feature_set['fut_obs']
feats.fut_obs = self._fext_fut_obs.build(data.fut_obs).process(data.fut_obs)
self._fext_fut_act = self._feature_set['fut_act']
feats.fut_act = self._fext_fut_act.build(data.fut_act).process(data.fut_act)
'''
Project future into subspace predictable by history
'''
if self._past_projection == 'svd':
self._fext_fut_obs, feats.fut_obs = self._project_on_past(self._fext_fut_obs, feats.fut_obs, feats.past)
self._fext_fut_act, feats.fut_act = self._project_on_past(self._fext_fut_act, feats.fut_act, feats.past)
else:
assert self._past_projection is None, 'past projection is not None'
# Shifted Future
print('Shifted Future')
feats.shfut_obs = self._fext_fut_obs.process(data.shfut_obs)
feats.shfut_act = self._fext_fut_act.process(data.shfut_act)
print('Derived Features')
# Derived Features:
# Extended Future
# Note that for exteded future observation, the current observation is
# the "lower order" factor. This makes filtering easier.
feats.exfut_obs = ula.khatri_rao_rowwise(feats.shfut_obs, feats.obs)
feats.exfut_act = ula.khatri_rao_rowwise(feats.act, feats.shfut_act)
self._U_efo, _, exfut_obs = ula.rand_svd_f(feats.exfut_obs.T, k=self._p, rng=self.rng)
feats.exfut_obs = exfut_obs.T
self._U_efa, _, exfut_act = ula.rand_svd_f(feats.exfut_act.T, k=self._p, rng=self.rng)
feats.exfut_act = exfut_act.T
# Observation Covariance
feats.oo = ula.khatri_rao_rowwise(feats.obs, feats.obs)
# Project lower triangle part of observation covariance
d_obs = feats.obs.shape[1]
lt_idx = np.array([i for i in xrange(d_obs**2) if (i//d_obs) >= (i%d_obs)])
ut_idx = (lt_idx%d_obs)*d_obs + lt_idx//d_obs
feats.oo = feats.oo[:,lt_idx]
self._U_oo,_,oo = ula.rand_svd_f(feats.oo.T, k=self._p, rng=self.rng)
feats.oo = oo.T
# Convert singular vectors back to represent symmetric matrices
Usym = np.empty((d_obs**2, self._U_oo.shape[1]))
Usym[lt_idx,:] = self._U_oo
Usym[ut_idx,:] = self._U_oo
self._U_oo = Usym
K = structtype()
K.obs = feats.obs.shape[1]
K.act = feats.act.shape[1]
K.past = feats.past.shape[1]
K.fut_obs = feats.fut_obs.shape[1]
K.fut_act = feats.fut_act.shape[1]
K.exfut_obs = feats.exfut_obs.shape[1]
K.exfut_act = feats.exfut_act.shape[1]
K.oo = feats.oo.shape[1]
self._feat_dim = K
for (k,v) in K.__dict__.items():
print('%s:%d' % (k,v))
return feats