def obj(w):
p = unpack(w)
p.update(fixed_params)
f = 0.0
# ln_p_a, ln_p_mix = self.class_prior()
ln_p_a = np.log(self.action(p)) # individual- and time-invariant
logits_mix = p[self.mixture_param_key]
ln_p_mix = logits_mix - logsumexp(logits_mix)
for y, x in samples:
# Outcome model
mixture = log_likelihood(p, y, x, self.mean, self.cov, self.tr,
ln_p_a, ln_p_mix)
f -= logsumexp(np.array(mixture))
# Action model
_, rx = x
f -= action_log_likelihood(rx, ln_p_a, self.tr_cont_flag)
# Regularizers
for k, _ in trainable_params.items():
if k.endswith('_F'):
f += np.sum(p[k]**2)
return f
def sinkhorn_logspace(logP, niters=10):
for _ in range(niters):
# Normalize columns and take the log again
logP = logP - logsumexp(logP, axis=0, keepdims=True)
# Normalize rows and take the log again
logP = logP - logsumexp(logP, axis=1, keepdims=True)
return logP
def hmm_expected_states(log_pi0, log_Ps, ll):
T, K = ll.shape
# Make sure everything is C contiguous
log_pi0 = to_c(log_pi0)
log_Ps = to_c(log_Ps)
ll = to_c(ll)
alphas = np.zeros((T, K))
forward_pass(log_pi0, log_Ps, ll, alphas)
normalizer = logsumexp(alphas[-1])
betas = np.zeros((T, K))
backward_pass(log_Ps, ll, betas)
expected_states = alphas + betas
expected_states -= logsumexp(expected_states, axis=1, keepdims=True)
expected_states = np.exp(expected_states)
expected_joints = alphas[:-1,:,None] + betas[1:,None,:] + ll[1:,None,:] + log_Ps
expected_joints -= expected_joints.max((1,2))[:,None, None]
expected_joints = np.exp(expected_joints)
expected_joints /= expected_joints.sum((1,2))[:,None,None]
return expected_states, expected_joints, normalizer
def hmm_expected_states(log_pi0, log_Ps, ll):
T, K = ll.shape
# Make sure everything is C contiguous
to_c = lambda arr: np.copy(arr, 'C') if not arr.flags['C_CONTIGUOUS'] else arr
log_pi0 = to_c(getval(log_pi0))
log_Ps = to_c(getval(log_Ps))
ll = to_c(getval(ll))
alphas = np.zeros((T, K))
forward_pass(log_pi0, log_Ps, ll, alphas)
normalizer = logsumexp(alphas[-1])
betas = np.zeros((T, K))
backward_pass(log_Ps, ll, betas)
expected_states = alphas + betas
expected_states -= logsumexp(expected_states, axis=1, keepdims=True)
expected_states = np.exp(expected_states)
expected_joints = alphas[:-1,:,None] + betas[1:,None,:] + ll[1:,None,:] + log_Ps
expected_joints -= expected_joints.max((1,2))[:,None, None]
expected_joints = np.exp(expected_joints)
expected_joints /= expected_joints.sum((1,2))[:,None,None]
return expected_states, expected_joints, normalizer
def location_mixture_logpdf(samps, locations, location_weights, distr_at_origin, contr_var = False, variant = 1):
# lpdfs = zeroprop.logpdf()
diff = samps - locations[:, np.newaxis, :]
lpdfs = distr_at_origin.logpdf(diff.reshape([np.prod(diff.shape[:2]), diff.shape[-1]])).reshape(diff.shape[:2])
logprop_weights = log(location_weights/location_weights.sum())[:, np.newaxis]
if not contr_var:
return logsumexp(lpdfs + logprop_weights, 0)
#time_m1 = np.hstack([time0[:,:-1],time0[:,-1:]])
else:
time0 = lpdfs + logprop_weights + log(len(location_weights))
if variant == 1:
time1 = np.hstack([time0[:,1:],time0[:,:1]])
cov = np.mean(time0**2-time0*time1)
var = np.mean((time0-time1)**2)
lpdfs = lpdfs - cov/var * (time0-time1)
return logsumexp(lpdfs - log(len(location_weights)), 0)
elif variant == 2:
cvar = (time0[:,:,np.newaxis] -
np.dstack([np.hstack([time0[:, 1:], time0[:, :1]]),
np.hstack([time0[:,-1:], time0[:,:-1]])]))
## self-covariance matrix of control variates
K_cvar = np.diag(np.mean(cvar**2, (0, 1)))
#add off diagonal
K_cvar = K_cvar + (1.-np.eye(2)) * np.mean(cvar[:,:,0]*cvar[:,:,1])
## covariance of control variates with random variable
cov = np.mean(time0[:,:,np.newaxis] * cvar, 0).mean(0)
optimal_comb = np.linalg.inv(K_cvar) @ cov
lpdfs = lpdfs - cvar @ optimal_comb
return logsumexp(lpdfs - log(len(location_weights)), 0)
def hmm_logZ_python(natparam):
init_params, pair_params, node_params = natparam
log_alpha = init_params + node_params[0]
for node_param in node_params[1:]:
log_alpha = logsumexp(log_alpha[:,None] + pair_params, axis=0) + node_param
return logsumexp(log_alpha)
def single_episode_log_partition_function(episode):
log_p_state = log_p_init
for action, rendering in episode:
log_p_state = (logsumexp(
log_p_state[:, None] + log_p_dynamics[action], axis=0) +
log_p_render[:, rendering])
return logsumexp(log_p_state)
def hmm_expected_states(log_pi0, log_Ps, ll, memlimit=2**31):
T, K = ll.shape
# Make sure everything is C contiguous
log_pi0 = to_c(log_pi0)
log_Ps = to_c(log_Ps)
ll = to_c(ll)
alphas = np.zeros((T, K))
forward_pass(log_pi0, log_Ps, ll, alphas)
normalizer = logsumexp(alphas[-1])
betas = np.zeros((T, K))
backward_pass(log_Ps, ll, betas)
# Compute E[z_t] for t = 1, ..., T
expected_states = alphas + betas
expected_states -= logsumexp(expected_states, axis=1, keepdims=True)
expected_states = np.exp(expected_states)
# Compute E[z_t, z_{t+1}] for t = 1, ..., T-1
# Note that this is an array of size T*K*K, which can be quite large.
# To be a bit more frugal with memory, first check if the given log_Ps
# are TxKxK. If so, instantiate the full expected joints as well, since
# we will need them for the M-step. However, if log_Ps is 1xKxK then we
# know that the transition matrix is stationary, and all we need for the
# M-step is the sum of the expected joints.
stationary = (log_Ps.shape[0] == 1)
if not stationary:
expected_joints = alphas[:-1, :, None] + betas[1:, None, :] + ll[
1:, None, :] + log_Ps
expected_joints -= expected_joints.max((1, 2))[:, None, None]
expected_joints = np.exp(expected_joints)
expected_joints /= expected_joints.sum((1, 2))[:, None, None]
else:
# Compute the sum over time axis of the expected joints
# Limit ourselves to approximately 1GB of memory, assuming
# the entries are float64's (8 bytes)
batch_size = int(memlimit / (8 * K * K))
assert batch_size > 0
expected_joints = np.zeros((1, K, K))
for start in range(0, T - 1, batch_size):
stop = min(T - 1, start + batch_size)
# Compute expectations in this batch
tmp = alphas[start:stop, :,
None] + betas[start + 1:stop + 1,
None, :] + ll[start + 1:stop + 1,
None, :] + log_Ps
tmp -= tmp.max((1, 2))[:, None, None]
tmp = np.exp(tmp)
tmp /= tmp.sum((1, 2))[:, None, None]
expected_joints += tmp.sum(axis=0)
return expected_states, expected_joints, normalizer
def log_partition_function(natural_params, data):
if isinstance(data, list):
return sum(map(partial(log_partition_function, natural_params), data))
log_pi, log_A, log_B = natural_params
log_alpha = log_pi
for y_t in data:
log_alpha = logsumexp(log_alpha[:,None] + log_A, axis=0) + log_B[:,y_t]
return logsumexp(log_alpha)
def energy(w, X, y, v_prior, m_prior, K, N, alpha):
"""Extract parameters"""
q = get_parameters_q(w, v_prior)
v_noise = np.exp(parser.get(w, 'log_v_noise')[ 0, 0 ])
"""Note: A-approx computes its own log_factor value inside the helper
function log_Z_likelihood, so we can shave of some computation time"""
if Dtype != "A-approx":
samples_q = draw_samples(q, K)
log_factor_value = 1.0 * N * log_likelihood_factor(samples_q, v_noise, X, y)
if Dtype == "KL":
"""I.e., standard VI"""
KL = np.sum(-0.5 * np.log(2 * math.pi * v_prior) - 0.5 * ((q[ 'm' ]-m_prior)**2 + q[ 'v' ]) / v_prior) - \
np.sum(-0.5 * np.log(2 * math.pi * q[ 'v' ] * np.exp(1)))
vfe = -(np.mean(log_factor_value) + KL)
elif Dtype == "AR-approx":
"""NOTE: Needs modification to be GVI"""
logp0 = log_prior(samples_q, v_prior, m_prior)
logq = log_q(samples_q, q)
logF = logp0 + log_factor_value - logq
logF = (1 - alpha) * logF
vfe = -(logsumexp(logF) - np.log(K))
vfe = vfe / (1 - alpha)
elif Dtype == "AB-approx":
logp0 = log_prior(samples_q, v_prior, m_prior)
logq = log_q(samples_q, q)
part1 = (alpha + beta_D) * (log_factor_value + logp0) - logq
part2 = (alpha + beta_D -1 ) * logq
part3 = (beta_D * (log_factor_value + logp0) + (alpha - 1) * logq)
vfe = ( (1.0 / (alpha * (alpha + beta_D))) * (logsumexp(part1) - np.log(K))
+ (1.0 / (beta_D * (alpha + beta_D))) * (logsumexp(part2) - np.log(K))
- (1.0 / (alpha * beta_D)) * (logsumexp(part3) - np.log(K)))
elif Dtype == "A-approx":
f_hat = get_parameters_f_hat(q, v_prior, m_prior, N)
vfe = -log_normalizer(q) - 1.0 * N / X.shape[ 0 ] / alpha * log_Z_likelihood(q, f_hat,
v_noise, X, y, K) + log_Z_prior(v_prior,m_prior)
elif Dtype == "AR":
prior_reg = (1/(alpha*(alpha-1))) * prior_regularizer(q,v_prior, m_prior,alpha)
vfe = -np.mean(log_factor_value) + prior_reg
#NOTE: While this should work, this is the alpha-divergence regularizer, which
# overconcentrates substantially. We refer to the appendix of our
# paper for some visuals on this phenomenon. The performance from this
# divergence should be expected to be much worse than that for the
# Alpha-renyi as Uncertainty Quantifier
elif Dtype == "A":
prior_reg = (1/(alpha*(alpha-1))) * (
np.exp(prior_regularizer(q,v_prior, m_prior,alpha))-1)
vfe = -np.mean(log_factor_value) + prior_reg
return vfe
def single_update_belief_log_probas(prev_belief_log_proba_K,
curr_data_log_proba_K, ltrans, a):
trans_log_proba_KK = ltrans[:, int(a), :]
curr_belief_log_proba_K = logsumexp(trans_log_proba_KK +
prev_belief_log_proba_K)
curr_belief_log_proba_K = curr_belief_log_proba_K + curr_data_log_proba_K
log_norm_const = logsumexp(curr_belief_log_proba_K)
cur_belief_log_proba_K = curr_belief_log_proba_K - log_norm_const
return curr_belief_log_proba_K
def predict_half(self, X_top): """ plot the top half the image concatenated with the marginal distribution over each pixel in the bottom half. """ X_bot = np.zeros((X_top.shape[0], X_top.shape[1])) theta_top, theta_bot = self.theta[:, :392].T, self.theta[:, 392:].T for i in range(392): constant = np.dot(X_top, np.log(theta_top)) + np.dot(1 - X_top, np.log(1 - theta_top)) X_bot[:, i] = logsumexp(np.add(constant, np.log(theta_bot[i])), axis=1) - logsumexp(constant, axis=1) save_images(np.concatenate((X_top, np.exp(X_bot)), axis=1), "predict_half.png")
def log_partition_function(natural_params, data):
if isinstance(data, list):
return sum(map(partial(log_partition_function, natural_params), data))
log_pi, log_A, log_B = natural_params
log_alpha = log_pi
for y_t in data:
log_alpha = logsumexp(log_alpha[:, None] + log_A, axis=0) + log_B[:,
y_t]
return logsumexp(log_alpha)
def build_pomdp(pi, trans, emission_mu, emission_std, data, fcpt, args):
lpi = pi - logsumexp(pi, axis=0)
ltrans = trans - logsumexp(trans, axis=-1, keepdims=True)
ll = 0
lbelief_state_set_TK = None
# collect the complete set of beliefs over all sequences
ll, lbelief_state_set_TK = calc_log_proba_for_many_sequences(
lpi, ltrans, emission_mu, emission_std, data, fcpt, args)
return lbelief_state_set_TK, ll
def label_meanfield(label_global, gaussian_globals, gaussian_stats):
partial_contract = lambda a, b: \
sum(np.tensordot(x, y, axes=np.ndim(y)) for x, y, in zip(a, b))
gaussian_local_natparams = map(niw.expectedstats, gaussian_globals)
node_params = np.array([
partial_contract(gaussian_stats, natparam) for natparam in gaussian_local_natparams]).T
local_natparam = dirichlet.expectedstats(label_global) + node_params
stats = normalize(np.exp(local_natparam - logsumexp(local_natparam, axis=1, keepdims=True)))
vlb = np.sum(logsumexp(local_natparam, axis=1)) - contract(stats, node_params)
return local_natparam, stats, vlb
def logpdf(self, x):
comp_logpdf = np.array([self.dist_cat.logpdf(i)+ self.comp_dist[i].logpdf(x)
for i in range(len(self.comp_dist))])
rval = logsumexp(comp_logpdf, 0)
if len(comp_logpdf.shape) > 1:
rval = rval.reshape((rval.size, 1))
return rval
def logpdf_grad(self, x):
rval = np.array([exp(self.dist_cat.logpdf(i))* self.comp_dist[i].logpdf_grad(x)
for i in range(len(self.comp_dist))])
rval = logsumexp(rval, 0)
return rval
def categorical_logpdf(data, logits, mask=None):
"""
Compute the log probability density of a categorical distribution.
This will broadcast as long as data and logits have the same
(or at least compatible) leading dimensions.
Parameters
----------
data : array_like (..., D) int (0 <= data < C)
The points at which to evaluate the log density
lambdas : array_like (..., D, C)
The logits of the categorical distribution(s) with C classes
mask : array_like (..., D) bool
Optional mask indicating which entries in the data are observed
Returns
-------
lps : array_like (...,)
Log probabilities under the categorical distribution(s).
"""
D = data.shape[-1]
C = logits.shape[-1]
assert data.dtype in (int, np.int8, np.int16, np.int32, np.int64)
assert np.all((data >= 0) & (data < C))
assert logits.shape[-2] == D
# Check mask
mask = mask if mask is not None else np.ones_like(data, dtype=bool)
assert mask.shape == data.shape
logits = logits - logsumexp(logits, axis=-1, keepdims=True) # (..., D, C)
x = one_hot(data, C) # (..., D, C)
lls = np.sum(x * logits, axis=-1) # (..., D)
return np.sum(lls * mask, axis=-1) # (...,)
def logloss(K_conj):
"""
K is a tensor of CONJUGATE Kraus Operators of dim s x y x x x x
s: dim of features
y: number of features
x: number of labels
"""
total_loss = 0.0
# Iterate over each sequence in batch
for i in range(labels.shape[0]):
features = feats_matrix[i, :]
label = labels[i] - 1
# Compute likelihood of the label generating the given features
conjKrausProduct = np.log(K_conj[features[0] - 1, 0, :, :])
for s in range(1, features.shape[0]):
conjKrausProduct = logdotexp(
np.log(K_conj[features[s] - 1, s, :, :]), conjKrausProduct)
eta = np.zeros([K_conj.shape[3], K_conj.shape[3]],
dtype='complex128')
eta[label, label] = 1
prod1 = logdotexp(np.conjugate(conjKrausProduct), np.log(eta))
prod2 = logdotexp(prod1, conjKrausProduct.T)
total_loss += np.real(logsumexp(np.diag(prod2)))
# total_loss += np.real(np.trace(np.kron(np.conjugate(conjKrausProduct)[:, label], conjKrausProduct.T[:, label]).reshape(K_conj.shape[2], K_conj.shape[3])))
return -total_loss / labels.shape[0]
def avg_log_likelihood(self, X, y, theta): ll = 0 for c in range(10): X_c = get_images_by_label(X, y, c) log_p_x = logsumexp(np.log(0.1) + np.dot(X_c, np.log(theta.T)) + np.dot((1. - X_c), np.log(1. - theta.T)), axis=1) ll += np.sum(np.dot(X_c, np.log(theta[c])) + np.dot((1. - X_c), np.log(1. - theta[c])) + np.log(0.1) - log_p_x) return ll / X.shape[0]
def _cost_with_vis(self,
inputs,
targets,
hprev,
weights,
disable_tqdm=True,
epoch=None):
if epoch is not None:
f = open('values' + str(epoch) + '.txt', "w+")
W_hh, W_xh, b_h, W_hy, b_y = weights
h = np.copy(hprev)
loss = 0
for t in tqdm(range(len(inputs)), disable=disable_tqdm):
x = char_to_one_hot(inputs[t])
h = np.tanh(W_hh @ h + W_xh @ x + b_h)
if epoch is not None:
f.write(','.join(h.astype(str)) + '\n')
y = W_hy @ h + b_y
target_index = char_to_index[targets[t]]
# ps_target[t] = np.exp(ys[t][target_index])/np.sum(np.exp(ys[t])) # probability for next chars being target
# loss += -np.log(ps_target[t])
loss += -(y[target_index] - logsumexp(y))
if epoch is not None:
f.close()
loss = loss / len(inputs)
return loss
def nn_predict_GCN(params, x):
# x: NSAMPLES x NFEATURES
U = hyper['U']
xf = np.matmul(x, U)
xf = np.expand_dims(xf, 1) # NSAMPLES x 1 x NFEATURES
xf = np.transpose(xf) # NFEATURES x 1 x NSAMPLES
# Filter
yf = np.matmul(params['W1'], xf) # for each feature
yf = np.transpose(yf) # NSAMPLES x NFILTERS x NFEATURES
yf = np.reshape(yf, [-1, hyper['NFEATURES']])
# Transform back to graph domain
Ut = np.transpose(U)
y = np.matmul(yf, Ut)
y = np.reshape(y, [-1, hyper['F'], hyper['NFEATURES']])
y += params['b1'] # NSAMPLES x NFILTERS x NFEATURES
# nonlinear layer
y = ReLU(y)
# y = np.tanh(y)
# dense layer
y = np.reshape(y, [-1, hyper['F']*hyper['NFEATURES']])
y = np.matmul(y, params['W2']) + params['b2']
outputs = y
return outputs - logsumexp(outputs, axis=1, keepdims=True)
def _cost(self, inputs, targets, hprev, Cprev, weights, disable_tqdm=True):
W_1, b_1, W_f, b_f, W_i, b_i, W_c, b_c, W_o, b_o, W_2, b_2 = weights
h = np.copy(hprev)
C = np.copy(Cprev)
loss = 0
for t in tqdm(range(len(inputs)), disable=disable_tqdm):
x = char_to_one_hot(inputs[t])
x = np.matmul(W_1, x) + b_1
f = sigmoid(np.matmul(W_f, np.concatenate((h, x))) + b_f)
i = sigmoid(np.matmul(W_i, np.concatenate((h, x))) + b_i)
C_hat = np.tanh(np.matmul(W_c, np.concatenate((h, x))) + b_c)
C = f * C + i * C_hat
o = sigmoid(np.matmul(W_o, np.concatenate((h, x))) + b_o)
h = o * np.tanh(C)
y = np.matmul(W_2, h) + b_2
target_index = char_to_index[targets[t]]
# ps_target[t] = np.exp(ys[t][target_index])/np.sum(np.exp(ys[t])) # probability for next chars being target
# loss += -np.log(ps_target[t])
loss += -(y[target_index] - logsumexp(y))
loss = loss / len(inputs)
return loss
def _cost_batched(self,
inputs,
targets,
hprev,
Cprev,
weights,
disable_tqdm=True):
W_1, b_1, W_f, b_f, W_i, b_i, W_c, b_c, W_o, b_o, W_2, b_2 = weights
h = np.copy(hprev)
C = np.copy(Cprev)
h = h.reshape((self.batch_size, self.h_size, 1))
C = C.reshape((self.batch_size, self.h_size, 1))
loss = 0
# W_sth_dropout = get_dropout_function((self.h_size, self.h_size + self.x_size), self.keep_prob)
# b_sth_dropout = get_dropout_function((self.h_size,), self.keep_prob)
# W_dropout = get_dropout_function((self.y_size, self.h_size), self.keep_prob)
# b_dropout = get_dropout_function((self.y_size,), self.keep_prob)
cell_dropout = get_dropout_function((self.batch_size, self.h_size, 1),
self.keep_prob)
y_dropout = get_dropout_function((self.batch_size, self.y_size, 1),
self.keep_prob)
for t in tqdm(range(len(inputs)), disable=disable_tqdm):
x = np.array([char_to_one_hot(c) for c in inputs[:, t]])
x = x.reshape((self.batch_size, -1, 1))
x = np.matmul(W_1, x) + np.reshape(b_1, (-1, 1))
x = cell_dropout(x)
f = sigmoid(
np.matmul(W_f, np.concatenate((h, x), axis=1)) +
np.reshape(b_f, (-1, 1)))
f = cell_dropout(f)
i = sigmoid(
np.matmul(W_i, np.concatenate((h, x), axis=1)) +
np.reshape(b_i, (-1, 1)))
i = cell_dropout(i)
C_hat = np.tanh(
np.matmul(W_c, np.concatenate((h, x), axis=1)) +
np.reshape(b_c, (-1, 1)))
C_hat = cell_dropout(C_hat)
C = f * C + i * C_hat
C = cell_dropout(C)
o = sigmoid(
np.matmul(W_o, np.concatenate((h, x), axis=1)) +
np.reshape(b_o, (-1, 1)))
o = cell_dropout(o)
h = o * np.tanh(C)
h = cell_dropout(h)
ys = np.matmul(W_2, h) + np.reshape(b_2, (-1, 1))
ys = y_dropout(ys)
target_indices = np.array(
[char_to_index[c] for c in targets[:, t]])
# ps_target[t] = np.exp(ys[t][target_index])/np.sum(np.exp(ys[t])) # probability for next chars being target
# loss += -np.log(ps_target[t])
loss += np.sum([
-(y[target_index] - logsumexp(y))
for y, target_index in zip(ys, target_indices)
]) / (self.number_of_steps * self.batch_size)
return loss
def _backprop_single(self, params, x, num_samples = 1, alpha = 1.0):
"""
Efficient training by computing k forward pass and only 1
backward pass (by sampling particles according to the weights).
For VI all the weights are equal.
"""
# compute weights
logF = self._comp_log_weights(params, x, num_samples)
batchsize = x.shape[1]
lowerbound = 0.0
logFa = (1 - alpha) * logF
for i in xrange(batchsize):
indl = int(i * num_samples); indr = int((i+1) * num_samples)
log_weights = logFa[indl:indr] - logsumexp(logFa[indl:indr])
prob = list(np.exp(log_weights))
# current autograd doesn't support np.random.choice!
sample_uniform = np.random.random()
for j in xrange(num_samples):
sample_uniform = sample_uniform - prob[j]
if sample_uniform <= 0.0:
break
ind_current = indl + j
lowerbound = lowerbound + logF[ind_current]
return lowerbound
def c_given_x(x):
p = np.ndarray(shape=(x.shape[0], 10))
for c in range(10):
p[:, c] = np.log(theta[c]**x * (1 - theta[c])**(1 - x)).sum(axis=1)
p = p - logsumexp(p, axis=1, keepdims=True)
p = np.exp(p)
return p
def log_transition_matrices(self, data, input, mask, tag):
T, D = data.shape
log_Ps = np.dot(input[1:], self.Ws.T)[:, None, :] # inputs
log_Ps = log_Ps + np.dot(data[:-1], self.Rs.T)[:, None, :] # past observations
log_Ps = log_Ps + self.r # bias
log_Ps = np.tile(log_Ps, (1, self.K, 1)) # expand
return log_Ps - logsumexp(log_Ps, axis=2, keepdims=True) # normalize
def grad_pred_ll(self, X, W, c): """ This function calculate the gradient of the predictive log-likelihood. return a 10 * 784 vector """ constant = np.exp(logsumexp(np.dot(X, W.T), axis=1)) return np.sum(X - (X.T * np.divide(np.exp(np.dot(X, W[c])), constant)).T, axis=0)
def calc_log_proba_for_one_seq(x_n_TD, a_n_T, lpi, ltrans, emission_mu,
emission_std):
n_timesteps = x_n_TD.shape[0]
n_states = lpi.shape[0]
belief_log_proba_TK = np.zeros((n_timesteps, n_states))
# Compute log proba array
x_n_log_proba_TK = calc_log_proba_arr_for_x(x_n_TD, emission_mu,
emission_std, n_states, a_n_T)
x_n_log_proba_TK = x_n_log_proba_TK.flatten()
# Initialise fwd belief vector at t = 0
curr_belief_log_proba_K = lpi + x_n_log_proba_TK[0]
curr_x_log_proba = logsumexp(curr_belief_log_proba_K)
curr_belief_log_proba_K = curr_belief_log_proba_K - curr_x_log_proba
belief_log_proba_TK[0, :] = curr_belief_log_proba_K
log_proba_x = curr_x_log_proba
for t in range(1, n_timesteps):
# Update the beliefs over time
curr_belief_log_proba_K, curr_x_log_proba = update_belief_log_probas(
curr_belief_log_proba_K, x_n_log_proba_TK[t], ltrans, a_n_T[t])
belief_log_proba_TK[t, :] = curr_belief_log_proba_K
log_proba_x += curr_x_log_proba
return log_proba_x, belief_log_proba_TK
def class_prior(self):
ln_p_a = np.log(self.action(
self.params)) # individual- and time-invariant
logits_mix = self.params[self.mixture_param_key]
ln_p_mix = logits_mix - logsumexp(logits_mix)
return ln_p_a, ln_p_mix
def lower_bound_MoG(theta,
s2min=1e-7,
return_dmu=False,
n=0,
return_ds2=False):
"""
Lower bound on the entropy of a mixture of Gaussians.
INPUT:
theta --- all MoG parameters in the [mu; lns2] format
s2min --- minimum variance
return_dmu --- returns gradient with respect to mu_n, for the input n
n --- see above
return_ds2 ---- returns grad with respect to all s2 params
"""
# unpack num components and dimensionality
N, Dpp = theta.shape
D = Dpp - 1
# unpack mean and variance parameters
mu = theta[:, :D]
s2 = np.exp(theta[:, -1]) + s2min
# compute lower bound to entropy, Eq (7) --- we compute the
# Normal Probability N(mu_n | mu_j, s2_n + s2_j)
S = sq_dist(mu)
s = s2[:, None] + s2[None, :]
lnP = (-.5 * S / s) - .5 * D * np.log(2 * np.pi) - .5 * D * np.log(s)
lnqn = scpm.logsumexp(lnP, 1) - np.log(N)
H = np.sum(lnqn) / float(N)
# TODO implement gradients in the same matlab style
return -1. * H
def avg_pred_log(w,images):
log_pc_x = 0
for i in range(0,images.shape[0]):
current_log_pc_x = np.dot(np.transpose(w),images[i,:]) - logsumexp(np.dot(np.transpose(w),images[i,:]))
log_pc_x = log_pc_x + current_log_pc_x
return np.sum(log_pc_x)/float(images.shape[0])
def _cost_batched(self,
inputs,
targets,
hprev,
weights,
disable_tqdm=True):
W_hh, W_xh, b_h, W_hy, b_y = weights
h = np.copy(hprev)
h = h.reshape((self.batch_size, self.hidden_size, 1))
loss = 0
for t in tqdm(range(self.number_of_steps), disable=disable_tqdm):
x = np.array([char_to_one_hot(c) for c in inputs[:, t]])
x = x.reshape((self.batch_size, -1, 1))
h = np.tanh(W_hh @ h + W_xh @ x + np.reshape(b_h, (-1, 1)))
ys = W_hy @ h + np.reshape(b_y, (-1, 1))
ys = np.squeeze(ys)
target_indices = np.array(
[char_to_index[c] for c in targets[:, t]])
# ps_target[t] = np.exp(ys[t][target_index])/np.sum(np.exp(ys[t])) # probability for next chars being target
# loss += -np.log(ps_target[t])
loss += np.sum([
-(y[target_index] - logsumexp(y))
for y, target_index in zip(ys, target_indices)
]) / (self.number_of_steps * self.batch_size)
return loss
def callback(params, t, g):
print("Iteration {} lower bound {}".format(t, -objective(params, t)))
# print (params.shape)
log_weights = params[:k] - logsumexp(params[:k])
print(np.exp(log_weights))
# params2 = np.reshape(params[10:], (10, -1))
# print (params2.shape)
# print (params2)
plt.cla()
target_distribution = lambda x: np.exp(log_density(x, t))
var_distribution = lambda x: np.exp(variational_log_density(params, x))
plot_isocontours(ax, target_distribution)
plot_isocontours(ax, var_distribution, cmap=plt.cm.bone)
ax.set_autoscale_on(False)
rs = npr.RandomState(0)
samples = variational_sampler(params, num_plotting_samples, rs)
plt.plot(samples[:, 0], samples[:, 1], 'x')
plt.draw()
plt.pause(1.0 / 30.0)
def loss(params, X, T):
W_vect = params[:-1]
alpha = params[-1]
log_prior = -L2_reg * np.dot(W_vect, W_vect)
preds = predictions(W_vect, X, alpha)
normalised_log_probs = preds - logsumexp(preds)
log_lik = np.sum(normalised_log_probs * T)
return -1.0 * (log_prior + log_lik)
def log_likelihood(all_params, X, y, n_samples):
rs = npr.RandomState(0)
samples = [sample_mean_cov_from_deep_gp(all_params, X, True, rs, FITC=True) for i in xrange(n_samples)]
return (
logsumexp(np.array([mvn.logpdf(y, mean, var) for mean, var in samples]))
- np.log(n_samples)
+ evaluate_prior(all_params)
)
def log_marginal_likelihood(params, data):
cluster_lls = []
for log_proportion, mean, chol in zip(*unpack_params(params)):
cov = np.dot(chol.T, chol) + 0.000001 * np.eye(D)
cluster_log_likelihood = log_proportion + mvn.logpdf(data, mean, cov)
cluster_lls.append(np.expand_dims(cluster_log_likelihood, axis=0))
cluster_lls = np.concatenate(cluster_lls, axis=0)
return np.sum(logsumexp(cluster_lls, axis=0))
def get_error_and_ll(w, v_prior, X, y, K, location, scale):
v_noise = np.exp(parser.get(w, 'log_v_noise')[ 0, 0 ]) * scale**2
q = get_parameters_q(w, v_prior)
samples_q = draw_samples(q, K)
outputs = predict(samples_q, X) * scale + location
log_factor = -0.5 * np.log(2 * math.pi * v_noise) - 0.5 * (np.tile(y, (1, K)) - np.array(outputs))**2 / v_noise
ll = np.mean(logsumexp(log_factor - np.log(K), 1))
error = np.sqrt(np.mean((y - np.mean(outputs, 1, keepdims = True))**2))
return error, ll
def neural_net_predict(params, inputs):
"""Implements a deep neural network for classification.
params is a list of (weights, bias) tuples.
inputs is an (N x D) matrix.
returns normalized class log-probabilities."""
for W, b in params:
outputs = np.dot(inputs, W) + b
inputs = np.tanh(outputs)
return outputs - logsumexp(outputs, axis=1, keepdims=True)
def predicted_class_logprobs(self, W_vect, inputs):
for W, b in self.unpack_layers(W_vect):
outputs = np.dot(inputs, W) + b
if self.activation_type == 'tanh':
inputs = np.tanh(outputs)
elif self.activation_type == 'relu':
inputs = relu(outputs)
else:
raise ValueException('unknown activation_type {}'.format(self.activation_type))
return outputs - logsumexp(outputs, axis=1, keepdims=True)
def predictions(self, W_vect, inputs):
'''For classsification, returns N*C matrix of log probabilities.
For rregression, returns N*K matrix of predicted means'''
for W, b in self.unpack_layers(W_vect):
outputs = np.dot(inputs, W) + b
inputs = self.nonlinearity(outputs)
if self.output_type == 'regression':
return outputs
if self.output_type == 'classification':
logprobs = outputs - logsumexp(outputs, axis=1, keepdims=True)
return logprobs
def unpack_params(params):
"""Unpacks parameter vector into the proportions, means and covariances
of each mixture component. The covariance matrices are parametrized by
their Cholesky decompositions."""
log_proportions = parser.get(params, "log proportions")
normalized_log_proportions = log_proportions - logsumexp(log_proportions)
means = parser.get(params, "means")
lower_tris = np.tril(parser.get(params, "lower triangles"), k=-1)
diag_chols = np.exp(parser.get(params, "log diagonals"))
chols = lower_tris + np.make_diagonal(diag_chols, axis1=-1, axis2=-2)
return normalized_log_proportions, means, chols
def unpack_params(params):
"""Unpacks parameter vector into the proportions, means and covariances
of each mixture component. The covariance matrices are parametrized by
their Cholesky decompositions."""
log_proportions = parser.get(params, 'log proportions')
normalized_log_proportions = log_proportions - logsumexp(log_proportions)
means = parser.get(params, 'means')
lower_tris = np.tril(parser.get(params, 'lower triangles'), k=-1)
diag_chols = np.exp( parser.get(params, 'log diagonals'))
chols = []
for lower_tri, diag in zip(lower_tris, diag_chols):
chols.append(np.expand_dims(lower_tri + np.diag(diag), 0))
chols = np.concatenate(chols, axis=0)
return normalized_log_proportions, means, chols
def _m_step(self):
assert(self.resp.shape[0] == self.num_samp)
pseud_lcount = logsumexp(self.resp, axis = 0).flat
r = exp(self.resp)
self.comp_dist = []
for c in range(self.num_components):
norm = exp(pseud_lcount[c])
mu = np.sum(r[:,c:c+1] * self.samples, axis=0) / norm
diff = self.samples - mu
scatter_matrix = np.zeros([self.samples.shape[1]]*2)
for i in range(diff.shape[0]):
scatter_matrix += r[i,c:c+1] *diff[i:i+1,:].T.dot(diff[i:i+1,:])
scatter_matrix /= norm
self.comp_dist.append(mvnorm(mu, scatter_matrix))
self.comp_lprior = pseud_lcount - log(self.num_samp)
def mog_like(x, means, icovs, dets, pis):
""" compute the log likelihood according to a mixture of gaussians
with means = [mu0, mu1, ... muk]
icovs = [C0^-1, ..., CK^-1]
dets = [|C0|, ..., |CK|]
pis = [pi1, ..., piK] (sum to 1)
at locations given by x = [x1, ..., xN]
"""
xx = np.atleast_2d(x)
centered = xx[:,:,np.newaxis] - means.T[np.newaxis,:,:]
solved = np.einsum('ijk,lji->lki', icovs, centered)
logprobs = -0.5*np.sum(solved * centered, axis=1) - np.log(2*np.pi) - 0.5*np.log(dets) + np.log(pis)
logprob = scpm.logsumexp(logprobs, axis=1)
if len(x.shape) == 1:
return np.exp(logprob[0])
else:
return np.exp(logprob)
def mog_logmarglike(x, means, covs, pis, ind=0):
""" marginal x or y (depending on ind) """
K = pis.shape[0]
xx = np.atleast_2d(x)
centered = xx.T - means[:,ind,np.newaxis].T
logprobs = []
for kk in xrange(K):
quadterm = centered[:,kk] * centered[:,kk] * (1./covs[kk,ind,ind])
logprobsk = -.5*quadterm - .5*np.log(2*np.pi) \
-.5*np.log(covs[kk,ind,ind]) + np.log(pis[kk])
logprobs.append(np.squeeze(logprobsk))
logprobs = np.array(logprobs)
logprob = scpm.logsumexp(logprobs, axis=0)
if np.isscalar(x):
return logprob[0]
else:
return logprob
def cost(theta):
# Unpack parameters
nu = np.concatenate([theta[1], [0]], axis=0)
S = theta[0]
logdetS = np.expand_dims(np.linalg.slogdet(S)[1], 1)
y = np.concatenate([samples.T, np.ones((1, N))], axis=0)
# Calculate log_q
y = np.expand_dims(y, 0)
# 'Probability' of y belonging to each cluster
log_q = -0.5 * (np.sum(y * np.linalg.solve(S, y), axis=1) + logdetS)
alpha = np.exp(nu)
alpha = alpha / np.sum(alpha)
alpha = np.expand_dims(alpha, 1)
loglikvec = logsumexp(np.log(alpha) + log_q, axis=0)
return -np.sum(loglikvec)
def gmm_logprob(x, ws, mus, sigs, invsigs=None, logdets=None):
""" Gaussian Mixture Model likelihood
Input:
- x = N x D array of data (N iid)
- ws = K length vector that sums to 1, mixing weights
- mus = K x D array of mixture component means
- sigs = K x D x D array of mixture component covariances
- invsigs = K x D x D array of mixture component covariance inverses
- logdets = K array of mixture component covariance logdets
Output:
- N length array of log likelihood values
TODO: speed this up
"""
if sigs is None:
assert invsigs is not None and logdets is not None, \
"need sigs if you don't include logdets and invsigs"
# compute invsigs if needed
if invsigs is None:
invsigs = np.array([np.linalg.inv(sig) for sig in sigs])
logdets = np.array([np.linalg.slogdet(sig)[1] for sig in sigs])
# compute each gauss component separately
xx = np.atleast_2d(x)
centered = xx[:,:,np.newaxis] - mus.T[np.newaxis,:,:]
solved = np.einsum('ijk,lji->lki', invsigs, centered)
logprobs = -0.5*np.sum(solved * centered, axis=1) - \
np.log(2*np.pi) - 0.5*logdets + np.log(ws)
logprob = scpm.logsumexp(logprobs, axis=1)
if len(x.shape) == 1:
return logprob[0]
else:
return logprob
def logsoftmax(v):
return v - logsumexp(v, 1).reshape(-1, 1)
def hiddens_to_output_probs(hiddens):
output = concat_and_multiply(params['predict'], hiddens)
return output - logsumexp(output, axis=1, keepdims=True)
def loss(W_vect, X, T):
log_prior = -L2_reg * np.dot(W_vect, W_vect)
preds = predictions(W_vect, X)
normalised_log_probs = preds - logsumexp(preds)
log_lik = np.sum(normalised_log_probs * T)
return -1.0 * (log_prior + log_lik)
def mixture_log_density(var_mixture_params, x):
"""Returns a weighted average over component densities."""
log_weights, var_params = unpack_mixture_params(var_mixture_params)
component_log_densities = np.vstack([component_log_density(params_k, x)
for params_k in var_params]).T
return logsumexp(component_log_densities + log_weights, axis=1, keepdims=False)
def _e_step(self):
lpdfs = np.array([d.logpdf(self.samples).flat[:]
for d in self.comp_dist]).T + self.comp_lprior
self.resp = lpdfs - logsumexp(lpdfs, axis = 1).reshape((self.num_samp, 1))
def log_likelihood(all_params): # implement mini batches later?
n_samples = 1
samples = [sample_mean_cov_from_deep_gp(all_params, X, True) for i in xrange(n_samples)]
return logsumexp(np.array([mvn.logpdf(y,mean,var+1e-6*np.eye(len(var))*np.max(np.diag(var))) for mean,var in samples])) - np.log(n_samples) \
+ evaluate_prior(all_params)
def log_Z_likelihood(q, f_hat, v_noise, X, y, K):
samples = draw_samples(q, K)
log_f_hat = np.sum(-0.5 / f_hat[ 'v' ] * samples**2 + f_hat[ 'm' ] / f_hat[ 'v' ] * samples, 1)
log_factor_value = alpha * (log_likelihood_factor(samples, v_noise, X, y) - log_f_hat)
return np.sum(logsumexp(log_factor_value, 1) + np.log(1.0 / K))
def predictions(W_vect, inputs):
for W, b in unpack_layers(W_vect):
outputs = np.dot(inputs, W) + b
inputs = np.tanh(outputs)
return outputs - logsumexp(outputs, axis=1, keepdims=True)
def gaussian_loglike(x, mu, log_sigmasq):
return np.mean(logsumexp(
-0.5*((np.log(2*np.pi) + log_sigmasq) + (x - mu)**2. / np.exp(log_sigmasq)),
axis=0))
def log_softmax(self, batch):
batch = batch - np.max(batch, axis=1, keepdims=True)
return batch - logsumexp(batch, axis=1).reshape((batch.shape[0], -1))
def logpdf(self, x):
rval = np.array([self.comp_lprior[i]+ self.comp_dist[i].logpdf(x)
for i in range(self.comp_lprior.size)])
rval = logsumexp(rval, 0).flatten()
return rval
