def unit(parent_x, child_h, child_c, child_exists):
(h_i, h_o,
h_u), _ = theano.map(fn=lambda Ui, Uo, Uu, h, exists:
(exists * T.dot(Ui, h), exists * T.dot(
Uo, h), exists * T.dot(Uu, h)),
sequences=[
self.U_i, self.U_o, self.U_u, child_h,
child_exists
])
i = T.nnet.sigmoid(
T.dot(self.W_i, parent_x) + h_i.sum(axis=0) + self.b_i)
o = T.nnet.sigmoid(
T.dot(self.W_o, parent_x) + h_o.sum(axis=0) + self.b_o)
u = T.tanh(T.dot(self.W_u, parent_x) + h_u.sum(axis=0) + self.b_u)
def _sub_f(U):
sub_h_f, _ = theano.map(
fn=lambda sub_U, h, exists: exists * T.dot(sub_U, h),
sequences=[U, child_h, child_exists])
return sub_h_f.sum(axis=0)
h_f, _ = theano.map(fn=lambda U: _sub_f(U), sequences=[self.U_f])
f = (T.nnet.sigmoid(
T.dot(self.W_f, parent_x).dimshuffle('x', 0) + h_f +
self.b_f.dimshuffle('x', 0)) * child_exists.dimshuffle(0, 'x'))
c = i * u + T.sum(f * child_c, axis=0)
h = o * T.tanh(c)
return h, c
def unit(parent_x, child_h, child_c, child_exists):
(h_i, h_o, h_u), _ = theano.map(
fn=lambda Ui, Uo, Uu, h, exists:
(exists * T.dot(Ui, h), exists * T.dot(Uo, h), exists * T.dot(Uu, h)),
sequences=[self.U_i, self.U_o, self.U_u, child_h, child_exists])
i = T.nnet.sigmoid(T.dot(self.W_i, parent_x) + h_i.sum(axis=0) + self.b_i)
o = T.nnet.sigmoid(T.dot(self.W_o, parent_x) + h_o.sum(axis=0) + self.b_o)
u = T.tanh(T.dot(self.W_u, parent_x) + h_u.sum(axis=0) + self.b_u)
def _sub_f(U):
sub_h_f, _ = theano.map(
fn=lambda sub_U, h, exists: exists * T.dot(sub_U, h),
sequences=[U, child_h, child_exists])
return sub_h_f.sum(axis=0)
h_f, _ = theano.map(
fn=lambda U: _sub_f(U),
sequences=[self.U_f])
f = (T.nnet.sigmoid(
T.dot(self.W_f, parent_x).dimshuffle('x', 0) + h_f +
self.b_f.dimshuffle('x', 0)) *
child_exists.dimshuffle(0, 'x'))
c = i * u + T.sum(f * child_c, axis=0)
h = o * T.tanh(c)
return h, c
def compute_cost_log_in_parallel(original_rnn_outputs, labels, func, x_ends,
y_ends):
mask = T.log(1 - T.or_(T.eq(labels, T.zeros_like(labels)),
T.eq(labels, shift_matrix(labels, 2))))
initial_state = T.log(T.zeros_like(labels))
initial_state = T.set_subtensor(initial_state[:, 0], 0)
def select_probabilities(rnn_outputs, label):
return rnn_outputs[:, label]
rnn_outputs, _ = theano.map(select_probabilities,
[original_rnn_outputs, labels])
rnn_outputs = T.log(rnn_outputs.dimshuffle((1, 0, 2)))
def forward_step(probabilities, last_probabilities):
all_forward_probabilities = T.stack(
last_probabilities + probabilities,
log_shift_matrix(last_probabilities, 1) + probabilities,
log_shift_matrix(last_probabilities, 2) + probabilities + mask,
)
result = func(all_forward_probabilities, 0)
return result
forward_probabilities, _ = theano.scan(fn=forward_step,
sequences=rnn_outputs,
outputs_info=initial_state)
forward_probabilities = forward_probabilities.dimshuffle((1, 0, 2))
def compute_cost(forward_probabilities, x_end, y_end):
return -func(forward_probabilities[x_end - 1, y_end - 2:y_end])
return theano.map(compute_cost, [forward_probabilities, x_ends, y_ends])[0]
def compute_cost_log_in_parallel(original_rnn_outputs, labels, func, x_ends, y_ends): mask = T.log(1 - T.or_(T.eq(labels, T.zeros_like(labels)), T.eq(labels, shift_matrix(labels, 2)))) initial_state = T.log(T.zeros_like(labels)) initial_state = T.set_subtensor(initial_state[:,0], 0) def select_probabilities(rnn_outputs, label): return rnn_outputs[:,label] rnn_outputs, _ = theano.map(select_probabilities, [original_rnn_outputs, labels]) rnn_outputs = T.log(rnn_outputs.dimshuffle((1,0,2))) def forward_step(probabilities, last_probabilities): all_forward_probabilities = T.stack( last_probabilities + probabilities, log_shift_matrix(last_probabilities, 1) + probabilities, log_shift_matrix(last_probabilities, 2) + probabilities + mask, ) result = func(all_forward_probabilities, 0) return result forward_probabilities, _ = theano.scan(fn = forward_step, sequences = rnn_outputs, outputs_info = initial_state) forward_probabilities = forward_probabilities.dimshuffle((1,0,2)) def compute_cost(forward_probabilities, x_end, y_end): return -func(forward_probabilities[x_end-1,y_end-2:y_end]) return theano.map(compute_cost, [forward_probabilities, x_ends, y_ends])[0]
def compute_cost_with_cross_entropy_in_parallel(original_rnn_outputs, labels,
x_ends, y_ends):
mask = T.log(1 - T.or_(T.eq(labels, T.zeros_like(labels)),
T.eq(labels, shift_matrix(labels, 2))))
arange = T.arange(labels.shape[1])
initial_state = T.log(T.zeros_like(labels))
initial_state = T.set_subtensor(initial_state[:, 0], 0)
def select_probabilities(rnn_outputs, label):
return rnn_outputs[:, label]
rnn_outputs, _ = theano.map(select_probabilities,
[original_rnn_outputs, labels])
rnn_outputs = T.log(rnn_outputs.dimshuffle((1, 0, 2)))
def forward_step(probabilities, last_probabilities):
all_forward_probabilities = T.stack(
last_probabilities + probabilities,
log_shift_matrix(last_probabilities, 1) + probabilities,
log_shift_matrix(last_probabilities, 2) + probabilities + mask,
)
max_probability, backlink = T.max_and_argmax(all_forward_probabilities,
0)
backlink = arange - backlink
return max_probability, backlink
results, _ = theano.scan(fn=forward_step,
sequences=rnn_outputs,
outputs_info=[initial_state, None])
forward_probabilities, backward_pointers = results
def compute_cost(rnn_outputs, forward_probabilities, backward_pointers,
x_end, y_end, label):
def backward_step(backlinks, position):
new_position = backlinks[position]
return new_position, position
initial_state = T.argmax(
forward_probabilities[x_end - 1, y_end - 2:y_end]) + y_end - 2
results, _ = theano.scan(fn=backward_step,
sequences=backward_pointers[0:x_end, :],
outputs_info=[initial_state, None],
go_backwards=True)
alignment = label[results[1][::-1]]
return aggregate(categorical_crossentropy(rnn_outputs[0:x_end],
alignment),
mode='sum')
forward_probabilities = forward_probabilities.dimshuffle((1, 0, 2))
backward_pointers = backward_pointers.dimshuffle((1, 0, 2))
return theano.map(compute_cost, [
original_rnn_outputs, forward_probabilities, backward_pointers, x_ends,
y_ends, labels
])[0]
def compute_objective_and_gradients(self, nSamp):
hsamp = self.mrec.getSample(self.Y, nSamp)
# evaluate the generative model density P_\theta(y_i , h_i)
p_yh, _ = theano.map(self.mprior.evaluateLogDensity, sequences=hsamp)
# evaluate the recognition model density Q_\phi(h_i | y_i)
q_hgy, _ = theano.map(self.mrec.evalLogDensity, sequences=hsamp)
ff = (p_yh - q_hgy)
sortidx = ff.argsort(axis=0)
fmax = ff[(sortidx[-1], T.arange(ff.shape[-1]))].dimshuffle('x', 0)
f_hy = T.exp(ff - fmax)
sum_across_samples = f_hy.sum(axis=0, keepdims=True)
Lhat = T.log(sum_across_samples / nSamp) + fmax
col_idx = T.arange(ff.shape[-1])
# This 1e-12 constant is for debugging nans
# in other parts of code. We know we'll get
# nans where we'll then overwrite. Use it with
# compute cross-validated estimates of Lhat # nanguard mode.
hold_out_except_last = T.log(
(sum_across_samples - f_hy) / (nSamp - 1)) + fmax #+1e-12) + fmax
f2max_vec = ff[(sortidx[-2], T.arange(ff.shape[-1]))]
f2max = f2max_vec.dimshuffle('x', 0)
# Do tricky things to keep the numerics in order (avoid a term being \approxeq 0)
ff_nolast = T.set_subtensor(ff[(sortidx[-1], col_idx)], f2max_vec)
f_hy_last = T.exp(ff_nolast - f2max)
# compute held-out sum when we hold out the maximum element
hold_out_last = T.log(
(f_hy_last.sum(axis=0, keepdims=True) - f_hy_last) /
(nSamp - 1)) + f2max
# compute final held-out estimates
hold_out = T.set_subtensor(
hold_out_except_last[(sortidx[-1], col_idx)],
hold_out_last[(sortidx[-1], col_idx)])
Lhat_cv = Lhat - hold_out
the_ws = f_hy / sum_across_samples
weighted_q = T.sum((Lhat_cv * q_hgy + the_ws * ff).mean(axis=1))
#weighted_q = T.sum((Lhat_cv*q_hgy + the_ws*(p_yh-q_hgy)).sum(axis=1))
# gradients for approximate posterior
dqhgy = T.grad(cost=weighted_q,
wrt=self.mrec.getParams(),
consider_constant=([the_ws, Lhat_cv] + hsamp),
disconnected_inputs='ignore')
# gradients for prior
dpyh = T.grad(cost=T.sum((the_ws * ff).mean(axis=1)),
wrt=self.mprior.getParams(),
consider_constant=hsamp + [the_ws],
disconnected_inputs='ignore')
#dpyh = T.grad(cost=T.sum((the_ws*(p_yh-q_hgy)).sum(axis=1)), wrt = self.mprior.getParams(), consider_constant=hsamp + [the_ws], disconnected_inputs='ignore')
return [Lhat.mean(), dpyh, dqhgy]
def make_nade(D, z_dim):
log("make_nade with D={},z_dim={},g={}".format(D, z_dim, g))
x = T.fmatrix('x')
c_vals = np.random.normal(0, 1, size=(1, z_dim)).astype('float32')
c = theano.shared(c_vals, name="c")
p_x = 1
def a_adder(W_col_T, x_i, acc):
W_col_T.name = "W_col_T"
prod = W_col_T * T.sum(x_i)
prod.name = "prod"
ret_T = acc.T + prod
return ret_T.T
"""
for i in range(D):
W_col_vals = np.random.normal(0,1,size=(z_dim,1)).astype('float32')
W_col = theano.shared(W_col_vals,name="W_col_%d"%(i+1))
W_cols.append(W_col)
"""
W_vals = np.random.normal(0, 1, size=(z_dim, D)).astype('float32')
W = theano.shared(W_vals, name="W")
a_s_W, _u = theano.scan(fn=a_adder,
outputs_info=c[0, :],
sequences=[W.T, x])
a_s_excess = T.concatenate([c, a_s_W], axis=0)
a_s = a_s_excess[:D, :]
V_vals = np.random.normal(0, 1, size=(D, z_dim)).astype('float32')
V = theano.shared(V_vals, name="V")
hs = g(a_s)
b_val = np.random.normal(0, 1, size=(D, 1)).astype('float32')
b = theano.shared(b_val, name="b")
def scan_p_x_cond(V_row, hi, b_i):
p_x_cond = g(T.dot(V_row, hi) + b_i)
return p_x_cond
p_x_cond, _u = theano.map(fn=scan_p_x_cond, sequences=[V, hs, b])
def scan_p_x_cond_obs(x_i, p):
ret = x_i * p + (1 - x_i) * (1 - p)
return ret
p_x_cond_obs, _u = theano.map(fn=scan_p_x_cond_obs,
sequences=[x, p_x_cond])
nll = -T.sum(T.log(p_x_cond_obs))
p_x = T.prod(p_x_cond_obs)
return (W, c, V, b), x, hs, p_x, nll, p_x_cond
def step(visible, filtered_hidden_mean_m1, filtered_hidden_cov_m1):
A, B = transition, emission # (h, h), (h, v)
# Shortcuts for the filtered mean and covariance from the previous
# time step.
f_m1 = filtered_hidden_mean_m1 # (n, h)
F_m1 = filtered_hidden_cov_m1 # (n, h, h)
# Calculate mean of joint.
hidden_mean = T.dot(f_m1, A) + hnm # (n, h)
visible_mean = T.dot(hidden_mean, B) + vnm # (n, v)
# Calculate covariance of joint.
hidden_cov = stacked_dot(A.T, stacked_dot(F_m1, A)) # (n, h, h)
hidden_cov += hnc
visible_cov = stacked_dot( # (n, v, v)
B.T, stacked_dot(hidden_cov, B))
visible_cov += vnc
visible_hidden_cov = stacked_dot(hidden_cov, B) # (n, h, v)
visible_error = visible - visible_mean # (n, v)
inv_visible_cov, _ = theano.map(lambda x: matrix_inverse(x),
visible_cov) # (n, v, v)
# I don't know a better name for this monster.
visible_hidden_cov_T = visible_hidden_cov.dimshuffle(0, 2,
1) # (n, v, h)
D = stacked_dot(inv_visible_cov, visible_hidden_cov_T)
f = (
D * visible_error.dimshuffle(0, 1, 'x') # (n, h)
).sum(axis=1)
f += hidden_mean
F = hidden_cov
F -= stacked_dot(visible_hidden_cov, D)
log_l = (
inv_visible_cov * # (n,)
visible_error.dimshuffle(0, 1, 'x') *
visible_error.dimshuffle(0, 'x', 1)).sum(axis=(1, 2))
log_l *= -.5
dets, _ = theano.map(lambda x: det(x), visible_cov)
log_l -= 0.5 * T.log(dets)
log_l -= np.log(2 * np.pi)
return f, F, log_l
def step(visible, filtered_hidden_mean_m1, filtered_hidden_cov_m1):
A, B = transition, emission # (h, h), (h, v)
# Shortcuts for the filtered mean and covariance from the previous
# time step.
f_m1 = filtered_hidden_mean_m1 # (n, h)
F_m1 = filtered_hidden_cov_m1 # (n, h, h)
# Calculate mean of joint.
hidden_mean = T.dot(f_m1, A) + hnm # (n, h)
visible_mean = T.dot(hidden_mean, B) + vnm # (n, v)
# Calculate covariance of joint.
hidden_cov = stacked_dot(
A.T, stacked_dot(F_m1, A)) # (n, h, h)
hidden_cov += hnc
visible_cov = stacked_dot( # (n, v, v)
B.T, stacked_dot(hidden_cov, B))
visible_cov += vnc
visible_hidden_cov = stacked_dot(hidden_cov, B) # (n, h, v)
visible_error = visible - visible_mean # (n, v)
inv_visible_cov, _ = theano.map(
lambda x: matrix_inverse(x), visible_cov) # (n, v, v)
# I don't know a better name for this monster.
visible_hidden_cov_T = visible_hidden_cov.dimshuffle(0, 2, 1) # (n, v, h)
D = stacked_dot(inv_visible_cov, visible_hidden_cov_T)
f = (D * visible_error.dimshuffle(0, 1, 'x') # (n, h)
).sum(axis=1)
f += hidden_mean
F = hidden_cov
F -= stacked_dot(visible_hidden_cov, D)
log_l = (inv_visible_cov * # (n,)
visible_error.dimshuffle(0, 1, 'x') *
visible_error.dimshuffle(0,'x', 1)).sum(axis=(1, 2))
log_l *= -.5
dets, _ = theano.map(lambda x: det(x), visible_cov)
log_l -= 0.5 * T.log(dets)
log_l -= np.log(2 * np.pi)
return f, F, log_l
def sym_histograms(self, X, masks=None):
"""
Encodes a set of objects (X is a tensor3)
:param X: tensor3 containing the feature vectors for each object
:return:
"""
if masks is None:
histograms, updates = theano.map(self.sym_histogram,
sequences=(X, ))
else:
histograms, updates = theano.map(self.sym_histogram,
sequences=(X, masks))
return histograms
def createObjectiveFunction(self):
'''
@escription: initialize objective function and minimization function
@X,y data matrix/vector
@u random noise for simulator
@v standard normal for reparametrization trick
'''
y = T.dvector("y")
W, U = T.dvectors("W", "U")
V = T.dscalar("V")
mu = self.params[0]
#logSigma = self.params[1]
logSigma = sharedX(0.6)
logLambda = sharedX(0)
#self.params[2]
negKL = 0.5 * self.dimTheta + 0.5 * T.sum(2 * logSigma - mu**2 -
T.exp(logSigma)**2)
results, updates = th.map(fn=self.alpha_stable,
sequences=[W, U],
non_sequences=[V])
f = results
results2, updates2 = th.map(fn=self.alpha_perfect, sequences=[W, U])
f2 = results2
#SSE = T.sum((y-f)**2)
logLike = -self.m * (
0.5 * np.log(2 * np.pi) + logLambda) - 0.5 * T.sum(
(T.flatten(y) - T.flatten(f))**2) / (T.exp(logLambda)**2)
#logLike2 = -self.m*(0.5 * np.log(2 * np.pi) + logLambda)-0.5*T.sum((y-f2)**2)/(T.exp(logLambda)**2)
elbo = (negKL + logLike)
#elbo2 = (negKL + logLike2)
obj = -elbo
#obj = SSE
self.f = th.function([y, W, U, V],
f,
updates=updates,
on_unused_input='ignore')
self.lowerboundfunction = th.function([y, W, U, V],
obj,
updates=updates,
on_unused_input='ignore')
derivatives = T.grad(obj, self.params)
self.gradientfunction = th.function([y, W, U, V],
derivatives,
updates=updates,
on_unused_input='ignore')
def set_output(self):
cshape = self.camera_params.shape # (c, batch, 11)
voxel = self._input
voxel_tiled = tensor.tile(voxel, (cshape[0], 1, 1, 1, 1))
cams = tensor.reshape(self.camera_params,
(cshape[0] * cshape[1], cshape[2]))
camlocs = theano.map(self.get_camloc, cams)[0].astype('float32')
raydirs = theano.map(self.get_raydirs, cams)[0].astype('float32')
rendered = self.op(voxel_tiled, camlocs, raydirs)
rendered_sub = theano.map(self.data_augmentation, [rendered, cams])[0]
output_shape = (cshape[0], cshape[1], self.img_h, self.img_w,
self.feat_d)
self._output = tensor.reshape(rendered_sub, output_shape)
self._output = self._output.dimshuffle(0, 1, 4, 2, 3)
def pv_function(self, tensor_input): indexf_matrix = theano.shared( np.zeros( [self.max_length, self.max_length], dtype=np.int32 ), name = 'indexf_matrix', borrow=True ) pf_matrix = theano.shared( np.zeros( [self.max_length, self.max_length], dtype=theano.config.floatX ), name = 'pf_matrix', borrow=True ) pf_matrix = T.set_subtensor(pf_matrix[0, 0:tensor_input.shape[0]], 1.0) vf_matrix = theano.shared( np.zeros( (self.max_length, self.max_length, self.size), dtype=theano.config.floatX ), name = 'vf_matrix', borrow=True ) results, updates = theano.map( fn = lambda i, L, t_tensor_input: L[t_tensor_input[i]], sequences=[T.arange(tensor_input.shape[0])], non_sequences=[self.L, tensor_input], name = 'vf_matrix prepare' ) vf_matrix = T.set_subtensor(vf_matrix[0, 0:tensor_input.shape[0]], results) for i in range(1,self.max_length): results, updates = theano.map( fn = self._pv_function, sequences=[T.arange(self.max_length-i)], non_sequences = [i, pf_matrix, vf_matrix], #name = 'pv function' ) indexf_matrix = T.set_subtensor(indexf_matrix[i, 0:self.max_length-i], results[0]) pf_matrix = T.set_subtensor(pf_matrix[i, 0:self.max_length-i], results[1]) vf_matrix = T.set_subtensor(vf_matrix[i, 0:self.max_length-i], results[2]) return indexf_matrix, pf_matrix, vf_matrix
def pv_function(self, tensor_input): indexf_matrix = theano.shared( np.zeros( [self.max_length, self.max_length], dtype=np.int32 ), name = 'indexf_matrix', borrow=True ) pf_matrix = theano.shared( np.zeros( [self.max_length, self.max_length], dtype=theano.config.floatX ), name = 'pf_matrix', borrow=True ) pf_matrix = T.set_subtensor(pf_matrix[0, 0:tensor_input.shape[0]], 1.0) vf_matrix = theano.shared( np.zeros( (self.max_length, self.max_length, self.size), dtype=theano.config.floatX ), name = 'vf_matrix', borrow=True ) results, updates = theano.map( fn = lambda i, L, t_tensor_input: L[t_tensor_input[i]], sequences=[T.arange(tensor_input.shape[0])], non_sequences=[self.L, tensor_input], name = 'vf_matrix prepare' ) vf_matrix = T.set_subtensor(vf_matrix[0, 0:tensor_input.shape[0]], results) for i in range(1,self.max_length): results, updates = theano.map( fn = self._pv_function, sequences=[T.arange(self.max_length-i)], non_sequences = [i, pf_matrix, vf_matrix], #name = 'pv function' ) indexf_matrix = T.set_subtensor(indexf_matrix[i, 0:self.max_length-i], results[0]) pf_matrix = T.set_subtensor(pf_matrix[i, 0:self.max_length-i], results[1]) vf_matrix = T.set_subtensor(vf_matrix[i, 0:self.max_length-i], results[2]) return indexf_matrix, pf_matrix, vf_matrix
def call(self, x, mask):
x_switched = K.switch(mask[:, :, None], x, 0.0)
activation_ranks = theano.map(rank_function, x_switched)[0]
activation_energies = K.switch(mask[:, None, :maxsents],
activation_ranks, -1e20)
activation_weights = theano.map(K.softmax, activation_energies)[0]
base_values = (mask * ((K.sum(mask[:, :maxsents] + 0.0, axis=-1))**
-1)[:, None])[:, None, :maxsents]
pad_weights = K.concatenate(
(base_values, activation_weights[:, :-1, :]), axis=1)
diff_weights = activation_weights - pad_weights
posi_diffs = K.switch(diff_weights > 0, diff_weights, 0.0)
norm_pds = (K.sum(posi_diffs, axis=-1) + K.epsilon())**-1
attentions = posi_diffs * norm_pds[:, :, None]
return attentions
def decode_to_probs(self, activations, relative_position, low_bound, high_bound):
squashed = T.reshape(activations, (-1,self.RAW_ENCODING_WIDTH))
n_parallel = squashed.shape[0]
probs = T.nnet.softmax(squashed)
def _scan_fn(cprobs, cpos):
if self.with_artic:
abs_probs = cprobs[:2]
rel_probs = cprobs[2:]
else:
rel_probs = cprobs
abs_probs = T.ones((2,))
aligned = T.roll(rel_probs, (cpos-low_bound)%12)
num_tile = int(math.ceil((high_bound-low_bound)/self.WINDOW_SIZE))
tiled = T.tile(aligned, (num_tile,))[:(high_bound-low_bound)]
full = T.concatenate([abs_probs, tiled], 0)
return full
# probs = theano.printing.Print("probs",['shape'])(probs)
# relative_position = theano.printing.Print("relative_position",['shape'])(relative_position)
from_scan, _ = theano.map(fn=_scan_fn, sequences=[probs, T.flatten(relative_position)])
# from_scan = theano.printing.Print("from_scan",['shape'])(from_scan)
newshape = T.concatenate([activations.shape[:-1],[2+high_bound-low_bound]],0)
fixed = T.reshape(from_scan, newshape, ndim=activations.ndim)
return fixed
def apply(self, image, image_shape, location, scale):
a, b = self.compute_hard_windows(image_shape, location, scale)
if self.batched_window:
patch = self.apply_inner(image, location, scale, a[0], b[0])
else:
def map_fn(image, image_shape, a, b, location, scale):
# apply_inner expects a batch axis
image = T.shape_padleft(image)
location = T.shape_padleft(location)
scale = T.shape_padleft(scale)
patch = self.apply_inner(image, location, scale, a, b)
# return without batch axis
return patch[0]
patch, _ = theano.map(map_fn,
sequences=[image, a, b, location, scale])
savings = (1 - T.cast(
(b - a).prod(axis=1), floatX) / image_shape.prod(axis=1))
self.add_auxiliary_variable(savings, name="savings")
return patch
def get_output_for(self, input, **kwargs):
def sample_one_image(img, y, x):
return theano.map(
lambda x, y, image: image[:, y:(y + self.patch_size[0]), x:
(x + self.patch_size[1])],
sequences=[x, y],
non_sequences=img)[0]
if self.pad:
shp = (input.shape[0], input.shape[1],
input.shape[2] + self.patch_size[0] * 2 - 2,
input.shape[3] + self.patch_size[1] * 2 - 2)
padded_input = T.zeros(shp)
padded_input = T.set_subtensor(
padded_input[:, :, (self.patch_size[0] -
1):(-self.patch_size[0] + 1),
(self.patch_size[1] - 1):(-self.patch_size[1] +
1)], input)
input = padded_input
y = self.rng.random_integers(size=(input.shape[0],
self.patches_per_example),
low=0,
high=input.shape[2] - self.patch_size[0])
x = self.rng.random_integers(size=(input.shape[0],
self.patches_per_example),
low=0,
high=input.shape[3] - self.patch_size[1])
return theano.map(sample_one_image,
sequences=[input, y, x])[0].reshape(
(-1, input.shape[1], self.patch_size[0],
self.patch_size[1]))
def get_reward_sequences(self, env_state_sessions, agent_action_sessions):
"""Computes the rewards given to agent at each time step for each batch.
:param env_state_sessions: Environment state [batch_i,seq_i,state_units] history for all sessions.
:type env_state_sessions: theano tensor [batch_i,seq_i,state_units]
:param agent_action_sessions: Actions chosen by agent at each tick for all sessions.
:type agent_action_sessions: int[batch_i,seq_i]
:return rewards: What reward was given to an agent for corresponding action from state in that batch.
:rtype: float[batch_i,seq_i]
"""
env_state_sessions = check_list(env_state_sessions)
n_states = len(env_state_sessions)
agent_action_sessions = check_list(agent_action_sessions)
n_actions = len(agent_action_sessions)
def compute_reward(batch_i, *args):
session_states, session_actions = unpack_list(args, [n_states, n_actions])
return self.get_reward(session_states, session_actions, batch_i)
sequences = [T.arange(agent_action_sessions[0].shape[0], ), ] + env_state_sessions + agent_action_sessions
rewards, updates = theano.map(compute_reward, sequences=sequences)
assert len(updates) == 0
return rewards.reshape(agent_action_sessions[0].shape) # reshape bach to original
def cosine_similarity(x, y, eps=1e-6):
r"""
Cosine similarity between a vector and each row of a base matrix.
Parameters
----------
x: a 1D Theano variable
Vector to compare to each row of the matrix y.
y: a 2D Theano variable
Matrix to be compared to
eps: float
Precision of the operation (necessary for differentiability).
Return
------
z: a 1D Theano variable
A vector whose components are the cosine similarities
between x and each row of y.
"""
def _cosine_similarity(x, y, eps=1e-6):
y = y.dimshuffle(1, 0)
z = T.dot(x, y)
z /= T.sqrt(T.sum(x * x) * T.sum(y * y, axis=0) + eps)
return z
def step(x_b, y_b):
return _cosine_similarity(x_b, y_b, eps)
z, _ = theano.map(step, sequences=[x, y])
return z
def gen_full_alignment(self):
# Get only the focus columns
for seq_name,sequence in self.seq_name_to_sequence.items():
# Replace periods with dashes (the uppercase equivalent)
sequence = sequence.replace(".","-")
#then get only the focus columns
self.seq_name_to_sequence[seq_name] = [sequence[ix].upper() for ix in self.focus_cols]
# Remove sequences that have bad characters
alphabet_set = set(list(self.alphabet))
seq_names_to_remove = []
for seq_name,sequence in self.seq_name_to_sequence.items():
for letter in sequence:
if letter not in alphabet_set and letter != "-":
seq_names_to_remove.append(seq_name)
seq_names_to_remove = list(set(seq_names_to_remove))
for seq_name in seq_names_to_remove:
del self.seq_name_to_sequence[seq_name]
# Encode the sequences
print ("Encoding sequences")
self.x_train = np.zeros((len(self.seq_name_to_sequence.keys()),len(self.focus_cols),len(self.alphabet)))
self.x_train_name_list = []
for i,seq_name in enumerate(self.seq_name_to_sequence.keys()):
sequence = self.seq_name_to_sequence[seq_name]
self.x_train_name_list.append(seq_name)
for j,letter in enumerate(sequence):
if letter in self.aa_dict:
k = self.aa_dict[letter]
self.x_train[i,j,k] = 1.0
# Fast sequence weights with Theano
if self.calc_weights:
print ("Computing sequence weights")
# Numpy version
# import scipy
# from scipy.spatial.distance import pdist, squareform
# self.weights = scale / np.sum(squareform(pdist(seq_index_array, metric="hamming")) < theta, axis=0)
#
# Theano weights
X = T.tensor3("x")
cutoff = T.scalar("theta")
X_flat = X.reshape((X.shape[0], X.shape[1]*X.shape[2]))
N_list, updates = theano.map(lambda x: 1.0 / T.sum(T.dot(X_flat, x) / T.dot(x, x) > 1 - cutoff), X_flat)
weightfun = theano.function(inputs=[X, cutoff], outputs=[N_list],allow_input_downcast=True)
#
self.weights = weightfun(self.x_train, self.theta)[0]
else:
# If not using weights, use an isotropic weight matrix
self.weights = np.ones(self.x_train.shape[0])
self.Neff = np.sum(self.weights)
print ("Neff =",str(self.Neff))
print ("Data Shape =",self.x_train.shape)
def get_reward_sequences(self,env_state_sessions,agent_action_sessions):
"""
computes the rewards given to agent at each time step for each batch
parameters:
env_state_seq - environment state [batch_i,seq_i,state_units] history for all sessions
agent_action_seq - int[batch_i,seq_i]
returns:
rewards float[batch_i,seq_i] - what reward was given to an agent for corresponding action from state in that batch
"""
def compute_reward(batch_i,session_states,session_actions):
return self.get_reward(session_states,session_actions,batch_i)
sequences = [
T.arange(env_state_sessions.shape[0],),
env_state_sessions,
agent_action_sessions,
]
rewards,updates = theano.map(compute_reward,
sequences=sequences)
assert len(updates)==0
return rewards.reshape(agent_action_sessions.shape) #reshape bach to original
def connect(self, S):
self.S = S
def step(s_current, h_prev):
h_t = self.activation(
T.dot(s_current, self.W_ih) +
T.dot(h_prev, self.W_hh)
)
y_t = self.activation(
T.dot(h_t, self.W_ho)
)
return h_t, y_t
[self.H, self.output], _ = theano.scan(
step,
sequences = self.S,
outputs_info = [self.h_init, None]
)
self.prediction, _ = theano.map(
lambda x: T.argmax(x),
sequences = self.output
)
self.final_state = self.H[self.H.shape[0] - 1]
self.outputter = theano.function([self.S], self.output)
self.predicter = theano.function([self.S], self.prediction)
self.CONNECTED = True
def gaussian_filter_2d_variable_sigma(input,
sigmas,
window_radius=None,
border_mode='zero'):
def filter_sigma(idx, kernel):
dimpattern_w = ('x', 'x', 'x', 0)
dimpattern_h = ('x', 'x', 0, 'x')
filter_w = kernel.dimshuffle(dimpattern_w)
blur_w = T.nnet.conv2d(padded_input[idx:idx + 1],
filter_w,
border_mode=_get_chained_w_h_conv_border(
conv_border, 'w'),
filter_shape=[1, 1, 1, None])
filter_h = kernel.dimshuffle(dimpattern_h)
return T.nnet.conv2d(blur_w,
filter_h,
border_mode=_get_chained_w_h_conv_border(
conv_border, 'h'),
filter_shape=[1, 1, None, 1])
ndim = 4
assert input.ndim == ndim, \
"there must be {} dimensions, got {}".format(ndim, input.ndim)
window_radius = gaussian_kernel_default_radius(sigmas, window_radius)
padded_input, conv_border = add_border(input, window_radius, border_mode)
kernel = gaussian_kernel_1d(sigmas, window_radius)
blur, _ = theano.map(filter_sigma,
sequences=[T.arange(sigmas.shape[0]), kernel])
return blur.reshape(input.shape)
def getSample(self, Y, nSamp = 1):
def get_layers(ii):
output = lasagne.layers.get_output(lasagne.layers.get_all_layers(self.sbn_nn), inputs = Y)
return output[::-1]
output,_ = theano.map(get_layers, T.arange(nSamp))
return output
def compile_vime_reward(l_prediction,l_prev_states,l_actions,weights,
get_loss = lambda pred,real: T.mean((pred-real)**2),
n_samples = 1,
delta=0.01,**kwargs):
"""compiles a function that predicts vime reward for each state in a batch"""
prev_states = T.matrix("previous states")
actions = T.ivector("actions")
next_states = T.matrix("next states")
if n_samples ==1:
get_bnn = lambda state,action: lasagne.layers.get_output(l_prediction,
inputs={l_prev_states:state[None,:],
l_actions:action[None]},**kwargs)
else:
get_bnn = lambda state,action: sample_output(l_prediction,
input_dict={l_prev_states:state[None,:],
l_actions:action[None]},
n_samples=n_samples,**kwargs)
vime_reward_per_state,auto_updates = theano.map(lambda s,a,s_next: get_r_vime_on_state(weights,
get_loss(get_bnn(s,a),s_next),
delta),
sequences=[prev_states,actions,next_states])
return theano.function([prev_states,actions,next_states],vime_reward_per_state,
updates=auto_updates,allow_input_downcast=True)
def test_function5(self):
w = theano.shared(1.0, name="w")
def joke(a, b):
k = w * a
# g = 0.01 * T.grad((k - 1)**2, w)
return k, {w: w - 1.0}
x = T.dscalar("x")
hs, _ = theano.scan(joke, sequences=[np.array([1.0, 2.0, 3.0])], outputs_info=[np.float64(1.0)] )
print hs, _
def upd(h):
return T.grad(hs[h], w)
gs, up = theano.map(upd, sequences=[T.arange(hs.shape[0])])
print gs, up
# print hs, _
# print gs, up
func = theano.function(inputs=[], outputs=gs, updates= [])
print func()
print w.get_value()
def build_validation(X, Y):
def myscanfunc(ind):
X_ = X[ind*megabatch_size:(ind+1)*megabatch_size]
Y_ = Y[ind*megabatch_size:(ind+1)*megabatch_size]
return simple_build_likelihood(X_, Y_, ind=T.mod(pool_ind+ind, param_pool_size))
result = theano.map (myscanfunc, sequences=[T.arange(T.max([1, X.shape[0]/megabatch_size]))]) [0]
return T.sum(result)*dataset_size/X.shape[0]-_KLD
def compute_tree(self, emb_x, tree):
self.recursive_unit = self.create_recursive_unit()
self.leaf_unit = self.create_leaf_unit()
num_nodes = tree.shape[0] # num internal nodes
num_leaves = self.num_words - num_nodes
# compute leaf hidden states
leaf_h, _ = theano.map(fn=self.leaf_unit,
sequences=[emb_x[:num_leaves]])
if self.irregular_tree:
init_node_h = T.concatenate([leaf_h, leaf_h], axis=0)
else:
init_node_h = leaf_h
# use recurrence to compute internal node hidden states
def _recurrence(cur_emb, node_info, t, node_h, last_h):
child_exists = node_info > -1
offset = num_leaves * int(self.irregular_tree) - child_exists * t
child_h = node_h[node_info + offset] * child_exists.dimshuffle(
0, 'x')
parent_h = self.recursive_unit(cur_emb, child_h, child_exists)
node_h = T.concatenate(
[node_h, parent_h.reshape([1, self.hidden_dim])])
return node_h[1:], parent_h
dummy = theano.shared(self.init_vector([self.hidden_dim]))
(_, parent_h), _ = theano.scan(
fn=_recurrence,
outputs_info=[init_node_h, dummy],
sequences=[emb_x[num_leaves:], tree,
T.arange(num_nodes)],
n_steps=num_nodes)
return T.concatenate([leaf_h, parent_h], axis=0)
def gaussian_filter_2d_variable_sigma(input, sigmas,
window_radius=None,
border_mode='zero'
):
def filter_sigma(idx, kernel):
dimpattern_w = ('x', 'x', 'x', 0)
dimpattern_h = ('x', 'x', 0, 'x')
filter_w = kernel.dimshuffle(dimpattern_w)
blur_w = T.nnet.conv2d(
padded_input[idx:idx+1], filter_w,
border_mode=_get_chained_w_h_conv_border(conv_border, 'w'),
filter_shape=[1, 1, 1, None])
filter_h = kernel.dimshuffle(dimpattern_h)
return T.nnet.conv2d(
blur_w, filter_h,
border_mode=_get_chained_w_h_conv_border(conv_border, 'h'),
filter_shape=[1, 1, None, 1])
ndim = 4
assert input.ndim == ndim, \
"there must be {} dimensions, got {}".format(ndim, input.ndim)
window_radius = gaussian_kernel_default_radius(sigmas, window_radius)
padded_input, conv_border = add_border(input, window_radius, border_mode)
kernel = gaussian_kernel_1d(sigmas, window_radius)
blur, _ = theano.map(
filter_sigma,
sequences=[T.arange(sigmas.shape[0]), kernel])
return blur.reshape(input.shape)
def gradSort(met, X):
# -----------Start Batch Loop------------
m = T.fvector()
x = T.fmatrix()
# Sort the input on the metric
z = T.argsort(m, axis=0)
out = x[z] + 0 * T.sum(m)
sort = theano.function([m, x], [out])
# Fix the gradient of the sort operation to be the sum of
# the gradients with respect to the input features
def grad_edit(inps, grads):
m, x = inps
g, = grads
z = T.argsort(m, axis=0)
s = T.sum(g, axis=-1)
am = T.max(abs(s), axis=-1)
s = 10 * (s - T.clip(s, -.90 * am, .90 * am))
out = s
return out, g
op = theano.OpFromGraph([m, x], [out])
op.grad = grad_edit
results, updates = theano.map(fn=op, sequences=[met, X], name='batch_sort')
# ---------END Batch Loop-----------------
r = results
return r
def cosine_similarity(x, y, eps=1e-6):
r"""
Cosine similarity between a vector and each row of a base matrix.
Parameters
----------
x: a 1D Theano variable
Vector to compare to each row of the matrix y.
y: a 2D Theano variable
Matrix to be compared to
eps: float
Precision of the operation (necessary for differentiability).
Return
------
z: a 1D Theano variable
A vector whose components are the cosine similarities
between x and each row of y.
"""
def _cosine_similarity(x, y, eps=1e-6):
y = y.dimshuffle(1, 0)
z = T.dot(x, y)
z /= T.sqrt(T.sum(x * x) * T.sum(y * y, axis=0) + eps)
return z
def step(x_b, y_b):
return _cosine_similarity(x_b, y_b, eps)
z, _ = theano.map(step, sequences=[x, y])
return z
def attend(self, y_p):
updates = self.default_updates()
for g in range(self.attrs['glimpse']):
for i in range(len(self.base)-1,-1,-1):
factor = T.constant(self.base[i].attrs['factor'][0], 'int32') if i > 0 else 1
B, C, I, h_p, _ = self.get(y_p, i, g)
if i == len(self.base) - 1:
z_i = self.distance(C, h_p)
else:
length = T.cast(T.max(T.sum(I,axis=0))+1,'int32')
ext = T.cast(T.minimum(ext/factor,T.min(length)),'int32')
def pick(i_t, ext):
pad = T.minimum(i_t+ext, B.shape[0]) - ext
return T.concatenate([T.zeros((pad,), 'int8'), T.ones((ext,), 'int8'), T.zeros((B.shape[0]-pad-ext+1,), 'int8')], axis=0)
idx, _ = theano.map(pick, sequences = [pos/factor], non_sequences = [ext])
idx = (idx.dimshuffle(1,0)[:-1].flatten() > 0).nonzero()
C = C.reshape((C.shape[0]*C.shape[1],C.shape[2]))[idx].reshape((ext,C.shape[1],C.shape[2]))
z_i = self.distance(C, h_p)
I = I.reshape((I.shape[0]*I.shape[1],))[idx].reshape((ext,I.shape[1]))
if i > 0:
pos = T.argmax(self.softmax(z_i,I),axis=0) * factor
ext = factor
else:
w_i = self.softmax(z_i,I)
B = B.reshape((B.shape[0]*B.shape[1],B.shape[2]))[idx].reshape((ext,B.shape[1],B.shape[2]))
proto = T.sum(B * w_i.dimshuffle(0,1,'x').repeat(B.shape[2],axis=2),axis=0)
for i in range(len(self.base)):
self.glimpses[i].append(proto)
return T.dot(proto, self.custom_vars['W_att_in_0']), updates
def cross_cpu(self, entities):
n, m = entities.shape
pop = T.reshape(entities, (2, n * m / 2))
if self.fast_rng is None:
xpoints = self.rng.random_integers(size=(n / 2, ),
low=0,
high=m - 1)
else:
xpoints = self.fast_rng.uniform(size=(n / 2, ), low=0, high=m - 1)
xpoints = xpoints.astype('int32')
def choice_vector(xpoint, nbits):
return T.concatenate([
T.zeros((xpoint, ), dtype='uint8'),
T.ones((nbits - xpoint, ), dtype='uint8')
])
values, updates = theano.map(fn=choice_vector,
sequences=[xpoints],
non_sequences=[m],
name='choice_vector_building')
a = T.reshape(values, (n * m / 2, ))
pop = T.concatenate([T.choose(a, pop), T.choose(1 - a, pop)])
pop = T.reshape(pop, (n, m))
return pop
def theano_scan_color(writer, draw_fn):
with writer as writer_buf:
writer_buf_reshaped = writer_buf.reshape((Screen.screen_vane_count, Screen.screen_max_magnitude, 3))
vane_matrix = [[[float(vane), float(vane), float(vane)] for px in range(Screen.screen_max_magnitude)]
for vane in range(Screen.screen_vane_count)]
px_matrix = [[[float(px),float(px),float(px)] for px in range(Screen.screen_max_magnitude)]
for vane in range(Screen.screen_vane_count)]
col_matrix = [[[float(0), float(1), float(2)] for px in range(Screen.screen_max_magnitude)]
for vane in range(Screen.screen_vane_count)]
vane_vec = T.as_tensor(vane_matrix)
px_vec = T.as_tensor(px_matrix)
col_vec = T.as_tensor(col_matrix)
step = T.fscalar('step')
draw_fn_with_step = draw_fn(step)
f, _ = theano.map(draw_fn_with_step, [vane_vec, px_vec, col_vec])
fn_actual = theano.function([step], f, allow_input_downcast=True, on_unused_input='ignore')
step_actual = 0
while True:
writer.frame_ready()
start = time.time()
writer_buf_reshaped[:] = fn_actual(step_actual)
step_actual -= 1
done = time.time()
fps = 1.0/(done - start)
if fps < TARGET_FPS:
logging.warning('Frame rate is %f, which is lower than target %d', fps, TARGET_FPS)
def compute_tree(self, emb_x, tree):
self.recursive_unit = self.create_recursive_unit()
self.leaf_unit = self.create_leaf_unit()
num_nodes = tree.shape[0] # num internal nodes
num_leaves = self.num_words - num_nodes
# compute leaf hidden states
leaf_h, _ = theano.map(
fn=self.leaf_unit,
sequences=[emb_x[:num_leaves]])
# use recurrence to compute internal node hidden states
def _recurrence(cur_emb, node_info, t, node_h, last_h):
child_exists = node_info > -1
child_h = node_h[node_info - child_exists * t] * child_exists.dimshuffle(0, 'x')
parent_h = self.recursive_unit(cur_emb, child_h, child_exists)
node_h = T.concatenate([node_h,
parent_h.reshape([1, self.hidden_dim])])
return node_h[1:], parent_h
dummy = theano.shared(self.init_vector([self.hidden_dim]))
(_, parent_h), _ = theano.scan(
fn=_recurrence,
outputs_info=[leaf_h, dummy],
sequences=[emb_x[num_leaves:], tree, T.arange(num_nodes)],
n_steps=num_nodes)
return T.concatenate([leaf_h, parent_h], axis=0)
def attend(self, y_p):
updates = self.default_updates()
for g in range(self.attrs['glimpse']):
for i in range(len(self.base)-1,-1,-1):
factor = T.constant(self.base[i].attrs['factor'][0], 'int32') if i > 0 else 1
B, C, I, H, W_att_in, b_att_in = self.get(y_p, i, g)
if i == len(self.base) - 1:
z_i = self.distance(C, H)
else:
length = T.cast(T.max(T.sum(I,axis=0))+1,'int32')
ext = T.cast(T.minimum(ext/factor,T.min(length)),'int32')
def pick(i_t, ext):
pad = T.minimum(i_t+ext, B.shape[0]) - ext
return T.concatenate([T.zeros((pad,), 'int8'), T.ones((ext,), 'int8'), T.zeros((B.shape[0]-pad-ext+1,), 'int8')], axis=0)
idx, _ = theano.map(pick, sequences = [pos/factor], non_sequences = [ext])
idx = (idx.dimshuffle(1,0)[:-1].flatten() > 0).nonzero()
C = C.reshape((C.shape[0]*C.shape[1],C.shape[2]))[idx].reshape((ext,C.shape[1],C.shape[2]))
z_i = self.distance(C, H)
I = I.reshape((I.shape[0]*I.shape[1],))[idx].reshape((ext,I.shape[1]))
if i > 0:
pos = T.argmax(self.softmax(z_i,I),axis=0) * factor
ext = factor
else:
w_i = self.softmax(z_i,I)
B = B.reshape((B.shape[0]*B.shape[1],B.shape[2]))[idx].reshape((ext,B.shape[1],B.shape[2]))
proto = T.sum(B * w_i.dimshuffle(0,1,'x').repeat(B.shape[2],axis=2),axis=0)
for i in range(len(self.base)):
self.glimpses[i].append(proto)
return T.dot(proto, self.custom_vars['W_att_in_0']), updates
def get_reward_sequences(self, env_state_sessions, agent_action_sessions):
"""
computes the rewards given to agent at each time step for each batch
parameters:
env_state_seq - environment state [batch_i,seq_i,state_units] history for all sessions
agent_action_seq - int[batch_i,seq_i]
returns:
rewards float[batch_i,seq_i] - what reward was given to an agent for corresponding action from state in that batch
"""
env_state_sessions = check_list(env_state_sessions)
n_states = len(env_state_sessions)
agent_action_sessions = check_list(agent_action_sessions)
n_actions = len(agent_action_sessions)
def compute_reward(batch_i, *args):
session_states, session_actions = unpack_list(
args, [n_states, n_actions])
return self.get_reward(session_states, session_actions, batch_i)
sequences = [
T.arange(agent_action_sessions[0].shape[0], ),
] + env_state_sessions + agent_action_sessions
rewards, updates = theano.map(compute_reward, sequences=sequences)
assert len(updates) == 0
return rewards.reshape(
agent_action_sessions[0].shape) # reshape bach to original
def get_reward_sequences(self, env_state_sessions, agent_action_sessions):
"""Computes the rewards given to agent at each time step for each batch.
:param env_state_sessions: Environment state [batch_i,seq_i,state_units] history for all sessions.
:type env_state_sessions: theano tensor [batch_i,seq_i,state_units]
:param agent_action_sessions: Actions chosen by agent at each tick for all sessions.
:type agent_action_sessions: int[batch_i,seq_i]
:return rewards: What reward was given to an agent for corresponding action from state in that batch.
:rtype: float[batch_i,seq_i]
"""
env_state_sessions = check_list(env_state_sessions)
n_states = len(env_state_sessions)
agent_action_sessions = check_list(agent_action_sessions)
n_actions = len(agent_action_sessions)
def compute_reward(batch_i, *args):
session_states, session_actions = unpack_list(
args, [n_states, n_actions])
return self.get_reward(session_states, session_actions, batch_i)
sequences = [
T.arange(agent_action_sessions[0].shape[0], ),
] + env_state_sessions + agent_action_sessions
rewards, updates = theano.map(compute_reward, sequences=sequences)
assert len(updates) == 0
return rewards.reshape(
agent_action_sessions[0].shape) # reshape bach to original
def compute_objective_and_gradients(self, nSamp):
hsamp = self.mrec.getSample(self.Y, nSamp)
# evaluate the generative model density P_\theta(y_i , h_i)
p_yh,_ = theano.map(self.mprior.evaluateLogDensity, sequences=hsamp)
# evaluate the recognition model density Q_\phi(h_i | y_i)
q_hgy,_ = theano.map(self.mrec.evalLogDensity, sequences=hsamp)
ff = (p_yh-q_hgy)
sortidx = ff.argsort(axis=0)
fmax = ff[(sortidx[-1],T.arange(ff.shape[-1]))].dimshuffle('x',0)
f_hy = T.exp(ff - fmax)
sum_across_samples = f_hy.sum(axis=0, keepdims = True)
Lhat = T.log(sum_across_samples/nSamp) + fmax
col_idx = T.arange(ff.shape[-1])
# This 1e-12 constant is for debugging nans
# in other parts of code. We know we'll get
# nans where we'll then overwrite. Use it with
# compute cross-validated estimates of Lhat # nanguard mode.
hold_out_except_last = T.log((sum_across_samples - f_hy)/(nSamp-1)) + fmax #+1e-12) + fmax
f2max_vec = ff[(sortidx[-2],T.arange(ff.shape[-1]))]
f2max = f2max_vec.dimshuffle('x',0)
# Do tricky things to keep the numerics in order (avoid a term being \approxeq 0)
ff_nolast = T.set_subtensor(ff[(sortidx[-1],col_idx)], f2max_vec)
f_hy_last = T.exp(ff_nolast - f2max)
# compute held-out sum when we hold out the maximum element
hold_out_last = T.log((f_hy_last.sum(axis=0, keepdims=True) - f_hy_last)/(nSamp-1)) + f2max
# compute final held-out estimates
hold_out = T.set_subtensor(hold_out_except_last[(sortidx[-1],col_idx)], hold_out_last[(sortidx[-1],col_idx)])
Lhat_cv = Lhat - hold_out
the_ws = f_hy / sum_across_samples
weighted_q = T.sum((Lhat_cv*q_hgy + the_ws*ff).mean(axis=1))
#weighted_q = T.sum((Lhat_cv*q_hgy + the_ws*(p_yh-q_hgy)).sum(axis=1))
# gradients for approximate posterior
dqhgy = T.grad(cost=weighted_q, wrt = self.mrec.getParams(), consider_constant=([the_ws,Lhat_cv]+hsamp), disconnected_inputs='ignore')
# gradients for prior
dpyh = T.grad(cost=T.sum((the_ws*ff).mean(axis=1)), wrt = self.mprior.getParams(), consider_constant=hsamp + [the_ws], disconnected_inputs='ignore')
#dpyh = T.grad(cost=T.sum((the_ws*(p_yh-q_hgy)).sum(axis=1)), wrt = self.mprior.getParams(), consider_constant=hsamp + [the_ws], disconnected_inputs='ignore')
return [Lhat.mean(), dpyh, dqhgy]
def _score_ao_tg(self, tag_ids, word_ids):
output, _ = theano.map(fn=order0_ll_score_given_word_and_tag,
sequences=[tag_ids, word_ids],
non_sequences=[self._tg_lp_tag_np_table,
self._tg_tag_emb,
self._tg_word_emb],
name="_score_ao_tg")
return T.sum(output)
def hessian_diag1(f, v):
g = gradient1(f, v)
idx = tt.arange(g.shape[0], dtype='int32')
def hess_ii(i):
return gradient1(g[i], v)[i]
return theano.map(hess_ii, idx)[0]
def call(self, inputs, mask=None):
l1 = inputs[0]
l2 = inputs[1]
def f(i, l1, l2):
return T.clip(T.batched_tensordot(l1[i], l2[i], 1), FLOAT_MIN, FLOAT_MAX).astype(FLOATX)
return theano.map(f, T.arange(l1.shape[0]), non_sequences=[l1, l2])[0]
def hessian_diag1(f, v):
g = gradient1(f, v)
idx = tt.arange(g.shape[0], dtype='int32')
def hess_ii(i):
return gradient1(g[i], v)[i]
return theano.map(hess_ii, idx)[0]
def __call__(self,X):
#out = self.W[:,X]
def step(x):
return self.W[:x]
stk = theano.map( lambda x: self.W[x],X)
out = T.stacklists(stk[0])
#return out.dimshuffle('x','x',0,1)
return out
def sym_histograms(self, X):
"""
Encodes a set of objects (X is a tensor3)
:param X: tensor3 containing the feature vectors for each object
:return:
"""
histograms, updates = theano.map(self.sym_histogram, X)
return histograms
def __init__(
self,
rng,
input,
vocab_size,
embed_dm,
embeddings=None,
):
"""
input: theano.tensor.dmatrix, (number of instances, sentence word number)
vocab_size: integer, the size of vocabulary,
embed_dm: integer, the dimension of word vector representation
embeddings: theano.tensor.TensorType
pretrained embeddings
"""
if embeddings:
print "Use pretrained embeddings: ON"
assert embeddings.get_value().shape == (
vocab_size,
embed_dm), "%r != %r" % (embeddings.get_value().shape,
(vocab_size, embed_dm))
self.embeddings = embeddings
else:
print "Use pretrained embeddings: OFF"
embedding_val = np.asarray(rng.normal(0,
0.05,
size=(vocab_size, embed_dm)),
dtype=theano.config.floatX)
embedding_val[
vocab_size -
1, :] = 0 # the <PADDING> character is intialized to 0
self.embeddings = theano.shared(np.asarray(
embedding_val, dtype=theano.config.floatX),
borrow=True,
name='embeddings')
self.params = [self.embeddings]
self.param_shapes = [(vocab_size, embed_dm)]
# Return:
# :type, theano.tensor.tensor4
# :param, dimension(1, 1, word embedding dimension, number of words in sentence)
# made to be 4D to fit into the dimension of convolution operation
sent_embedding_list, updates = theano.map(
lambda sent: self.embeddings[sent], input)
sent_embedding_tensor = T.stacklists(
sent_embedding_list) # make it into a 3D tensor
self.output = sent_embedding_tensor.dimshuffle(
0, 'x', 2, 1) # make it a 4D tensor
def jacobian1(f, v):
"""jacobian of f wrt v"""
f = tt.flatten(f)
idx = tt.arange(f.shape[0], dtype='int32')
def grad_i(i):
return gradient1(f[i], v)
return theano.map(grad_i, idx)[0]
def _pv_function1(self, tensor_left, tensor_right, pf_matrix, vf_matrix): results, updates = theano.map( fn = lambda i, tensor_left, tensor_right, pf_matrix, vf_matrix: self.get_new_p(tensor_left, tensor_right, i, pf_matrix, vf_matrix), sequences=[T.arange(tensor_left, tensor_right)], non_sequences = [tensor_left, tensor_right, pf_matrix, vf_matrix], name = 'pv function' ) max_pf, index = T.max_and_argmax(results, axis=0) return [index + tensor_left, max_pf, self.get_new_v(tensor_left, tensor_right, index + tensor_left, vf_matrix)]
def call(self, inputs, mask=None):
def f(i, embedding, text_input):
mask = T.neq(text_input[i], 0).astype(FLOATX)
vec = T.dot(mask, embedding[i])
vec /= T.maximum(vec.norm(2, 0), K.epsilon())
return T.dot(vec, self.W) + self.b
return theano.map(f, T.arange(inputs[0].shape[0]), non_sequences=inputs)[0]
def read(self, img, center_x, center_y):
loc_x, loc_y = center_x, center_y
if img.ndim == 2:
img = self.matrix2tensor4(img)
batch_size = img.shape[0]
img_paded = self.padding_img(img)
retina, _ = theano.map(self.do_glimpes, sequences=[img_paded, loc_x,
loc_y])
return retina.reshape((batch_size, self.get_dim('glimpse')))
def jacobian1(f, v):
"""jacobian of f wrt v"""
f = tt.flatten(f)
idx = tt.arange(f.shape[0], dtype='int32')
def grad_i(i):
return gradient1(f[i], v)
return theano.map(grad_i, idx)[0]
def compute_cost_with_cross_entropy_in_parallel(original_rnn_outputs, labels, x_ends, y_ends): mask = T.log(1 - T.or_(T.eq(labels, T.zeros_like(labels)), T.eq(labels, shift_matrix(labels, 2)))) arange = T.arange(labels.shape[1]) initial_state = T.log(T.zeros_like(labels)) initial_state = T.set_subtensor(initial_state[:,0], 0) def select_probabilities(rnn_outputs, label): return rnn_outputs[:,label] rnn_outputs, _ = theano.map(select_probabilities, [original_rnn_outputs, labels]) rnn_outputs = T.log(rnn_outputs.dimshuffle((1,0,2))) def forward_step(probabilities, last_probabilities): all_forward_probabilities = T.stack( last_probabilities + probabilities, log_shift_matrix(last_probabilities, 1) + probabilities, log_shift_matrix(last_probabilities, 2) + probabilities + mask, ) max_probability, backlink = T.max_and_argmax(all_forward_probabilities, 0) backlink = arange - backlink return max_probability, backlink results, _ = theano.scan(fn = forward_step, sequences = rnn_outputs, outputs_info = [initial_state, None]) forward_probabilities, backward_pointers = results def compute_cost(rnn_outputs, forward_probabilities, backward_pointers, x_end, y_end, label): def backward_step(backlinks, position): new_position = backlinks[position] return new_position, position initial_state = T.argmax(forward_probabilities[x_end-1,y_end-2:y_end]) + y_end - 2 results, _ = theano.scan(fn = backward_step, sequences = backward_pointers[0:x_end,:], outputs_info = [initial_state, None], go_backwards = True) alignment = label[results[1][::-1]] return aggregate(categorical_crossentropy(rnn_outputs[0:x_end], alignment), mode='sum') forward_probabilities = forward_probabilities.dimshuffle((1,0,2)) backward_pointers = backward_pointers.dimshuffle((1,0,2)) return theano.map(compute_cost, [original_rnn_outputs, forward_probabilities, backward_pointers, x_ends, y_ends, labels])[0]
def __init__(self, rng,
input,
vocab_size,
embed_dm,
embeddings = None,
):
"""
input: theano.tensor.dmatrix, (number of instances, sentence word number)
vocab_size: integer, the size of vocabulary,
embed_dm: integer, the dimension of word vector representation
embeddings: theano.tensor.TensorType
pretrained embeddings
"""
if embeddings:
print "Use pretrained embeddings: ON"
assert embeddings.get_value().shape == (vocab_size, embed_dm), "%r != %r" %(
embeddings.get_value().shape,
(vocab_size, embed_dm)
)
self.embeddings = embeddings
else:
print "Use pretrained embeddings: OFF"
embedding_val = np.asarray(
rng.normal(0, 0.05, size = (vocab_size, embed_dm)),
dtype = theano.config.floatX
)
embedding_val[vocab_size-1,:] = 0 # the <PADDING> character is intialized to 0
self.embeddings = theano.shared(
np.asarray(embedding_val,
dtype = theano.config.floatX),
borrow = True,
name = 'embeddings'
)
self.params = [self.embeddings]
self.param_shapes = [(vocab_size, embed_dm)]
# Return:
# :type, theano.tensor.tensor4
# :param, dimension(1, 1, word embedding dimension, number of words in sentence)
# made to be 4D to fit into the dimension of convolution operation
sent_embedding_list, updates = theano.map(lambda sent: self.embeddings[sent],
input)
sent_embedding_tensor = T.stacklists(sent_embedding_list) # make it into a 3D tensor
self.output = sent_embedding_tensor.dimshuffle(0, 'x', 2, 1) # make it a 4D tensor
def pv_function(self, tensor_input): indexf_matrix = theano.shared( np.ones( (self.max_length, self.max_length), dtype=theano.config.floatX ), name = 'indexf_matrix', borrow=True ) pf_matrix = theano.shared( np.eye( self.max_length, dtype=theano.config.floatX ), name = 'pf_matrix', borrow=True ) vf_matrix = theano.shared( np.zeros( (self.max_length, self.max_length, self.size), dtype=theano.config.floatX ), name = 'vf_matrix', borrow=True ) results, updates = theano.reduce( fn = lambda i, t_vf_matrix, L, t_tensor_input: T.set_subtensor(t_vf_matrix[i, i], L[t_tensor_input[i]]), outputs_info = vf_matrix, sequences=[T.arange(tensor_input.shape[0])], non_sequences=[self.L, tensor_input], name = 'vf_matrix prepare' ) vf_matrix = results for i in range(1,self.max_length): ''' for j in range(self.max_length-i): new_index, new_pf, new_vf = self._pv_function1(j, j+i, pf_matrix, vf_matrix) indexf_matrix = T.set_subtensor(indexf_matrix[j, j+i], new_index) pf_matrix = T.set_subtensor(pf_matrix[j, j+i], new_pf) vf_matrix = T.set_subtensor(vf_matrix[j, j+i], new_vf) ''' results, updates = theano.map( fn = lambda j, pf_matrix, vf_matrix, i: self._pv_function1(j, j+i, pf_matrix, vf_matrix), sequences=[T.arange(self.max_length-i)], non_sequences = [pf_matrix, vf_matrix, i], #name = 'pv function' ) for j in range(self.max_length-i): indexf_matrix = T.set_subtensor(indexf_matrix[j, j+i], results[0][j]) pf_matrix = T.set_subtensor(pf_matrix[j, j+i], results[1][j]) vf_matrix = T.set_subtensor(vf_matrix[j, j+i], results[2][j]) return indexf_matrix, pf_matrix, vf_matrix
def make_gradlogps(mdp,agent):
o = TT.matrix("o",mdp.output_dtype("o"))
b = TT.matrix("b",agent.output_dtype("b"))
newa = agent.ponder({"o":o})["a"]
logp_n = agent.cpd().logliks(newa, b)
def onegrad(i):
logp1 = theano.clone(logp_n, replace = {b:b[i:i+1],o:o[i:i+1]})[0]
return symbolic.flatten(TT.grad(logp1, agent.policy_vars()))
gradlogps,_ = theano.map(onegrad, TT.arange(logp_n.shape[0]))
Glf.ftheano_gradlogp = theano.function([o,b],gradlogps)
Glf.f_gradlogp = staticmethod(lambda : G.pool.gather(gradlogpmapper,np.concatenate,None))
def order0_ll_score_given_word_only(tg_word_id,
tg_lp_tag_np_table, tg_tag_emb,
tg_word_emb,num_tag):
# Return the total
return log_sum_exp(theano.map(fn=order0_ll_score_given_word_and_tag,
sequences=[T.arange(num_tag)],
non_sequences=[tg_word_id,
tg_lp_tag_np_table,
tg_tag_emb,
tg_word_emb],
name="order0_ll_score_map")[0])
def _pv_function(self, tensor_left, length, pf_matrix, vf_matrix): tensor_right = tensor_left + length results, updates = theano.map( fn = self.get_new_p, sequences=[T.arange(tensor_left, tensor_right)], non_sequences = [tensor_left, tensor_right, pf_matrix, vf_matrix], name = 'pv function' ) max_pf, index = T.max_and_argmax(results, axis=0) max_vf = self.get_new_v(tensor_left, tensor_right, index + tensor_left, vf_matrix) return [index + tensor_left, max_pf, max_vf]
