def advance(u, u_1, u_2, f_a, Cx2, Cy2, dt2, V=None, step1=False):
u_in, u_1_in, u_2_in = T.fmatrices('u_in','u_1_in','u_2_in')
f_a_in, V_in = T.fmatrices('f_in','V_in')
step1_in = T.lscalar('step1_in')
Cx2_in, Cy2_in, dt2_in = T.fscalars('Cx2_in','Cy2_in','dt2_in')
u_out = T.fmatrix('u_out')
if V is None:
V = np.zeros_like(f_a)
step_f = theano.function([u_in, u_1_in, u_2_in, f_a_in, Cx2_in, Cy2_in, dt2_in, V_in, step1_in], u_out, step, on_unused_input='ignore')
u_out = step_f(u, u_1, u_2, f_a, Cx2, Cy2, dt2, V, step1)
return u_out
def get_samples(): # get samples from the model
X, Y = T.fmatrices(2)
givens_train_samples = {X: train_x[0:50000], Y: train_y[0:50000]}
H1, H2 = iteration(X, 15, 0.1)
# get prior statistics (100 mean and std)
H2_mean = T.mean(H2, axis=0)
H2_std = T.std(H2, axis=0)
# sampling h2 from prior
H2_ = RNG.normal((10000, 100),
avg=H2_mean,
std=4 * H2_std,
ndim=None,
dtype=H2.dtype,
nstreams=None)
# iterative sampling from samples h2
X_ = G1(G2(H2_))
for i in range(3):
H1_, H2_ = iteration(X_, 15, 0.1, 3)
X_ = G1(H1_)
#H1_, H2_ = iteration(X_, 1, 0.1, 3)
#X_ = G1(H1_)
sampling = theano.function([],
X_,
on_unused_input='ignore',
givens=givens_train_samples)
samples = sampling()
np.save('samples', samples)
return samples
def __init__(self,
steps = 1,
num_layers = 2,
num_units = 32,
eps = 1e-2):
self.X, self.Z = T.fvectors('X','Z')
self.P, self.Q, self.R = T.fmatrices('P','Q','R')
self.dt = T.scalar('dt')
self.matrix_inv = T.nlinalg.MatrixInverse()
self.ar = AutoRegressiveModel(steps = steps,
num_layers = num_layers,
num_units = num_units,
eps = eps)
l = InputLayer(input_var = self.X,
shape = (steps,))
l = ReshapeLayer(l, shape = (1,steps,))
l = self.ar.network(l)
l = ReshapeLayer(l, shape=(1,))
self.l_ = l
self.f_ = get_output(self.l_)
self.X_ = T.concatenate([self.f_, T.dot(T.eye(steps)[:-1], self.X)], axis=0)
self.fX_ = G.jacobian(self.X_.flatten(), self.X)
self.P_ = T.dot(T.dot(self.fX_, self.P), T.transpose(self.fX_)) + \
T.dot(T.dot(T.eye(steps)[:,0:1], self.dt * self.Q), T.eye(steps)[0:1,:])
self.h = T.dot(T.eye(steps)[0:1], self.X_)
self.y = self.Z - self.h
self.hX_ = G.jacobian(self.h, self.X_)
self.S = T.dot(T.dot(self.hX_, self.P_), T.transpose(self.hX_)) + self.R
self.K = T.dot(T.dot(self.P_, T.transpose(self.hX_)), self.matrix_inv(self.S))
self.X__ = self.X_ + T.dot(self.K, self.y)
self.P__ = T.dot(T.identity_like(self.P) - T.dot(self.K, self.hX_), self.P_)
self.prediction = theano.function(inputs = [self.X,
self.P,
self.Q,
self.dt],
outputs = [self.X_,
self.P_],
allow_input_downcast = True)
self.update = theano.function(inputs = [self.X,
self.Z,
self.P,
self.Q,
self.R,
self.dt],
outputs = [self.X__,
self.P__],
allow_input_downcast = True)
def SGD(eta, n_epochs, valid_steps, momentum, low, high, init, random_init='gaussian'):
t0 = time.time()
index = T.iscalar('index')
x, y, z, alpha = T.fmatrices('x', 'y', 'z', 'alpha')
n_minibatch = max_minibatch - 2
model = Model(n_tree, n_nodes, low, high, init, random_init)
model_op, auto_upd = model.op(x)
valid_op, valid_upd = model.valid_op(z, valid_steps)
loss = model.loss(y, model_op)
valid_loss = model.loss(alpha, valid_op)
print "Updation to be compiled yet"
params = model.params
train_upd = gradient_updates_momentum(loss, params, eta, momentum) + auto_upd
train_output = [model_op, loss]
valid_output = [valid_op, valid_loss]
print "Train function to be compiled"
train_fn = theano.function([index], train_output, updates=train_upd,
givens={x: train_x[:, n_in * index: n_in * (index + 1)],
y: train_x[:, (n_in * index + n_tree): (n_in * (index + 1) + 1)]}, name='train_fn')
valid_fn = theano.function([index], valid_output, updates=valid_upd,
givens={z: train_x[:, n_tree * index: n_tree * (index + 1)],
alpha: train_x[:, (n_in * index + n_tree): (n_in * index + n_tree + valid_steps)]}, name='valid_fn')
print "Train function compiled"
# Compilation over
#################
## TRAIN MODEL ##
#################
print 'The compilation time is', time.time() - t0
loss_list = []
for i in range(n_epochs):
epoch_loss = 0
t1 = time.time()
for idx in range(n_minibatch):
print 'The current idx is ', idx,' and the epoch number is ', i
output, loss_ = train_fn(idx)[:-1], train_fn(idx)[-1]
if idx%500 == 0:
v_output, v_loss = valid_fn(idx/500)[:-1][0], valid_fn(idx/500)[-1]
print 'v_pred is', ' '.join([mappings_words[prediction(abc)] for abc in v_output])
print 'v_loss is', np.array(v_loss)
print 'The loss is', loss_
epoch_loss +=loss_
loss_list.append(loss_)
print '=='*20
print 'The mean loss for the epoch was', epoch_loss/float(n_minibatch)
print 'Time taken by this epoch is', time.time()-t1
print '-'*50
pyplot.plot(loss_list)
pyplot.show()
def get_cost_updates(self, corruption_level, learning_rate, sample_method, enc_function):
""" This function computes the cost and the updates for one trainng
step of the dA """
tilde_x = self.get_corrupted_input(self.x, corruption_level)
y = self.get_hidden_values(tilde_x, enc_function)
z = self.get_reconstructed_input(y, enc_function)
L = T.fmatrices()
# if only encoding but not sample
if sample_method == -1:
if self.error_type == 1:
L = - T.sum(tilde_x * T.log(z) + (1 - tilde_x) * T.log(1 - z), axis=1)
#square error, added by feng
#print 'using'
if self.error_type == 0:
L = T.sum((tilde_x - z)**2, axis=1)
else:
sampled_x = self.get_sampled(tilde_x)
sampled_z = self.get_sampled(z)
#sampled version
if self.error_type == 1:
L = - T.sum(sampled_x * T.log(sampled_z) + (1 - sampled_x) * T.log(1 - sampled_z), axis=1)
#square error, added by feng
#print 'using'
if self.error_type == 0:
L = T.sum((sampled_x - sampled_z)**2, axis = 1)
# note : L is now a vector, where each element is the
# cross-entropy cost of the reconstruction of the
# corresponding example of the minibatch. We need to
# compute the average of all these to get the cost of
# the minibatch
cost = T.mean(L)
# compute the gradients of the cost of the `dA` with respect
# to its parameters
gparams = T.grad(cost, self.params)
# generate the list of updates
updates = [
(param, param - learning_rate * gparam)
for param, gparam in zip(self.params, gparams)
]
return (cost, updates)
def test_advinc_subtensor1():
""" Test the second case in the opt local_gpu_advanced_incsubtensor1 """
shared = cuda.shared_constructor
# shared = tensor.shared
xval = numpy.asarray([[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype="float32")
yval = numpy.asarray([[10, 10, 10], [10, 10, 10]], dtype="float32")
x = shared(xval, name="x")
y = T.fmatrices("y")
expr = T.advanced_inc_subtensor1(x, y, [0, 2])
f = theano.function([y], expr, mode=mode_with_gpu)
assert sum([isinstance(node.op, cuda.GpuAdvancedIncSubtensor1) for node in f.maker.env.toposort()]) == 1
assert numpy.allclose(f(yval), [[11.0, 12.0, 13.0], [4.0, 5.0, 6.0], [17.0, 18.0, 19.0]])
def test_set_subtensor():
shared = cuda.shared_constructor
#shared = tensor.shared
x,y = T.fmatrices('x','y')
xval = numpy.asarray([[1,2,3], [4,5,6], [7,8,9]],
dtype='float32')
yval = numpy.asarray([[10,10,10], [10,10,10], [10,10,10]],
dtype='float32')
expr = T.set_subtensor(x[:,1:3], y[:,1:3])
f=theano.function([x,y], expr, mode=mode_with_gpu)
assert sum([isinstance(node.op,cuda.GpuSubtensor) for node in f.maker.env.toposort() ])==1
assert sum([isinstance(node.op,cuda.GpuIncSubtensor) and node.op.set_instead_of_inc==True for node in f.maker.env.toposort() ])==1
print f(xval,yval)
def test_advinc_subtensor1():
""" Test the second case in the opt local_gpu_advanced_incsubtensor1 """
shared = cuda.shared_constructor
#shared = tensor.shared
xval = numpy.asarray([[1,2,3], [4,5,6], [7,8,9]],
dtype='float32')
yval = numpy.asarray([[10,10,10], [10,10,10]],
dtype='float32')
x = shared(xval, name = 'x')
y = T.fmatrices('y')
expr = T.advanced_inc_subtensor1(x,y,[0,2])
f=theano.function([y], expr, mode=mode_with_gpu)
assert sum([isinstance(node.op,cuda.GpuAdvancedIncSubtensor1) for node in f.maker.env.toposort() ])==1
assert numpy.allclose(f(yval),[[11.,12.,13.], [4.,5.,6.], [17.,18.,19.]])
def test_abs_cost():
ySym,yhatSym = T.fmatrices('y','yhat')
ac = theano.function([yhatSym,ySym],
outputs=absoluteError(yhatSym,ySym))
yhat = np.asarray([[1],[2],[3]],dtype=theano.config.floatX)
y = np.asarray([[1],[2],[3]],dtype=theano.config.floatX)
assert np.abs(ac(yhat,y)) < 1e-5
yhat = np.asarray([[1],[2.1],[3]],dtype=theano.config.floatX)
assert np.abs(ac(yhat,y) - 0.1/3) < 1e-5
def test_squared_error_cost():
ySym,yhatSym = T.fmatrices('y','yhat')
sqerr = theano.function([yhatSym,ySym],
outputs=squaredError(yhatSym,ySym))
yhat = np.asarray([[1],[2],[3]],dtype=theano.config.floatX)
y = np.asarray([[1],[2],[3]],dtype=theano.config.floatX)
assert np.abs(sqerr(yhat,y)) < 1e-5
yhat = np.asarray([[1],[2.1],[3]],dtype=theano.config.floatX)
assert np.abs(sqerr(yhat,y) - 0.01/3) < 1e-5
def test_inc_subtensor():
shared = cuda.shared_constructor
#shared = tensor.shared
x, y = T.fmatrices('x', 'y')
xval = numpy.asarray([[1, 2, 3], [4, 5, 6], [7, 8, 9]],
dtype='float32')
yval = numpy.asarray([[10, 10, 10], [10, 10, 10], [10, 10, 10]],
dtype='float32')
expr = T.inc_subtensor(x[:, 1:3], y[:, 1:3])
f = theano.function([x, y], expr, mode=mode_with_gpu)
assert sum([isinstance(node.op, cuda.GpuSubtensor)
for node in f.maker.fgraph.toposort()]) == 1
assert sum([isinstance(node.op, cuda.GpuIncSubtensor) and
node.op.set_instead_of_inc==False
for node in f.maker.fgraph.toposort()]) == 1
assert numpy.allclose(f(xval, yval), [[1., 12., 13.],
[4., 15., 16.], [7., 18., 19.]])
def test_inc_subtensor():
shared = cuda.shared_constructor
#shared = tensor.shared
x, y = T.fmatrices('x', 'y')
xval = numpy.asarray([[1, 2, 3], [4, 5, 6], [7, 8, 9]],
dtype='float32')
yval = numpy.asarray([[10, 10, 10], [10, 10, 10], [10, 10, 10]],
dtype='float32')
expr = T.inc_subtensor(x[:, 1:3], y[:, 1:3])
f = theano.function([x, y], expr, mode=mode_with_gpu)
assert sum([isinstance(node.op, cuda.GpuSubtensor)
for node in f.maker.fgraph.toposort()]) == 1
assert sum([isinstance(node.op, cuda.GpuIncSubtensor) and
node.op.set_instead_of_inc==False
for node in f.maker.fgraph.toposort()]) == 1
assert numpy.allclose(f(xval, yval), [[1., 12., 13.],
[4., 15., 16.], [7., 18., 19.]])
from theano import *
import theano.tensor as T
import numpy as np
# Logistic function
x = T.matrix('x', 'float32')
op = 1 / (1 + T.exp(-x))
logistic = function([x], op)
mat1 = [[0, 1], [-1, -2]]
print(logistic(mat1))
# Multiple outputs
a, b = T.fmatrices('a', 'b')
diff = a - b
absDiff = abs(diff)
sqrDiff = diff**2
f = function([a, b], [diff, absDiff, sqrDiff])
mat2 = [[10, 5], [5, 10]]
mat3 = [[5, 10], [10, 5]]
print(f(mat2, mat3))
# Default values
x, y = T.fscalars('x', 'y')
z = x + y
def __init__(self, We_initial, char_embedd_table_initial, params):
We = theano.shared(We_initial)
# initial embedding for the InfNet
We_inf = theano.shared(We_initial)
embsize = We_initial.shape[1]
hidden = params.hidden
self.en_hidden_size = params.hidden_inf
self.num_labels = 17
self.de_hidden_size = params.de_hidden_size
char_embedd_dim = params.char_embedd_dim
char_dic_size = len(params.char_dic)
char_embedd_table = theano.shared(char_embedd_table_initial)
char_embedd_table_inf = theano.shared(char_embedd_table_initial)
input_var = T.imatrix(name='inputs')
target_var = T.imatrix(name='targets')
target_var_in = T.imatrix(name='targets')
mask_var = T.fmatrix(name='masks')
mask_var1 = T.fmatrix(name='masks1')
char_input_var = T.itensor3(name='char-inputs')
length = T.iscalar()
length0 = T.iscalar()
t_t = T.fscalar()
t_t0 = T.fscalar()
use_dropout = T.fscalar()
use_dropout0 = T.fscalar()
Wyy0 = np.random.uniform(-0.02, 0.02, (self.num_labels +1 , self.num_labels + 1)).astype('float32')
Wyy = theano.shared(Wyy0)
l_in_word = lasagne.layers.InputLayer((None, None))
l_mask_word = lasagne.layers.InputLayer(shape=(None, None))
if params.emb ==1:
l_emb_word = lasagne.layers.EmbeddingLayer(l_in_word, input_size= We_initial.shape[0] , output_size = embsize, W =We, name='word_embedding')
else:
l_emb_word = lasagne_embedding_layer_2(l_in_word, embsize, We)
layer_char_input = lasagne.layers.InputLayer(shape=(None, None, Max_Char_Length ),
input_var=char_input_var, name='char-input')
layer_char = lasagne.layers.reshape(layer_char_input, (-1, [2]))
layer_char_embedding = lasagne.layers.EmbeddingLayer(layer_char, input_size=char_dic_size,
output_size=char_embedd_dim, W=char_embedd_table,
name='char_embedding')
layer_char = lasagne.layers.DimshuffleLayer(layer_char_embedding, pattern=(0, 2, 1))
# first get some necessary dimensions or parameters
conv_window = 3
num_filters = params.num_filters
# construct convolution layer
cnn_layer = lasagne.layers.Conv1DLayer(layer_char, num_filters=num_filters, filter_size=conv_window, pad='full',
nonlinearity=lasagne.nonlinearities.tanh, name='cnn')
# infer the pool size for pooling (pool size should go through all time step of cnn)
_, _, pool_size = cnn_layer.output_shape
# construct max pool layer
pool_layer = lasagne.layers.MaxPool1DLayer(cnn_layer, pool_size=pool_size)
# reshape the layer to match lstm incoming layer [batch * sent_length, num_filters, 1] --> [batch, sent_length, num_filters]
output_cnn_layer = lasagne.layers.reshape(pool_layer, (-1, length, [1]))
# finally, concatenate the two incoming layers together.
incoming = lasagne.layers.concat([output_cnn_layer, l_emb_word], axis=2)
l_lstm_wordf = lasagne.layers.LSTMLayer(incoming, hidden, mask_input=l_mask_word)
l_lstm_wordb = lasagne.layers.LSTMLayer(incoming, hidden, mask_input=l_mask_word, backwards = True)
concat = lasagne.layers.concat([l_lstm_wordf, l_lstm_wordb], axis=2)
l_reshape_concat = lasagne.layers.ReshapeLayer(concat,(-1,2*hidden))
l_local = lasagne.layers.DenseLayer(l_reshape_concat, num_units= self.num_labels + 1, nonlinearity=lasagne.nonlinearities.linear)
network_params = lasagne.layers.get_all_params(l_local, trainable=True)
network_params.append(Wyy)
print len(network_params)
f = open('NER_BiLSTM_CNN_CRF_.Batchsize_10_dropout_1_LearningRate_0.005_0.0_50_hidden_200.pickle','r')
data = pickle.load(f)
f.close()
for idx, p in enumerate(network_params):
p.set_value(data[idx])
self.params = []
self.hos = []
self.Cos = []
self.encoder_lstm_layers = []
self.decoder_lstm_layers = []
self.lstm_layers_num = 1
ei, di, dt = T.imatrices(3) #place holders
decoderInputs0 ,em, em1, dm, tf, di0 =T.fmatrices(6)
ci = T.itensor3()
#### the last one is for the stary symbole
self.de_lookuptable = theano.shared(name="Decoder LookUpTable", value=init_xavier_uniform(self.num_labels +1, self.de_hidden_size), borrow=True)
self.linear = theano.shared(name="Linear", value = init_xavier_uniform(self.de_hidden_size + 2*self.en_hidden_size, self.num_labels), borrow= True)
self.linear_bias = theano.shared(name="Hidden to Bias", value=np.asarray(np.random.randn(self.num_labels, )*0., dtype=theano.config.floatX), borrow=True)
#self.hidden_decode = theano.shared(name="Hidden to Decode", value= init_xavier_uniform(2*hidden, self.de_hidden_size), borrow = True)
#self.hidden_bias = theano.shared(
# name="Hidden to Bias",
# value=np.asarray(np.random.randn(self.de_hidden_size, )*0., dtype=theano.config.floatX) ,
# borrow=True
# )
input_var_shuffle = input_var.dimshuffle(1, 0)
mask_var_shuffle = mask_var.dimshuffle(1, 0)
target_var_in_shuffle = target_var_in.dimshuffle(1,0)
target_var_shuffle = target_var.dimshuffle(1,0)
self.params += [We_inf, self.linear, self.de_lookuptable, self.linear_bias]
######[batch, sent_length, embsize]
state_below = We_inf[input_var_shuffle.flatten()].reshape((input_var_shuffle.shape[0], input_var_shuffle.shape[1], embsize))
###### character word embedding
layer_char_input_inf = lasagne.layers.InputLayer(shape=(None, None, Max_Char_Length ),
input_var=char_input_var, name='char-input')
layer_char_inf = lasagne.layers.reshape(layer_char_input_inf, (-1, [2]))
layer_char_embedding_inf = lasagne.layers.EmbeddingLayer(layer_char_inf, input_size=char_dic_size,
output_size=char_embedd_dim, W=char_embedd_table_inf,
name='char_embedding_inf')
layer_char_inf = lasagne.layers.DimshuffleLayer(layer_char_embedding_inf, pattern=(0, 2, 1))
#layer_char_inf = lasagne.layers.DropoutLayer(layer_char_inf, p=0.5)
cnn_layer_inf = lasagne.layers.Conv1DLayer(layer_char_inf, num_filters=num_filters, filter_size=conv_window, pad='full',
nonlinearity=lasagne.nonlinearities.tanh, name='cnn_inf')
pool_layer_inf = lasagne.layers.MaxPool1DLayer(cnn_layer_inf, pool_size=pool_size)
output_cnn_layer_inf = lasagne.layers.reshape(pool_layer_inf, (-1, length, [1]))
char_params = lasagne.layers.get_all_params(output_cnn_layer_inf, trainable=True)
self.params += char_params
###### [batch, sent_length, num_filters]
#char_state_below = lasagne.layers.get_output(output_cnn_layer_inf, {layer_char_input_inf:char_input_var})
char_state_below = lasagne.layers.get_output(output_cnn_layer_inf)
char_state_below = dropout_layer(char_state_below, use_dropout, trng)
char_state_shuff = char_state_below.dimshuffle(1,0, 2)
state_below = T.concatenate([state_below, char_state_shuff], axis=2)
state_below = dropout_layer(state_below, use_dropout, trng)
enclstm_f = LSTM(embsize+num_filters, self.en_hidden_size)
enclstm_b = LSTM(embsize+num_filters, self.en_hidden_size, True)
self.encoder_lstm_layers.append(enclstm_f) #append
self.encoder_lstm_layers.append(enclstm_b) #append
self.params += enclstm_f.params + enclstm_b.params #concatenate
hs_f, Cs_f = enclstm_f.forward(state_below, mask_var_shuffle)
hs_b, Cs_b = enclstm_b.forward(state_below, mask_var_shuffle)
hs = T.concatenate([hs_f, hs_b], axis=2)
Cs = T.concatenate([Cs_f, Cs_b], axis=2)
hs0 = T.concatenate([hs_f[-1], hs_b[0]], axis=1)
Cs0 = T.concatenate([Cs_f[-1], Cs_b[0]], axis=1)
#self.hos += T.tanh(tensor.dot(hs0, self.hidden_decode) + self.hidden_bias),
#self.Cos += T.tanh(tensor.dot(Cs0, self.hidden_decode) + self.hidden_bias),
self.hos += T.alloc(np.asarray(0., dtype=theano.config.floatX), input_var_shuffle.shape[1], self.de_hidden_size),
self.Cos += T.alloc(np.asarray(0., dtype=theano.config.floatX), input_var_shuffle.shape[1], self.de_hidden_size),
Encoder = hs
state_below = self.de_lookuptable[target_var_in_shuffle.flatten()].reshape((target_var_in_shuffle.shape[0], target_var_in_shuffle.shape[1], self.de_hidden_size))
for i in range(self.lstm_layers_num):
declstm = LSTM(self.de_hidden_size, self.de_hidden_size)
self.decoder_lstm_layers += declstm, #append
self.params += declstm.params #concatenate
ho, Co = self.hos[i], self.Cos[i]
state_below, Cs = declstm.forward(state_below, mask_var_shuffle, ho, Co)
decoder_lstm_outputs = T.concatenate([state_below, Encoder], axis=2)
linear_outputs = T.dot(decoder_lstm_outputs, self.linear) + self.linear_bias[None, None, :]
softmax_outputs, updates = theano.scan(
fn=lambda x: T.nnet.softmax(x),
sequences=[linear_outputs],
)
def _NLL(pred, y, m):
return -m * T.log(pred[T.arange(input_var.shape[0]), y])
"""
costs, _ = theano.scan(fn=_NLL, sequences=[softmax_outputs, target_var_shuffle, mask_var_shuffle])
#loss = costs.sum() / mask_var.sum() + params.L2*sum(lasagne.regularization.l2(x) for x in self.params)
loss = costs.sum() / mask_var.sum()
updates = lasagne.updates.sgd(loss, self.params, self.eta)
updates = lasagne.updates.apply_momentum(updates, self.params, momentum=0.9)
###################################################
#### using the ground truth when training
##################################################
self._train = theano.function(
inputs=[ei, em, di, dm, dt],
outputs=[loss, softmax_outputs],
updates=updates,
givens={input_var:ei, mask_var:em, target_var_in:di, decoderMask:dm, target_var:dt}
)
"""
def _step2(ctx_, state_, hs_, Cs_):
hs, Cs = [], []
token_idxs = T.cast(state_.argmax(axis=-1), "int32" )
msk_ = T.fill( (T.zeros_like(token_idxs, dtype="float32")), 1.)
msk_ = msk_.dimshuffle('x', 0)
state_below0 = self.de_lookuptable[token_idxs].reshape((1, ctx_.shape[0], self.de_hidden_size))
for i, lstm in enumerate(self.decoder_lstm_layers):
h, C = lstm.forward(state_below0, msk_, hs_[i], Cs_[i]) #mind msk
hs += h[-1],
Cs += C[-1],
state_below0 = h
hs, Cs = T.as_tensor_variable(hs), T.as_tensor_variable(Cs)
state_below0 = state_below0.reshape((ctx_.shape[0], self.de_hidden_size))
state_below0 = T.concatenate([ctx_, state_below0], axis =1)
newpred = T.dot(state_below0, self.linear) + self.linear_bias[None, :]
state_below = T.nnet.softmax(newpred)
##### the beging symbole probablity is 0
extra_p = T.zeros_like(hs[:,:,0])
state_below = T.concatenate([state_below, extra_p.T], axis=1)
return state_below, hs, Cs
hs0, Cs0 = T.as_tensor_variable(self.hos, name="hs0"), T.as_tensor_variable(self.Cos, name="Cs0")
train_outputs, _ = theano.scan(
fn=_step2,
sequences = [Encoder],
outputs_info=[decoderInputs0, hs0, Cs0],
n_steps=input_var_shuffle.shape[0]
)
predy = train_outputs[0].dimshuffle(1, 0 , 2)
predy = predy[:,:,:-1]*mask_var[:,:,None]
predy0 = predy.reshape((-1, 17))
def inner_function( targets_one_step, mask_one_step, prev_label, tg_energy):
"""
:param targets_one_step: [batch_size, t]
:param prev_label: [batch_size, t]
:param tg_energy: [batch_size]
:return:
"""
new_ta_energy = T.dot(prev_label, Wyy[:-1,:-1])
new_ta_energy_t = tg_energy + T.sum(new_ta_energy*targets_one_step, axis =1)
tg_energy_t = T.switch(mask_one_step, new_ta_energy_t, tg_energy)
return [targets_one_step, tg_energy_t]
local_energy = lasagne.layers.get_output(l_local, {l_in_word: input_var, l_mask_word: mask_var, layer_char_input:char_input_var})
local_energy = local_energy.reshape((-1, length, 17))
local_energy = local_energy*mask_var[:,:,None]
#####################
# for the end symbole of a sequence
####################
end_term = Wyy[:-1,-1]
local_energy = local_energy + end_term.dimshuffle('x', 'x', 0)*mask_var1[:,:, None]
#predy0 = lasagne.layers.get_output(l_local_a, {l_in_word_a:input_var, l_mask_word_a:mask_var})
predy_in = T.argmax(predy0, axis=1)
A = T.extra_ops.to_one_hot(predy_in, 17)
A = A.reshape((-1, length, 17))
#predy = predy0.reshape((-1, length, 25))
#predy = predy*mask_var[:,:,None]
targets_shuffled = predy.dimshuffle(1, 0, 2)
target_time0 = targets_shuffled[0]
masks_shuffled = mask_var.dimshuffle(1, 0)
initial_energy0 = T.dot(target_time0, Wyy[-1,:-1])
initials = [target_time0, initial_energy0]
[ _, target_energies], _ = theano.scan(fn=inner_function, outputs_info=initials, sequences=[targets_shuffled[1:], masks_shuffled[1:]])
cost11 = target_energies[-1] + T.sum(T.sum(local_energy*predy, axis=2)*mask_var, axis=1)
cost = T.mean(-cost11)
from momentum import momentum
updates_a = momentum(cost, self.params, params.eta, momentum=0.9)
self.train_fn = theano.function(
inputs=[ei, ci, em, em1, length0, di0, use_dropout0],
outputs=[cost],
updates=updates_a,
on_unused_input='ignore',
givens={input_var:ei, char_input_var:ci, mask_var:em, mask_var1:em1, length: length0, decoderInputs0:di0, use_dropout:use_dropout0}
)
prediction = T.argmax(predy, axis=2)
corr = T.eq(prediction, target_var)
corr_train = (corr * mask_var).sum(dtype=theano.config.floatX)
num_tokens = mask_var.sum(dtype=theano.config.floatX)
self.eval_fn = theano.function(
inputs=[ei, ci, em, em1, length0, di0, use_dropout0],
outputs=[prediction, -cost11],
on_unused_input='ignore',
givens={input_var:ei, char_input_var:ci, mask_var:em, mask_var1:em1, length: length0, decoderInputs0:di0, use_dropout:use_dropout0}
)
from theano import *
import theano.tensor as T
import numpy as np
# Logistic function
x = T.matrix('x', 'float32')
op = 1 / (1 + T.exp(-x))
logistic = function([x], op)
mat1 = [[0, 1], [-1, -2]]
print(logistic(mat1))
# Multiple outputs
a, b = T.fmatrices('a', 'b')
diff = a - b
absDiff = abs(diff)
sqrDiff = diff ** 2
f = function([a, b], [diff, absDiff, sqrDiff])
mat2 = [[10, 5], [5, 10]]
mat3 = [[5, 10], [10, 5]]
print(f(mat2, mat3))
# Default values
x, y = T.fscalars('x', 'y')
z = x + y
def __init__(self,
steps = 1,
num_layers = 2,
num_units = 32,
eps = 1e-2,
alpha = 1e-2,
beta = 2.0,
kappa = 0.0):
lam = alpha * alpha * (steps + kappa) - steps + beta
self.X, self.Z = T.fvectors('X','Z')
self.P, self.Q, self.R = T.fmatrices('P','Q','R')
self.dt = T.scalar('dt')
sqrtm = MatrixSqrt()
self.matrix_inv = T.nlinalg.MatrixInverse()
self.ar = AutoRegressiveModel(steps = steps,
num_layers = num_layers,
num_units = num_units,
eps = eps)
def weighted_mean(A,w):
mu = T.zeros((steps, 1))
for i in range(2 * steps + 1):
mu += w[i] * A[:,i:i+1]
return mu
def weighted_covariance(A,B,a,b,w):
sigma = T.zeros((steps,steps))
for i in range(2 * steps + 1):
sigma += w[i] * T.dot((A[:,i:i+1] - a), (B[:,i:i+1] - b).T)
return sigma
self.sqrtP = sqrtm(self.P)
self.XB = T.dot(T.stack(self.X).T, T.ones((1, 2 * steps +1))) + T.concatenate([T.zeros((steps,1)),
T.sqrt(steps + lam) * self.sqrtP,
-T.sqrt(steps + lam) * self.sqrtP], axis=1)
l = InputLayer(input_var = self.XB.T,
shape = (2 * steps + 1, steps))
l = self.ar.network(l)
l = ReshapeLayer(l, shape=(1, 2 * steps + 1))
self.l_ = l
self.f_ = get_output(self.l_)
self.XC = T.concatenate([self.f_, T.dot(T.eye(steps)[:-1], self.XB)], axis=0)
W_m = T.concatenate([(lam / (steps + lam)) * T.ones(1),
(1.0 / (2.0 * (steps + lam))) * T.ones(2 * steps)], axis=0)
W_c = T.concatenate([(lam / (steps + lam) + (1.0 - alpha * alpha + beta)) * T.ones(1),
(1.0 / (2.0 * (steps + lam))) * T.ones(2 * steps)], axis=0)
self.X_ = weighted_mean(self.XC, W_m)
self.P_ = weighted_covariance(self.XC, self.XC, self.X_, self.X_, W_c) + \
T.dot(T.dot(T.eye(steps)[:,0:1], self.dt * self.Q), T.eye(steps)[0:1,:])
self.ZB = T.dot(T.eye(steps)[0:1,:], self.XC)
self.Z_ = weighted_mean(self.ZB, W_m)
self.S = weighted_covariance(self.ZB, self.ZB, self.Z_, self.Z_, W_c) + self.R
self.K = T.dot(weighted_covariance(self.XC, self.ZB, self.X_, self.Z_, W_c),
self.matrix_inv(self.S))
self.X__ = self.X_ + T.dot(self.K, self.Z - self.Z_)
self.P__ = self.P_ - T.dot(T.dot(self.K, self.S), self.K.T)
self.prediction = theano.function(inputs = [self.X,
self.P,
self.Q,
self.dt],
outputs = [self.X_,
self.P_],
allow_input_downcast = True)
self.update = theano.function(inputs = [self.X,
self.Z,
self.P,
self.Q,
self.R,
self.dt],
outputs = [self.X__,
self.P__],
allow_input_downcast = True)
def exp(__lr):
max_epochs, batch_size, n_batches = 1000, 100, 500 # = 50000/100
nX, nH1, nH2 = 784, 1000, 100
W1 = rand_ortho((nX, nH1), np.sqrt(6. / (nX + nH1)))
B1 = zeros((nH1, ))
W2 = rand_ortho((nH1, nH2), np.sqrt(6. / (nH1 + nH2)))
B2 = zeros((nH2, ))
V1 = rand_ortho((nH1, nX), np.sqrt(6. / (nH1 + nX)))
C1 = zeros((nX, ))
V2 = rand_ortho((nH2, nH1), np.sqrt(6. / (nH2 + nH1)))
C2 = zeros((nH1, ))
# layer definitions - functions of layers
F1 = lambda x: softplus(T.dot(x, W1) + B1)
G1 = lambda h1: sigm(T.dot(h1, V1) + C1)
F2 = lambda h1: sigm(T.dot(h1, W2) + B2)
G2 = lambda h2: softplus(T.dot(h2, V2) + C2)
i, e = T.lscalar(), T.fscalar()
X, Y = T.fmatrices(2)
givens_train = lambda i: {
X: train_x[i * batch_size:(i + 1) * batch_size],
Y: train_y[i * batch_size:(i + 1) * batch_size]
}
givens_valid, givens_test = {
X: valid_x,
Y: valid_y
}, {
X: test_x,
Y: test_y
}
givens_empty = {
X: sharedX(np.zeros((10000, 784))),
Y: sharedX(np.zeros((10000, 10)))
}
def iteration(X, k, alpha, beta=0.01): # infer h1 and h2 from x
H1 = F1(X)
H2 = F2(H1)
for i in range(k):
H2 = H2 + alpha * (F2(H1) - F2(G2(H2)))
H1 = H1 + alpha * (F1(X) - F1(G1(H1))) + alpha * beta * (G2(H2) -
H1)
return H1, H2
H1, H2 = F1(X), F2(F1(X))
H1_, H2_ = iteration(X, 15, 0.1)
def avg_bin(x, k): # average of sampled random binary values
S = 0. * x
for i in range(k):
S = S + samp(x)
return S / k
# get gradients
g_V1, g_C1 = T.grad(mse(G1(gaussian(H1_, 0.3)), X), [V1, C1],
consider_constant=[H1_, X])
g_W1, g_B1 = T.grad(mse(F1(gaussian(X, 0.5)), H1_), [W1, B1],
consider_constant=[X, H1_])
g_V2, g_C2 = T.grad(mse(G2(avg_bin(H2_, 3)), H1_), [V2, C2],
consider_constant=[H2_, H1_])
g_W2, g_B2 = T.grad(mse(F2(gaussian(H1_, 0.5)), H2_), [W2, B2],
consider_constant=[H1_, H2_])
cost = mse(G1(G2(F2(F1(X)))), X)
# training
train_sync = theano.function([i, e], [cost],
givens=givens_train(i),
on_unused_input='ignore',
updates=rms_prop(
{
W1: g_W1,
B1: g_B1,
V1: g_V1,
C1: g_C1,
W2: g_W2,
B2: g_B2,
V2: g_V2,
C2: g_C2
}, __lr))
def get_samples(): # get samples from the model
X, Y = T.fmatrices(2)
givens_train_samples = {X: train_x[0:50000], Y: train_y[0:50000]}
H1, H2 = iteration(X, 15, 0.1)
# get prior statistics (100 mean and std)
H2_mean = T.mean(H2, axis=0)
H2_std = T.std(H2, axis=0)
# sampling h2 from prior
H2_ = RNG.normal((10000, 100),
avg=H2_mean,
std=4 * H2_std,
ndim=None,
dtype=H2.dtype,
nstreams=None)
# iterative sampling from samples h2
X_ = G1(G2(H2_))
for i in range(3):
H1_, H2_ = iteration(X_, 15, 0.1, 3)
X_ = G1(H1_)
#H1_, H2_ = iteration(X_, 1, 0.1, 3)
#X_ = G1(H1_)
sampling = theano.function([],
X_,
on_unused_input='ignore',
givens=givens_train_samples)
samples = sampling()
np.save('samples', samples)
return samples
# get test log-likelihood
def test_ll(sigma):
samples = get_samples()
return get_ll(np_test_x, theano_parzen(samples, sigma), batch_size=10)
test_cost = theano.function([i, e], [cost],
on_unused_input='ignore',
givens=givens_test)
print('epochs test_loglikelihood time')
# training loop
t = time.time()
monitor = {
'train': [],
'valid': [],
'test': [],
'test_ll': [],
'test_ll_base': []
}
for e in range(1, max_epochs + 1):
monitor['train'].append(
np.array([train_sync(i, e)
for i in range(n_batches)]).mean(axis=0))
if e % 5 == 0:
monitor['test'].append(test_cost(0, 0))
monitor['test_ll'].append(np.mean(test_ll(0.2)))
print(e, monitor['test_ll'][-1], time.time() - t)
# Input Layer
l_in = InputLayer((batch_size, n_in), input_var=input_var)
# Recurrent EI Net
l_in_hid = DenseLayer(l_in, n_hid, nonlinearity=lasagne.nonlinearities.rectify)
# Output Layer
l_shp = ReshapeLayer(l_in_hid, (-1, n_hid))
l_dense = DenseLayer(l_shp, num_units=n_out, nonlinearity=lasagne.nonlinearities.sigmoid)
# To reshape back to our original shape, we can use the symbolic shape variables we retrieved above.
l_out = ReshapeLayer(l_dense, (batch_size, n_out))
return l_out, l_in_hid
if __name__ == '__main__':
# Define the input and expected output variable
input_var, target_var = T.fmatrices('input', 'target')
# The generator to sample examples from
tr_cond = 'two_gains'
test_cond = 'all_gains'
generator = CausalInferenceTaskFFWD(max_iter=250001, batch_size=100, n_in=50, n_out=1, sigma_sq=100.0, tr_cond=tr_cond)
test_generator = CausalInferenceTaskFFWD(max_iter=2501, batch_size=100, n_in=50, n_out=1, sigma_sq=100.0, tr_cond=test_cond)
l_out, l_rec = model(input_var, batch_size=generator.batch_size, n_in=2*generator.n_in, n_out=generator.n_out, n_hid=200)
# The generated output variable and the loss function
# all_layers = lasagne.layers.get_all_layers(l_out)
# l2_penalty = lasagne.regularization.regularize_layer_params(all_layers, lasagne.regularization.l2) * 1e-6
pred_var = T.clip(lasagne.layers.get_output(l_out), 1e-6, 1. - 1e-6)
loss = T.mean(lasagne.objectives.squared_error(pred_var, target_var)) # + l2_penalty
def RnnEvaluator(weights):
"""Build a Theano function that computes the internal state of the network
when called.
"""
numInputs = 0
numOutputs = 0
for neurons, activator, isInput, isOutput, weightFrame in weights:
if isInput:
numInputs += 1
if isOutput:
numOutputs += 1
def evaluate_net(*states):
activations = T.fvectors(len(weights))
idx = 0
for neurons, activator, isInput, isOutput, weightFrame in weights:
sumParts = []
for i, info in enumerate(weightFrame):
srcIdx, w = info
sumParts.append(T.dot(states[srcIdx], w.transpose()))
if len(sumParts):
sumParts = T.stack(*sumParts)
activity = T.sum(sumParts, axis=0)
if activator == TIDENTITY:
activation = activity
elif activator == TLOGISTIC:
activation = 1. / (1. + T.exp(-activity))
elif activator == THYPERBOLIC:
activation = T.tanh(activity)
elif activator == TTHRESHOLD:
activation = T.sgn(activity)
elif activator == TBIAS:
activation = T.ones_like(activity, dtype='float32')
elif activator == TRADIAL:
activation = T.exp(-activity*activity/2.0)
else:
raise Exception("Unknown activation function for layer {0}" + layer.id)
else:
activation = T.zeros_like(states[idx])#states[idx]
activations[idx] = activation
idx += 1
checklist = [T.all(T.eq(a,s)) for a,s in zip(activations, states)]
condition = T.all(T.as_tensor_variable(checklist))
return activations, {}, theano.scan_module.until(condition )
def make_states(*inputs):
states = []
idx = 0
numPoints = len(inputs) and inputs[0].shape[0] or 1
for neurons, activator, isInput, isOutput, weightFrame in weights:
if isInput:
states.append(inputs[idx])
idx += 1
else:
states.append(T.ones((numPoints,neurons), dtype='float32'))
return states
def project_output(states):
outputs = []
idx = 0
for neurons, activator, isInput, isOutput, weightFrame in weights:
if isOutput:
outputs.append(states[idx])
idx += 1
return outputs
inputs = T.fmatrices(numInputs)
times = T.iscalar()
netValue, updates = theano.scan(
fn=evaluate_net,
outputs_info=make_states(*inputs),
n_steps=times
)
result = [n[-1] for n in netValue]
outputs = project_output(result)
net = theano.function(inputs + [times], outputs)
def fix_inputs(inputs, times=5):
reshape = False
if len(inputs) and (len(np.shape(inputs[0])) == 1):
reshape = True
inputs = [np.reshape(i, (1,i.shape[0])) for i in inputs]
args = list(inputs) + [times]
outputs = net(*args)
if reshape:
return [o[0] for o in outputs]
return outputs
return fix_inputs
def __init__(self, We_initial, params):
self.textfile = open(params.outfile, 'w')
We = theano.shared(We_initial)
embsize = We_initial.shape[1]
hidden = params.hidden
self.num_labels = params.num_labels
self.de_hidden_size = params.de_hidden_size
self.en_hidden_size = params.en_hidden_size
print params.de_hidden_size, hidden, params.num_labels
self.lstm_layers_num = 1
input_var = T.imatrix(name='inputs')
target_var = T.imatrix(name='targets')
target_var_in = T.imatrix(name='in_targets')
mask_var = T.fmatrix(name='masks')
mask_var1 = T.fmatrix(name='masks1')
length = T.iscalar()
length0 = T.iscalar()
t_t = T.fscalar()
t_t0 = T.fscalar()
Wyy0 = np.random.uniform(
-0.02, 0.02,
(self.num_labels + 1, self.num_labels + 1)).astype('float32')
Wyy = theano.shared(Wyy0)
l_in_word = lasagne.layers.InputLayer((None, None))
l_mask_word = lasagne.layers.InputLayer(shape=(None, None))
if params.emb == 1:
l_emb_word = lasagne.layers.EmbeddingLayer(
l_in_word,
input_size=We_initial.shape[0],
output_size=embsize,
W=We)
else:
l_emb_word = lasagne_embedding_layer_2(l_in_word, embsize, We)
l_lstm_wordf = lasagne.layers.LSTMLayer(l_emb_word,
512,
mask_input=l_mask_word)
l_lstm_wordb = lasagne.layers.LSTMLayer(l_emb_word,
512,
mask_input=l_mask_word,
backwards=True)
concat = lasagne.layers.concat([l_lstm_wordf, l_lstm_wordb], axis=2)
l_reshape_concat = lasagne.layers.ReshapeLayer(concat, (-1, 2 * 512))
l_local = lasagne.layers.DenseLayer(
l_reshape_concat,
num_units=self.num_labels,
nonlinearity=lasagne.nonlinearities.linear)
network_params = lasagne.layers.get_all_params(l_local, trainable=True)
network_params.append(Wyy)
print len(network_params)
f = open(
'ccctag_CRF_Bilstm_Viterbi_.Batchsize_10_dropout_0_LearningRate_0.01_0.0512_tagversoin_2.pickle',
'r')
data = pickle.load(f)
f.close()
for idx, p in enumerate(network_params):
#print data[idx].shape
p.set_value(data[idx])
self.params = []
self.hos = []
self.Cos = []
self.encoder_lstm_layers = []
self.decoder_lstm_layers = []
ei, di, dt = T.imatrices(3) #place holders
decoderInputs0, em, em1, dm, tf, di0 = T.fmatrices(6)
#### the last one is for the stary symbole
self.de_lookuptable = theano.shared(name="Decoder LookUpTable",
value=init_xavier_uniform(
self.num_labels + 1,
self.de_hidden_size),
borrow=True)
self.linear = theano.shared(
name="Linear",
value=init_xavier_uniform(
self.de_hidden_size + 2 * self.en_hidden_size,
self.num_labels),
borrow=True)
self.linear_bias = theano.shared(
name="Hidden to Bias",
value=np.asarray(np.random.randn(self.num_labels, ) * 0.,
dtype=theano.config.floatX),
borrow=True)
#self.hidden_decode = theano.shared(name="Hidden to Decode", value= init_xavier_uniform(2*hidden, self.de_hidden_size), borrow = True)
#self.hidden_bias = theano.shared(
# name="Hidden to Bias",
# value=np.asarray(np.random.randn(self.de_hidden_size, )*0., dtype=theano.config.floatX) ,
# borrow=True
# )
input_var_shuffle = input_var.dimshuffle(1, 0)
mask_var_shuffle = mask_var.dimshuffle(1, 0)
target_var_in_shuffle = target_var_in.dimshuffle(1, 0)
target_var_shuffle = target_var.dimshuffle(1, 0)
self.params += [self.linear, self.linear_bias,
self.de_lookuptable] #concatenate
state_below = We[input_var_shuffle.flatten()].reshape(
(input_var_shuffle.shape[0], input_var_shuffle.shape[1], embsize))
enclstm_f = LSTM(embsize, self.en_hidden_size)
enclstm_b = LSTM(embsize, self.en_hidden_size, True)
self.encoder_lstm_layers.append(enclstm_f) #append
self.encoder_lstm_layers.append(enclstm_b) #append
self.params += enclstm_f.params + enclstm_b.params #concatenate
hs_f, Cs_f = enclstm_f.forward(state_below, mask_var_shuffle)
hs_b, Cs_b = enclstm_b.forward(state_below, mask_var_shuffle)
hs = T.concatenate([hs_f, hs_b], axis=2)
Cs = T.concatenate([Cs_f, Cs_b], axis=2)
hs0 = T.concatenate([hs_f[-1], hs_b[0]], axis=1)
Cs0 = T.concatenate([Cs_f[-1], Cs_b[0]], axis=1)
#self.hos += T.tanh(tensor.dot(hs0, self.hidden_decode) + self.hidden_bias),
#self.Cos += T.tanh(tensor.dot(Cs0, self.hidden_decode) + self.hidden_bias),
self.hos += T.alloc(np.asarray(0., dtype=theano.config.floatX),
input_var_shuffle.shape[1], self.de_hidden_size),
self.Cos += T.alloc(np.asarray(0., dtype=theano.config.floatX),
input_var_shuffle.shape[1], self.de_hidden_size),
Encoder = hs
ei, di, dt = T.imatrices(3) #place holders
em, dm, tf, di0 = T.fmatrices(4)
self.encoder_function = theano.function(inputs=[ei, em],
outputs=Encoder,
givens={
input_var: ei,
mask_var: em
})
state_below = self.de_lookuptable[
target_var_in_shuffle.flatten()].reshape(
(target_var_in_shuffle.shape[0],
target_var_in_shuffle.shape[1], self.de_hidden_size))
for i in range(self.lstm_layers_num):
declstm = LSTM(self.de_hidden_size, self.de_hidden_size)
self.decoder_lstm_layers += declstm, #append
self.params += declstm.params #concatenate
ho, Co = self.hos[i], self.Cos[i]
state_below, Cs = declstm.forward(state_below, mask_var_shuffle,
ho, Co)
decoder_lstm_outputs = T.concatenate([Encoder, state_below], axis=2)
linear_outputs = T.dot(decoder_lstm_outputs,
self.linear) + self.linear_bias[None, None, :]
softmax_outputs, updates = theano.scan(
fn=lambda x: T.nnet.softmax(x),
sequences=[linear_outputs],
)
def _NLL(pred, y, m):
return -m * T.log(pred[T.arange(input_var.shape[0]), y])
def _step2(ctx_, state_, hs_, Cs_):
#print ctx_.shape, state_.shape, hs_.shape, Cs_.shape
hs, Cs = [], []
token_idxs = T.cast(state_.argmax(axis=-1), "int32")
msk_ = T.fill((T.zeros_like(token_idxs, dtype="float32")), 1)
msk_ = msk_.dimshuffle('x', 0)
state_below0 = self.de_lookuptable[token_idxs].reshape(
(1, ctx_.shape[0], self.de_hidden_size))
for i, lstm in enumerate(self.decoder_lstm_layers):
h, C = lstm.forward(state_below0, msk_, hs_[i],
Cs_[i]) #mind msk
hs += h[-1],
Cs += C[-1],
state_below0 = h
hs, Cs = T.as_tensor_variable(hs), T.as_tensor_variable(Cs)
state_below0 = state_below0.reshape(
(ctx_.shape[0], self.de_hidden_size))
state_below0 = T.concatenate([ctx_, state_below0], axis=1)
newpred = T.dot(state_below0,
self.linear) + self.linear_bias[None, :]
state_below = T.nnet.softmax(newpred)
extra_p = T.zeros_like(hs[:, :, 0])
state_below = T.concatenate([state_below, extra_p.T], axis=1)
return state_below, hs, Cs
ctx_0, state_0 = T.fmatrices(2)
hs_0 = T.ftensor3()
Cs_0 = T.ftensor3()
state_below_tmp, hs_tmp, Cs_tmp = _step2(ctx_0, state_0, hs_0, Cs_0)
self.f_next = theano.function([ctx_0, state_0, hs_0, Cs_0],
[state_below_tmp, hs_tmp, Cs_tmp],
name='f_next')
hs0, Cs0 = T.as_tensor_variable(
self.hos, name="hs0"), T.as_tensor_variable(self.Cos, name="Cs0")
train_outputs, _ = theano.scan(fn=_step2,
sequences=[Encoder],
outputs_info=[decoderInputs0, hs0, Cs0],
n_steps=input_var_shuffle.shape[0])
predy = train_outputs[0].dimshuffle(1, 0, 2)
predy = predy[:, :, :-1] * mask_var[:, :, None]
predy0 = predy.reshape((-1, self.num_labels))
def inner_function(targets_one_step, mask_one_step, prev_label,
tg_energy):
"""
:param targets_one_step: [batch_size, t]
:param prev_label: [batch_size, t]
:param tg_energy: [batch_size]
:return:
"""
new_ta_energy = T.dot(prev_label, Wyy[:-1, :-1])
new_ta_energy_t = tg_energy + T.sum(
new_ta_energy * targets_one_step, axis=1)
tg_energy_t = T.switch(mask_one_step, new_ta_energy_t, tg_energy)
return [targets_one_step, tg_energy_t]
local_energy = lasagne.layers.get_output(l_local, {
l_in_word: input_var,
l_mask_word: mask_var
})
local_energy = local_energy.reshape((-1, length, self.num_labels))
local_energy = local_energy * mask_var[:, :, None]
#####################
# for the end symbole of a sequence
####################
end_term = Wyy[:-1, -1]
local_energy = local_energy + end_term.dimshuffle(
'x', 'x', 0) * mask_var1[:, :, None]
#predy0 = lasagne.layers.get_output(l_local_a, {l_in_word_a:input_var, l_mask_word_a:mask_var})
predy_in = T.argmax(predy0, axis=1)
A = T.extra_ops.to_one_hot(predy_in, self.num_labels)
A = A.reshape((-1, length, self.num_labels))
#predy = predy0.reshape((-1, length, 25))
#predy = predy*mask_var[:,:,None]
targets_shuffled = predy.dimshuffle(1, 0, 2)
target_time0 = targets_shuffled[0]
masks_shuffled = mask_var.dimshuffle(1, 0)
initial_energy0 = T.dot(target_time0, Wyy[-1, :-1])
initials = [target_time0, initial_energy0]
[_, target_energies], _ = theano.scan(
fn=inner_function,
outputs_info=initials,
sequences=[targets_shuffled[1:], masks_shuffled[1:]])
cost11 = target_energies[-1] + T.sum(
T.sum(local_energy * predy, axis=2) * mask_var, axis=1)
# compute the ground-truth energy
targets_shuffled0 = A.dimshuffle(1, 0, 2)
target_time00 = targets_shuffled0[0]
initial_energy00 = T.dot(target_time00, Wyy[-1, :-1])
initials0 = [target_time00, initial_energy00]
[_, target_energies0], _ = theano.scan(
fn=inner_function,
outputs_info=initials0,
sequences=[targets_shuffled0[1:], masks_shuffled[1:]])
cost110 = target_energies0[-1] + T.sum(
T.sum(local_energy * A, axis=2) * mask_var, axis=1)
#predy_f = predy.reshape((-1, 25))
y_f = target_var.flatten()
if (params.annealing == 0):
lamb = params.L3
elif (params.annealing == 1):
lamb = params.L3 * (1 - 0.01 * t_t)
if (params.regutype == 0):
ce_hinge = lasagne.objectives.categorical_crossentropy(
predy0 + eps, y_f)
ce_hinge = ce_hinge.reshape((-1, length))
ce_hinge = T.sum(ce_hinge * mask_var, axis=1)
cost = T.mean(-cost11) + lamb * T.mean(ce_hinge)
else:
entropy_term = -T.sum(predy0 * T.log(predy0 + eps), axis=1)
entropy_term = entropy_term.reshape((-1, length))
entropy_term = T.sum(entropy_term * mask_var, axis=1)
cost = T.mean(-cost11) - lamb * T.mean(entropy_term)
"""
f = open('F0_simple.pickle')
PARA = pickle.load(f)
f.close()
l2_term = sum(lasagne.regularization.l2(x-PARA[index]) for index, x in enumerate(a_params))
cost = T.mean(-cost11) + params.L2*l2_term
"""
##from adam import adam
##updates_a = adam(cost, self.params, params.eta)
#updates_a = lasagne.updates.sgd(cost, self.params, params.eta)
#updates_a = lasagne.updates.apply_momentum(updates_a, self.params, momentum=0.9)
from momentum import momentum
updates_a = momentum(cost, self.params, params.eta, momentum=0.9)
if (params.regutype == 0):
self.train_fn = theano.function(
inputs=[ei, dt, em, em1, length0, t_t0, di0],
outputs=[cost, ce_hinge],
updates=updates_a,
on_unused_input='ignore',
givens={
input_var: ei,
target_var: dt,
mask_var: em,
mask_var1: em1,
length: length0,
t_t: t_t0,
decoderInputs0: di0
})
#self.train_fn = theano.function([input_var, target_var, mask_var, mask_var1, length, t_t], [cost, ce_hinge], updates = updates_a, on_unused_input='ignore')
else:
self.train_fn = theano.function(
inputs=[ei, dt, em, em1, length0, t_t0, di0],
outputs=[cost, entropy_term],
updates=updates_a,
on_unused_input='ignore',
givens={
input_var: ei,
target_var: dt,
mask_var: em,
mask_var1: em1,
length: length0,
t_t: t_t0,
decoderInputs0: di0
})
#self.train_fn = theano.function([input_var, target_var, mask_var, mask_var1, length, t_t], [cost, entropy_term], updates = updates_a, on_unused_input='ignore')
prediction = T.argmax(predy, axis=2)
corr = T.eq(prediction, target_var)
corr_train = (corr * mask_var).sum(dtype=theano.config.floatX)
num_tokens = mask_var.sum(dtype=theano.config.floatX)
self.eval_fn = theano.function(
inputs=[ei, dt, em, em1, length0, di0],
outputs=[cost11, cost110, corr_train, num_tokens, prediction],
on_unused_input='ignore',
givens={
input_var: ei,
target_var: dt,
mask_var: em,
mask_var1: em1,
length: length0,
decoderInputs0: di0
})
def __init__(self,Ne,Ni,n_inp,W_inp=None,W_inner=None):
'''class SNNgroup's self Parameters:
self.A: update matrix
self.S: neuron state varaibles
self.W_inner: inner-connect weights in the group
self.W_inp: input weights
self.spikes: the spikes matrix in the time t
self.SpkC : spike containers
input : '''
self.number = Ne+Ni
self.Ne = Ne
self.Ni = Ni
self.mV=self.ms=1e-3 # units
dt=1*self.ms # timestep
self.dt = dt
taum=20*self.ms # membrane time constant
taue=5*self.ms
taui=10*self.ms
#self.Vt=-1*self.mV # threshold = -50+49
self.Vt = 15*self.mV #threshold = -55+70
#self.Vr=-11*self.mV # reset = -60+49
self.Vr = 0*self.mV # reset = -70+70
self.Vi = -10*self.mV # VI = -80+70
self.dApre = .0001
#self.dApre = .95 #changed into .95
self.dApost = -self.dApre*1.05
self.tauP = 20*self.ms
#self.input = input
self.n_inp = n_inp
self.weight = .001
self.weightIn = 1.
self.wmax = 200*self.weight
zero = np.array([0]).astype(theano.config.floatX)
self.zero = theano.shared(zero,name='zero',borrow=True)
"""
Equations
---------
eqs='''
dv/dt = (ge*70mV-gi*10-(v+70*mV))/(20*ms) : volt
dge/dt = -ge/(5*self.ms) : volt
dgi/dt = -gi/(10*self.ms) : volt
'''
"""
# Update matrix
A = np.array([[np.exp(-dt/taum),0,0],
[taue/(taum-taue)*(np.exp(-dt/taum)-np.exp(-dt/taue)),np.exp(-dt/taue),0],
[-taui/(taum-taui)*(np.exp(-dt/taum)-np.exp(-dt/taui)),0,np.exp(-dt/taui)]
],dtype=theano.config.floatX).T
A = theano.shared(value=A,name='A',borrow=True)
self.A = A
# State varible : [v;ge;gi] (size=3*self.number)
S = np.ones((1,self.number),dtype=theano.config.floatX)*self.Vr
S = np.vstack((S,np.zeros((2,self.number),dtype=theano.config.floatX)))
self.S_init = S
S = theano.shared(value=S,name='S',borrow=True)
self.S = S
if W_inner == None:
# weights of inner connections (size= self.number*self.number)
self.W_inner_ini = np.ones((self.number,self.number),dtype=theano.config.floatX)*self.weight
#self.W_inner_ini[Ne:,:] = self.weightIn
self.W_inner_ini[Ne:,:] = self.weight
wtmp = np.eye(self.number)
ind = wtmp.nonzero()
self.W_inner_ini[ind]=0
W_inner = theano.shared(value=self.W_inner_ini,name='W_inner',borrow=True)
self.W_inner = W_inner
else:
self.W_inner = theano.shared(W_inner,name='W_inner',borrow=True)
# weights of input connections (size=n_inp*self.number
rng = np.random.RandomState(1234)
if W_inp ==None:
#W_inp = np.ones((self.n_inp,self.number)).astype(theano.config.floatX) #needs specification later
#W_inp = np.random.rand(self.n_inp,self.number).astype(theano.config.floatX)*.00001*self.ms #needs specification later
self.W_inp_ini = np.ones((self.n_inp,self.number)).astype(theano.config.floatX)*self.weight
self.W_inp_ini[:,self.Ne:] = self.weightIn
W_inp = theano.shared(self.W_inp_ini,name='W_inp',borrow=True)
self.W_inp = W_inp
else:
self.W_inp = theano.shared(W_inp,name='W_inp',borrow=True)
# Spike Container
#spkC = theano.shared(value=np.empty((1,self.number)).astype(theano.config.floatX),name='spkC',borrow=True)
spkC = np.empty((1,self.number)).astype(theano.config.floatX)
self.spkC = spkC
#spikes=np.empty((self.number,1),dtype=theano.config.floatX)
#self.spikes = theano.shared(value=spikes,name='spikes',borrow=True)
# not sure the dtype of sp_history
self.sp_history = np.array([])
#output = np.empty(self.number,dtype=theano.config.floatX)
#self.output = theano.shared(value=output,name='output',borrow=True)
self.V_record = np.empty((1,self.number))
self.ge_record = np.empty((1,self.number))
self.gi_record = np.empty((1,self.number))
#================================================
# Process Function Initial
# input:: 0-1 vector
'''Update Schedule:
1.Update state variables of SNNgroup: dot(A,S)
1.Update state variables of Synapses: dot(exp(-dt/tau),Ssynapse), including W_inp and W_inner
2.Call thresholding function: S[0,:]>Vt
3.Push spikes into SpikeContainer
4.Propagate spikes via Connection(possibly with delays)
5.Update state variables of Synapses (STDP)
6.Call reset function on neurons which has spiked'''
Ne = self.Ne
Ni = self.Ni
m = T.fmatrix(name='m')
#self.Vt = T.as_tensor_variable(self.Vt,'Vt')
# "Update state function:: stat()"
# return np array
# shape(stat()) = shape(self.S)
S_update = T.dot(self.A,self.S)
self.stat = theano.function(
inputs = [],
outputs = [],
updates = {self.S : S_update})
#============================================================
# Update state of Synapses
# Update matrix of Synapse
A_STDP = np.array([[np.exp(-self.dt/self.tauP),0],[0,np.exp(-self.dt/self.tauP)]],dtype=theano.config.floatX)
# Spre_inner :: pre synapse of inner connections
# Spost_inner:: post synapse of inner connections
# Spre_inp :: pre synapse of input conenctions
# Spost_inp :: post synapse of input connections
self.Spre_inner_ini = np.zeros((self.number,self.number),dtype=theano.config.floatX)
Spre_inner = theano.shared(self.Spre_inner_ini,name='Spre_inner',borrow=True)
self.Spre_inner = Spre_inner
self.Spost_inner_ini = np.zeros((self.number,self.number),dtype=theano.config.floatX)
Spost_inner = theano.shared(value=self.Spost_inner_ini,name='Spost_inner',borrow=True)
self.Spost_inner = Spost_inner
self.Spre_inp_ini = np.zeros((self.n_inp,self.number)).astype(theano.config.floatX) #needs specification later
Spre_inp = theano.shared(value=self.Spre_inp_ini,name='Spre_inp',borrow=True)
self.Spre_inp = Spre_inp
self.Spost_inp_ini = np.zeros((self.n_inp,self.number)).astype(theano.config.floatX) #needs specification later
Spost_inp = theano.shared(value=self.Spost_inp_ini,name='Spost_inp',borrow=True)
self.Spost_inp = Spost_inp
U = T.fscalar('U')
UM = T.fmatrix('UM')
#UpreV = theano.shared(A_STDP[0,0],name='UpreV',borrow=True) # Wpre = UpreV*Wpre
#UpostV = theano.shared(A_STDP[1,1],name='UpostV',borrow=True)
self.tmp = np.array(np.exp(-self.dt/self.tauP).astype(theano.config.floatX))
self.SynFresh = theano.shared(self.tmp,name='SynFresh',borrow=True)
self.UpdateSpre_inner = theano.function(inputs=[],outputs=None,updates={self.Spre_inner:T.dot(self.SynFresh,self.Spre_inner)},allow_input_downcast=True)
self.UpdateSpost_inner = theano.function(inputs=[],outputs=None,updates={self.Spost_inner:T.dot(self.SynFresh,self.Spost_inner)},allow_input_downcast=True)
self.UpdateSpre_inp = theano.function(inputs=[],outputs=None,updates={self.Spre_inp:T.dot(self.SynFresh,self.Spre_inp)},allow_input_downcast=True)
self.UpdateSpost_inp = theano.function(inputs=[],outputs=None,updates={self.Spost_inp:T.dot(self.SynFresh,self.Spost_inp)},allow_input_downcast=True)
#------------------------------------------
#tmp = math.exp(-self.dt/self.tauP)
#tmp = T.as_tensor(0.95122945)
#================================================================
#------------------------------------------
# "thresholding function:: spike_fun()"
# type return :: np.ndarray list
# shape return:: shape(spike_fun()) = (self.number,)
self.spike_fun = theano.function(
inputs = [U], #[self.S]
outputs = (T.gt(self.S[0,:],U))) #type outputs: np.ndarray,shape::(nL,)
#'outputs = (self.S[0,:]>Vt).astype(theano.config.floatX)), #type outputs: list'
#'updates={self.spikes:(self.S[0,:]>Vt).astype(theano.config.floatX)}'
#------------------------------------
#------------------------------------
#=================================================================
# "Push spike into Container function:: spCfun(vector)"
# type vector :: np.array([],dtype=theano.config.floatX)!!!
# type return :: np array
# shape return:: shape(spCfun()) = ( shape(self.spkC)[0]+1 , shape(self.spkC)[1] )
#updates={self.spkC:T.stack(self.spkC,sp)})
'''spike_prop = theano.function( #wrong
inputs = [],
outputs =[],
updates = {self.S:np.dot(self.W_inner,self.spikes)+self.S})#wrong'''
#-------------------------------
#--------------------------------
#====================================================================
# Propagate spikes
# inner connection:
# S_inner = f(inputs, outputs, updates)
# Param:: inputs: spike 0-1 vector
# Param:: inputs: spike is from function-> spike_fun
# S_inner(spk)::-> for i in spk[0:Ne].nonzero()[0]:
# S[1,:] = Winner[i,:]+S[1,:] (excitatory conenction)
# for j in spk[Ne,:].nonzero()[0]:
# S[2,:] = Winner[j,:]+S[2,:] (inhibitory connection)
vinner = T.fvector(name='vinner') # vinner = spk :: np.array((1,self.number)
def add_f1(i,p,q):
np = T.inc_subtensor(p[1,:],q[i,:]) #ge
return {p:np}
def add_f2(i,p,q):
np = T.inc_subtensor(p[2,:],q[i,:]) #gi
return {p:np}
#deltaWinner1,updates1 = theano.scan(fn=lambda i: self.W_inner[i,:]*i+self.S[1,:], sequences=vinner[0:Ne])
deltaWinner1,updates1 = theano.scan(fn=add_f1, sequences=vinner[0:Ne].nonzero()[0],non_sequences=[self.S,self.W_inner])
#deltaWinner2,updates2 = theano.scan(fn=lambda i: self.W_inner[i,:]*i+self.S[2,:], sequences=vinner[Ne:])
deltaWinner2,updates2 = theano.scan(fn=add_f2, sequences=vinner[Ne:].nonzero()[0]+self.Ne,non_sequences=[self.S,self.W_inner])
# S = S+W
self.S_inner1 = theano.function(inputs=[vinner],outputs=None,updates=updates1,allow_input_downcast=True)
self.S_inner2 = theano.function(inputs=[vinner],outputs=None,updates=updates2,allow_input_downcast=True)
#------------------------------------------
#------------------------------------------
# outter connection (input spikes):
# type input: index list
voutter = T.fvector(name='voutter')
#deltaWoutter = theano.scan(fn=lambda j: self.W_inp[j,:]+self.S[1,:],sequences=voutter)
deltaWoutter,updatesout1 = theano.scan(fn=add_f1,sequences=voutter.nonzero()[0],non_sequences=[self.S,self.W_inp])
self.S_inp = theano.function(inputs=[voutter],outputs=None,updates=updatesout1,allow_input_downcast=True)
#------------------------------------
#-------------------------------------
#=====================================================================
# Update Synapses (STDP | STDC)
# Pre:: Apre += self.dApre, w+=Apost
# Post:: Apost+=self.dApost, w+=Apre
#
# USpreInner :: Perform Pre function No.1 in inner connections
# UWInner :: Perform Pre function No.2 in inner connections
# UpreInner :: Function
def add_synap_pre(i,p,po,s,q):
# i :: sequence
# p :: pre | post
# s :: dApre | dApost
# q :: W
index = T.nonzero(q[i,:self.Ne])
np = T.inc_subtensor(p[i,index],s)
## tmp = p[i,:]
## tmp=T.inc_subtensor(tmp[index],s)
## np=T.set_subtensor(p[i,:],tmp)
#np = T.inc_subtensor(p[i,:],s)
nw = T.inc_subtensor(q[i,:],po[i,:])
nw=T.clip(nw,0,self.wmax)
return {p:np,q:nw}
def add_synap_pre_inp(i,p,po,s,q):
# i :: sequence
# p :: pre | post
# s :: dApre | dApost
# q :: W
index = T.nonzero(q[i,:self.Ne])
np = T.inc_subtensor(p[i,index],s)
## tmp = p[i,:]
## tmp=T.inc_subtensor(tmp[index],s)
## np=T.set_subtensor(p[i,:],tmp)
#np = T.inc_subtensor(p[i,:],s)
nw = T.inc_subtensor(q[i,:],po[i,:])
nw=T.clip(nw,0,self.wmax)
return {p:np,q:nw}
def add_synap_post(i,po,p,s,q):
# i:: sequence
# po:: post
# p:: pre
# s:: dA
# q:: W
index = T.nonzero(q[:self.Ne,i])
npo = T.inc_subtensor(po[index,i],s)
nw = T.inc_subtensor(q[:,i],p[:,i])
nw = T.clip(nw,0,self.wmax)
return {po:npo,q:nw}
def add_synap_post_inp(i,po,p,s,q):
# i:: sequence
# po:: post
# p:: pre
# s:: dA
# q:: W
index = T.nonzero(q[:self.Ne,i])
npo = T.inc_subtensor(po[index,i],s)
nw = T.inc_subtensor(q[:,i],p[:,i])
nw = T.clip(nw,0,self.wmax)
return {po:npo,q:nw}
add_dA = T.fscalar('add_dA')
add_p,add_po,add_q = T.fmatrices('add_p','add_po','add_q')
#-------------------------------------------------------------------------
#USinner,updatesUinner = theano.scan(fn=add_synap_pre,sequences=vinner,non_sequences=[self.Spre_inner,self.Spost_inp,self.dApre,self.W_inner])
'USinner,updatesUinner = theano.scan(fn=add_synap_pre,sequences=vinner.nonzero()[0],non_sequences=[add_p,add_po,add_dA,add_q])'
#USinner1,updatesUinner1 = theano.scan(fn=add_synap_pre,sequences=vinner,non_sequences=[self.Spost_inner,self.Spre_inner,self.dApost,self.W_inner])
#-------------------------------------------------------------------------
#UpostInner = theano.function(inputs[vinner],updates={self.Spost_inner:USpostInner})
#UpostInp = theano.function(inputs=[vinner],updates={self.W_inner:UWInnerpost})
'USinner_f = theano.function(inputs=[vinner,add_p,add_po,add_dA,add_q],outputs=None,updates=updatesUinner)'
#USinner_step2 = theano.function(inputs=[vinner,add_p,add_po,add_dA,add_q],outputs=None,updates=updatesUinner)
USinner_inner_pre,updatesUinner_inner_pre = theano.scan(fn=add_synap_pre,sequences=vinner[:self.Ne].nonzero()[0],non_sequences=[self.Spre_inner,self.Spost_inner,add_dA,self.W_inner])
self.USinner_f_inner_pre = theano.function(inputs=[vinner,add_dA],outputs=None,updates=updatesUinner_inner_pre,allow_input_downcast=True)
USinner_innerpost,updatesUinner_inner_post = theano.scan(fn=add_synap_post,sequences=vinner[:self.Ne].nonzero()[0],non_sequences=[self.Spost_inner,self.Spre_inner,add_dA,self.W_inner])
self.USinner_f_inner_post = theano.function(inputs=[vinner,add_dA],outputs=None,updates=updatesUinner_inner_post,allow_input_downcast=True)
USinner_inp_pre,updatesUSinner_inp_pre =theano.scan(fn=add_synap_pre_inp,sequences=vinner.nonzero()[0],non_sequences=[self.Spre_inp,self.Spost_inp,add_dA,self.W_inp])
self.USinner_f_inp_pre = theano.function(inputs=[vinner,add_dA],outputs=None,updates=updatesUSinner_inp_pre,allow_input_downcast=True)
USinner_inp_post,updatesUSinner_inp_post =theano.scan(fn=add_synap_post_inp,sequences=vinner[:self.Ne].nonzero()[0],non_sequences=[self.Spost_inp,self.Spre_inp,add_dA,self.W_inp])
self.USinner_f_inp_post = theano.function(inputs=[vinner,add_dA],outputs=None,updates=updatesUSinner_inp_post,allow_input_downcast=True)
# Call reset function
def reset_v(index,vr):
nv = T.set_subtensor(self.S[0,index],vr)
return{self.S:nv}
resetV,resetV_update = theano.scan(fn=reset_v,sequences=vinner.nonzero()[0],non_sequences=[U])
self.resetV_f = theano.function(inputs=[vinner,U],outputs=None,updates=resetV_update,allow_input_downcast=True)
setvalue = T.fscalar('setvalue')
iv = T.ivector('iv')
def reset_state(i,value,state):
nstate = T.set_subtensor(state[i,:],value)
return {state:nstate}
reset_S_state,Upreset_S_state = theano.scan(fn=reset_state,sequences=iv,non_sequences=[setvalue,self.S])
self.reset_S_fn = theano.function(inputs=[iv,setvalue],outputs=None,updates=Upreset_S_state)
def __init__(self, We, char_embedd_table_initial, params):
lstm_layers_num = 1
emb_size = We.shape[1]
self.eta = params.eta
self.num_labels = params.num_labels
self.en_hidden_size = params.en_hidden_size
self.de_hidden_size = params.de_hidden_size
self.lstm_layers_num = params.lstm_layers_num
self._train = None
self._utter = None
self.params = []
self.encoder_lstm_layers = []
self.decoder_lstm_layers = []
self.hos = []
self.Cos = []
char_embedd_dim = params.char_embedd_dim
char_dic_size = len(params.char_dic)
char_embedd_table = theano.shared(char_embedd_table_initial)
encoderInputs = tensor.imatrix()
decoderInputs, decoderTarget = tensor.imatrices(2)
encoderMask, TF, decoderMask, decoderInputs0 = tensor.fmatrices(4)
char_input_var = tensor.itensor3(name='char-inputs')
ci = tensor.itensor3()
use_dropout = tensor.fscalar()
use_dropout0 = tensor.fscalar()
self.lookuptable = theano.shared(We)
#### the last one is for the stary symbole
self.de_lookuptable = theano.shared(name="Decoder LookUpTable",
value=init_xavier_uniform(
self.num_labels + 1,
self.de_hidden_size),
borrow=True)
self.linear = theano.shared(
name="Linear",
value=init_xavier_uniform(
self.de_hidden_size + 2 * self.en_hidden_size,
self.num_labels),
borrow=True)
self.linear_bias = theano.shared(
name="Hidden to Bias",
value=np.asarray(np.random.randn(self.num_labels, ) * 0.,
dtype=theano.config.floatX),
borrow=True)
#self.hidden_decode = theano.shared(name="Hidden to Decode", value= init_xavier_uniform(2*en_hidden_size, self.de_hidden_size), borrow = True)
#self.hidden_bias = theano.shared(
# name="Hidden to Bias",
# value=np.asarray(np.random.randn(self.de_hidden_size, )*0., dtype=theano.config.floatX) ,
# borrow=True
# )
#self.params += [self.linear, self.de_lookuptable, self.hidden_decode, self.hidden_bias] #concatenate
self.params += [
self.lookuptable, self.linear, self.linear_bias,
self.de_lookuptable
] #the initial hidden state of decoder lstm is zeros
#(max_sent_size, batch_size, hidden_size)
state_below = self.lookuptable[encoderInputs.flatten()].reshape(
(encoderInputs.shape[0], encoderInputs.shape[1], emb_size))
layer_char_input = lasagne.layers.InputLayer(shape=(None, None,
Max_Char_Length),
input_var=char_input_var,
name='char-input')
layer_char = lasagne.layers.reshape(layer_char_input, (-1, [2]))
layer_char_embedding = lasagne.layers.EmbeddingLayer(
layer_char,
input_size=char_dic_size,
output_size=char_embedd_dim,
W=char_embedd_table,
name='char_embedding')
layer_char = lasagne.layers.DimshuffleLayer(layer_char_embedding,
pattern=(0, 2, 1))
# first get some necessary dimensions or parameters
conv_window = 3
num_filters = params.num_filters
# construct convolution layer
cnn_layer = lasagne.layers.Conv1DLayer(
layer_char,
num_filters=num_filters,
filter_size=conv_window,
pad='full',
nonlinearity=lasagne.nonlinearities.tanh,
name='cnn')
# infer the pool size for pooling (pool size should go through all time step of cnn)
_, _, pool_size = cnn_layer.output_shape
# construct max pool layer
pool_layer = lasagne.layers.MaxPool1DLayer(cnn_layer,
pool_size=pool_size)
# reshape the layer to match lstm incoming layer [batch * sent_length, num_filters, 1] --> [batch, sent_length, num_filters]
output_cnn_layer = lasagne.layers.reshape(
pool_layer, (-1, encoderInputs.shape[0], [1]))
char_params = lasagne.layers.get_all_params(output_cnn_layer,
trainable=True)
self.params += char_params
char_state_below = lasagne.layers.get_output(output_cnn_layer)
char_state_below = dropout_layer(char_state_below, use_dropout, trng)
char_state_shuff = char_state_below.dimshuffle(1, 0, 2)
state_below = tensor.concatenate([state_below, char_state_shuff],
axis=2)
state_below = dropout_layer(state_below, use_dropout, trng)
for _ in range(self.lstm_layers_num):
enclstm_f = LSTM(emb_size + num_filters, self.en_hidden_size)
enclstm_b = LSTM(emb_size + num_filters, self.en_hidden_size, True)
self.encoder_lstm_layers.append(enclstm_f) #append
self.encoder_lstm_layers.append(enclstm_b) #append
self.params += enclstm_f.params + enclstm_b.params #concatenate
hs_f, Cs_f = enclstm_f.forward(state_below, encoderMask)
hs_b, Cs_b = enclstm_b.forward(state_below, encoderMask)
hs = tensor.concatenate([hs_f, hs_b], axis=2)
Cs = tensor.concatenate([Cs_f, Cs_b], axis=2)
hs0 = tensor.concatenate([hs_f[-1], hs_b[0]], axis=1)
Cs0 = tensor.concatenate([Cs_f[-1], Cs_b[0]], axis=1)
#self.hos += tensor.tanh(tensor.dot(hs0, self.hidden_decode) + self.hidden_bias),
#self.Cos += tensor.tanh(tensor.dot(Cs0, self.hidden_decode) + self.hidden_bias),
self.hos += tensor.alloc(
np.asarray(0., dtype=theano.config.floatX),
encoderInputs.shape[1], self.de_hidden_size),
self.Cos += tensor.alloc(
np.asarray(0., dtype=theano.config.floatX),
encoderInputs.shape[1], self.de_hidden_size),
state_below = hs
Encoder = state_below
state_below = self.de_lookuptable[decoderInputs.flatten()].reshape(
(decoderInputs.shape[0], decoderInputs.shape[1],
self.de_hidden_size))
for i in range(self.lstm_layers_num):
declstm = LSTM(self.de_hidden_size, self.de_hidden_size)
self.decoder_lstm_layers += declstm, #append
self.params += declstm.params #concatenate
ho, Co = self.hos[i], self.Cos[i]
state_below, Cs = declstm.forward(state_below, decoderMask, ho, Co)
##### Here we include the representation from the decoder
decoder_lstm_outputs = tensor.concatenate([state_below, Encoder],
axis=2)
ei, di, dt = tensor.imatrices(3) #place holders
em, dm, tf, di0 = tensor.fmatrices(4)
#####################################################
#####################################################
linear_outputs = tensor.dot(decoder_lstm_outputs,
self.linear) + self.linear_bias[None,
None, :]
softmax_outputs, _ = theano.scan(
fn=lambda x: tensor.nnet.softmax(x),
sequences=[linear_outputs],
)
def _NLL(pred, y, m):
return -m * tensor.log(pred[tensor.arange(encoderInputs.shape[1]),
y])
costs, _ = theano.scan(
fn=_NLL, sequences=[softmax_outputs, decoderTarget, decoderMask])
loss = costs.sum() / decoderMask.sum() + params.L2 * sum(
lasagne.regularization.l2(x) for x in self.params)
#updates = lasagne.updates.adam(loss, self.params, self.eta)
#updates = lasagne.updates.apply_momentum(updates, self.params, momentum=0.9)
###################################################
#### using the ground truth when training
##################################################
#self._train = theano.function(
# inputs=[ei, em, di, dm, dt],
# outputs=[loss, softmax_outputs],
# updates=updates,
# givens={encoderInputs:ei, encoderMask:em, decoderInputs:di, decoderMask:dm, decoderTarget:dt}
# )
#########################################################################
### For schedule sampling
#########################################################################
###### always use privous predict as next input
def _step2(ctx_, state_, hs_, Cs_):
hs, Cs = [], []
token_idxs = tensor.cast(state_.argmax(axis=-1), "int32")
msk_ = tensor.fill(
(tensor.zeros_like(token_idxs, dtype="float32")), 1.)
msk_ = msk_.dimshuffle('x', 0)
state_below0 = self.de_lookuptable[token_idxs].reshape(
(1, encoderInputs.shape[1], self.de_hidden_size))
for i, lstm in enumerate(self.decoder_lstm_layers):
h, C = lstm.forward(state_below0, msk_, hs_[i],
Cs_[i]) #mind msk
hs += h[-1],
Cs += C[-1],
state_below0 = h
hs, Cs = tensor.as_tensor_variable(hs), tensor.as_tensor_variable(
Cs)
state_below0 = state_below0.reshape(
(encoderInputs.shape[1], self.de_hidden_size))
state_below0 = tensor.concatenate([ctx_, state_below0], axis=1)
newpred = tensor.dot(state_below0,
self.linear) + self.linear_bias[None, :]
state_below = tensor.nnet.softmax(newpred)
##### the beging symbole probablity is 0
extra_p = tensor.zeros_like(hs[:, :, 0])
state_below = tensor.concatenate([state_below, extra_p.T], axis=1)
return state_below, hs, Cs
hs0, Cs0 = tensor.as_tensor_variable(
self.hos, name="hs0"), tensor.as_tensor_variable(self.Cos,
name="Cs0")
train_outputs, _ = theano.scan(fn=_step2,
sequences=[Encoder],
outputs_info=[decoderInputs0, hs0, Cs0],
n_steps=encoderInputs.shape[0])
train_predict = train_outputs[0]
train_costs, _ = theano.scan(
fn=_NLL, sequences=[train_predict, decoderTarget, decoderMask])
train_loss = train_costs.sum() / decoderMask.sum() + params.L2 * sum(
lasagne.regularization.l2(x) for x in self.params)
#from adam import adam
#train_updates = adam(train_loss, self.params, self.eta)
#train_updates = lasagne.updates.apply_momentum(train_updates, self.params, momentum=0.9)
#train_updates = lasagne.updates.sgd(train_loss, self.params, self.eta)
#train_updates = lasagne.updates.apply_momentum(train_updates, self.params, momentum=0.9)
from momentum import momentum
train_updates = momentum(train_loss,
self.params,
params.eta,
momentum=0.9)
self._train2 = theano.function(
inputs=[ei, ci, em, di0, dm, dt, use_dropout0],
outputs=[train_loss, train_predict],
updates=train_updates,
givens={
encoderInputs: ei,
char_input_var: ci,
encoderMask: em,
decoderInputs0: di0,
decoderMask: dm,
decoderTarget: dt,
use_dropout: use_dropout0
}
#givens={encoderInputs:ei, encoderMask:em, decoderInputs:di, decoderMask:dm, decoderTarget:dt, TF:tf}
)
listof_token_idx = train_predict.argmax(axis=-1)
self._utter = theano.function(inputs=[ei, ci, em, di0, use_dropout0],
outputs=listof_token_idx,
givens={
encoderInputs: ei,
char_input_var: ci,
encoderMask: em,
decoderInputs0: di0,
use_dropout: use_dropout0
})
def UnitTest_OnestepAttend():
N = 2 #number of sample
D = 5 #dimension of input
H = 4 #dimension of hidden
T_new = 1 #length of per each sample
context_dim = 3
K = 5
x = np.linspace(-0.4, 0.6, num=N*T_new*D, dtype = theano.config.floatX).reshape(T_new, N, D)
h0= np.linspace(-0.4, 0.8, num=N*H, dtype = theano.config.floatX).reshape(N, H)
Wx= np.linspace(-0.2, 0.9, num=4*D*H, dtype = theano.config.floatX).reshape(D, 4*H)
Wh= np.linspace(-0.3,0.6, num =4*H*H, dtype = theano.config.floatX).reshape(H,4*H)
b = np.linspace(0.0, 0.0, num = 4*H, dtype = theano.config.floatX)
Wz= np.linspace(-0.3, 0.6, num=4*H*context_dim, dtype = theano.config.floatX).reshape(context_dim, 4*H)
Hcontext = np.linspace(-0.2, 0.6, num=H*K, dtype = theano.config.floatX).reshape(H, K)
Zcontext = np.linspace(-0.2, 0.5, num=context_dim*K, dtype= theano.config.floatX).reshape(context_dim, K)
Va= np.linspace(0.1, 0.4, num=K, dtype = theano.config.floatX)
Va_reshape = Va.reshape(K,1)
image_feature_3D = np.linspace(-0.2, 0.5, num=10*N*context_dim, dtype = theano.config.floatX).reshape(N,10, context_dim)
h0_theano = h0.reshape(1, N, H)
# h0_symb = theano.tensor.ftensor3("h_symb")
# lstm_theano_layer.h_m1.set_value(h0_theano)
c0_theano = np.zeros((1, N, H), dtype = theano.config.floatX)
# c0_symb = theano.tensor.ftensor3("c_symb")
# lstm_theano_layer.c_m1.set_value(c0_theano)
z0_theano = np.zeros((1, N, context_dim), dtype = theano.config.floatX)
x_theano = x.reshape(T_new, N, D, 1)
image_feature_input = image_feature_3D
weight_y_in_value = np.zeros(( 10, context_dim) , dtype= theano.config.floatX)
b_theano= b.reshape(1, 1, 4*H)
pdb.set_trace()
#symbolic variables
initial_h0_layer_out = theano.tensor.tensor3(name = 'h0_initial', dtype = theano.config.floatX)
initial_c0_layer_out = theano.tensor.tensor3(name = 'c0_initial', dtype = theano.config.floatX)
initial_z0 = T.tensor3(name= 'z0_initial', dtype = theano.config.floatX)
weight_y_in = theano.tensor.fmatrix("weight_y")
input_data = theano.tensor.tensor3(name ='x', dtype=theano.config.floatX)
image_feature_region = theano.tensor.tensor3(name = 'feature_region', dtype = theano.config.floatX)
Wi_sym, Wf_sym, Wc_sym, Wo_sym, Ui_sym, Uf_sym, Uc_sym, Uo_sym, Zi_sym, Zf_sym, Zc_sym, Zo_sym = T.fmatrices(12)
Zcontext_sym, Hcontext_sym = T.fmatrices(2)
bi = T.ftensor3("bi")
bf = T.ftensor3("bf")
bc = T.ftensor3("bc")
bo = T.ftensor3("bo")
Va_sym = T.fcol("Va")
out_sym = onestep_attend_tell(input_data, initial_h0_layer_out, initial_c0_layer_out, initial_z0,
Wi_sym, Wf_sym, Wc_sym, Wo_sym, Ui_sym, Uf_sym, Uc_sym, Uo_sym, Zi_sym, Zf_sym, Zc_sym, Zo_sym,
Zcontext_sym, Hcontext_sym, Va_sym,
bi, bf, bc, bo, image_feature_region, weight_y_in)
onestep_func = theano.function([input_data, initial_h0_layer_out, initial_c0_layer_out, initial_z0,
Wi_sym, Wf_sym, Wc_sym, Wo_sym, Ui_sym, Uf_sym, Uc_sym, Uo_sym, Zi_sym, Zf_sym, Zc_sym, Zo_sym,
Zcontext_sym, Hcontext_sym, Va_sym,
bi, bf, bc, bo, image_feature_region, weight_y_in], out_sym)
list_output = onestep_func(x, h0_theano, c0_theano, z0_theano,
Wx[:, :H], Wx[:, H:2*H], Wx[:, 2*H:3*H], Wx[:, 3*H:],
Wh[:, :H], Wh[:, H:2*H], Wh[:, 2*H:3*H], Wh[:, 3*H:],
Wz[:, :H], Wz[:, H:2*H], Wz[:, 2*H:3*H], Wz[:, 3*H:],
Zcontext,Hcontext,
Va_reshape,
b_theano[:,: , :H], b_theano[:, :, H:2*H], b_theano[:, :, 2*H:3*H], b_theano[:, :, 3*H:],
image_feature_input, weight_y_in_value)
pdb.set_trace()
print(list_output[0].shape)
print(list_output[1].shape)
print(list_output[2].shape)
pdb.set_trace()
def find_all_step_visible_data(all_step_original_data_shared,
all_step_visible_data_shared,
sigmas_shared,
N,
steps,
output_dims,
n_epochs,
initial_lr,
final_lr,
lr_switch,
initial_momentum,
final_momentum,
momentum_switch,
penalty_lambda,
metric,
verbose=0):
"""Optimize cost wrt all_step_visible_data[t], simultaneously for all t"""
# Optimization hyper-parameters
initial_lr = np.array(initial_lr, dtype=floath)
final_lr = np.array(final_lr, dtype=floath)
initial_momentum = np.array(initial_momentum, dtype=floath)
final_momentum = np.array(final_momentum, dtype=floath)
lr = T.fscalar('lr')
lr_shared = theano.shared(initial_lr)
momentum = T.fscalar('momentum')
momentum_shared = theano.shared(initial_momentum)
# Penalty hyper-parameter
penalty_lambda_var = T.fscalar('penalty_lambda')
penalty_lambda_shared = theano.shared(
np.array(penalty_lambda, dtype=floath))
# Yv velocities
all_step_visible_progress_shared = []
zero_velocities = np.zeros((N, output_dims), dtype=floath)
for t in range(steps):
all_step_visible_progress_shared.append(
theano.shared(np.array(zero_velocities)))
# Cost
all_step_original_data_vars = T.fmatrices(steps)
all_step_visible_data_vars = T.fmatrices(steps)
all_step_visible_progress_vars = T.fmatrices(steps)
sigmas_vars = T.fvectors(steps)
c_vars = []
for t in range(steps):
c_vars.append(
cost_var(all_step_original_data_vars[t],
all_step_visible_data_vars[t], sigmas_vars[t], metric))
cost = T.sum(c_vars) + penalty_lambda_var * movement_penalty(
all_step_visible_data_vars, N)
# Setting update for all_step_visible_data velocities
grad_Y = T.grad(cost, all_step_visible_data_vars)
givens = {
lr: lr_shared,
momentum: momentum_shared,
penalty_lambda_var: penalty_lambda_shared
}
updates = []
for t in range(steps):
updates.append(
(all_step_visible_progress_shared[t],
momentum * all_step_visible_progress_vars[t] - lr * grad_Y[t]))
givens[
all_step_original_data_vars[t]] = all_step_original_data_shared[t]
givens[all_step_visible_data_vars[t]] = all_step_visible_data_shared[t]
givens[all_step_visible_progress_vars[
t]] = all_step_visible_progress_shared[t]
givens[sigmas_vars[t]] = sigmas_shared[t]
update_Yvs = theano.function([], cost, givens=givens, updates=updates)
# Setting update for all_step_visible_data positions
updates = []
givens = dict()
for t in range(steps):
updates.append(
(all_step_visible_data_shared[t], all_step_visible_data_vars[t] +
all_step_visible_progress_vars[t]))
givens[all_step_visible_data_vars[t]] = all_step_visible_data_shared[t]
givens[all_step_visible_progress_vars[
t]] = all_step_visible_progress_shared[t]
update_all_step_visible_data = theano.function([], [],
givens=givens,
updates=updates)
# Momentum-based gradient descent
for epoch in range(n_epochs):
if epoch == lr_switch:
lr_shared.set_value(final_lr)
if epoch == momentum_switch:
momentum_shared.set_value(final_momentum)
c = update_Yvs()
update_all_step_visible_data()
if verbose:
print('Epoch: {0}. Cost: {1:.6f}.'.format(epoch + 1, float(c)))
all_step_visible_data = []
for t in range(steps):
all_step_visible_data.append(
np.array(all_step_visible_data_shared[t].get_value(),
dtype=floath))
return all_step_visible_data
def SGD(eta,
n_epochs,
valid_steps,
momentum,
low,
high,
init,
random_init='gaussian'):
t0 = time.time()
index = T.iscalar('index')
x, y, z, alpha = T.fmatrices('x', 'y', 'z', 'alpha')
n_minibatch = max_minibatch - 2
model = Model(n_tree, n_nodes, low, high, init, random_init)
model_op, auto_upd = model.op(x)
valid_op, valid_upd = model.valid_op(z, valid_steps)
loss = model.loss(y, model_op)
valid_loss = model.loss(alpha, valid_op)
print "Updation to be compiled yet"
params = model.params
train_upd = gradient_updates_momentum(loss, params, eta,
momentum) + auto_upd
train_output = [model_op, loss]
valid_output = [valid_op, valid_loss]
print "Train function to be compiled"
train_fn = theano.function(
[index],
train_output,
updates=train_upd,
givens={
x: train_x[:, n_in * index:n_in * (index + 1)],
y: train_x[:, (n_in * index + n_tree):(n_in * (index + 1) + 1)]
},
name='train_fn')
valid_fn = theano.function(
[index],
valid_output,
updates=valid_upd,
givens={
z:
train_x[:, n_tree * index:n_tree * (index + 1)],
alpha:
train_x[:, (n_in * index + n_tree):(n_in * index + n_tree +
valid_steps)]
},
name='valid_fn')
print "Train function compiled"
# Compilation over
#################
## TRAIN MODEL ##
#################
print 'The compilation time is', time.time() - t0
loss_list = []
for i in range(n_epochs):
epoch_loss = 0
t1 = time.time()
for idx in range(n_minibatch):
print 'The current idx is ', idx, ' and the epoch number is ', i
output, loss_ = train_fn(idx)[:-1], train_fn(idx)[-1]
if idx % 500 == 0:
v_output, v_loss = valid_fn(idx / 500)[:-1][0], valid_fn(
idx / 500)[-1]
print 'v_pred is', ' '.join(
[mappings_words[prediction(abc)] for abc in v_output])
print 'v_loss is', np.array(v_loss)
print 'The loss is', loss_
epoch_loss += loss_
loss_list.append(loss_)
print '==' * 20
print 'The mean loss for the epoch was', epoch_loss / float(
n_minibatch)
print 'Time taken by this epoch is', time.time() - t1
print '-' * 50
pyplot.plot(loss_list)
pyplot.show()
'''
A theano implementation of the T-LSTM
'''
import theano.tensor as T
from theano import function
import numpy as np
import collections
import pdb
import os
#np.seterr(under='warn')
h, b = T.fvectors('h', 'b')
W, X = T.fmatrices('W', 'X')
dotvec = function([h,b], T.dot(h,b))
dot = function([W, h], T.dot(W, h))
#dotF = function([W, h], T.dot(W, h))
#dot = lambda W, h: dotF(W, h.squeeze())
dotW = function([W, X], T.dot(W,X))
layer = function([W, h, b], T.dot(W, h) + b)
#layerF = function([W, h, b], T.dot(W, h) + b)
#layer = lambda W, h, b: layerF(W, h.squeeze(), b.squeeze())
sigmoid = function([h], T.nnet.ultra_fast_sigmoid(h))
#sigmoidF = function([h], T.nnet.ultra_fast_sigmoid(h))
#sigmoid = lambda h: sigmoidF(h.squeeze())
tanh = function([h], T.tanh(h))
#tanhF = function([h], T.tanh(h))
#tanh = lambda h: tanhF(h.squeeze())
add = function([h, b], h+b)
def __init__(self, We, params):
lstm_layers_num = 1
en_hidden_size = We.shape[1]
self.eta = params.eta
self.num_labels = params.num_labels
self.en_hidden_size = en_hidden_size
self.de_hidden_size = params.de_hidden_size
self.lstm_layers_num = params.lstm_layers_num
self._train = None
self._utter = None
self.params = []
self.encoder_lstm_layers = []
self.decoder_lstm_layers = []
self.hos = []
self.Cos = []
encoderInputs = tensor.imatrix()
decoderInputs, decoderTarget = tensor.imatrices(2)
encoderMask, TF, decoderMask, decoderInputs0 = tensor.fmatrices(4)
self.lookuptable = theano.shared(We)
#### the last one is for the stary symbole
self.de_lookuptable = theano.shared(name="Decoder LookUpTable",
value=init_xavier_uniform(
self.num_labels + 1,
self.de_hidden_size),
borrow=True)
self.linear = theano.shared(name="Linear",
value=init_xavier_uniform(
self.de_hidden_size, self.num_labels),
borrow=True)
self.hidden_decode = theano.shared(name="Hidden to Decode",
value=init_xavier_uniform(
2 * en_hidden_size,
self.de_hidden_size),
borrow=True)
self.hidden_bias = theano.shared(
name="Hidden to Bias",
value=np.asarray(np.random.randn(self.de_hidden_size, ) * 0.,
dtype=theano.config.floatX),
borrow=True)
self.params += [
self.linear, self.de_lookuptable, self.hidden_decode,
self.hidden_bias
] #concatenate
#(max_sent_size, batch_size, hidden_size)
state_below = self.lookuptable[encoderInputs.flatten()].reshape(
(encoderInputs.shape[0], encoderInputs.shape[1],
self.en_hidden_size))
for _ in range(self.lstm_layers_num):
enclstm_f = LSTM(self.en_hidden_size)
enclstm_b = LSTM(self.en_hidden_size, True)
self.encoder_lstm_layers.append(enclstm_f) #append
self.encoder_lstm_layers.append(enclstm_b) #append
self.params += enclstm_f.params + enclstm_b.params #concatenate
hs_f, Cs_f = enclstm_f.forward(state_below, encoderMask)
hs_b, Cs_b = enclstm_b.forward(state_below, encoderMask)
hs = tensor.concatenate([hs_f, hs_b], axis=2)
Cs = tensor.concatenate([Cs_f, Cs_b], axis=2)
self.hos += tensor.tanh(
tensor.dot(hs[-1], self.hidden_decode) + self.hidden_bias),
self.Cos += tensor.tanh(
tensor.dot(Cs[-1], self.hidden_decode) + self.hidden_bias),
state_below = hs
state_below = self.de_lookuptable[decoderInputs.flatten()].reshape(
(decoderInputs.shape[0], decoderInputs.shape[1],
self.de_hidden_size))
for i in range(self.lstm_layers_num):
declstm = LSTM(self.de_hidden_size)
self.decoder_lstm_layers += declstm, #append
self.params += declstm.params #concatenate
ho, Co = self.hos[i], self.Cos[i]
state_below, Cs = declstm.forward(state_below, decoderMask, ho, Co)
decoder_lstm_outputs = state_below
ei, di, dt = tensor.imatrices(3) #place holders
em, dm, tf, di0 = tensor.fmatrices(4)
#####################################################
#####################################################
linear_outputs = tensor.dot(decoder_lstm_outputs, self.linear)
softmax_outputs, updates = theano.scan(
fn=lambda x: tensor.nnet.softmax(x),
sequences=[linear_outputs],
)
def _NLL(pred, y, m):
return -m * tensor.log(pred[tensor.arange(encoderInputs.shape[1]),
y])
costs, _ = theano.scan(
fn=_NLL, sequences=[softmax_outputs, decoderTarget, decoderMask])
loss = costs.sum() / decoderMask.sum()
updates = lasagne.updates.adam(loss, self.params, self.eta)
#updates = lasagne.updates.apply_momentum(updates, self.params, momentum=0.9)
###################################################
#### using the ground truth when training
##################################################
self._train = theano.function(inputs=[ei, em, di, dm, dt],
outputs=[loss, softmax_outputs],
updates=updates,
givens={
encoderInputs: ei,
encoderMask: em,
decoderInputs: di,
decoderMask: dm,
decoderTarget: dt
})
#########################################################################
### For schedule sampling
#########################################################################
###### always use privous predict as next input
def _step2(state_, hs_, Cs_):
hs, Cs = [], []
token_idxs = tensor.cast(state_.argmax(axis=-1), "int32")
msk_ = tensor.fill(
(tensor.zeros_like(token_idxs, dtype="float32")), 1)
msk_ = msk_.dimshuffle('x', 0)
state_below0 = self.de_lookuptable[token_idxs].reshape(
(1, encoderInputs.shape[1], self.de_hidden_size))
for i, lstm in enumerate(self.decoder_lstm_layers):
h, C = lstm.forward(state_below0, msk_, hs_[i],
Cs_[i]) #mind msk
hs += h[-1],
Cs += C[-1],
state_below0 = h
hs, Cs = tensor.as_tensor_variable(hs), tensor.as_tensor_variable(
Cs)
newpred = tensor.dot(state_below0, self.linear).reshape(
(encoderInputs.shape[1], self.num_labels))
state_below = tensor.nnet.softmax(newpred)
return state_below, hs, Cs
hs0, Cs0 = tensor.as_tensor_variable(
self.hos, name="hs0"), tensor.as_tensor_variable(self.Cos,
name="Cs0")
train_outputs, _ = theano.scan(fn=_step2,
outputs_info=[decoderInputs0, hs0, Cs0],
n_steps=encoderInputs.shape[0])
train_predict = train_outputs[0]
train_costs, _ = theano.scan(
fn=_NLL, sequences=[train_predict, decoderTarget, decoderMask])
train_loss = train_costs.sum() / decoderMask.sum()
train_updates = lasagne.updates.adam(train_loss, self.params, self.eta)
#train_updates = lasagne.updates.apply_momentum(train_updates, self.params, momentum=0.9)
self._train2 = theano.function(
inputs=[ei, em, di0, dm, dt],
outputs=[train_loss, train_predict],
updates=train_updates,
givens={
encoderInputs: ei,
encoderMask: em,
decoderInputs0: di0,
decoderMask: dm,
decoderTarget: dt
}
#givens={encoderInputs:ei, encoderMask:em, decoderInputs:di, decoderMask:dm, decoderTarget:dt, TF:tf}
)
listof_token_idx = train_predict.argmax(axis=-1)
self._utter = theano.function(inputs=[ei, em, di0],
outputs=listof_token_idx,
givens={
encoderInputs: ei,
encoderMask: em,
decoderInputs0: di0
})
import mnist
def init_weights(n_in, n_out):
weights = np.random.randn(n_in, n_out) / np.sqrt(n_in)
return theano.shared(np.asarray(weights, dtype=theano.config.floatX))
def feed_forward(X, w_h, w_o):
h = T.nnet.sigmoid(T.dot(X, w_h))
return T.nnet.softmax(T.dot(h, w_o))
trX, trY, teX, teY = mnist.load_data(one_hot=True)
w_h, w_o = init_weights(28*28, 100), init_weights(100, 10)
num_epochs, batch_size, learn_rate = 30, 10, 0.2
X, Y = T.fmatrices('X', 'Y')
y_ = feed_forward(X, w_h, w_o)
weights = [w_h, w_o]
grads = T.grad(cost=T.nnet.categorical_crossentropy(y_, Y).mean(), wrt=weights)
train = theano.function(
inputs=[X, Y],
updates=[[w, w - g * learn_rate] for w, g in zip(weights, grads)],
allow_input_downcast=True)
predict = theano.function(inputs=[X], outputs=T.argmax(y_, axis=1))
for i in range(num_epochs):
for j in xrange(0, len(trX), batch_size):
train(trX[j:j+batch_size], trY[j:j+batch_size])
print i, np.mean(predict(teX) == np.argmax(teY, axis=1))
def __init__(self,
state = 'x',
measurement = 'z',
motion_transition = None,
measurement_transition = None):
self.N = len(state.split(' '))
self.M = len(measurement.split(' '))
self.X, self.Z = T.fvectors('X','Z')
self.P, self.Q, self.R = T.fmatrices('P','Q','R')
self.F, self.H = T.matrices('F','H')
self.dt = T.scalar('dt')
self.X_ = T.dot(self.F, self.X)
self.fX_ = G.jacobian(T.flatten(self.X_), self.X)
self.P_ = T.dot(T.dot(self.fX_, self.P), T.transpose(self.fX_)) + self.dt * self.Q
self.h = T.dot(self.H, self.X_)
self.y = self.Z - self.h
self.hX_ = G.jacobian(self.h, self.X_)
self.matrix_inv = T.nlinalg.MatrixInverse()
self.S = T.dot(T.dot(self.hX_, self.P_), T.transpose(self.hX_)) + self.R
self.K = T.dot(T.dot(self.P_, T.transpose(self.hX_)), self.matrix_inv(self.S))
self.X__ = self.X_ + T.dot(self.K, self.y)
self.P__ = T.dot(T.identity_like(self.P) - T.dot(self.K, self.hX_), self.P_)
self.prediction = theano.function(inputs = [self.X,
self.P,
self.Q,
self.F,
self.dt],
outputs = [self.X_,
self.P_],
allow_input_downcast = True)
self.update = theano.function(inputs = [self.X,
self.Z,
self.P,
self.Q,
self.R,
self.F,
self.H,
self.dt],
outputs = [self.X__,
self.P__],
allow_input_downcast = True)
if motion_transition == None:
self.motion_transition = np.eye(self.N)
else:
self.motion_transition = np.array(motion_transition)
if measurement_transition == None:
self.measurement_transition = np.eye(self.M)
else:
self.measurement_transition = np.array(motion_transition)
def find_Ys(Xs_shared,
Ys_shared,
sigmas_shared,
N,
steps,
output_dims,
n_epochs,
initial_lr,
final_lr,
lr_switch,
init_stdev,
initial_momentum,
final_momentum,
momentum_switch,
lmbda,
metric,
verbose=0):
"""Optimize cost wrt Ys[t], simultaneously for all t"""
# Optimization hyperparameters
initial_lr = np.array(initial_lr, dtype=floath)
final_lr = np.array(final_lr, dtype=floath)
initial_momentum = np.array(initial_momentum, dtype=floath)
final_momentum = np.array(final_momentum, dtype=floath)
lr = T.fscalar('lr')
lr_shared = theano.shared(initial_lr)
momentum = T.fscalar('momentum')
momentum_shared = theano.shared(initial_momentum)
# Penalty hyperparameter
lmbda_var = T.fscalar('lmbda')
lmbda_shared = theano.shared(np.array(lmbda, dtype=floath))
# Yv velocities
Yvs_shared = []
zero_velocities = np.zeros((N, output_dims), dtype=floath)
for t in range(steps):
Yvs_shared.append(theano.shared(np.array(zero_velocities)))
# Cost
Xvars = T.fmatrices(steps)
Yvars = T.fmatrices(steps)
Yv_vars = T.fmatrices(steps)
sigmas_vars = T.fvectors(steps)
c_vars = []
for t in range(steps):
c_vars.append(cost_var(Xvars[t], Yvars[t], sigmas_vars[t], metric))
cost = T.sum(c_vars) + lmbda_var * movement_penalty(Yvars, N)
# Setting update for Ys velocities
grad_Y = T.grad(cost, Yvars)
givens = {
lr: lr_shared,
momentum: momentum_shared,
lmbda_var: lmbda_shared
}
updates = []
for t in range(steps):
updates.append((Yvs_shared[t], momentum * Yv_vars[t] - lr * grad_Y[t]))
givens[Xvars[t]] = Xs_shared[t]
givens[Yvars[t]] = Ys_shared[t]
givens[Yv_vars[t]] = Yvs_shared[t]
givens[sigmas_vars[t]] = sigmas_shared[t]
update_Yvs = theano.function([], cost, givens=givens, updates=updates)
# Setting update for Ys positions
updates = []
givens = dict()
for t in range(steps):
updates.append((Ys_shared[t], Yvars[t] + Yv_vars[t]))
givens[Yvars[t]] = Ys_shared[t]
givens[Yv_vars[t]] = Yvs_shared[t]
update_Ys = theano.function([], [], givens=givens, updates=updates)
# Momentum-based gradient descent
for epoch in range(n_epochs):
if epoch == lr_switch:
lr_shared.set_value(final_lr)
if epoch == momentum_switch:
momentum_shared.set_value(final_momentum)
c = update_Yvs()
update_Ys()
if verbose:
print('Epoch: {0}. Cost: {1:.6f}.'.format(epoch + 1, float(c)))
Ys = []
for t in range(steps):
Ys.append(np.array(Ys_shared[t].get_value(), dtype=floath))
return Ys
def __init__(self, inf=1e37):
pos, vel = T.fmatrices(['pos', 'vel'])
nc, N, n_steps = T.iscalars(['nc', 'N', 'n_steps'])
ra, rb, re, r0 = T.fscalars(['ra', 'rb', 're', 'r0'])
v0, j, b = T.fscalars(['v0', 'J', 'b'])
nu = trng.uniform(size=(N, 2), low=0.0, high=3.14159, dtype='floatX')
def distance_tensor(X):
E = X.reshape((X.shape[0], 1, -1)) - X.reshape((1, X.shape[0], -1))
D = T.sqrt(T.sum(T.square(E), axis=2))
return D
def direction_tensor(X):
E = X.reshape((X.shape[0], 1, -1)) - X.reshape((1, X.shape[0], -1))
L = T.sqrt(T.sum(T.square(E), axis=2))
L = T.pow(L + T.identity_like(L), -1)
L = T.stack([L, L, L], axis=2)
return L * E
def neighbourhood(X):
D = distance_tensor(X)
N = T.argsort(D, axis=0)
mask = T.cast(T.lt(N, nc), 'float32')
return N[1:nc + 1], mask
def alignment(X, Y):
n, d = neighbourhood(X)
return T.sum(Y[n], axis=0)
def cohesion(X, inf=100.0):
D = distance_tensor(X)
E = direction_tensor(X)
n, d = neighbourhood(X)
F = T.zeros_like(E)
D = T.stack([D, D, D], axis=2)
d = T.stack([d, d, d], axis=2)
c1 = T.lt(D, rb)
c2 = T.and_(T.gt(D, rb), T.lt(D, ra))
c3 = T.and_(T.gt(D, ra), T.lt(D, r0))
F = T.set_subtensor(F[c1], -E[c1])
F = T.set_subtensor(F[c2], 0.25 * (D[c2] - re) / (ra - re) * E[c2])
F = T.set_subtensor(F[c3], E[c3])
return T.sum(d * F, axis=0)
def perturbation(nu=nu):
phi = nu[:, 0]
theta = 2.0 * nu[:, 1]
return T.stack([
T.sin(theta) * T.sin(phi),
T.cos(theta) * T.sin(phi),
T.cos(phi)
],
axis=1)
def step(X, dX):
X_ = X + dX
V_ = j * nc / v0 * (alignment(
X, dX)) + b * (cohesion(X)) + nc * (perturbation())
dV = T.sqrt(T.sum(T.square(V_), axis=1)).reshape(V_.shape[0], 1)
dV = T.stack([dV, dV, dV], axis=1)
V = v0 * V_ / dV
return T.cast(X_, 'float32'), T.cast(V, 'float32')
def probability(X, Y):
n, d = neighbourhood(X)
vDv = T.batched_dot(Y[n].swapaxes(0, 1), Y)
p = T.exp((j / 2.0) * T.sum(vDv, axis=1))
return p / T.sum(p)
sim, update = theano.scan(step,
outputs_info=[pos, vel],
n_steps=n_steps)
pos_, vel_ = sim
mean_final_velocity = 1 / (N * v0) * T.sqrt(
T.sum(T.square(T.sum(vel_[-1], axis=0))))
particle_probability = probability(pos_[-1], vel_[-1])
self.f = theano.function(
[pos, vel, nc, ra, rb, r0, re, j, v0, b, N, n_steps], [pos_, vel_],
allow_input_downcast=True)
self.g = theano.function(
[pos, vel, nc, ra, rb, r0, re, j, v0, b, N, n_steps],
mean_final_velocity,
allow_input_downcast=True)
self.h = theano.function(
[pos, vel, nc, ra, rb, r0, re, j, v0, b, N, n_steps],
particle_probability,
allow_input_downcast=True)
def __init__(self, Ne, Ni, n_inp, W_inp=None, W_inner=None):
'''class SNNgroup's self Parameters:
self.A: update matrix
self.S: neuron state varaibles
self.W_inner: inner-connect weights in the group
self.W_inp: input weights
self.spikes: the spikes matrix in the time t
self.SpkC : spike containers
input : '''
self.number = Ne + Ni
self.Ne = Ne
self.Ni = Ni
self.mV = self.ms = 1e-3 # units
dt = 1 * self.ms # timestep
self.dt = dt
taum = 20 * self.ms # membrane time constant
taue = 5 * self.ms
taui = 10 * self.ms
#self.Vt=-1*self.mV # threshold = -50+49
self.Vt = 15 * self.mV #threshold = -55+70
#self.Vr=-11*self.mV # reset = -60+49
self.Vr = 0 * self.mV # reset = -70+70
self.Vi = -10 * self.mV # VI = -80+70
self.dApre = .0001
#self.dApre = .95 #changed into .95
self.dApost = -self.dApre * 1.05
self.tauP = 20 * self.ms
#self.input = input
self.n_inp = n_inp
self.weight = .001
self.weightIn = 1.
self.wmax = 200 * self.weight
zero = np.array([0]).astype(theano.config.floatX)
self.zero = theano.shared(zero, name='zero', borrow=True)
"""
Equations
---------
eqs='''
dv/dt = (ge*70mV-gi*10-(v+70*mV))/(20*ms) : volt
dge/dt = -ge/(5*self.ms) : volt
dgi/dt = -gi/(10*self.ms) : volt
'''
"""
# Update matrix
A = np.array([[np.exp(-dt / taum), 0, 0],
[
taue / (taum - taue) *
(np.exp(-dt / taum) - np.exp(-dt / taue)),
np.exp(-dt / taue), 0
],
[
-taui / (taum - taui) *
(np.exp(-dt / taum) - np.exp(-dt / taui)), 0,
np.exp(-dt / taui)
]],
dtype=theano.config.floatX).T
A = theano.shared(value=A, name='A', borrow=True)
self.A = A
# State varible : [v;ge;gi] (size=3*self.number)
S = np.ones((1, self.number), dtype=theano.config.floatX) * self.Vr
S = np.vstack((S, np.zeros((2, self.number),
dtype=theano.config.floatX)))
self.S_init = S
S = theano.shared(value=S, name='S', borrow=True)
self.S = S
if W_inner == None:
# weights of inner connections (size= self.number*self.number)
self.W_inner_ini = np.ones(
(self.number, self.number),
dtype=theano.config.floatX) * self.weight
#self.W_inner_ini[Ne:,:] = self.weightIn
self.W_inner_ini[Ne:, :] = self.weight
wtmp = np.eye(self.number)
ind = wtmp.nonzero()
self.W_inner_ini[ind] = 0
W_inner = theano.shared(value=self.W_inner_ini,
name='W_inner',
borrow=True)
self.W_inner = W_inner
else:
self.W_inner = theano.shared(W_inner, name='W_inner', borrow=True)
# weights of input connections (size=n_inp*self.number
rng = np.random.RandomState(1234)
if W_inp == None:
#W_inp = np.ones((self.n_inp,self.number)).astype(theano.config.floatX) #needs specification later
#W_inp = np.random.rand(self.n_inp,self.number).astype(theano.config.floatX)*.00001*self.ms #needs specification later
self.W_inp_ini = np.ones((self.n_inp, self.number)).astype(
theano.config.floatX) * self.weight
self.W_inp_ini[:, self.Ne:] = self.weightIn
W_inp = theano.shared(self.W_inp_ini, name='W_inp', borrow=True)
self.W_inp = W_inp
else:
self.W_inp = theano.shared(W_inp, name='W_inp', borrow=True)
# Spike Container
#spkC = theano.shared(value=np.empty((1,self.number)).astype(theano.config.floatX),name='spkC',borrow=True)
spkC = np.empty((1, self.number)).astype(theano.config.floatX)
self.spkC = spkC
#spikes=np.empty((self.number,1),dtype=theano.config.floatX)
#self.spikes = theano.shared(value=spikes,name='spikes',borrow=True)
# not sure the dtype of sp_history
self.sp_history = np.array([])
#output = np.empty(self.number,dtype=theano.config.floatX)
#self.output = theano.shared(value=output,name='output',borrow=True)
self.V_record = np.empty((1, self.number))
self.ge_record = np.empty((1, self.number))
self.gi_record = np.empty((1, self.number))
#================================================
# Process Function Initial
# input:: 0-1 vector
'''Update Schedule:
1.Update state variables of SNNgroup: dot(A,S)
1.Update state variables of Synapses: dot(exp(-dt/tau),Ssynapse), including W_inp and W_inner
2.Call thresholding function: S[0,:]>Vt
3.Push spikes into SpikeContainer
4.Propagate spikes via Connection(possibly with delays)
5.Update state variables of Synapses (STDP)
6.Call reset function on neurons which has spiked'''
Ne = self.Ne
Ni = self.Ni
m = T.fmatrix(name='m')
#self.Vt = T.as_tensor_variable(self.Vt,'Vt')
# "Update state function:: stat()"
# return np array
# shape(stat()) = shape(self.S)
S_update = T.dot(self.A, self.S)
self.stat = theano.function(inputs=[],
outputs=[],
updates={self.S: S_update})
#============================================================
# Update state of Synapses
# Update matrix of Synapse
A_STDP = np.array([[np.exp(-self.dt / self.tauP), 0],
[0, np.exp(-self.dt / self.tauP)]],
dtype=theano.config.floatX)
# Spre_inner :: pre synapse of inner connections
# Spost_inner:: post synapse of inner connections
# Spre_inp :: pre synapse of input conenctions
# Spost_inp :: post synapse of input connections
self.Spre_inner_ini = np.zeros((self.number, self.number),
dtype=theano.config.floatX)
Spre_inner = theano.shared(self.Spre_inner_ini,
name='Spre_inner',
borrow=True)
self.Spre_inner = Spre_inner
self.Spost_inner_ini = np.zeros((self.number, self.number),
dtype=theano.config.floatX)
Spost_inner = theano.shared(value=self.Spost_inner_ini,
name='Spost_inner',
borrow=True)
self.Spost_inner = Spost_inner
self.Spre_inp_ini = np.zeros((self.n_inp, self.number)).astype(
theano.config.floatX) #needs specification later
Spre_inp = theano.shared(value=self.Spre_inp_ini,
name='Spre_inp',
borrow=True)
self.Spre_inp = Spre_inp
self.Spost_inp_ini = np.zeros((self.n_inp, self.number)).astype(
theano.config.floatX) #needs specification later
Spost_inp = theano.shared(value=self.Spost_inp_ini,
name='Spost_inp',
borrow=True)
self.Spost_inp = Spost_inp
U = T.fscalar('U')
UM = T.fmatrix('UM')
#UpreV = theano.shared(A_STDP[0,0],name='UpreV',borrow=True) # Wpre = UpreV*Wpre
#UpostV = theano.shared(A_STDP[1,1],name='UpostV',borrow=True)
self.tmp = np.array(
np.exp(-self.dt / self.tauP).astype(theano.config.floatX))
self.SynFresh = theano.shared(self.tmp, name='SynFresh', borrow=True)
self.UpdateSpre_inner = theano.function(
inputs=[],
outputs=None,
updates={self.Spre_inner: T.dot(self.SynFresh, self.Spre_inner)},
allow_input_downcast=True)
self.UpdateSpost_inner = theano.function(
inputs=[],
outputs=None,
updates={self.Spost_inner: T.dot(self.SynFresh, self.Spost_inner)},
allow_input_downcast=True)
self.UpdateSpre_inp = theano.function(
inputs=[],
outputs=None,
updates={self.Spre_inp: T.dot(self.SynFresh, self.Spre_inp)},
allow_input_downcast=True)
self.UpdateSpost_inp = theano.function(
inputs=[],
outputs=None,
updates={self.Spost_inp: T.dot(self.SynFresh, self.Spost_inp)},
allow_input_downcast=True)
#------------------------------------------
#tmp = math.exp(-self.dt/self.tauP)
#tmp = T.as_tensor(0.95122945)
#================================================================
#------------------------------------------
# "thresholding function:: spike_fun()"
# type return :: np.ndarray list
# shape return:: shape(spike_fun()) = (self.number,)
self.spike_fun = theano.function(
inputs=[U], #[self.S]
outputs=(T.gt(self.S[0, :],
U))) #type outputs: np.ndarray,shape::(nL,)
#'outputs = (self.S[0,:]>Vt).astype(theano.config.floatX)), #type outputs: list'
#'updates={self.spikes:(self.S[0,:]>Vt).astype(theano.config.floatX)}'
#------------------------------------
#------------------------------------
#=================================================================
# "Push spike into Container function:: spCfun(vector)"
# type vector :: np.array([],dtype=theano.config.floatX)!!!
# type return :: np array
# shape return:: shape(spCfun()) = ( shape(self.spkC)[0]+1 , shape(self.spkC)[1] )
#updates={self.spkC:T.stack(self.spkC,sp)})
'''spike_prop = theano.function( #wrong
inputs = [],
outputs =[],
updates = {self.S:np.dot(self.W_inner,self.spikes)+self.S})#wrong'''
#-------------------------------
#--------------------------------
#====================================================================
# Propagate spikes
# inner connection:
# S_inner = f(inputs, outputs, updates)
# Param:: inputs: spike 0-1 vector
# Param:: inputs: spike is from function-> spike_fun
# S_inner(spk)::-> for i in spk[0:Ne].nonzero()[0]:
# S[1,:] = Winner[i,:]+S[1,:] (excitatory conenction)
# for j in spk[Ne,:].nonzero()[0]:
# S[2,:] = Winner[j,:]+S[2,:] (inhibitory connection)
vinner = T.fvector(
name='vinner') # vinner = spk :: np.array((1,self.number)
def add_f1(i, p, q):
np = T.inc_subtensor(p[1, :], q[i, :]) #ge
return {p: np}
def add_f2(i, p, q):
np = T.inc_subtensor(p[2, :], q[i, :]) #gi
return {p: np}
#deltaWinner1,updates1 = theano.scan(fn=lambda i: self.W_inner[i,:]*i+self.S[1,:], sequences=vinner[0:Ne])
deltaWinner1, updates1 = theano.scan(
fn=add_f1,
sequences=vinner[0:Ne].nonzero()[0],
non_sequences=[self.S, self.W_inner])
#deltaWinner2,updates2 = theano.scan(fn=lambda i: self.W_inner[i,:]*i+self.S[2,:], sequences=vinner[Ne:])
deltaWinner2, updates2 = theano.scan(
fn=add_f2,
sequences=vinner[Ne:].nonzero()[0] + self.Ne,
non_sequences=[self.S, self.W_inner])
# S = S+W
self.S_inner1 = theano.function(inputs=[vinner],
outputs=None,
updates=updates1,
allow_input_downcast=True)
self.S_inner2 = theano.function(inputs=[vinner],
outputs=None,
updates=updates2,
allow_input_downcast=True)
#------------------------------------------
#------------------------------------------
# outter connection (input spikes):
# type input: index list
voutter = T.fvector(name='voutter')
#deltaWoutter = theano.scan(fn=lambda j: self.W_inp[j,:]+self.S[1,:],sequences=voutter)
deltaWoutter, updatesout1 = theano.scan(
fn=add_f1,
sequences=voutter.nonzero()[0],
non_sequences=[self.S, self.W_inp])
self.S_inp = theano.function(inputs=[voutter],
outputs=None,
updates=updatesout1,
allow_input_downcast=True)
#------------------------------------
#-------------------------------------
#=====================================================================
# Update Synapses (STDP | STDC)
# Pre:: Apre += self.dApre, w+=Apost
# Post:: Apost+=self.dApost, w+=Apre
#
# USpreInner :: Perform Pre function No.1 in inner connections
# UWInner :: Perform Pre function No.2 in inner connections
# UpreInner :: Function
def add_synap_pre(i, p, po, s, q):
# i :: sequence
# p :: pre | post
# s :: dApre | dApost
# q :: W
index = T.nonzero(q[i, :self.Ne])
np = T.inc_subtensor(p[i, index], s)
## tmp = p[i,:]
## tmp=T.inc_subtensor(tmp[index],s)
## np=T.set_subtensor(p[i,:],tmp)
#np = T.inc_subtensor(p[i,:],s)
nw = T.inc_subtensor(q[i, :], po[i, :])
nw = T.clip(nw, 0, self.wmax)
return {p: np, q: nw}
def add_synap_pre_inp(i, p, po, s, q):
# i :: sequence
# p :: pre | post
# s :: dApre | dApost
# q :: W
index = T.nonzero(q[i, :self.Ne])
np = T.inc_subtensor(p[i, index], s)
## tmp = p[i,:]
## tmp=T.inc_subtensor(tmp[index],s)
## np=T.set_subtensor(p[i,:],tmp)
#np = T.inc_subtensor(p[i,:],s)
nw = T.inc_subtensor(q[i, :], po[i, :])
nw = T.clip(nw, 0, self.wmax)
return {p: np, q: nw}
def add_synap_post(i, po, p, s, q):
# i:: sequence
# po:: post
# p:: pre
# s:: dA
# q:: W
index = T.nonzero(q[:self.Ne, i])
npo = T.inc_subtensor(po[index, i], s)
nw = T.inc_subtensor(q[:, i], p[:, i])
nw = T.clip(nw, 0, self.wmax)
return {po: npo, q: nw}
def add_synap_post_inp(i, po, p, s, q):
# i:: sequence
# po:: post
# p:: pre
# s:: dA
# q:: W
index = T.nonzero(q[:self.Ne, i])
npo = T.inc_subtensor(po[index, i], s)
nw = T.inc_subtensor(q[:, i], p[:, i])
nw = T.clip(nw, 0, self.wmax)
return {po: npo, q: nw}
add_dA = T.fscalar('add_dA')
add_p, add_po, add_q = T.fmatrices('add_p', 'add_po', 'add_q')
#-------------------------------------------------------------------------
#USinner,updatesUinner = theano.scan(fn=add_synap_pre,sequences=vinner,non_sequences=[self.Spre_inner,self.Spost_inp,self.dApre,self.W_inner])
'USinner,updatesUinner = theano.scan(fn=add_synap_pre,sequences=vinner.nonzero()[0],non_sequences=[add_p,add_po,add_dA,add_q])'
#USinner1,updatesUinner1 = theano.scan(fn=add_synap_pre,sequences=vinner,non_sequences=[self.Spost_inner,self.Spre_inner,self.dApost,self.W_inner])
#-------------------------------------------------------------------------
#UpostInner = theano.function(inputs[vinner],updates={self.Spost_inner:USpostInner})
#UpostInp = theano.function(inputs=[vinner],updates={self.W_inner:UWInnerpost})
'USinner_f = theano.function(inputs=[vinner,add_p,add_po,add_dA,add_q],outputs=None,updates=updatesUinner)'
#USinner_step2 = theano.function(inputs=[vinner,add_p,add_po,add_dA,add_q],outputs=None,updates=updatesUinner)
USinner_inner_pre, updatesUinner_inner_pre = theano.scan(
fn=add_synap_pre,
sequences=vinner[:self.Ne].nonzero()[0],
non_sequences=[
self.Spre_inner, self.Spost_inner, add_dA, self.W_inner
])
self.USinner_f_inner_pre = theano.function(
inputs=[vinner, add_dA],
outputs=None,
updates=updatesUinner_inner_pre,
allow_input_downcast=True)
USinner_innerpost, updatesUinner_inner_post = theano.scan(
fn=add_synap_post,
sequences=vinner[:self.Ne].nonzero()[0],
non_sequences=[
self.Spost_inner, self.Spre_inner, add_dA, self.W_inner
])
self.USinner_f_inner_post = theano.function(
inputs=[vinner, add_dA],
outputs=None,
updates=updatesUinner_inner_post,
allow_input_downcast=True)
USinner_inp_pre, updatesUSinner_inp_pre = theano.scan(
fn=add_synap_pre_inp,
sequences=vinner.nonzero()[0],
non_sequences=[self.Spre_inp, self.Spost_inp, add_dA, self.W_inp])
self.USinner_f_inp_pre = theano.function(
inputs=[vinner, add_dA],
outputs=None,
updates=updatesUSinner_inp_pre,
allow_input_downcast=True)
USinner_inp_post, updatesUSinner_inp_post = theano.scan(
fn=add_synap_post_inp,
sequences=vinner[:self.Ne].nonzero()[0],
non_sequences=[self.Spost_inp, self.Spre_inp, add_dA, self.W_inp])
self.USinner_f_inp_post = theano.function(
inputs=[vinner, add_dA],
outputs=None,
updates=updatesUSinner_inp_post,
allow_input_downcast=True)
# Call reset function
def reset_v(index, vr):
nv = T.set_subtensor(self.S[0, index], vr)
return {self.S: nv}
resetV, resetV_update = theano.scan(fn=reset_v,
sequences=vinner.nonzero()[0],
non_sequences=[U])
self.resetV_f = theano.function(inputs=[vinner, U],
outputs=None,
updates=resetV_update,
allow_input_downcast=True)
setvalue = T.fscalar('setvalue')
iv = T.ivector('iv')
def reset_state(i, value, state):
nstate = T.set_subtensor(state[i, :], value)
return {state: nstate}
reset_S_state, Upreset_S_state = theano.scan(
fn=reset_state, sequences=iv, non_sequences=[setvalue, self.S])
self.reset_S_fn = theano.function(inputs=[iv, setvalue],
outputs=None,
updates=Upreset_S_state)
def init_weights(n_in, n_out):
weights = np.random.randn(n_in, n_out) / np.sqrt(n_in)
return theano.shared(np.asarray(weights, dtype=theano.config.floatX))
def feed_forward(X, w_h, w_o):
h = T.nnet.sigmoid(T.dot(X, w_h))
return T.nnet.softmax(T.dot(h, w_o))
trX, trY, teX, teY = mnist.load_data(one_hot=True)
w_h, w_o = init_weights(28 * 28, 100), init_weights(100, 10)
num_epochs, batch_size, learn_rate = 30, 10, 0.2
X, Y = T.fmatrices('X', 'Y')
y_ = feed_forward(X, w_h, w_o)
weights = [w_h, w_o]
grads = T.grad(cost=T.nnet.categorical_crossentropy(y_, Y).mean(), wrt=weights)
train = theano.function(inputs=[X, Y],
updates=[[w, w - g * learn_rate]
for w, g in zip(weights, grads)],
allow_input_downcast=True)
predict = theano.function(inputs=[X], outputs=T.argmax(y_, axis=1))
for i in range(num_epochs):
for j in xrange(0, len(trX), batch_size):
train(trX[j:j + batch_size], trY[j:j + batch_size])
print i, np.mean(predict(teX) == np.argmax(teY, axis=1))
def __init__(self, We, params):
lstm_layers_num = 1
en_hidden_size = We.shape[1]
self.eta = params.eta
self.num_labels = params.num_labels
self.en_hidden_size = en_hidden_size
self.de_hidden_size = params.de_hidden_size
self.lstm_layers_num = params.lstm_layers_num
self._train = None
self._utter = None
self.params = []
self.encoder_lstm_layers = []
self.decoder_lstm_layers = []
self.hos = []
self.Cos = []
encoderInputs = tensor.imatrix()
decoderInputs, decoderTarget = tensor.imatrices(2)
encoderMask, TF, decoderMask, decoderInputs0 = tensor.fmatrices(4)
self.lookuptable = theano.shared(We)
#### the last one is for the stary symbole
self.de_lookuptable = theano.shared(name="Decoder LookUpTable",
value=init_xavier_uniform(
self.num_labels + 1,
self.de_hidden_size),
borrow=True)
self.linear = theano.shared(
name="Linear",
value=init_xavier_uniform(self.de_hidden_size + 2 * en_hidden_size,
self.num_labels),
borrow=True)
self.linear_bias = theano.shared(
name="Hidden to Bias",
value=np.asarray(np.random.randn(self.num_labels, ) * 0.,
dtype=theano.config.floatX),
borrow=True)
#self.hidden_decode = theano.shared(name="Hidden to Decode", value= init_xavier_uniform(2*en_hidden_size, self.de_hidden_size), borrow = True)
#self.hidden_bias = theano.shared(
# name="Hidden to Bias",
# value=np.asarray(np.random.randn(self.de_hidden_size, )*0., dtype=theano.config.floatX) ,
# borrow=True
# )
#self.params += [self.linear, self.de_lookuptable, self.hidden_decode, self.hidden_bias] #concatenate
self.params += [self.linear, self.linear_bias, self.de_lookuptable
] #the initial hidden state of decoder lstm is zeros
#(max_sent_size, batch_size, hidden_size)
state_below = self.lookuptable[encoderInputs.flatten()].reshape(
(encoderInputs.shape[0], encoderInputs.shape[1],
self.en_hidden_size))
for _ in range(self.lstm_layers_num):
enclstm_f = LSTM(self.en_hidden_size)
enclstm_b = LSTM(self.en_hidden_size, True)
self.encoder_lstm_layers.append(enclstm_f) #append
self.encoder_lstm_layers.append(enclstm_b) #append
self.params += enclstm_f.params + enclstm_b.params #concatenate
hs_f, Cs_f = enclstm_f.forward(state_below, encoderMask)
hs_b, Cs_b = enclstm_b.forward(state_below, encoderMask)
hs = tensor.concatenate([hs_f, hs_b], axis=2)
Cs = tensor.concatenate([Cs_f, Cs_b], axis=2)
hs0 = tensor.concatenate([hs_f[-1], hs_b[0]], axis=1)
Cs0 = tensor.concatenate([Cs_f[-1], Cs_b[0]], axis=1)
#self.hos += tensor.tanh(tensor.dot(hs0, self.hidden_decode) + self.hidden_bias),
#self.Cos += tensor.tanh(tensor.dot(Cs0, self.hidden_decode) + self.hidden_bias),
self.hos += tensor.alloc(
np.asarray(0., dtype=theano.config.floatX),
encoderInputs.shape[1], self.de_hidden_size),
self.Cos += tensor.alloc(
np.asarray(0., dtype=theano.config.floatX),
encoderInputs.shape[1], self.de_hidden_size),
state_below = hs
Encoder = state_below
ei, di, dt = tensor.imatrices(3) #place holders
em, dm, tf, di0 = tensor.fmatrices(4)
self.encoder_function = theano.function(inputs=[ei, em],
outputs=Encoder,
givens={
encoderInputs: ei,
encoderMask: em
})
#####################################################
#####################################################
state_below = self.de_lookuptable[decoderInputs.flatten()].reshape(
(decoderInputs.shape[0], decoderInputs.shape[1],
self.de_hidden_size))
for i in range(self.lstm_layers_num):
declstm = LSTM(self.de_hidden_size)
self.decoder_lstm_layers += declstm, #append
self.params += declstm.params #concatenate
ho, Co = self.hos[i], self.Cos[i]
state_below, Cs = declstm.forward(state_below, decoderMask, ho, Co)
##### Here we include the representation from the decoder
decoder_lstm_outputs = tensor.concatenate([state_below, Encoder],
axis=2)
linear_outputs = tensor.dot(decoder_lstm_outputs,
self.linear) + self.linear_bias[None,
None, :]
softmax_outputs, _ = theano.scan(
fn=lambda x: tensor.nnet.softmax(x),
sequences=[linear_outputs],
)
def _NLL(pred, y, m):
return -m * tensor.log(pred[tensor.arange(encoderInputs.shape[1]),
y])
costs, _ = theano.scan(
fn=_NLL, sequences=[softmax_outputs, decoderTarget, decoderMask])
loss = costs.sum() / decoderMask.sum() + params.L2 * sum(
lasagne.regularization.l2(x) for x in self.params)
updates = lasagne.updates.adam(loss, self.params, self.eta)
#updates = lasagne.updates.apply_momentum(updates, self.params, momentum=0.9)
###################################################
#### using the ground truth when training
##################################################
self._train = theano.function(inputs=[ei, em, di, dm, dt],
outputs=[loss, softmax_outputs],
updates=updates,
givens={
encoderInputs: ei,
encoderMask: em,
decoderInputs: di,
decoderMask: dm,
decoderTarget: dt
})
#########################################################################
### For schedule sampling
#########################################################################
###### always use privous predict as next input
def _step2(ctx_, state_, hs_, Cs_):
### ctx_: b x h
### state_ : b x h
### hs_ : 1 x b x h the first dimension is the number of the decoder layers
### Cs_ : 1 x b x h the first dimension is the number of the decoder layers
hs, Cs = [], []
token_idxs = tensor.cast(state_.argmax(axis=-1), "int32")
msk_ = tensor.fill(
(tensor.zeros_like(token_idxs, dtype="float32")), 1)
msk_ = msk_.dimshuffle('x', 0)
state_below0 = self.de_lookuptable[token_idxs].reshape(
(1, ctx_.shape[0], self.de_hidden_size))
for i, lstm in enumerate(self.decoder_lstm_layers):
h, C = lstm.forward(state_below0, msk_, hs_[i],
Cs_[i]) #mind msk
hs += h[-1],
Cs += C[-1],
state_below0 = h
hs, Cs = tensor.as_tensor_variable(hs), tensor.as_tensor_variable(
Cs)
state_below0 = state_below0.reshape(
(ctx_.shape[0], self.de_hidden_size))
state_below0 = tensor.concatenate([ctx_, state_below0], axis=1)
newpred = tensor.dot(state_below0,
self.linear) + self.linear_bias[None, :]
state_below = tensor.nnet.softmax(newpred)
##### the beging symbole probablity is 0
extra_p = tensor.zeros_like(hs[:, :, 0])
state_below = tensor.concatenate([state_below, extra_p.T], axis=1)
return state_below, hs, Cs
ctx_0, state_0 = tensor.fmatrices(2)
hs_0 = tensor.ftensor3()
Cs_0 = tensor.ftensor3()
state_below_tmp, hs_tmp, Cs_tmp = _step2(ctx_0, state_0, hs_0, Cs_0)
self.f_next = theano.function([ctx_0, state_0, hs_0, Cs_0],
[state_below_tmp, hs_tmp, Cs_tmp],
name='f_next')
hs0, Cs0 = tensor.as_tensor_variable(
self.hos, name="hs0"), tensor.as_tensor_variable(self.Cos,
name="Cs0")
train_outputs, _ = theano.scan(fn=_step2,
sequences=[Encoder],
outputs_info=[decoderInputs0, hs0, Cs0],
n_steps=encoderInputs.shape[0])
train_predict = train_outputs[0]
train_costs, _ = theano.scan(
fn=_NLL, sequences=[train_predict, decoderTarget, decoderMask])
train_loss = train_costs.sum() / decoderMask.sum() + params.L2 * sum(
lasagne.regularization.l2(x) for x in self.params)
##from adam import adam
##train_updates = adam(train_loss, self.params, self.eta)
#train_updates = lasagne.updates.apply_momentum(train_updates, self.params, momentum=0.9)
#train_updates = lasagne.updates.sgd(train_loss, self.params, self.eta)
#train_updates = lasagne.updates.apply_momentum(train_updates, self.params, momentum=0.9)
from momentum import momentum
train_updates = momentum(train_loss,
self.params,
params.eta,
momentum=0.9)
self._train2 = theano.function(
inputs=[ei, em, di0, dm, dt],
outputs=[train_loss, train_predict],
updates=train_updates,
givens={
encoderInputs: ei,
encoderMask: em,
decoderInputs0: di0,
decoderMask: dm,
decoderTarget: dt
}
#givens={encoderInputs:ei, encoderMask:em, decoderInputs:di, decoderMask:dm, decoderTarget:dt, TF:tf}
)
listof_token_idx = train_predict.argmax(axis=-1)
self._utter = theano.function(inputs=[ei, em, di0],
outputs=listof_token_idx,
givens={
encoderInputs: ei,
encoderMask: em,
decoderInputs0: di0
})
def find_Ys(Xs_shared, Ys_shared, sigmas_shared, N, steps, output_dims,
n_epochs, initial_lr, final_lr, lr_switch, init_stdev,
initial_momentum, final_momentum, momentum_switch, lmbda, metric,
verbose=0):
"""Optimize cost wrt Ys[t], simultaneously for all t"""
# Optimization hyperparameters
initial_lr = np.array(initial_lr, dtype=floath)
final_lr = np.array(final_lr, dtype=floath)
initial_momentum = np.array(initial_momentum, dtype=floath)
final_momentum = np.array(final_momentum, dtype=floath)
lr = T.fscalar('lr')
lr_shared = theano.shared(initial_lr)
momentum = T.fscalar('momentum')
momentum_shared = theano.shared(initial_momentum)
# Penalty hyperparameter
lmbda_var = T.fscalar('lmbda')
lmbda_shared = theano.shared(np.array(lmbda, dtype=floath))
# Yv velocities
Yvs_shared = []
zero_velocities = np.zeros((N, output_dims), dtype=floath)
for t in range(steps):
Yvs_shared.append(theano.shared(np.array(zero_velocities)))
# Cost
Xvars = T.fmatrices(steps)
Yvars = T.fmatrices(steps)
Yv_vars = T.fmatrices(steps)
sigmas_vars = T.fvectors(steps)
c_vars = []
for t in range(steps):
c_vars.append(cost_var(Xvars[t], Yvars[t], sigmas_vars[t], metric))
cost = T.sum(c_vars) + lmbda_var*movement_penalty(Yvars, N)
# Setting update for Ys velocities
grad_Y = T.grad(cost, Yvars)
givens = {lr: lr_shared, momentum: momentum_shared,
lmbda_var: lmbda_shared}
updates = []
for t in range(steps):
updates.append((Yvs_shared[t], momentum*Yv_vars[t] - lr*grad_Y[t]))
givens[Xvars[t]] = Xs_shared[t]
givens[Yvars[t]] = Ys_shared[t]
givens[Yv_vars[t]] = Yvs_shared[t]
givens[sigmas_vars[t]] = sigmas_shared[t]
update_Yvs = theano.function([], cost, givens=givens, updates=updates)
# Setting update for Ys positions
updates = []
givens = dict()
for t in range(steps):
updates.append((Ys_shared[t], Yvars[t] + Yv_vars[t]))
givens[Yvars[t]] = Ys_shared[t]
givens[Yv_vars[t]] = Yvs_shared[t]
update_Ys = theano.function([], [], givens=givens, updates=updates)
# Momentum-based gradient descent
for epoch in range(n_epochs):
if epoch == lr_switch:
lr_shared.set_value(final_lr)
if epoch == momentum_switch:
momentum_shared.set_value(final_momentum)
c = update_Yvs()
update_Ys()
if verbose:
print('Epoch: {0}. Cost: {1:.6f}.'.format(epoch + 1, float(c)))
Ys = []
for t in range(steps):
Ys.append(np.array(Ys_shared[t].get_value(), dtype=floath))
return Ys
