def _compute_log_moment(self, sigma, q, moment_order):
"""Compute high moment of privacy loss.
Args:
sigma: the noise sigma, in the multiples of the sensitivity.
q: the sampling ratio.
moment_order: the order of moment.
Returns:
log E[exp(moment_order * X)]
"""
assert moment_order <= self._max_moment_order, ("The order of %d is out "
"of the upper bound %d."
% (moment_order,
self._max_moment_order))
binomial_table = tf.slice(self._binomial_table, [moment_order, 0],
[1, moment_order + 1])
# qs = [1 q q^2 ... q^L] = exp([0 1 2 ... L] * log(q))
qs = tf.exp(tf.constant([i * 1.0 for i in range(moment_order + 1)],
dtype=tf.float64) * tf.cast(
tf.log(q), dtype=tf.float64))
moments0 = self._differential_moments(sigma, 0.0, moment_order)
term0 = tf.reduce_sum(binomial_table * qs * moments0)
moments1 = self._differential_moments(sigma, 1.0, moment_order)
term1 = tf.reduce_sum(binomial_table * qs * moments1)
return tf.squeeze(tf.log(tf.cast(q * term0 + (1.0 - q) * term1,
tf.float64)))
def cross_entropy(output, target):
"""Returns the cost function of Cross-entropy of two distributions, implement
softmax internally.
Parameters
----------
output : Tensorflow variable
A distribution with shape: [None, n_feature].
target : Tensorflow variable
A distribution with shape: [None, n_feature].
Examples
--------
>>> ce = tf.cost.cross_entropy(y_logits, y_target_logits)
Notes
-----
About cross-entropy: `wiki <https://en.wikipedia.org/wiki/Cross_entropy>`_.\n
The code is borrowed from: `here <https://en.wikipedia.org/wiki/Cross_entropy>`_.
"""
with tf.name_scope("cross_entropy_loss"):
net_output_tf = output
target_tf = target
cross_entropy = tf.add(tf.mul(tf.log(net_output_tf, name=None),target_tf),
tf.mul(tf.log(1 - net_output_tf), (1 - target_tf)))
return -1 * tf.reduce_mean(tf.reduce_sum(cross_entropy, 1), name='cross_entropy_mean')
def inverse_transform_box(bbox, height, width):
""" Transform the bounding box format
Args:
bbox: [N X 4] input N bbox
format = [left top right bottom]
height: height of original image
width: width of original image
Return:
bbox: [N X 4] output rounded N bbox
fromat = [cx, cy, log(w/W), log(h/H)]
"""
x1, y1, x2, y2 = tf.split(1, 4, bbox)
w = x2 - x1
h = y2 - y1
x = x1 + w / 2
y = y1 + h / 2
x /= width / 2
y /= height / 2
x -= 1
y -= 1
w = tf.log(w / width)
h = tf.log(h / height)
bbox_out = tf.concat(1, [x, y, h, w])
return bbox_out
def _create_loss_optimizer(self):
# The loss is composed of two terms:
# 1.) The reconstruction loss (the negative log probability
# of the input under the reconstructed Bernoulli distribution
# induced by the decoder in the data space).
# This can be interpreted as the number of "nats" required
# for reconstructing the input when the activation in latent
# is given.
# Adding 1e-10 to avoid evaluatio of log(0.0)
reconstr_loss = \
-tf.reduce_sum(self.x * tf.log(1e-10 + self.x_reconstr_mean)
+ (1 - self.x) * tf.log(1e-10 + 1 - self.x_reconstr_mean),
1)
# 2.) The latent loss, which is defined as the Kullback Leibler divergence
# between the distribution in latent space induced by the encoder on
# the data and some prior. This acts as a kind of regularizer.
# This can be interpreted as the number of "nats" required
# for transmitting the the latent space distribution given
# the prior.
latent_loss = -0.5 * tf.reduce_sum(1 + self.z_log_sigma_sq
- tf.square(self.z_mean)
- tf.exp(self.z_log_sigma_sq), 1)
self.cost = tf.reduce_mean(reconstr_loss + latent_loss) # average over batch
# Use ADAM optimizer
self.optimizer = \
tf.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(self.cost)
tf.multiply(tf.tile(query_rnn_output, [NEG + 1, 1]), doc_y), 1, True)
norm_prod = tf.multiply(query_norm, doc_norm)
# cos_sim_raw = query * doc / (||query|| * ||doc||)
cos_sim_raw = tf.truediv(prod, norm_prod)
# gamma = 20
cos_sim = tf.transpose(
tf.reshape(tf.transpose(cos_sim_raw), [NEG + 1, query_BS])) * 20
with tf.name_scope('Loss'):
# Train Loss
# 转化为softmax概率矩阵。
prob = tf.nn.softmax(cos_sim)
# 只取第一列，即正样本列概率。
hit_prob = tf.slice(prob, [0, 0], [-1, 1])
loss = -tf.reduce_sum(tf.log(hit_prob))
tf.summary.scalar('loss', loss)
with tf.name_scope('Training'):
# Optimizer
train_step = tf.train.AdamOptimizer(conf.learning_rate).minimize(loss)
# with tf.name_scope('Accuracy'):
# correct_prediction = tf.equal(tf.argmax(prob, 1), 0)
# accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# tf.summary.scalar('accuracy', accuracy)
merged = tf.summary.merge_all()
with tf.name_scope('Test'):
average_loss = tf.placeholder(tf.float32)
def train_neural_networks():
decoded, predict_output = neural_networks()
us_cost_function = tf.reduce_mean(tf.pow(X - decoded, 2))
s_cost_function = -tf.reduce_sum(Y * tf.log(predict_output))
us_optimizer = tf.train.GradientDescentOptimizer(
learning_rate=0.01).minimize(us_cost_function)
s_optimizer = tf.train.GradientDescentOptimizer(
learning_rate=0.01).minimize(s_cost_function)
correct_prediction = tf.equal(tf.argmax(predict_output, 1),
tf.argmax(Y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
training_epochs = 20
batch_size = 10
total_batches = training_data.shape[0]
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
# ------------ Training Autoencoders - Unsupervised Learning ----------- #
# autoencoder是一种非监督学习算法，他利用反向传播算法，让目标值等于输入值
for epoch in range(training_epochs):
epoch_costs = np.empty(0)
for b in range(total_batches):
offset = (b * batch_size) % (train_x.shape[0] - batch_size)
batch_x = train_x[offset:(offset + batch_size), :]
_, c = sess.run([us_optimizer, us_cost_function],
feed_dict={X: batch_x})
epoch_costs = np.append(epoch_costs, c)
print("Epoch: ", epoch, " Loss: ", np.mean(epoch_costs))
print(
"------------------------------------------------------------------"
)
# ---------------- Training NN - Supervised Learning ------------------ #
for epoch in range(training_epochs):
epoch_costs = np.empty(0)
for b in range(total_batches):
offset = (b * batch_size) % (train_x.shape[0] - batch_size)
batch_x = train_x[offset:(offset + batch_size), :]
batch_y = train_y[offset:(offset + batch_size), :]
_, c = sess.run([s_optimizer, s_cost_function],
feed_dict={
X: batch_x,
Y: batch_y
})
epoch_costs = np.append(epoch_costs, c)
accuracy_in_train_set = sess.run(accuracy,
feed_dict={
X: train_x,
Y: train_y
})
accuracy_in_test_set = sess.run(accuracy,
feed_dict={
X: test_x,
Y: test_y
})
print("Epoch: ", epoch, " Loss: ", np.mean(epoch_costs),
" Accuracy: ", accuracy_in_train_set, ' ',
accuracy_in_test_set)
def createModel():
# Create the model
x = tf.placeholder(tf.float32, [None, 784])
y_ = tf.placeholder(tf.float32, [None, 10])
W = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
y = tf.nn.softmax(tf.matmul(x, W) + b)
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])
x_image = tf.reshape(x, [-1,28,28,1])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)
W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)
W_fc1 = weight_variable([7 * 7 * 64, 1024])
b_fc1 = bias_variable([1024])
h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])
y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
# Define loss and optimizer
cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
"""
Train the model and save the model to disk as a model.ckpt file
file is stored in the same directory as this python script is started
Based on the documentatoin at
https://www.tensorflow.org/versions/master/how_tos/variables/index.html
"""
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
for i in range(20000):
batch = mnist.train.next_batch(50)
if i%100 == 0:
train_accuracy = accuracy.eval(feed_dict={
x:batch[0], y_: batch[1], keep_prob: 1.0})
print("step %d, training accuracy %g"%(i, train_accuracy))
train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})
save_path = saver.save(sess, os.path.join(FLAGS.log_dir, 'model.ckpt'))
print ("Model saved in file: ", save_path)
def _inc_loss_psi(self, psi, signal, t):
return - tf.log(1. + self._expectation(psi, t) * signal / self.A)
#将2D图片结构转换为1D向量结构，并连接一个全连接层
h_pool2_flat=tf.reshape(h_pool2,[-1,7*7*64]) #对第二个卷积层的输出tensor进行变形，转换为1D向量
#全连接层
W_fc1=weight_variable([7*7*64,1024])
b_fc1=bias_variable([1024])
h_fc1=tf.nn.relu(tf.matmul(h_pool2_flat,W_fc1)+b_fc1) #隐含节点1024，使用relu激活函数
#dropout层，减轻过拟合
keep_prob=tf.placeholder(tf.float32)
h_fc1_drop=tf.nn.dropout(h_fc1,keep_prob)
#softmax层
W_fc2=weight_variable([1024,10])
b_fc2=bias_variable([10])
y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop,W_fc2)+b_fc2)
#定义损失函数和优化器
cross_entropy=tf.reduce_mean(-tf.reduce_sum(y_*tf.log(y_conv),reduction_indices=[1]))
train_step=tf.train.AdamOptimizer(1e-4).minimize(cross_entropy) #给予一个较小的学习速率1e-4
#定义评测准确率
correct_prediction=tf.equal(tf.argmax(y_conv,1),tf.argmax(y_,1))
accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
#训练
tf.global_variables_initializer().run() #初始化所有参数
for i in range(20000):
batch=mnist.train.next_batch(50)
if i%100==0: #训练中评测时的keep_prob设为1，实时监测模型的性能
train_accuracy=accuracy.eval(feed_dict={x:batch[0],y_:batch[1],keep_prob:1.0})
print("step %d,training accuracy %g"%(i,train_accuracy))
train_step.run(feed_dict={x:batch[0],y_:batch[1],keep_prob:0.5})
def log(x):
return tf.log(x)
division = - tf.divide( pairwise_sq, tf.constant(2*sigma**2, dtype=tf.float32))
match_ab = tf.exp(division, name='match_ab')
p_ab = tf.nn.softmax(match_ab, name='p_ab')
p_ba = tf.nn.softmax(tf.transpose(match_ab), name='p_ba')
p_aba = tf.matmul(p_ab, p_ba, name='p_aba')
model.create_walk_statistics(p_aba, equality_matrix)
loss_aba = tf.losses.softmax_cross_entropy(
p_target,
tf.log(1e-8 + p_aba),
weights=walker_weight,
scope='loss_aba')
model.add_visit_loss(p_ab, visit_weight)
mab_dt, pab_dt, paba_dt, semb_dt, uemb_dt = tf.gradients([loss_aba], [match_ab, p_ab, p_aba, t_sup_emb, t_unsup_emb])
tf.summary.scalar('Loss_aba', loss_aba)
t_learning_rate = tf.train.exponential_decay(
learning_rate,
model.step,
decay_steps,
decay_factor,
staircase = True
)
w9 = tf.get_variable("w9", shape=[4*4*256,256],initializer=tf.contrib.layers.xavier_initializer())
b9 = tf.Variable(tf.random_normal([256]), tf.float32)
L9 = tf.nn.elu(tf.matmul(L8,w9) + b9)
L9 = tf.nn.dropout(L9, keep_prob=keep_prob)
w10 = tf.get_variable("w10", shape=[256,32],initializer=tf.contrib.layers.xavier_initializer())
b10 = tf.Variable(tf.random_normal([32]), tf.float32)
L10 = tf.nn.elu(tf.matmul(L9,w10) + b10)
L10 = tf.nn.dropout(L10, keep_prob=keep_prob)
w11 = tf.get_variable("w11", shape=[32,10],initializer=tf.contrib.layers.xavier_initializer())
b11 = tf.Variable(tf.random_normal([10]), tf.float32)
hypothesis = tf.nn.softmax(tf.matmul(L10,w11) + b11)
cost = tf.reduce_mean(-tf.reduce_sum(y * tf.log(hypothesis), axis = 1)) ## categorical_crossentropy
train = tf.train.AdamOptimizer(learning_rate=lr).minimize(cost)
predicted = tf.cast(tf.argmax(hypothesis, axis=1), dtype = tf.float32)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for step in range(training_epochs):
avg_cost = 0
for batch in range(total_batch):
batch_xs, batch_ys = x_train[batch*batch_size : (batch+1)*batch_size], y_train[batch*batch_size : (batch+1)*batch_size]
feed_dict = {x_imag:batch_xs, y:batch_ys, keep_prob:0.8}
cost_val, _ = sess.run([cost, train], feed_dict=feed_dict)
print(step, 'cost :', cost_val)#, '\n',hyp_val)
avg_cost += cost_val/total_batch
def _get_loss(self, eps):
return -tf.reduce_mean(
self._tfy * tf.log(self._output + eps) + (1 - self._tfy) * tf.log(1 - self._output + eps)
)
def build(self):
hparam = self.hparam
code = tf.placeholder("float32", [None, hparam.code_dim], name='code')
real_seq = tf.placeholder("int32", [None, hparam.timesteps],
name='real_seq')
real_seq_img = tf.one_hot(real_seq, hparam.vocab_size)
dis_train = tf.placeholder('bool', name='is_train')
bs = tf.shape(code)[0]
# generator
with tf.variable_scope('generator'):
cells = rnn.MultiRNNCell(
[hparam.basic_cell(c) for c in hparam.cells])
# play with code
# code_with_timestep = tf.concat(
# [
# tf.tile(tf.expand_dims(code, 1),
# (1, hparam.timesteps, 1)),
# tf.tile(tf.expand_dims(tf.eye(hparam.timesteps), 0),
# (bs, 1, 1)),
# ], -1
# )
code_with_timestep = tf.concat([
tf.tile(tf.expand_dims(code, 1), (1, hparam.timesteps, 1)),
tf.tile(
tf.expand_dims(
tf.expand_dims(tf.lin_space(0., 1., hparam.timesteps),
0), -1), (bs, 1, 1)),
], -1)
outputs, states = tf.nn.dynamic_rnn(
cells,
code_with_timestep,
dtype=tf.float32,
)
fake_seq_img = tf.nn.softmax(
layers.linear(outputs[:, :, :], hparam.vocab_size), -1)
# fake_seq_img = outputs
# fake_seq_img = tf.nn.softmax(fake_seq_img)
fake_seq = tf.argmax(fake_seq_img, -1)
# discriminator
def dis(seq_img, bn_scope, reuse=False):
with tf.variable_scope('discriminator', reuse=reuse):
# x = tf.nn.embedding_lookup(embeddings, seq)
x = tf.expand_dims(seq_img, -1)
x = ResNetBuilder(dis_train,
bn_scopes=['fake', 'real'],
bn_scope=bn_scope).\
resnet(x, structure=[2, 2, 2, 2], filters=4, nb_class=1)
x = tf.nn.sigmoid(x)
return x
# opt
# problematic with the reuse bn
fake_dis_pred = dis(fake_seq_img, bn_scope='fake')
real_dis_pred = dis(real_seq_img, bn_scope='real', reuse=True)
G_loss = tf.reduce_mean(-tf.log(fake_dis_pred))
D_loss = tf.reduce_mean(-tf.log(real_dis_pred)) +\
tf.reduce_mean(-tf.log(1-fake_dis_pred))
G_opt = tf.train.AdamOptimizer(learning_rate=hparam.G_lr,
beta1=0.5,
beta2=0.9)
D_opt = tf.train.AdamOptimizer(learning_rate=hparam.D_lr,
beta1=0.5,
beta2=0.9)
D_iter = tf.Variable(0, name='D_iter')
G_iter = tf.Variable(0, name='G_iter')
G_train_op = slim.learning.create_train_op(
G_loss,
G_opt,
variables_to_train=tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, "generator"),
global_step=G_iter,
clip_gradient_norm=hparam.G_clipnorm)
D_train_op = slim.learning.create_train_op(
D_loss,
D_opt,
variables_to_train=tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, "discriminator"),
global_step=D_iter,
)
iter_step = tf.Variable(0, name='iter_step')
iter_step_op = iter_step.assign_add(1)
# input
self.code = code
self.real_seq = real_seq
# summary
self.summary_fake_img = tf.summary.image(
'fake_img', tf.expand_dims(fake_seq_img, -1))
self.summary_real_img = tf.summary.image(
'real_img', tf.expand_dims(real_seq_img, -1))
self.summary_G_loss = tf.summary.scalar('G_loss', G_loss)
self.summary_D_loss = tf.summary.scalar('D_loss', D_loss)
self.summary_fake_dis_pred = tf.summary.scalar(
'fake_dis_pred', tf.reduce_mean(fake_dis_pred))
self.summary_real_dis_pred = tf.summary.scalar(
'real_dis_pred', tf.reduce_mean(real_dis_pred))
# debug
self.fake_seq_img = fake_seq_img
# train
self.dis_train = dis_train
self.G_train_op = G_train_op
self.D_train_op = D_train_op
self.iter_step = iter_step
self.iter_step_op = iter_step_op
# output
self.fake_seq = fake_seq
self.built = True
def log2(x):
with tf.name_scope('Log2'):
return tf.log(x) * np.float32(1.0 / np.log(2.0))
def __init__(self, dropout=1.0):
tf.reset_default_graph()
with tf.variable_scope(NAMESPACE):
config = tf.ConfigProto(allow_soft_placement=True)
self.sess = tf.Session(config=config)
# Input variables
self.item = tf.placeholder(tf.float32, shape=(None, self._nodes_vocab_size), name='item')
self.question_vectors_fw = tf.placeholder(tf.float32, shape=(None, None, self._question_vocab_size),
name='question_vectors_inp_fw')
self.question_vectors_bw = tf.placeholder(tf.float32, shape=(None, None, self._question_vocab_size),
name='question_vectors_inp_nw')
self.question_mask = tf.placeholder(tf.float32, shape=(None, None, self._mask_size),
name='question_mask')
# The question is pre-processed by a bi-GRU
self.Wq = tf.Variable(tf.random_uniform([self._question_vocab_size,
self._word_proj_size_for_rnn], -_rw, _rw))
self.bq = tf.Variable(tf.random_uniform([self._word_proj_size_for_rnn], -_rw, _rw))
self.internal_projection = lambda x: tf.nn.relu(tf.matmul(x, self.Wq) + self.bq)
self.question_int_fw = tf.map_fn(self.internal_projection, self.question_vectors_fw)
self.question_int_bw = tf.map_fn(self.internal_projection, self.question_vectors_bw)
self.rnn_cell_fw = rnn.MultiRNNCell([rnn.GRUCell(self._memory_dim) for _ in range(self._stack_dimension)],
state_is_tuple=True)
self.rnn_cell_bw = rnn.MultiRNNCell([rnn.GRUCell(self._memory_dim) for _ in range(self._stack_dimension)],
state_is_tuple=True)
with tf.variable_scope('fw'):
output_fw, state_fw = tf.nn.dynamic_rnn(self.rnn_cell_fw, self.question_int_fw, time_major=True,
dtype=tf.float32)
with tf.variable_scope('bw'):
output_bw, state_bw = tf.nn.dynamic_rnn(self.rnn_cell_bw, self.question_int_bw, time_major=True,
dtype=tf.float32)
self.states = tf.concat(values=[output_fw, tf.reverse(output_bw, [0])], axis=2)
self.question_vector_pre = tf.reduce_mean(tf.multiply(self.question_mask, self.states), axis=0)
self.Wqa = tf.Variable(
tf.random_uniform([2 * self._memory_dim, self._question_vector_size], -_rw, _rw),
name='Wqa')
self.bqa = tf.Variable(tf.random_uniform([self._question_vector_size], -_rw, _rw), name='bqa')
self.question_vector = tf.nn.relu(tf.matmul(self.question_vector_pre, self.Wqa) + self.bqa)
# Item
self.Wit = tf.Variable(tf.random_uniform([self._nodes_vocab_size,
self._word_proj_size_for_item], -_rw, _rw))
self.bit = tf.Variable(tf.random_uniform([self._word_proj_size_for_item], -_rw, _rw))
self.item_proj = tf.nn.relu(tf.matmul(self.item, self.Wit) + self.bit)
# Concatenate
self.concatenated = tf.concat(values=[self.question_vector, self.item_proj], axis=1)
# Final feedforward layers
self.Ws1 = tf.Variable(
tf.random_uniform([self._question_vector_size
+ self._word_proj_size_for_item,
self._hidden_layer2_size], -_rw, _rw),
name='Ws1')
self.bs1 = tf.Variable(tf.random_uniform([self._hidden_layer2_size], -_rw, _rw), name='bs1')
self.first_hidden = tf.nn.relu(tf.matmul(self.concatenated, self.Ws1) + self.bs1)
self.first_hidden_dropout = tf.nn.dropout(self.first_hidden, dropout)
self.Wf = tf.Variable(
tf.random_uniform([self._hidden_layer2_size, self._output_size], -_rw,
_rw),
name='Wf')
self.bf = tf.Variable(tf.random_uniform([self._output_size], -_rw, _rw), name='bf')
self.outputs = tf.nn.softmax(tf.matmul(self.first_hidden_dropout, self.Wf) + self.bf)
# Loss function and training
self.y_ = tf.placeholder(tf.float32, shape=(None, self._output_size), name='y_')
self.outputs2 = tf.squeeze(self.outputs)
self.y2_ = tf.squeeze(self.y_)
self.one = tf.ones_like(self.outputs)
self.tiny = self.one * TINY
self.cross_entropy = (tf.reduce_mean(
-tf.reduce_sum(self.y_ * tf.log(self.outputs + self.tiny) * _weight_for_positive_matches
+ (self.one - self.y_) * tf.log(
self.one - self.outputs + self.tiny))
))
# Clipping the gradient
optimizer = tf.train.AdamOptimizer(1e-4)
gvs = optimizer.compute_gradients(self.cross_entropy)
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs if var.name.find(NAMESPACE) != -1]
self.train_step = optimizer.apply_gradients(capped_gvs)
self.sess.run(tf.global_variables_initializer())
# Adding the summaries
tf.summary.scalar('cross_entropy', self.cross_entropy)
self.merged = tf.summary.merge_all()
self.train_writer = tf.summary.FileWriter('./train', self.sess.graph)
# 乘出来之后每一列就是对应的一组weight值。
# shape of tf.matmul(input, Weights): [batch_size, layer_size]
# shape of Biases: [layer_size]
output = tf.matmul(input, Weights) + Biases
# 用sigmoid不行，应该是在BP更新参数的时候会非常受影响。
output = tf.nn.softmax(output)
# [batch_size, layer_size]
return output
x = tf.placeholder(shape=[None, 784], dtype=tf.float32)
y = tf.placeholder(shape=[None, 10], dtype=tf.float32)
prediction = hidden_layer(x)
cross_entropy = tf.reduce_mean(
-tf.reduce_sum(y * tf.log(prediction), reduction_indices=[1]))
optimizer = tf.train.GradientDescentOptimizer(0.5)
train = optimizer.minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(prediction, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
init = tf.initialize_all_variables()
with tf.Session() as sess:
sess.run(init)
for i in range(epoch):
batch_xs, batch_ys = mnist.train.next_batch(100)
# 以下这个形式是错的，因为每次next batch 会随机产生不同的数据
# batch_xs = mnist.train.next_batch(batch_size)[0]
# batch_ys = mnist.train.next_batch(batch_size)[1]
b_enc = tf.Variable(tf.zeros([625]))
b_dec = tf.Variable(tf.zeros([784]))
# Create the model
def model(X, w_e, b_e, w_d, b_d):
encoded = tf.sigmoid(tf.matmul(X, w_e) + b_e)
decoded = tf.sigmoid(tf.matmul(encoded, w_d) + b_d)
return encoded, decoded
encoded, decoded = model(x, w_enc, b_enc, w_dec, b_dec)
# Cost Function basic term
cross_entropy = -1. * x * tf.log(decoded) - (1. - x) * tf.log(1. - decoded)
loss = tf.reduce_mean(cross_entropy)
train_step = tf.train.AdagradOptimizer(0.1).minimize(loss)
# Train
init = tf.initialize_all_variables()
#Create a summary to monitor cost tensor
summa = tf.summary.scalar("loss", loss)
with tf.Session() as sess:
sess.run(init)
summary_writer = tf.summary.FileWriter('./graphs/simsc',
graph=tf.get_default_graph())
print('Training...')
for i in range(1001):
batch_xs, batch_ys = mnist.train.next_batch(128)
def setup_model(self):
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the ACER model must be " \
"an instance of common.policies.ActorCriticPolicy."
if isinstance(self.action_space, Discrete):
self.n_act = self.action_space.n
continuous = False
elif isinstance(self.action_space, Box):
# self.n_act = self.action_space.shape[-1]
# continuous = True
raise NotImplementedError(
"WIP: Acer does not support Continuous actions yet.")
else:
raise ValueError(
"Error: ACER does not work with {} actions space.".format(
self.action_space))
self.n_batch = self.n_envs * self.n_steps
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess,
graph=self.graph)
self.set_random_seed(self.seed)
n_batch_step = None
if issubclass(self.policy, RecurrentActorCriticPolicy):
n_batch_step = self.n_envs
n_batch_train = self.n_envs * (self.n_steps + 1)
step_model = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
n_batch_step,
reuse=False,
**self.policy_kwargs)
self.params = tf_util.get_trainable_vars("model")
with tf.variable_scope(
"train_model",
reuse=True,
custom_getter=tf_util.outer_scope_getter(
"train_model")):
train_model = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
self.n_steps + 1,
n_batch_train,
reuse=True,
**self.policy_kwargs)
with tf.variable_scope("moving_average"):
# create averaged model
ema = tf.train.ExponentialMovingAverage(self.alpha)
ema_apply_op = ema.apply(self.params)
def custom_getter(getter, name, *args, **kwargs):
name = name.replace("polyak_model/", "")
val = ema.average(getter(name, *args, **kwargs))
return val
with tf.variable_scope("polyak_model",
reuse=True,
custom_getter=custom_getter):
self.polyak_model = polyak_model = self.policy(
self.sess,
self.observation_space,
self.action_space,
self.n_envs,
self.n_steps + 1,
self.n_envs * (self.n_steps + 1),
reuse=True,
**self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
self.done_ph = tf.placeholder(tf.float32,
[self.n_batch]) # dones
self.reward_ph = tf.placeholder(
tf.float32, [self.n_batch]) # rewards, not returns
self.mu_ph = tf.placeholder(
tf.float32, [self.n_batch, self.n_act]) # mu's
self.action_ph = train_model.pdtype.sample_placeholder(
[self.n_batch])
self.learning_rate_ph = tf.placeholder(tf.float32, [])
eps = 1e-6
# Notation: (var) = batch variable, (var)s = sequence variable,
# (var)_i = variable index by action at step i
# shape is [n_envs * (n_steps + 1)]
if continuous:
value = train_model.value_flat
else:
value = tf.reduce_sum(train_model.policy_proba *
train_model.q_value,
axis=-1)
rho, rho_i_ = None, None
if continuous:
action_ = strip(
train_model.proba_distribution.sample(),
self.n_envs, self.n_steps)
distribution_f = tf.contrib.distributions.MultivariateNormalDiag(
loc=strip(train_model.proba_distribution.mean,
self.n_envs, self.n_steps),
scale_diag=strip(
train_model.proba_distribution.logstd,
self.n_envs, self.n_steps))
f_polyak = tf.contrib.distributions.MultivariateNormalDiag(
loc=strip(polyak_model.proba_distribution.mean,
self.n_envs, self.n_steps),
scale_diag=strip(
polyak_model.proba_distribution.logstd,
self.n_envs, self.n_steps))
f_i = distribution_f.prob(self.action_ph)
f_i_ = distribution_f.prob(action_)
f_polyak_i = f_polyak.prob(self.action_ph)
phi_i = strip(train_model.proba_distribution.mean,
self.n_envs, self.n_steps)
q_value = strip(train_model.value_fn, self.n_envs,
self.n_steps)
q_i = q_value[:, 0]
rho_i = tf.reshape(f_i, [-1, 1]) / (self.mu_ph + eps)
rho_i_ = tf.reshape(f_i_, [-1, 1]) / (self.mu_ph + eps)
qret = q_retrace(self.reward_ph, self.done_ph, q_i,
value, tf.pow(rho_i, 1 / self.n_act),
self.n_envs, self.n_steps, self.gamma)
else:
# strip off last step
# f is a distribution, chosen to be Gaussian distributions
# with fixed diagonal covariance and mean \phi(x)
# in the paper
distribution_f, f_polyak, q_value = \
map(lambda variables: strip(variables, self.n_envs, self.n_steps),
[train_model.policy_proba, polyak_model.policy_proba, train_model.q_value])
# Get pi and q values for actions taken
f_i = get_by_index(distribution_f, self.action_ph)
f_i_ = distribution_f
phi_i = distribution_f
f_polyak_i = f_polyak
q_i = get_by_index(q_value, self.action_ph)
# Compute ratios for importance truncation
rho = distribution_f / (self.mu_ph + eps)
rho_i = get_by_index(rho, self.action_ph)
# Calculate Q_retrace targets
qret = q_retrace(self.reward_ph, self.done_ph, q_i,
value, rho_i, self.n_envs,
self.n_steps, self.gamma)
# Calculate losses
# Entropy
entropy = tf.reduce_sum(
train_model.proba_distribution.entropy())
# Policy Gradient loss, with truncated importance sampling & bias correction
value = strip(value, self.n_envs, self.n_steps, True)
# check_shape([qret, value, rho_i, f_i], [[self.n_envs * self.n_steps]] * 4)
# check_shape([rho, distribution_f, q_value], [[self.n_envs * self.n_steps, self.n_act]] * 2)
# Truncated importance sampling
adv = qret - value
log_f = tf.log(f_i + eps)
# [n_envs * n_steps]
gain_f = log_f * tf.stop_gradient(
adv * tf.minimum(self.correction_term, rho_i))
loss_f = -tf.reduce_mean(gain_f)
# Bias correction for the truncation
adv_bc = (
q_value -
tf.reshape(value, [self.n_envs * self.n_steps, 1])
) # [n_envs * n_steps, n_act]
# check_shape([adv_bc, log_f_bc], [[self.n_envs * self.n_steps, self.n_act]] * 2)
if continuous:
gain_bc = tf.stop_gradient(
adv_bc * tf.nn.relu(1.0 - (self.correction_term /
(rho_i_ + eps))) * f_i_)
else:
log_f_bc = tf.log(f_i_ + eps) # / (f_old + eps)
gain_bc = tf.reduce_sum(log_f_bc * tf.stop_gradient(
adv_bc * tf.nn.relu(1.0 - (self.correction_term /
(rho + eps))) * f_i_),
axis=1)
# IMP: This is sum, as expectation wrt f
loss_bc = -tf.reduce_mean(gain_bc)
loss_policy = loss_f + loss_bc
# Value/Q function loss, and explained variance
check_shape([qret, q_i],
[[self.n_envs * self.n_steps]] * 2)
explained_variance = q_explained_variance(
tf.reshape(q_i, [self.n_envs, self.n_steps]),
tf.reshape(qret, [self.n_envs, self.n_steps]))
loss_q = tf.reduce_mean(
tf.square(tf.stop_gradient(qret) - q_i) * 0.5)
# Net loss
check_shape([loss_policy, loss_q, entropy], [[]] * 3)
loss = loss_policy + self.q_coef * loss_q - self.ent_coef * entropy
tf.summary.scalar('entropy_loss', entropy)
tf.summary.scalar('policy_gradient_loss', loss_policy)
tf.summary.scalar('value_function_loss', loss_q)
tf.summary.scalar('loss', loss)
norm_grads_q, norm_grads_policy, avg_norm_grads_f = None, None, None
avg_norm_k, avg_norm_g, avg_norm_k_dot_g, avg_norm_adj = None, None, None, None
if self.trust_region:
# [n_envs * n_steps, n_act]
grad = tf.gradients(
-(loss_policy - self.ent_coef * entropy) *
self.n_steps * self.n_envs, phi_i)
# [n_envs * n_steps, n_act] # Directly computed gradient of KL divergence wrt f
kl_grad = -f_polyak_i / (f_i_ + eps)
k_dot_g = tf.reduce_sum(kl_grad * grad, axis=-1)
adj = tf.maximum(
0.0, (tf.reduce_sum(kl_grad * grad, axis=-1) -
self.delta) /
(tf.reduce_sum(tf.square(kl_grad), axis=-1) +
eps)) # [n_envs * n_steps]
# Calculate stats (before doing adjustment) for logging.
avg_norm_k = avg_norm(kl_grad)
avg_norm_g = avg_norm(grad)
avg_norm_k_dot_g = tf.reduce_mean(tf.abs(k_dot_g))
avg_norm_adj = tf.reduce_mean(tf.abs(adj))
grad = grad - tf.reshape(
adj, [self.n_envs * self.n_steps, 1]) * kl_grad
# These are turst region adjusted gradients wrt f ie statistics of policy pi
grads_f = -grad / (self.n_envs * self.n_steps)
grads_policy = tf.gradients(f_i_, self.params, grads_f)
grads_q = tf.gradients(loss_q * self.q_coef,
self.params)
grads = [
gradient_add(g1, g2, param, verbose=self.verbose)
for (g1, g2, param
) in zip(grads_policy, grads_q, self.params)
]
avg_norm_grads_f = avg_norm(grads_f) * (self.n_steps *
self.n_envs)
norm_grads_q = tf.global_norm(grads_q)
norm_grads_policy = tf.global_norm(grads_policy)
else:
grads = tf.gradients(loss, self.params)
norm_grads = None
if self.max_grad_norm is not None:
grads, norm_grads = tf.clip_by_global_norm(
grads, self.max_grad_norm)
grads = list(zip(grads, self.params))
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('rewards',
tf.reduce_mean(self.reward_ph))
tf.summary.scalar('learning_rate',
tf.reduce_mean(self.learning_rate))
tf.summary.scalar('advantage', tf.reduce_mean(adv))
tf.summary.scalar('action_probability',
tf.reduce_mean(self.mu_ph))
if self.full_tensorboard_log:
tf.summary.histogram('rewards', self.reward_ph)
tf.summary.histogram('learning_rate',
self.learning_rate)
tf.summary.histogram('advantage', adv)
tf.summary.histogram('action_probability', self.mu_ph)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', train_model.obs_ph)
else:
tf.summary.histogram('observation',
train_model.obs_ph)
trainer = tf.train.RMSPropOptimizer(
learning_rate=self.learning_rate_ph,
decay=self.rprop_alpha,
epsilon=self.rprop_epsilon)
_opt_op = trainer.apply_gradients(grads)
# so when you call _train, you first do the gradient step, then you apply ema
with tf.control_dependencies([_opt_op]):
_train = tf.group(ema_apply_op)
# Ops/Summaries to run, and their names for logging
assert norm_grads is not None
run_ops = [
_train, loss, loss_q, entropy, loss_policy, loss_f,
loss_bc, explained_variance, norm_grads
]
names_ops = [
'loss', 'loss_q', 'entropy', 'loss_policy', 'loss_f',
'loss_bc', 'explained_variance', 'norm_grads'
]
if self.trust_region:
self.run_ops = run_ops + [
norm_grads_q, norm_grads_policy, avg_norm_grads_f,
avg_norm_k, avg_norm_g, avg_norm_k_dot_g, avg_norm_adj
]
self.names_ops = names_ops + [
'norm_grads_q', 'norm_grads_policy',
'avg_norm_grads_f', 'avg_norm_k', 'avg_norm_g',
'avg_norm_k_dot_g', 'avg_norm_adj'
]
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.proba_step = step_model.proba_step
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=self.sess)
self.summary = tf.summary.merge_all()
def _MixN(fractions, Xs, name=None):
# Convert Xs to be iterable incase we are dealing with the 1D case
Xs = [(x, ) if isinstance(x, tf.Tensor) else tuple(x) for x in Xs]
# Ensure all subdistributions have the same dimensionality
if len(set(len(x) for x in Xs)) != 1:
raise DistributionError("All components passed to 'MixN' must have "
"the same dimensionality.")
n_dims = len(Xs[0])
nd_X = tuple(
tf.placeholder(config.dtype, name=name) for i in range(n_dims))
nd_mix_bounds = Distribution.bounds(n_dims)
current_model = Model._current_model
full_pdf = []
all_integrals = []
for dists, f_scale in zip(Xs, fractions):
nd_logps = []
nd_bounds = []
nd_integrals = []
nd_normalisation_1 = []
for dist, mix_bounds, X in zip(dists, nd_mix_bounds, nd_X):
logp, integral, bounds, frac, _ = current_model._description[dist]
bounds = find_common_bounds(mix_bounds, bounds)
normalisation_1 = _integrate_component(bounds, integral)
nd_logps.append(logp - tf.log(normalisation_1))
nd_bounds.append(bounds)
nd_integrals.append(integral)
nd_normalisation_1.append(normalisation_1)
# Modify the current model to recognize that 'deps' has been removed
if dist in current_model._silently_replace.values():
# We need to copy the items to a list as we're adding items
for key, value in list(
current_model._silently_replace.items()):
if value != dist:
continue
current_model._silently_replace[value] = X
current_model._silently_replace[key] = X
if dist in current_model._description:
del current_model._description[dist]
else:
current_model._silently_replace[dist] = X
del current_model._description[dist]
_recurse_deps(dist, f_scale, bounds)
full_pdf.append(f_scale * tf.exp(tf.add_n(nd_logps)))
all_integrals.append(
(f_scale, nd_bounds, nd_integrals, nd_normalisation_1))
# Set properties on Distribution
Distribution.logp = tf.log(tf.add_n(full_pdf))
def _integral(lower, upper):
result = []
for f_scale, nd_bounds, nd_integral, normalisation_1 in all_integrals:
nd_normalisation_2 = []
for bounds, integral in zip(nd_bounds, nd_integral):
integral_bounds = find_common_bounds([Region(lower, upper)],
bounds)
nd_normalisation_2.append(
_integrate_component(integral_bounds, integral))
result.append(f_scale / tf.add_n(normalisation_1) *
tf.add_n(nd_normalisation_2))
return tf.add_n(result)
Distribution.integral = _integral
if n_dims == 1:
Distribution.depends = [x[0] for x in Xs]
else:
Distribution.depends = Xs
return nd_X
def build_graph(self):
para_embeddings_op = self.elmo_bilm(self.para)
que_embeddings_op = self.elmo_bilm(self.que)
with tf.variable_scope('elmo_encodings_input'):
elmo_para_input = weight_layers('input', para_embeddings_op, l2_coef=0.)['weighted_op']
with tf.variable_scope('elmo_encodings_input', reuse=True):
elmo_que_input = weight_layers('input', que_embeddings_op, l2_coef=0.)['weighted_op']
with tf.variable_scope('elmo_encodings_output'):
elmo_para_output = weight_layers('output', para_embeddings_op, l2_coef=0.)['weighted_op']
with tf.variable_scope('elmo_encodings_output', reuse=True):
elmo_que_output = weight_layers('output', que_embeddings_op, l2_coef=0.)['weighted_op']
with tf.variable_scope("embedding"):
with tf.device("/cpu:0"):
ch_emb = tf.reshape(tf.nn.embedding_lookup(
self.char_mat, self.para_char), [self.N * self.PL, self.CL, self.dc])
qh_emb = tf.reshape(tf.nn.embedding_lookup(
self.char_mat, self.que_char), [self.N * self.QL, self.CL, self.dc])
# Bidaf style conv-highway encoder
ch_emb = conv(ch_emb, self.dc, bias=True, activation=tf.nn.relu,
kernel_size=5, name="char_conv", reuse=None)
qh_emb = conv(qh_emb, self.dc, bias=True, activation=tf.nn.relu,
kernel_size=5, name="char_conv", reuse=True)
ch_emb = tf.reduce_max(ch_emb, axis=1)
qh_emb = tf.reduce_max(qh_emb, axis=1)
ch_emb = tf.reshape(ch_emb, [self.N, self.PL, ch_emb.shape[-1]])
qh_emb = tf.reshape(qh_emb, [self.N, self.QL, qh_emb.shape[-1]])
with tf.device("/cpu:0"):
c_emb = tf.nn.dropout(tf.nn.embedding_lookup(self.word_mat, self.para1), 1.0 - self.dropout)
q_emb = tf.nn.dropout(tf.nn.embedding_lookup(self.word_mat, self.que1), 1.0 - self.dropout)
para_emb = tf.concat([c_emb, ch_emb], axis=2)
que_emb = tf.concat([q_emb, qh_emb], axis=2)
# add elmo
embedded_para = tf.concat([para_emb, elmo_para_input], -1)
embedded_que = tf.concat([que_emb, elmo_que_input], -1)
print("word embedding:", embedded_para.get_shape().as_list(), embedded_que.get_shape().as_list())
# input encoder
with tf.variable_scope("para_encoder"):
para_rep, _ = bi_cudnn_rnn_encoder('gru', self.d, 1, self.dropout,
embedded_para, self.para_len, self.trainable)
with tf.variable_scope("para_encoder", reuse=True):
que_rep, _ = bi_cudnn_rnn_encoder('gru', self.d, 1, self.dropout,
embedded_que, self.que_len, self.trainable)
print("input encoder:", para_rep.get_shape().as_list(), que_rep.get_shape().as_list())
# add elmo
para_rep = tf.concat([para_rep, elmo_para_output], -1)
que_rep = tf.concat([que_rep, elmo_que_output], -1)
print("contextual word embedding:", para_rep.get_shape().as_list(), que_rep.get_shape().as_list())
# bi attention
with tf.variable_scope("bi_attention"):
joint_mask = tf.to_float(tf.expand_dims(self.para_mask, 2)) * tf.to_float(tf.expand_dims(self.que_mask, 1))
para_rep = bidaf_attention(para_rep, que_rep, joint_mask, tri_linear_attention)
para_rep = tf.nn.dropout(para_rep, 1.0 - self.dropout)
para_proj = tf.layers.dense(inputs=para_rep, units=self.d * 2, activation=tf.nn.relu,
name="self_attn_input_proj")
print("bi attention:", para_rep.get_shape().as_list())
# self attneiton
with tf.variable_scope("self_attention"):
with tf.variable_scope("input_proj"):
self_attn_para_input, _ = bi_cudnn_rnn_encoder('gru', self.d, 1, self.dropout,
para_proj, self.para_len,
self.trainable)
diag_mask = 1.0 - tf.diag(tf.ones((self.PL)))
context_mask = tf.to_float(tf.expand_dims(self.para_mask, 2)) * tf.to_float(
tf.expand_dims(self.para_mask, 1))
self_attn_para = self_attention(self_attn_para_input, context_mask * diag_mask, tri_linear_attention)
self_attn_para = tf.nn.dropout(self_attn_para, 1.0 - self.dropout)
self_attn_para = tf.layers.dense(inputs=self_attn_para, units=self.d * 2, activation=tf.nn.relu,
name="self_attn_output_proj")
self_attn_para += para_proj
print("self attention:", self_attn_para.get_shape().as_list())
# model encoder
with tf.variable_scope("model_encoder"):
with tf.variable_scope("start_pointer_proj"):
start_pointer_input, _ = bi_cudnn_rnn_encoder('gru', self.d, 1, self.dropout,
self_attn_para, self.para_len, self.trainable)
start_pointer_input = tf.nn.dropout(start_pointer_input, 1.0 - self.dropout)
logits1 = tf.layers.dense(inputs=start_pointer_input, units=1, use_bias=False,
name="start_pointer")
logits1 = mask_logits(tf.reshape(logits1, [self.N, -1]), tf.reshape(self.para_mask, [self.N, -1]))
with tf.variable_scope("end_pointer_proj"):
end_pointer_input, _ = bi_cudnn_rnn_encoder('gru', self.d, 1, self.dropout,
tf.concat([self_attn_para, start_pointer_input], axis=-1),
self.para_len, self.trainable)
end_pointer_input = tf.nn.dropout(end_pointer_input, 1.0 - self.dropout)
logits2 = tf.layers.dense(inputs=end_pointer_input, units=1, use_bias=False,
name="end_pointer")
logits2 = mask_logits(tf.reshape(logits2, [self.N, -1]), tf.reshape(self.para_mask, [self.N, -1]))
self.probs1 = tf.nn.softmax(logits1)
self.probs2 = tf.nn.softmax(logits2)
# get loss
losses = tf.nn.softmax_cross_entropy_with_logits(logits=logits1, labels=tf.reshape(self.y1, [self.N, -1]))
losses2 = tf.nn.softmax_cross_entropy_with_logits(logits=logits2, labels=tf.reshape(self.y2, [self.N, -1]))
self.batch_loss = losses + losses2
self.loss = tf.reduce_sum(self.batch_loss * tf.to_float(self.labels)) / (tf.reduce_sum(tf.to_float(self.labels)) + 1e-10)
# get prediction
outer = tf.matmul(tf.expand_dims(tf.nn.softmax(logits1), axis=2),
tf.expand_dims(tf.nn.softmax(logits2), axis=1))
outer = tf.matrix_band_part(outer, 0, self.AL)
bprobs, bindex = tf.nn.top_k(tf.reshape(outer, [-1, self.PL * self.PL]), k=self.config.beam_size)
self.byp1 = bindex // self.PL
self.byp2 = bindex % self.PL
self.bprobs = -tf.log(bprobs)
def _init_graph(self):
with self.graph.as_default():
# Model.
u_ids = self.train_features[:, 0]
i_ids = self.train_features[:, 1]
# cold sampling
drop_u_pos = tf.cast(tf.multinomial(
tf.log([[1 - self.cs_ratio, self.cs_ratio]]),
tf.shape(u_ids)[0]),
dtype=self.d_type)
drop_i_pos = tf.cast(tf.multinomial(
tf.log([[1 - self.cs_ratio, self.cs_ratio]]),
tf.shape(i_ids)[0]),
dtype=self.d_type)
drop_u_pos = tf.reshape(drop_u_pos, shape=[-1])
drop_i_pos = tf.reshape(drop_i_pos, shape=[-1])
drop_u_pos_zero = tf.zeros(shape=tf.shape(drop_u_pos),
dtype=self.d_type)
drop_i_pos_zero = tf.zeros(shape=tf.shape(drop_i_pos),
dtype=self.d_type)
drop_u_pos = tf.cond(self.train_phase, lambda: drop_u_pos,
lambda: drop_u_pos_zero)
drop_i_pos = tf.cond(self.train_phase, lambda: drop_i_pos,
lambda: drop_i_pos_zero)
drop_u_pos_v = tf.reshape(drop_u_pos, shape=[-1, 1])
drop_i_pos_v = tf.reshape(drop_i_pos, shape=[-1, 1])
# bias
self.u_bias = tf.nn.embedding_lookup(self.weights['user_bias'],
u_ids) * (1 - drop_u_pos)
self.i_bias = tf.nn.embedding_lookup(self.weights['item_bias'],
i_ids) * (1 - drop_i_pos)
self.cross_bias = tf.reduce_sum(tf.reduce_sum(
tf.nn.embedding_lookup(self.weights['cross_bias'],
self.train_features[:, 2:]),
axis=1),
axis=1)
self.bias = self.cross_bias + self.weights['global_bias']
self.bias += self.u_bias + self.i_bias
# cf part
cf_u_vectors = tf.nn.embedding_lookup(
self.weights['uid_embeddings'], u_ids)
cf_i_vectors = tf.nn.embedding_lookup(
self.weights['iid_embeddings'], i_ids)
random_u_vectors = tf.random_normal(tf.shape(cf_u_vectors),
0,
0.01,
dtype=self.d_type)
random_i_vectors = tf.random_normal(tf.shape(cf_i_vectors),
0,
0.01,
dtype=self.d_type)
cf_u_vectors = random_u_vectors * drop_u_pos_v + cf_u_vectors * (
1 - drop_u_pos_v)
cf_i_vectors = random_i_vectors * drop_i_pos_v + cf_i_vectors * (
1 - drop_i_pos_v)
cf_u_vectors = tf.nn.dropout(cf_u_vectors, self.dropout_keep)
cf_i_vectors = tf.nn.dropout(cf_i_vectors, self.dropout_keep)
self.cf_prediction = tf.reduce_sum(tf.multiply(
cf_u_vectors, cf_i_vectors),
axis=1)
# cb part
u_fs = self.train_features[:, 2:2 + self.user_feature_num]
i_fs = self.train_features[:, 2 + self.user_feature_num:]
uf_vectors = tf.nn.embedding_lookup(
self.weights['feature_embeddings'], u_fs)
if_vectors = tf.nn.embedding_lookup(
self.weights['feature_embeddings'], i_fs)
summed_u_features_emb = tf.reduce_sum(uf_vectors, axis=1)
summed_u_features_emb_square = tf.square(summed_u_features_emb)
squared_u_features_emb = tf.square(uf_vectors)
squared_sum_u_features_emb = tf.reduce_sum(squared_u_features_emb,
axis=1)
u_fm = 0.5 * tf.subtract(summed_u_features_emb_square,
squared_sum_u_features_emb)
summed_i_features_emb = tf.reduce_sum(if_vectors, axis=1)
summed_i_features_emb_square = tf.square(summed_i_features_emb)
squared_i_features_emb = tf.square(if_vectors)
squared_sum_i_features_emb = tf.reduce_sum(squared_i_features_emb,
axis=1)
i_fm = 0.5 * tf.subtract(summed_i_features_emb_square,
squared_sum_i_features_emb)
uf_layer = tf.reshape(
uf_vectors, (-1, self.f_vector_size * self.user_feature_num))
uf_layer = tf.concat([uf_layer, u_fm], axis=1)
uf_layer = tf.layers.batch_normalization(uf_layer,
training=self.train_phase,
name='u_bn_fs')
uf_layer = tf.nn.dropout(uf_layer, self.dropout_keep)
if_layer = tf.reshape(
if_vectors, (-1, self.f_vector_size * self.item_feature_num))
if_layer = tf.concat([if_layer, i_fm], axis=1)
if_layer = tf.layers.batch_normalization(if_layer,
training=self.train_phase,
name='i_bn_fs')
if_layer = tf.nn.dropout(if_layer, self.dropout_keep)
self.lrp_layers_u, self.lrp_layers_i = [uf_layer], [if_layer]
for i in range(0, len(self.cb_hidden_layers) + 1):
uf_layer = tf.add(
tf.matmul(uf_layer, self.weights['cb_user_layer_%d' % i]),
self.weights['cb_user_bias_%d' % i])
if_layer = tf.add(
tf.matmul(if_layer, self.weights['cb_item_layer_%d' % i]),
self.weights['cb_item_bias_%d' % i])
uf_layer = tf.layers.batch_normalization(
uf_layer, training=self.train_phase, name='u_bn_%d' % i)
if_layer = tf.layers.batch_normalization(
if_layer, training=self.train_phase, name='i_bn_%d' % i)
if i < len(self.cb_hidden_layers):
uf_layer = tf.nn.relu(uf_layer)
uf_layer = tf.nn.dropout(uf_layer, self.dropout_keep)
if_layer = tf.nn.relu(if_layer)
if_layer = tf.nn.dropout(if_layer, self.dropout_keep)
self.lrp_layers_u.append(uf_layer)
self.lrp_layers_i.append(if_layer)
cb_u_vectors, cb_i_vectors = uf_layer, if_layer
self.cb_prediction = tf.reduce_sum(tf.multiply(
cb_u_vectors, cb_i_vectors),
axis=1)
# attention
ah_cf_u = tf.add(
tf.matmul(cf_u_vectors, self.weights['attention_weights']),
self.weights['attention_bias'])
ah_cf_u = tf.tanh(ah_cf_u)
# ah_cf_u = tf.nn.relu(ah_cf_u)
ah_cf_u = tf.nn.dropout(ah_cf_u, self.dropout_keep)
a_cf_u = tf.reduce_sum(tf.multiply(ah_cf_u,
self.weights['attention_pre']),
axis=1)
a_cf_u = tf.exp(a_cf_u)
ah_cb_u = tf.add(
tf.matmul(cb_u_vectors, self.weights['attention_weights']),
self.weights['attention_bias'])
ah_cb_u = tf.tanh(ah_cb_u)
# ah_cb_u = tf.nn.relu(ah_cb_u)
ah_cb_u = tf.nn.dropout(ah_cb_u, self.dropout_keep)
a_cb_u = tf.reduce_sum(tf.multiply(ah_cb_u,
self.weights['attention_pre']),
axis=1)
a_cb_u = tf.exp(a_cb_u)
a_sum = a_cf_u + a_cb_u
self.a_cf_u = tf.reshape(a_cf_u / a_sum, shape=[-1, 1])
self.a_cb_u = tf.reshape(a_cb_u / a_sum, shape=[-1, 1])
ah_cf_i = tf.add(
tf.matmul(cf_i_vectors, self.weights['attention_weights']),
self.weights['attention_bias'])
ah_cf_i = tf.tanh(ah_cf_i)
# ah_cf_i = tf.nn.relu(ah_cf_i)
ah_cf_i = tf.nn.dropout(ah_cf_i, self.dropout_keep)
a_cf_i = tf.reduce_sum(tf.multiply(ah_cf_i,
self.weights['attention_pre']),
axis=1)
a_cf_i = tf.exp(a_cf_i)
ah_cb_i = tf.add(
tf.matmul(cb_i_vectors, self.weights['attention_weights']),
self.weights['attention_bias'])
ah_cb_i = tf.tanh(ah_cb_i)
# ah_cb_i = tf.nn.relu(ah_cb_i)
ah_cb_i = tf.nn.dropout(ah_cb_i, self.dropout_keep)
a_cb_i = tf.reduce_sum(tf.multiply(ah_cb_i,
self.weights['attention_pre']),
axis=1)
a_cb_i = tf.exp(a_cb_i)
a_sum = a_cf_i + a_cb_i
self.a_cf_i = tf.reshape(a_cf_i / a_sum, shape=[-1, 1])
self.a_cb_i = tf.reshape(a_cb_i / a_sum, shape=[-1, 1])
# prediction
self.u_vector = self.a_cf_u * cf_u_vectors + self.a_cb_u * cb_u_vectors
self.i_vector = self.a_cf_i * cf_i_vectors + self.a_cb_i * cb_i_vectors
self.prediction = self.bias + tf.reduce_sum(
tf.multiply(self.u_vector, self.i_vector), axis=1)
def _inc_loss_rho(self, rho, signal, t):
return - tf.log(1. + self._expectation(rho, t) * signal / self.A)
def binary_crossentropy(t,o):
return -(t*tf.log(o+eps) + (1.0-t)*tf.log(1.0-o+eps))
y_data = xy[:, [-1]]
print(x_data.shape, y_data.shape)
# placeholders for a tensor that will be always fed.
X = tf.placeholder(tf.float32, shape=[None, 8])
Y = tf.placeholder(tf.float32, shape=[None, 1])
W = tf.Variable(tf.random_normal([8, 1]), name='weight')
b = tf.Variable(tf.random_normal([1]), name='bias')
# Hypothesis using sigmoid: tf.div(1., 1. + tf.exp(tf.matmul(X, W)))
hypothesis = tf.sigmoid(tf.matmul(X, W) + b)
# cost/loss function
cost = -tf.reduce_mean(Y * tf.log(hypothesis) +
(1 - Y) * tf.log(1 - hypothesis))
train = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(cost)
# Accuracy computation
# True if hypothesis>0.5 else False
predicted = tf.cast(hypothesis > 0.5, dtype=tf.float32)
accuracy = tf.reduce_mean(tf.cast(tf.equal(predicted, Y), dtype=tf.float32))
# Launch graph
with tf.Session() as sess:
# Initialize TensorFlow variables
sess.run(tf.global_variables_initializer())
for step in range(10001):
c = apply_depthwise_conv(X,kernel_size,num_channels,depth)
p = apply_max_pool(c,20,2)
c = apply_depthwise_conv(p,6,depth*num_channels,depth//10)
shape = c.get_shape().as_list()
c_flat = tf.reshape(c, [-1, shape[1] * shape[2] * shape[3]])
f_weights_l1 = weight_variable([shape[1] * shape[2] * depth * num_channels * (depth//10), num_hidden])
f_biases_l1 = bias_variable([num_hidden])
f = tf.nn.tanh(tf.add(tf.matmul(c_flat, f_weights_l1),f_biases_l1))
out_weights = weight_variable([num_hidden, num_labels])
out_biases = bias_variable([num_labels])
y_ = tf.nn.softmax(tf.matmul(f, out_weights) + out_biases)
loss = -tf.reduce_sum(Y * tf.log(y_))
optimizer = tf.train.GradientDescentOptimizer(learning_rate = learning_rate).minimize(loss)
correct_prediction = tf.equal(tf.argmax(y_,1), tf.argmax(Y,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
cost_history = np.empty(shape=[1], dtype=float)
# 开始训练
with tf.Session() as session:
tf.initialize_all_variables().run()
# 开始迭代
for epoch in range(training_epochs):
for b in range(total_batchs):
offset = (b * batch_size) % (train_y.shape[0] - batch_size)
batch_x = train_x[offset:(offset + batch_size), :, :, :]
def tf_psnr(im1, im2):
# assert pixel value range is 0-1
#mse = tf.losses.mean_squared_error(labels=im2 * 255.0, predictions=im1 * 255.0)
mse = tf.compat.v1.losses.mean_squared_error(labels=im2 * 255.0,
predictions=im1 * 255.0)
return 10.0 * (tf.log(255.0**2 / mse) / tf.log(10.0))
# padding = 'SAME' 时，输出并不一定和原图size一致
# 但会保证覆盖原图所有像素，不会舍弃边上的某些元素
## fully connection 1 layer ##
W_fc1 = weight_variable([7 * 7 * 64, 1024])
b_fc1 = biase_variable([1024])
h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
## fully connection 2 layer ##
W_fc2 = weight_variable([1024, 10])
b_fc2 = biase_variable([10])
prediction = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
# 2 定义损失函数与反向传播方法
cross_entropy = tf.reduce_mean(-tf.reduce_sum(
y * tf.log(prediction), reduction_indices=[1])) # sum(- y * log(y0))/n 交叉熵
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
#3 生成会话,训练
with tf.Session() as sess:
# 变量初始化
init = tf.global_variables_initializer()
sess.run(init)
# 训练
for i in range(5):
# 提取 batch
batch_x, batch_y = mnist.train.next_batch(100)
sess.run(train_step,
feed_dict={
x: batch_x,
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("mnist_data/", one_hot=True)
# print (mnist.train.next_batch(1))
x = tf.placeholder(tf.float32, [None, 784])
w = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
y = tf.nn.softmax(tf.matmul(x, w) + b)
y_ = tf.placeholder(tf.float32, [None, 10])
cross_entropy = tf.reduce_mean(
-tf.reduce_sum(y_ * tf.log(y), reduction_indices=[1]))
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for xx in range(1000):
bx, by = mnist.train.next_batch(500)
# bx=mnist.train.images
# by=mnist.train.labels
sess.run(train_step, feed_dict={x: bx, y_: by})
correct_prediction = tf.equal(tf.argmax(y_, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
print(
def create_model(self):
g = tf.Graph()
with g.as_default():
# Define model variables
var_linear = tf.get_variable('linear',
[self.feature_dim, 1],
initializer=tf.random_uniform_initializer(
-self.args.init_mean, self.args.init_mean))
var_emb_factors = tf.get_variable('emb_factors',
[self.feature_dim, self.args.num_dims],
initializer=tf.random_uniform_initializer(
-self.args.init_mean, self.args.init_mean))
# Sparse placeholders
pl_user_list = tf.placeholder(tf.int64, shape=[None], name='pos_list')
pl_pos_indices = tf.placeholder(tf.int64, shape=[None, 2], name='pos_indices')
pl_pos_values = tf.placeholder(tf.float32, shape=[None], name='pos_values')
pl_pos_shape = tf.placeholder(tf.int64, shape=[2], name='pos_shape')
pl_neg_indices = tf.placeholder(tf.int64, shape=[None, 2], name='neg_indices')
pl_neg_values = tf.placeholder(tf.float32, shape=[None], name='neg_values')
pl_neg_shape = tf.placeholder(tf.int64, shape=[2], name='neg_shape')
placeholders = {
'pl_user_list': pl_user_list,
'pl_pos_indices': pl_pos_indices,
'pl_pos_values': pl_pos_values,
'pl_pos_shape': pl_pos_shape,
'pl_neg_indices': pl_neg_indices,
'pl_neg_values': pl_neg_values,
'pl_neg_shape': pl_neg_shape
}
# Input positive features, shape = (batch_size * feature_dim)
sparse_pos_feats = tf.SparseTensor(pl_pos_indices, pl_pos_values, pl_pos_shape)
# Input negative features, shape = (batch_size * feature_dim)
sparse_neg_feats = tf.SparseTensor(pl_neg_indices, pl_neg_values, pl_neg_shape)
pos_preds, neg_preds = self.get_preds(var_linear, var_emb_factors,
sparse_pos_feats, sparse_neg_feats)
l2_reg = tf.add_n([
self.args.linear_reg * tf.reduce_sum(tf.square(var_linear)),
self.args.emb_reg * tf.reduce_sum(tf.square(var_emb_factors)),
])
# BPR training op (add 1e-10 to help numerical stability)
bprloss_op = tf.reduce_sum(tf.log(1e-10 + tf.sigmoid(pos_preds - neg_preds))) - l2_reg
bprloss_op = -bprloss_op
global_step = tf.Variable(0, trainable=False)
learning_rate = tf.train.exponential_decay(self.args.starting_lr,
global_step, self.args.lr_decay_freq,
self.args.lr_decay_factor, staircase=False)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
train_op = optimizer.minimize(bprloss_op, global_step=global_step)
# AUC
binary_ranks = tf.to_float((pos_preds - neg_preds) > 0)
auc_per_user = tf.segment_mean(binary_ranks, pl_user_list)
auc_op = tf.divide(tf.reduce_sum(auc_per_user),
tf.to_float(tf.size(tf.unique(pl_user_list)[0])))
self.var_linear = var_linear
self.var_emb_factors = var_emb_factors
return (g, bprloss_op, optimizer, train_op, auc_op, l2_reg, placeholders)
learning_rate = 0.01
training_epochs = 15
batch_size = 100
display_step = 1
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
# tf Graph Input
x = tf.placeholder("float", [None, 784]) # mnist data image of shape 28*28=784 (흑백이므로 color를 위한 차원은 없음)
y = tf.placeholder("float", [None, 10]) # 0-9 digits recognition => 10 classes
# set model weight
W = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
activation = tf.nn.softmax(tf.matmul(x, W) + b)
cost = tf.reduce_mean(-tf.reduce_sum(y*tf.log(activation), reduction_indices=1))
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
# Initializing the variables
init = tf.global_variables_initializer()
# 4) 학습 실행 (Launch the graph)
with tf.Session() as sess:
sess.run(init)
for epoch in range(training_epochs):
avg_cost = 0.
# mnist.train.num_examples = 55000
# batch_size = 100
# total_batch = 550 ==> 1번의 training_epoch에 550개의 이미지를 학습한다
total_batch = int(mnist.train.num_examples / batch_size)
for i in range(total_batch):
return G_prob
# 判别网络
def discriminator(x):
D_h1 = tf.nn.relu(tf.matmul(x, D_W1) + D_b1)
D_logit = tf.matmul(D_h1, D_W2) + D_b2
D_prob = tf.nn.sigmoid(D_logit)
return D_prob, D_logit
G_sample = generator(Z)
D_real, D_logit_real = discriminator(X)
D_fake, D_logit_fake = discriminator(G_sample)
# 损失函数
D_loss = -tf.reduce_mean(tf.log(D_real) + tf.log(1. - D_fake))
G_loss = -tf.reduce_mean(tf.log(D_fake))
# 更新D(x)的参数
D_solver = tf.train.AdamOptimizer().minimize(D_loss, var_list=theta_D)
# 更新G(Z)的参数
G_solver = tf.train.AdamOptimizer().minimize(G_loss, var_list=theta_G)
def sample_Z(m, n):
return np.random.uniform(-1., 1., size=[m, n])
batch_size = 128
Z_dim = 100
sess = tf.Session()
w1 = tf.Variable(tf.random_normal(shape=[10, 6]))
b1 = tf.Variable(tf.random_normal(shape=[6]))
w2 = tf.Variable(tf.random_normal(shape=[6, 2]))
b2 = tf.Variable(tf.random_normal(shape=[2]))
interval = 50
epoch = 4000
LR = .005
Y1 = tf.nn.relu(tf.add(tf.matmul(X, w1), b1))
Y2 = tf.nn.softmax(tf.add(tf.matmul(Y1, w2), b2))
with tf.name_scope("loss"):
loss = -tf.reduce_sum(y_ * tf.log(Y2))
optimizer = tf.train.AdamOptimizer(LR).minimize(
loss) # Add scalar summary for cost tensor
tf.summary.scalar("loss", loss)
with tf.name_scope("accuracy"):
correct_pred = tf.equal(tf.argmax(y_, 1),
tf.argmax(Y2, 1)) # Count correct predictions
acc_op = tf.reduce_mean(tf.cast(
correct_pred, "float")) # Cast boolean to float to average
# Add scalar summary for accuracy tensor
tf.summary.scalar("accuracy", acc_op)
with tf.Session() as sess:
#tensorboard config ... use: tensorboard --logdir=/logs/churn/
fw = tf.summary.FileWriter("churn_prediction/logs/churn", sess.graph)
def __init__(self,
num_emb,
batch_size,
emb_dim,
hidden_dim,
sequence_length,
start_token,
good_id,
pos,
learning_rate=0.01,
reward_gamma=0.95):
self.num_emb = num_emb
self.batch_size = batch_size
self.emb_dim = emb_dim
self.hidden_dim = hidden_dim
self.sequence_length = sequence_length
self.start_token = tf.constant([start_token] * self.batch_size,
dtype=tf.int32)
self.learning_rate = tf.Variable(float(learning_rate), trainable=False)
self.reward_gamma = reward_gamma
self.g_params = []
self.d_params = []
self.good_id = tf.constant(good_id * self.batch_size, dtype=tf.int32)
self.temperature = 2
self.grad_clip = 5.0
self.pos = pos
self.expected_reward = tf.Variable(tf.zeros([self.sequence_length]))
with tf.variable_scope('generator'):
self.g_embeddings = tf.Variable(
self.init_matrix_embedding([self.num_emb, self.emb_dim],
self.pos))
self.g_params.append(self.g_embeddings)
self.g_recurrent_unit = self.create_recurrent_unit(
self.g_params) # maps h_tm1 to h_t for generator
self.g_output_unit = self.create_output_unit(
self.g_params) # maps h_t to o_t (output token logits)
# placeholder definition
self.x = tf.placeholder(tf.int32,
shape=[self.batch_size, self.sequence_length])
# sequence of indices of true data, not including start token
self.rewards = tf.placeholder(
tf.float32, shape=[self.batch_size, self.sequence_length])
# get from rollout policy and discriminator
# processed for batch
with tf.device("/cpu:0"):
inputs = tf.split(
1, self.sequence_length,
tf.nn.embedding_lookup(self.g_embeddings, self.x))
self.processed_x = tf.pack([
tf.squeeze(input_, [1]) for input_ in inputs
]) # seq_length x batch_size x emb_dim
with tf.device("/cpu:0"):
inputs = tf.split(1, self.sequence_length, self.x)
self.processed_token_x = tf.pack(
[tf.squeeze(input_, [1]) for input_ in inputs])
self.h0 = tf.zeros([self.batch_size, self.hidden_dim])
self.h0 = tf.pack([self.h0, self.h0])
gen_o = tensor_array_ops.TensorArray(dtype=tf.float32,
size=self.sequence_length,
dynamic_size=False,
infer_shape=True)
gen_x = tensor_array_ops.TensorArray(dtype=tf.int32,
size=self.sequence_length,
dynamic_size=False,
infer_shape=True)
ta_emb_x2 = tensor_array_ops.TensorArray(dtype=tf.int32,
size=self.sequence_length)
ta_emb_x2 = ta_emb_x2.unpack(self.processed_token_x)
x_t = tf.nn.embedding_lookup(self.g_embeddings, self.start_token)
h_t1 = self.g_recurrent_unit(x_t, self.h0) # hidden_memory_tuple
o_t = self.g_output_unit(h_t1) # batch x vocab , logits not prob
log_prob = tf.divide(tf.log(tf.nn.softmax(o_t)), self.temperature)
next_token = tf.cast(tf.reshape(ta_emb_x2.read(0), [self.batch_size]),
tf.int32)
#next_token = tf.cast(tf.reshape(tf.multinomial(log_prob, 1), [self.batch_size]), tf.int32)
x_tp1 = tf.nn.embedding_lookup(self.g_embeddings,
next_token) # batch x emb_dim
gen_o = gen_o.write(
tf.constant(0, dtype=tf.int32),
tf.reduce_sum(
tf.mul(tf.one_hot(next_token, self.num_emb, 1.0, 0.0),
tf.nn.softmax(o_t)), 1)) # [batch_size] , prob
gen_x = gen_x.write(tf.constant(0, dtype=tf.int32),
next_token) # indices, batch_size
#random sampling:
'''
x_t = tf.nn.embedding_lookup(self.g_embeddings,self.start_token)
h_t1 = self.g_recurrent_unit(x_t, self.h0) # hidden_memory_tuple
o_t = self.g_output_unit(h_t1) # batch x vocab , logits not prob
next_token = tf.multinomial(tf.nn.softmax(o_t),1)
next_token = tf.cast(tf.reshape(tf.multinomial(tf.nn.softmax(o_t),1),[self.batch_size]),tf.int32)
x_tp1 = tf.nn.embedding_lookup(self.g_embeddings, next_token) # batch x emb_dim
gen_o = gen_o.write(tf.constant(0,dtype=tf.int32), tf.reduce_sum(tf.mul(tf.one_hot(next_token, self.num_emb, 1.0, 0.0),tf.nn.softmax(o_t)), 1)) # [batch_size] , prob
gen_x = gen_x.write(tf.constant(0,dtype=tf.int32), next_token) # indices, batch_size
'''
def _g_recurrence(i, x_t, h_tm1, gen_o, gen_x):
h_t = self.g_recurrent_unit(x_t, h_tm1) # hidden_memory_tuple
o_t = self.g_output_unit(h_t) # batch x vocab , logits not prob
print('now here')
next_token = tf.argmax(tf.nn.softmax(o_t), 1)
next_token = tf.cast(
tf.reshape(tf.argmax(tf.nn.softmax(o_t), 1),
[self.batch_size]), tf.int32)
x_tp1 = tf.nn.embedding_lookup(self.g_embeddings, next_token)
gen_o = gen_o.write(
i,
tf.reduce_sum(
tf.mul(tf.one_hot(next_token, self.num_emb, 1.0, 0.0),
tf.nn.softmax(o_t)), 1)) # [batch_size] , prob
gen_x = gen_x.write(i, next_token) # indices, batch_size
return i + 1, x_tp1, h_t, gen_o, gen_x
_, _, _, self.gen_o, self.gen_x = tf.while_loop(
cond=lambda i, _1, _2, _3, _4: i < self.sequence_length,
body=_g_recurrence,
loop_vars=(tf.constant(1,
dtype=tf.int32), x_tp1, h_t1, gen_o, gen_x))
self.gen_x = self.gen_x.pack() # seq_length x batch_size
self.gen_x = tf.transpose(self.gen_x,
perm=[1, 0]) # batch_size x seq_length
self.h0 = tf.zeros([self.batch_size, self.hidden_dim])
self.h0 = tf.pack([self.h0, self.h0])
gen_o_argmax = tensor_array_ops.TensorArray(dtype=tf.float32,
size=self.sequence_length,
dynamic_size=False,
infer_shape=True)
gen_x_argmax = tensor_array_ops.TensorArray(dtype=tf.int32,
size=self.sequence_length,
dynamic_size=False,
infer_shape=True)
def _g_recurrence_argmax(i, x_t, h_tm1, gen_o, gen_x):
h_t = self.g_recurrent_unit(x_t, h_tm1) # hidden_memory_tuple
o_t = self.g_output_unit(h_t) # batch x vocab , logits not prob
next_token = tf.argmax(tf.nn.softmax(o_t), 1)
next_token = tf.cast(
tf.reshape(tf.argmax(tf.nn.softmax(o_t), 1),
[self.batch_size]), tf.int32)
x_tp1 = tf.nn.embedding_lookup(self.g_embeddings,
next_token) # batch x emb_dim
print(x_tp1)
gen_o = gen_o.write(
i,
tf.reduce_sum(
tf.mul(tf.one_hot(next_token, self.num_emb, 1.0, 0.0),
tf.nn.softmax(o_t)), 1)) # [batch_size] , prob
gen_x = gen_x.write(i, next_token) # indices, batch_size
return i + 1, x_tp1, h_t, gen_o, gen_x
_, _, _, self.gen_o_argmax, self.gen_x_argmax = tf.while_loop(
cond=lambda i, _1, _2, _3, _4: i < self.sequence_length,
body=_g_recurrence_argmax,
loop_vars=(tf.constant(0, dtype=tf.int32),
tf.nn.embedding_lookup(self.g_embeddings,
self.start_token), self.h0,
gen_o_argmax, gen_x_argmax))
self.gen_x_argmax = self.gen_x_argmax.pack() # seq_length x batch_size
self.gen_x_argmax = tf.transpose(self.gen_x_argmax,
perm=[1,
0]) # batch_size x seq_length
###################################################################
# wgan pretraining for generator
g_predictions_wgan = tensor_array_ops.TensorArray(
dtype=tf.float32,
size=self.sequence_length,
dynamic_size=False,
infer_shape=True)
def _wgantrain_recurrence(i, x_t, h_tm1, g_predictions_wgan):
h_t = self.g_recurrent_unit(x_t, h_tm1)
o_t = self.g_output_unit(h_t)
g_predictions_wgan = g_predictions_wgan.write(
i, tf.nn.softmax(o_t)) # batch x vocab_size
next_token = tf.argmax(tf.nn.softmax(o_t), 1)
x_tp1 = tf.nn.embedding_lookup(self.g_embeddings, next_token)
#x_tp1 = ta_emb_x.read(i)
return i + 1, x_tp1, h_t, g_predictions_wgan
_, _, _, self.g_predictions_wgan = tf.while_loop(
cond=lambda i, _1, _2, _3: i < self.sequence_length,
body=_wgantrain_recurrence,
loop_vars=(tf.constant(0, dtype=tf.int32),
tf.nn.embedding_lookup(self.g_embeddings,
self.start_token), self.h0,
g_predictions_wgan))
self.g_predictions_wgan = tf.transpose(
self.g_predictions_wgan.pack(),
perm=[1, 0, 2]) # batch_size x seq_length x vocab_size
###################################################################
# supervised pretraining for generator
g_predictions = tensor_array_ops.TensorArray(dtype=tf.float32,
size=self.sequence_length,
dynamic_size=False,
infer_shape=True)
ta_emb_x = tensor_array_ops.TensorArray(dtype=tf.float32,
size=self.sequence_length)
ta_emb_x = ta_emb_x.unpack(self.processed_x)
def _pretrain_recurrence(i, x_t, h_tm1, g_predictions):
h_t = self.g_recurrent_unit(x_t, h_tm1)
o_t = self.g_output_unit(h_t)
g_predictions = g_predictions.write(
i, tf.nn.softmax(o_t)) # batch x vocab_size
x_tp1 = ta_emb_x.read(i)
return i + 1, x_tp1, h_t, g_predictions
_, _, _, self.g_predictions = tf.while_loop(
cond=lambda i, _1, _2, _3: i < self.sequence_length,
body=_pretrain_recurrence,
loop_vars=(tf.constant(0, dtype=tf.int32),
tf.nn.embedding_lookup(self.g_embeddings,
self.start_token), self.h0,
g_predictions))
self.g_predictions = tf.transpose(
self.g_predictions.pack(),
perm=[1, 0, 2]) # batch_size x seq_length x vocab_size
# pretraining loss
self.pretrain_loss = -tf.reduce_sum(
tf.one_hot(tf.to_int32(tf.reshape(
self.x, [-1])), self.num_emb, 1.0, 0.0) * tf.log(
tf.clip_by_value(
tf.reshape(self.g_predictions, [-1, self.num_emb]),
1e-20, 1.0))) / (self.sequence_length *
self.batch_size)
# training updates
pretrain_opt = self.g_optimizer(self.learning_rate)
self.pretrain_grad, _ = tf.clip_by_global_norm(
tf.gradients(self.pretrain_loss, self.g_params), self.grad_clip)
self.pretrain_updates = pretrain_opt.apply_gradients(
zip(self.pretrain_grad, self.g_params))
#######################################################################################################
# Unsupervised Training
#######################################################################################################
self.g_loss = -tf.reduce_sum(
tf.reduce_sum(
tf.one_hot(tf.to_int32(tf.reshape(
self.x, [-1])), self.num_emb, 1.0, 0.0) * tf.log(
tf.clip_by_value(
tf.reshape(self.g_predictions, [-1, self.num_emb]),
1e-20, 1.0)), 1) * tf.reshape(self.rewards, [-1]))
g_opt = self.g_optimizer(self.learning_rate)
self.g_grad, _ = tf.clip_by_global_norm(
tf.gradients(self.g_loss, self.g_params), self.grad_clip)
self.g_updates = g_opt.apply_gradients(zip(self.g_grad, self.g_params))