def build(self):
print "build PGS_NET_WB"
params = self.params
N = params.batch_size # number of groups/display_ids per batch
A = params.max_ads # maximum number of Ads per display_id
F = params.meta_features # number of meta features per Ad
yb = tf.placeholder('float32', shape=[N, A], name='yb')
X = tf.placeholder('float32', shape=[N, A, F],
name='x') # zero padding
Xmask = tf.placeholder(
'float32', shape=[N, A], name='xmask'
) # in {-e10, 1}, 1 for real Ads and -e10 for padding Ads
#Xads = tf.placeholder('float32', shape=[N], name='xads') # number of Ads per display_id
y = tf.placeholder('float32', shape=[N, A],
name='y') # y in {0, 1} with zero padding
is_training = tf.placeholder(dtype=tf.bool)
if self.params.softmax_transform:
print("softmax_transform")
Xtmp = X + tf.reshape(Xmask, [N, A, 1])
Xtmp = tf.exp(Xtmp)
stmp = tf.reduce_sum(Xtmp, 1, keep_dims=True) + 1e-5
Xtmp = Xtmp / stmp
else:
Xtmp = X
with tf.name_scope("Fully-connected"):
Xtmp = tf.reshape(Xtmp, [N * A, F])
with tf.variable_scope("Layer1"):
ytmp = fully_connected(Xtmp,
num_neurons=50,
name='W1',
is_training=is_training,
use_batch_norm=True,
use_drop_out=False,
keep_prob=0.7,
activation='relu',
default_batch=params.default_batch)
with tf.variable_scope("Layer2"):
ytmp = fully_connected(ytmp,
num_neurons=25,
name='W2',
is_training=is_training,
use_batch_norm=True,
use_drop_out=False,
keep_prob=1.0,
activation='relu',
default_batch=params.default_batch)
#with tf.variable_scope("Layer4"):
# ytmp = fully_connected(ytmp, num_neurons=10, name='W4', is_training = is_training, use_batch_norm=True, use_drop_out=False, keep_prob = 1.0, activation = 'relu', default_batch = params.default_batch)
with tf.variable_scope("Layer3"):
ytmp = fully_connected(ytmp,
num_neurons=1,
name='W3',
is_training=is_training,
use_batch_norm=True,
use_drop_out=False,
keep_prob=1,
activation='None',
default_batch=params.default_batch)
# ytmp is [N*A, 1] now
yp = tf.reshape(ytmp,
[N, A]) * yb + Xmask #+ yb# masking the padding Ads
# yp is [N, A] now
with tf.name_scope('Loss'):
# Cross-Entropy loss
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(yp, y)
loss = tf.reduce_mean(cross_entropy)
total_loss = loss + params.weight_decay * tf.add_n(
tf.get_collection('l2'))
with tf.name_scope('Predict'):
pred = tf.nn.softmax(yp)
if self.params.opt == 'adam':
optimizer = tf.train.AdamOptimizer(params.learning_rate)
elif self.params.opt == 'sgd':
optimizer = tf.train.GradientDescentOptimizer(params.learning_rate)
elif self.params.opt == 'ada':
optimizer = tf.train.AdagradOptimizer(params.learning_rate)
elif self.params.opt == 'rmsprop':
optimizer = tf.train.RMSPropOptimizer(params.learning_rate)
opt_op = optimizer.minimize(total_loss, global_step=self.global_step)
self.predictions = pred
self.loss = cross_entropy
self.total_loss = total_loss
self.opt_op = opt_op
self.yb = yb
self.x = X
self.y = y
self.xmask = Xmask
self.is_train = is_training
def build(self):
params = self.params
N = params.batch_size # number of groups/display_ids per batch
A = params.max_ads # maximum number of Ads per display_id
F = params.meta_features # number of meta features per Ad
X = tf.placeholder('float32', shape=[N, A, F],
name='x') # zero padding
Xmask = tf.placeholder(
'float32', shape=[N, A], name='xmask'
) # in {-e10, 1}, 1 for real Ads and -e10 for padding Ads
#Xads = tf.placeholder('float32', shape=[N], name='xads') # number of Ads per display_id
y = tf.placeholder('float32', shape=[N, A],
name='y') # y in {0, 1} with zero padding
is_training = tf.placeholder(tf.bool)
if self.params.softmax_transform:
print("softmax_transform")
Xtmp = X + tf.reshape(Xmask, [N, A, 1])
Xtmp = tf.exp(Xtmp)
stmp = tf.reduce_sum(Xtmp, 1, keep_dims=True) + 1e-5
Xtmp = Xtmp / stmp
else:
Xtmp = X
with tf.name_scope("Fully-connected"):
Xtmp = tf.reshape(Xtmp, [N * A, F])
with tf.variable_scope("Layer1"):
ytmp = fully_connected(Xtmp,
num_neurons=50,
name='W1',
is_training=is_training,
use_batch_norm=True,
use_drop_out=False,
keep_prob=0.7,
activation='sigmoid')
with tf.variable_scope("Layer2"):
ytmp = fully_connected(ytmp,
num_neurons=25,
name='W2',
is_training=is_training,
use_batch_norm=True,
use_drop_out=False,
keep_prob=1.0,
activation='sigmoid')
with tf.variable_scope("Layer3"):
ytmp = fully_connected(ytmp,
num_neurons=1,
name='W3',
is_training=is_training,
use_batch_norm=True,
use_drop_out=False,
keep_prob=1,
activation='None')
# ytmp is [N*A, 1] now
yp = tf.reshape(ytmp, [N, A]) * Xmask # masking the padding Ads
# yp is [N, A] now
with tf.name_scope('Loss'):
# Cross-Entropy loss
#yp = ytmp * tf.reshape(Xmask,[N*A, 1])
#y = tf.reshape(y,)
#cross_entropy = tf.nn.softmax_cross_entropy_with_logits(yp, y)
yp = tf.maximum(yp, 1e-5)
yp = tf.minimum(yp, 1 - 1e-5)
cross_entropy = -(y * tf.log(yp) + (1 - y) * tf.log(1 - yp))
loss = tf.reduce_mean(cross_entropy)
total_loss = loss + params.weight_decay * tf.add_n(
tf.get_collection('l2'))
with tf.name_scope('Predict'):
pred = yp #tf.nn.softmax(yp)
optimizer = tf.train.AdamOptimizer(params.learning_rate)
opt_op = optimizer.minimize(total_loss, global_step=self.global_step)
self.predictions = pred
self.loss = cross_entropy
self.total_loss = total_loss
self.opt_op = opt_op
self.x = X
self.y = y
self.xmask = Xmask
self.is_train = is_training