def forward(self, X):
conv_out = tf.nn.conv2d(X, self.W, strides=self.s, padding=self.p)
# add a bias vector:
conv_out = tf.nn.bias_add(conv_out, self.b)
p1, p2 = self.poolsz
pool_out = tf.nn.max_pool(conv_out,
ksize=[1, p1, p2, 1],
strides=[1, p1, p2, 1],
padding='SAME')
return swish(pool_out, beta=10, tensorflow=True)
def activate(self, linear, name):
if name == 'sigmoid':
return tf.nn.sigmoid(linear, name='encoded')
elif name == 'softmax':
return tf.nn.softmax(linear, name='encoded')
elif name == 'linear':
return linear
elif name == 'tanh':
return tf.nn.tanh(linear, name='encoded')
elif name == 'relu':
return tf.nn.relu(linear, name='encoded')
elif name == 'swish':
return util.swish(linear, name='encoded')
def forward(self, X):
conv_out = conv2d(
input=X,
filters=self.W,
border_mode=self.border_mode,
subsample=self.s,
)
pooled_out = pool_2d(
input=conv_out,
ws=self.poolsz,
ignore_border=True,
mode='max',
)
return swish(pooled_out + self.b.dimshuffle('x', 0, 'x', 'x'))
def forward(self, X, is_training, decay=0.9):
Z = tf.matmul(X, self.W)
if is_training:
batch_mean, batch_var = tf.nn.moments(Z, [0])
# update the running mean and running variance:
update_rn_mean = tf.assign(
self.rn_mean, self.rn_mean * decay + batch_mean * (1 - decay))
update_rn_var = tf.assign(
self.rn_var, self.rn_var * decay + batch_var * (1 - decay))
# to make sure the aforementioned updates are calculated
# every time we call the train function,
# we have to use next function:
with tf.control_dependencies([update_rn_mean, update_rn_var]):
Z = tf.nn.batch_normalization(Z, batch_mean, batch_var,
self.betta, self.gamma, 1e-4)
else:
Z = tf.nn.batch_normalization(Z, self.rn_mean, self.rn_var,
self.betta, self.gamma, 1e-4)
return swish(Z, tensorflow=True)
def forward(self, X): return swish(X.dot(self.W) + self.b)
def forward(self, X):
return swish(tf.matmul(X, self.W) + self.b, tensorflow=True)