def cnn2d(input, training, height):
"""
Model function for CNN.
accepts input shape: [step_size, time_shift*features]
transformed to: [step_size, time_shift(height), features(width), channel]
"""
# with tf.variable_scope("conv2d_parent"):
print("shape of cnn input: {}".format(input.get_shape()))
width = int(input.get_shape()[1]) // height
input2d = tf.reshape(input, [-1, height, width, 1])
# Transforms into 2D compatible format [batch(step), height, width, channel]
print("input transformed to 2D pre-conv shape: {}".format(
input2d.get_shape()))
nlayer = numLayers(height, width)
print("height:{} width:{} #conv layers: {}".format(
height, width, nlayer))
filters = max(2, 2**(math.ceil(math.log(max(height, width), 2))))
# krange = 3
# drange = 3
convlayer = input2d
for i in range(nlayer):
filters *= 2
convlayer = conv2d(convlayer, filters, i)
print("final conv2d: {}".format(convlayer.get_shape()))
convlayer = tf.squeeze(convlayer, [1, 2])
print("squeeze: {}".format(convlayer.get_shape()))
dense = tf.compat.v1.layers.dense(
# name="cnn2d_dense",
inputs=convlayer,
units=convlayer.get_shape()[1] * 2,
kernel_initializer=tf.compat.v1.truncated_normal_initializer(
stddev=0.01),
bias_initializer=tf.compat.v1.constant_initializer(0.1),
activation=tf.nn.elu)
print("dense: {}".format(dense.get_shape()))
dropout = tf.compat.v1.layers.dropout(
# name="cnn2d_dropout",
inputs=dense,
rate=0.5,
training=training)
return dropout
def cnn(self, input):
with tf.compat.v1.variable_scope("cnn"):
# Transforms into 2D compatible format [batch, step, feature, channel]
convlayer = tf.expand_dims(input, 3)
height = int(input.get_shape()[1])
width = int(input.get_shape()[2])
nlayer = min(5, numLayers(height, width))
filters = max(
16, 2**(math.floor(math.log(self._num_hidden // nlayer))))
for i in range(nlayer):
filters = min(filters * 2, self._num_hidden * 2)
convlayer = self.conv2d(convlayer, int(filters), i,
tf.nn.relu6)
convlayer = tf.compat.v1.layers.flatten(convlayer)
# convlayer = tf.contrib.layers.layer_norm(inputs=convlayer)
# convlayer = tf.layers.dense(
# inputs=convlayer,
# units=math.ceil(
# math.sqrt(float(int(convlayer.get_shape()[1])))),
# kernel_initializer=tf.truncated_normal_initializer(
# stddev=0.01),
# bias_initializer=tf.constant_initializer(0.1)
# )
return convlayer
def cnn2d(input, training, height):
"""
Model function for CNN.
accepts input shape: [step_size, time_shift*features]
transformed to: [step_size, time_shift(height), features(width), channel]
"""
with tf.compat.v1.variable_scope("conv2d_parent"):
print("shape of cnn input: {}".format(input.get_shape()))
width = int(input.get_shape()[1]) // height
input2d = tf.reshape(input, [-1, height, width, 1])
# Transforms into 2D compatible format [batch(step), height, width, channel]
print("input transformed to 2D pre-conv shape: {}".format(
input2d.get_shape()))
nlayer = numLayers(height, width)
print("height:{} width:{} #conv layers: {}".format(
height, width, nlayer))
filters = max(16, 2**(math.ceil(math.log(max(height, width), 2))))
# krange = 3
# drange = 3
convlayer = input2d
for i in range(nlayer):
filters *= 2
convlayer = conv2d(convlayer, filters, i)
# conv = tf.layers.conv2d(
# name="conv_lv{}".format(i),
# inputs=convlayer,
# filters=filters,
# kernel_size=2,
# kernel_initializer=tf.truncated_normal_initializer(
# stddev=0.01),
# bias_initializer=tf.constant_initializer(0.1),
# padding="same")
# h_stride = 2 if int(conv.get_shape()[1]) >= 2 else 1
# w_stride = 2 if int(conv.get_shape()[2]) >= 2 else 1
# pool = tf.layers.max_pooling2d(
# name="pool_lv{}".format(i),
# inputs=conv, pool_size=2, strides=[h_stride, w_stride],
# padding="same")
# print("#{} conv:{} pool: {}".format(
# i+1, conv.get_shape(), pool.get_shape()))
# can't use tf.nn.batch_normalization in a mapped function
# tf.contrib.layers.batch_norm seems ineffective
# norm = tf.contrib.layers.batch_norm(
# inputs=pool,
# scale=True,
# # updates_collections=None,
# # param_initializers=tf.truncated_normal_initializer(
# # stddev=0.01),
# is_training=training
# )
# norm = tf.nn.lrn(pool, name="lrn_lv{}".format(i))
# convlayer = pool
print("final conv2d: {}".format(convlayer.get_shape()))
convlayer = tf.squeeze(convlayer, [1, 2])
print("squeezed: {}".format(convlayer.get_shape()))
# use tf.contrib.layers.fully_connected?
output = tf.compat.v1.layers.dense(
name="cnn2d_dense",
inputs=convlayer,
units=convlayer.get_shape()[1] * 2,
kernel_initializer=tf.compat.v1.truncated_normal_initializer(
stddev=0.01),
bias_initializer=tf.compat.v1.constant_initializer(0.1),
activation=tf.nn.relu6)
print("dense: {}".format(output.get_shape()))
output = tf.compat.v1.layers.dropout(inputs=output,
rate=0.5,
training=training)
return output