def model(activation='relu'):
# input layer
inputs = keras.layers.Input(shape=(400, 4))
# layer 1
activation = utils.activation_fn(activation)
nn = layers.conv_layer(inputs,
num_filters=24,
kernel_size=19,
padding='same',
activation=activation,
dropout=0.1,
l2=1e-6,
bn=True)
# layer 2
nn = layers.residual_block(nn, filter_size=5, activation='relu', l2=1e-6)
nn = keras.layers.MaxPool1D(pool_size=10)(nn)
# layer 3
nn = layers.conv_layer(nn,
num_filters=48,
kernel_size=7,
padding='same',
activation='relu',
dropout=0.3,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(pool_size=5)(nn)
# layer 4
nn = layers.conv_layer(nn,
num_filters=64,
kernel_size=3,
padding='valid',
activation='relu',
dropout=0.4,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(pool_size=4)(nn)
# layer 5
nn = keras.layers.Flatten()(nn)
nn = layers.dense_layer(nn,
num_units=96,
activation='relu',
dropout=0.5,
l2=1e-6,
bn=True)
# Output layer
logits = keras.layers.Dense(1, activation='linear', use_bias=True)(nn)
outputs = keras.layers.Activation('sigmoid')(logits)
# compile model
model = keras.Model(inputs=inputs, outputs=outputs)
return model
def model(pool_size=[25, 4], activation='relu', input_shape=200):
if input_shape == 1000:
multiplier = 2
else:
multiplier = 1
# input layer
inputs = keras.layers.Input(shape=(input_shape, 4))
# layer 1
activation = utils.activation_fn(activation)
nn = layers.conv_layer(inputs,
num_filters=32 * multiplier,
kernel_size=19,
padding='same',
activation=activation,
dropout=0.1,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(pool_size=pool_size[0],
strides=pool_size[0],
padding='same')(nn)
# layer 2
nn = layers.conv_layer(nn,
num_filters=124 * multiplier,
kernel_size=5,
padding='same',
activation='relu',
dropout=0.1,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(pool_size=pool_size[1],
strides=pool_size[1],
padding='same')(nn)
# layer 3
nn = keras.layers.Flatten()(nn)
nn = layers.dense_layer(nn,
num_units=512 * multiplier,
activation='relu',
dropout=0.5,
l2=1e-6,
bn=True)
# Output layer
logits = keras.layers.Dense(12, activation='linear', use_bias=True)(nn)
outputs = keras.layers.Activation('sigmoid')(logits)
# compile model
model = keras.Model(inputs=inputs, outputs=outputs)
return model
def model(activation='log_relu', l2_norm=True):
def l2_reg(weight_matrix):
return 0.1 * K.sum(K.square(weight_matrix))
if l2_norm:
l2_first = l2_reg
else:
l2_first = None
# input layer
inputs = keras.layers.Input(shape=(200, 4))
# layer 1
nn = keras.layers.Conv1D(
filters=32,
kernel_size=19,
strides=1,
activation=None,
use_bias=False,
padding='same',
kernel_regularizer=l2_first,
)(inputs)
nn = keras.layers.BatchNormalization()(nn)
activation = utils.activation_fn(activation)
nn = keras.layers.Activation(activation)(nn)
nn = keras.layers.Dropout(0.1)(nn)
# layer 2
nn = layers.conv_layer(
nn,
num_filters=48,
kernel_size=7, #176
padding='same',
activation='relu',
dropout=0.2,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(pool_size=4)(nn)
# layer 3
nn = layers.conv_layer(
nn,
num_filters=96,
kernel_size=7, # 44
padding='valid',
activation='relu',
dropout=0.3,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(pool_size=4)(nn)
# layer 4
nn = layers.conv_layer(
nn,
num_filters=128,
kernel_size=3, # 9
padding='valid',
activation='relu',
dropout=0.4,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(
pool_size=3, # 3
strides=3,
padding='same')(nn)
# layer 5
nn = keras.layers.Flatten()(nn)
nn = layers.dense_layer(nn,
num_units=512,
activation='relu',
dropout=0.5,
l2=1e-6,
bn=True)
# Output layer
logits = keras.layers.Dense(12, activation='linear', use_bias=True)(nn)
outputs = keras.layers.Activation('sigmoid')(logits)
model = keras.Model(inputs=inputs, outputs=outputs)
return model
def model(activation='log_relu', l2_norm=True):
def l2_reg(weight_matrix):
return 0.1 * K.sum(K.square(weight_matrix))
l2 = 1e-6
bn = True
dropout_block0 = 0.1
dropout_block1 = 0.2
dropout_block2 = 0.3
dropout_block3 = 0.4
dropout_block4 = 0.5
if l2_norm:
l2_first = l2_reg
else:
l2_first = None
# input layer
inputs = keras.layers.Input(shape=(200,4))
activation = utils.activation_fn(activation)
# block 1
nn = keras.layers.Conv1D(filters=24,
kernel_size=19,
strides=1,
activation=None,
use_bias=False,
padding='same',
#kernel_initializer=keras.initializers.RandomNormal(mean=0.1, stddev=0.05),
kernel_regularizer=l2_first,
)(inputs)
nn = keras.layers.BatchNormalization()(nn)
activation = utils.activation_fn(activation)
nn = keras.layers.Activation(activation)(nn)
nn = keras.layers.Dropout(dropout_block0)(nn)
nn = layers.conv_layer(nn,
num_filters=32,
kernel_size=7,
padding='same',
activation='relu',
dropout=dropout_block1,
l2=l2,
bn=bn)
nn = keras.layers.MaxPool1D(pool_size=4,
strides=4,
padding='same'
)(nn)
nn = layers.conv_layer(nn,
num_filters=48,
kernel_size=7,
padding='valid',
activation='relu',
dropout=dropout_block2,
l2=l2,
bn=bn)
nn = keras.layers.MaxPool1D(pool_size=4,
strides=4,
padding='same'
)(nn)
# layer 2
nn = layers.conv_layer(nn,
num_filters=64,
kernel_size=3,
padding='valid',
activation='relu',
dropout=dropout_block3,
l2=l2,
bn=bn)
nn = keras.layers.MaxPool1D(pool_size=3,
strides=3,
padding='same'
)(nn)
# Fully-connected NN
nn = keras.layers.Flatten()(nn)
nn = layers.dense_layer(nn, num_units=96, activation='relu', dropout=dropout_block4, l2=l2, bn=bn)
#nn = layers.dense_layer(nn, num_units=512, activation='relu', dropout=dropout_block4, l2=l2, bn=bn)
# Output layer - additive + learned non-linearity
logits = keras.layers.Dense(1, activation='linear', use_bias=True,
kernel_initializer='glorot_normal',
bias_initializer='zeros')(nn)
outputs = keras.layers.Activation('sigmoid')(logits)
model = keras.Model(inputs=inputs, outputs=outputs)
return model
def model(activation='relu'):
# input layer
inputs = keras.layers.Input(shape=(600, 4))
activation = utils.activation_fn(activation)
# layer 1
nn = layers.conv_layer(
inputs,
num_filters=300,
kernel_size=19, # 192
padding='same',
activation=activation,
dropout=0.2,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(pool_size=3)(nn)
# layer 2
nn = layers.conv_layer(
nn,
num_filters=200,
kernel_size=11, # 56
padding='valid',
activation='relu',
dropout=0.2,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(pool_size=4)(nn)
# layer 3
nn = layers.conv_layer(
nn,
num_filters=200,
kernel_size=7, # 56
padding='valid',
activation='relu',
dropout=0.2,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(pool_size=4)(nn)
# layer 4
nn = keras.layers.Flatten()(nn)
nn = layers.dense_layer(nn,
num_units=1000,
activation='relu',
dropout=0.5,
l2=1e-6,
bn=True)
# layer 5
nn = layers.dense_layer(nn,
num_units=1000,
activation='relu',
dropout=0.5,
l2=1e-6,
bn=True)
# Output layer
logits = keras.layers.Dense(164, activation='linear', use_bias=True)(nn)
outputs = keras.layers.Activation('sigmoid')(logits)
model = keras.Model(inputs=inputs, outputs=outputs)
return model
def model(activation='relu', input_shape=200, initialization=None):
if not initialization:
initialization == 'glorot_uniform'
if input_shape == 1000:
multiplier = 2
else:
multiplier = 1
# input layer
inputs = keras.layers.Input(shape=(input_shape,4))
# layer 1
activation = utils.activation_fn(activation)
nn = layers.conv_layer(inputs,
num_filters=32*multiplier,
kernel_size=19, #200
padding='same',
activation=activation,
kernel_initializer=initialization,
dropout=0.1,
l2=1e-6,
bn=True)
# layer 2
nn = layers.conv_layer(nn,
num_filters=48*multiplier,
kernel_size=7, #176
padding='same',
activation='relu',
dropout=0.2,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(pool_size=4)(nn)
# layer 3
nn = layers.conv_layer(nn,
num_filters=96*multiplier,
kernel_size=7, # 44
padding='valid',
activation='relu',
dropout=0.3,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(pool_size=4)(nn)
# layer 4
nn = layers.conv_layer(nn,
num_filters=128*multiplier,
kernel_size=3, # 9
padding='valid',
activation='relu',
dropout=0.4,
l2=1e-6,
bn=True)
nn = keras.layers.MaxPool1D(pool_size=3)(nn)
# layer 5
nn = keras.layers.Flatten()(nn)
nn = layers.dense_layer(nn, num_units=512*multiplier, activation='relu',
dropout=0.5, l2=1e-6, bn=True)
# Output layer
logits = keras.layers.Dense(12, activation='linear', use_bias=True)(nn)
outputs = keras.layers.Activation('sigmoid')(logits)
# compile model
model = keras.Model(inputs=inputs, outputs=outputs)
return model