def create_model(input_size, num_classes, batch_size, min_value, max_value,
combining_weight, weight_flag, regularization_weight):
## Build models based on Basset
from tensorflow.keras.layers import Conv1D, Dense, MaxPooling1D, Flatten, Dropout, Activation, BatchNormalization
from tensorflow.keras.models import Sequential
model = Sequential()
model.add(
Conv1D(
filters=300,
kernel_size=19,
padding='same', # , activation='relu'
input_shape=(input_size, 4)))
model.add(Activation('relu'))
model.add(MaxPooling1D(pool_size=3, strides=3, padding='same'))
model.add(BatchNormalization())
model.add(Conv1D(filters=200, kernel_size=11, padding='same'))
model.add(Activation('relu'))
model.add(MaxPooling1D(pool_size=4, strides=4, padding='same'))
model.add(BatchNormalization())
model.add(Conv1D(filters=200, kernel_size=7, padding='same'))
model.add(Activation('relu'))
model.add(MaxPooling1D(pool_size=4, strides=4, padding='same'))
model.add(BatchNormalization())
model.add(Flatten())
model.add(Dense(1000))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(Dropout(0.3))
model.add(Dense(1000))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(Dropout(0.3))
model.add(Dense(num_classes))
opt = tensorflow.keras.optimizers.Adam(lr=learning_rate,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-8,
decay=0,
amsgrad=False)
model.compile(loss=losses.combine_loss_by_sample(combining_weight),
optimizer=opt,
metrics=['mse', losses.pearson_loss, losses.r2_score])
return model
def create_model(input_size, num_classes, learning_rate, combined_loss_weight):
## Build models based on Basset
from tensorflow.keras.layers import Input, Conv1D, Dense, MaxPooling1D, Flatten, Dropout, Activation, BatchNormalization # BatchNormalization(different in keras standalone)
feature_input = Input(shape=(input_size, 4), name='input')
nn = Conv1D(filters=300, kernel_size=19, padding='same',
name='conv_1')(feature_input)
nn = Activation('relu', name='relu_1')(nn)
nn = MaxPooling1D(pool_size=3, strides=3, padding='same',
name='maxpool_1')(nn)
nn = BatchNormalization(name='BN_1')(nn)
nn = Conv1D(filters=200, kernel_size=11, padding='same', name='conv_2')(nn)
nn = Activation('relu', name='relu_2')(nn)
nn = MaxPooling1D(pool_size=4, strides=4, padding='same',
name='maxpool_2')(nn)
nn = BatchNormalization(name='BN_2')(nn)
nn = Conv1D(filters=200, kernel_size=7, padding='same', name='conv_3')(nn)
nn = Activation('relu', name='relu_3')(nn)
nn = MaxPooling1D(pool_size=4, strides=4, padding='same',
name='maxpool_3')(nn)
nn = BatchNormalization(name='BN_3')(nn)
nn = Flatten(name='flatten')(nn)
nn = Dense(1000, name='dense_1')(nn)
nn = Activation('relu', name='relu_4')(nn)
nn = Dropout(0.3, name='dropout_1')(nn)
nn = Dense(1000, name='dense_2')(nn)
nn = Activation('relu', name='relu_5')(nn)
nn = Dropout(0.3, name='dropout_2')(nn)
result = Dense(num_classes, name='dense_out')(nn)
model = tf.keras.models.Model(inputs=feature_input, outputs=result)
opt = tf.keras.optimizers.Adam(lr=learning_rate,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-8,
decay=0,
amsgrad=False)
model.compile(loss=losses.combine_loss_by_sample(combined_loss_weight),
optimizer=opt,
metrics=['mse', losses.pearson_loss])
return model
def create_model(input_size,
num_classes_1,
num_classes_2,
batch_size,
combining_weight,
weight_flag,
ratio=1.0,
percent_share=None): #, input_tag
from tensorflow.keras.layers import Input, Conv1D, Dense, MaxPooling1D, Flatten, Dropout, Activation, BatchNormalization
from tensorflow.keras.layers import Masking, Add, Multiply, RepeatVector, Reshape, Permute, Lambda
conv1_dim, conv2_dim, conv3_dim, fc1_dim, fc2_dim = 300, 200, 200, 1000, 1000
def cross_stitch(inputs):
in_1 = inputs[0]
in_2 = inputs[1]
share_rate = inputs[2]
in_1 = Lambda(lambda x: x * share_rate)(in_1)
in_2 = Lambda(lambda x: x * (1.0 - share_rate))(in_2)
return Add()([in_1, in_2])
inputs = Input(shape=(input_size, 4), name='inputs')
inputs1, inputs2 = inputs, inputs
nn1_in = Conv1D(filters=conv1_dim,
kernel_size=19,
padding='same',
name='conv_1_task1')(inputs1)
nn2_in = Conv1D(filters=conv1_dim,
kernel_size=19,
padding='same',
name='conv_1_task2')(inputs2)
# Add in cross-stitch operation
nn1 = Lambda(cross_stitch, name='cross_stitch_1_task1')(
[nn1_in, nn2_in, percent_share[0]])
nn2 = Lambda(cross_stitch, name='cross_stitch_1_task2')(
[nn2_in, nn1_in, percent_share[0]])
nn1 = Activation('relu', name='relu_1_task1')(nn1)
nn1 = MaxPooling1D(pool_size=3,
strides=3,
padding='same',
name='maxpool_1_task1')(nn1)
nn1 = BatchNormalization(name='BN_1_task1')(nn1)
nn2 = Activation('relu', name='relu_1_task2')(nn2)
nn2 = MaxPooling1D(pool_size=3,
strides=3,
padding='same',
name='maxpool_1_task2')(nn2)
nn2 = BatchNormalization(name='BN_1_task2')(nn2)
nn1_in = Conv1D(filters=conv2_dim,
kernel_size=11,
padding='same',
name='conv_2_task1')(nn1)
nn2_in = Conv1D(filters=conv2_dim,
kernel_size=11,
padding='same',
name='conv_2_task2')(nn2)
nn1 = Lambda(cross_stitch, name='cross_stitch_2_task1')(
[nn1_in, nn2_in, percent_share[1]])
nn2 = Lambda(cross_stitch, name='cross_stitch_2_task2')(
[nn2_in, nn1_in, percent_share[1]])
nn1 = Activation('relu', name='relu_2_task1')(nn1)
nn1 = MaxPooling1D(pool_size=4,
strides=4,
padding='same',
name='maxpool_2_task1')(nn1)
nn1 = BatchNormalization(name='BN_2_task1')(nn1)
nn2 = Activation('relu', name='relu_2_task2')(nn2)
nn2 = MaxPooling1D(pool_size=4,
strides=4,
padding='same',
name='maxpool_2_task2')(nn2)
nn2 = BatchNormalization(name='BN_2_task2')(nn2)
nn1_in = Conv1D(filters=conv3_dim,
kernel_size=7,
padding='same',
name='conv_3_task1')(nn1)
nn2_in = Conv1D(filters=conv3_dim,
kernel_size=7,
padding='same',
name='conv_3_task2')(nn2)
nn1 = Lambda(cross_stitch, name='cross_stitch_3_task1')(
[nn1_in, nn2_in, percent_share[2]])
nn2 = Lambda(cross_stitch, name='cross_stitch_3_task2')(
[nn2_in, nn1_in, percent_share[2]])
nn1 = Activation('relu', name='relu_3_task1')(nn1)
nn1 = MaxPooling1D(pool_size=4,
strides=4,
padding='same',
name='maxpool_3_task1')(nn1)
nn1 = BatchNormalization(name='BN_3_task1')(nn1)
nn2 = Activation('relu', name='relu_3_task2')(nn2)
nn2 = MaxPooling1D(pool_size=4,
strides=4,
padding='same',
name='maxpool_3_task2')(nn2)
nn2 = BatchNormalization(name='BN_3_task2')(nn2)
nn1 = Flatten(name='flatten_task1')(nn1)
nn2 = Flatten(name='flatten_task2')(nn2)
nn1_in = Dense(fc1_dim, name='dense_1_task1')(nn1)
nn2_in = Dense(fc1_dim, name='dense_1_task2')(nn2)
nn1 = Lambda(cross_stitch, name='cross_stitch_4_task1')(
[nn1_in, nn2_in, percent_share[3]])
nn2 = Lambda(cross_stitch, name='cross_stitch_4_task2')(
[nn2_in, nn1_in, percent_share[3]])
nn1 = BatchNormalization(name='BN_4_task1')(nn1)
nn1 = Activation('relu', name='relu_4_task1')(nn1)
nn1 = Dropout(0.3, name='dropout_1_task1')(nn1)
nn2 = BatchNormalization(name='BN_4_task2')(nn2)
nn2 = Activation('relu', name='relu_4_task2')(nn2)
nn2 = Dropout(0.3, name='dropout_1_task2')(nn2)
nn1_in = Dense(fc2_dim, name='dense_2_task1')(nn1)
nn2_in = Dense(fc2_dim, name='dense_2_task2')(nn2)
nn1 = Lambda(cross_stitch, name='cross_stitch_5_task1')(
[nn1_in, nn2_in, percent_share[4]])
nn2 = Lambda(cross_stitch, name='cross_stitch_5_task1')(
[nn2_in, nn1_in, percent_share[4]])
nn1 = BatchNormalization(name='BN_5_task1')(nn1)
nn1 = Activation('relu', name='relu_5_task1')(nn1)
nn1 = Dropout(0.3, name='dropout_2_task1')(nn1)
nn2 = BatchNormalization(name='BN_5_task2')(nn2)
nn2 = Activation('relu', name='relu_5_task2')(nn2)
nn2 = Dropout(0.3, name='dropout_2_task2')(nn2)
## two head version:
labels_mask_1 = Input(shape=([num_classes_1]),
dtype='float32',
name='labels_mask_1')
mask_1 = Masking(mask_value=0.0, name='mask_1')(labels_mask_1)
labels_mask_2 = Input(shape=([num_classes_2]),
dtype='float32',
name='labels_mask_2')
mask_2 = Masking(mask_value=0.0, name='mask_2')(labels_mask_2)
result_1_ori = Dense(num_classes_1, name='dense_out_1_ori')(nn1)
result_2_ori = Dense(num_classes_2, name='dense_out_2_ori')(nn2)
result_1 = Multiply(name='dense_out_1')([result_1_ori, mask_1])
result_2 = Multiply(name='dense_out_2')([result_2_ori, mask_2])
model = tf.keras.models.Model(
inputs=[inputs, labels_mask_1, labels_mask_2],
outputs=[result_1, result_2])
loss_combined = {
'dense_out_1': losses.combine_loss_by_sample(0.01),
'dense_out_2': losses.combine_loss_by_sample(0.0052)
}
lossWeights = {'dense_out_1': ratio, 'dense_out_2': 1.0}
metric = {
'dense_out_1': ['mse', losses.pearson_loss, losses.r2_score],
'dense_out_2': ['mse', losses.pearson_loss, losses.r2_score]
}
opt = tf.keras.optimizers.Adam(lr=learning_rate,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-8,
decay=0,
amsgrad=False)
model.compile(optimizer=opt,
loss=loss_combined,
loss_weights=lossWeights,
metrics=metric)
return model
def create_model(input_size, num_classes_1, num_classes_2, batch_size, combining_weight, weight_flag, ratio=1.0, percent_share=None): #, input_tag
from tensorflow.keras.layers import Input, Conv1D, Dense, MaxPooling1D, Flatten, Dropout, Activation, BatchNormalization
from tensorflow.keras.layers import Masking, Multiply, RepeatVector, Reshape, Permute, Lambda
conv1_dim, conv2_dim, conv3_dim, fc1_dim, fc2_dim = 300, 200, 200, 1000, 1000
inputs = Input(shape=(input_size, 4), name='inputs')
nn = Conv1D(filters=conv1_dim, kernel_size=19, padding='same', name='conv_1')(inputs)
custom_mask_in = Input(shape=([1]), name='custom_mask_in')
# Split into task1 + task2 + share (always = 1, Masking layer does not work on conv, so use Lambda layer)
conv1_mask_base = Reshape((input_size,), name='conv1_mask_base')(RepeatVector(input_size)(custom_mask_in))
conv1_mask_share = RepeatVector(int(conv1_dim * percent_share[0]), name='conv1_share')(conv1_mask_base)
conv1_mask_share = Permute((2, 1), name='conv1_mask_share_reordered')(conv1_mask_share)
conv1_mask_task1 = RepeatVector(int(conv1_dim * round(1.0 - percent_share[0], 10)/2.0), name='conv1_task1')(conv1_mask_base)
conv1_mask_task1_masking = Lambda(lambda x: K.cast(K.equal(x, 1), dtype='float32'), name='conv1_task1_converted')(conv1_mask_task1)
conv1_mask_task1 = Multiply(name='conv1_mask_task1')([conv1_mask_task1, conv1_mask_task1_masking])
conv1_mask_task1 = Permute((2, 1), name='conv1_mask_task1_reordered')(conv1_mask_task1)
conv1_mask_task2 = RepeatVector(int(conv1_dim * round(1.0 - percent_share[0], 10)/2.0), name='conv1_task2')(conv1_mask_base)
conv1_mask_task2_masking = Lambda(lambda x: K.cast(K.equal(x, 2), dtype='float32'), name='conv1_task2_converted')(conv1_mask_task2)
conv1_mask_task2 = Multiply(name='conv1_mask_task2')([conv1_mask_task2, conv1_mask_task2_masking])
conv1_mask_task2 = Permute((2, 1), name='conv1_mask_task2_reordered')(conv1_mask_task2)
conv1_mask = tf.keras.layers.concatenate([conv1_mask_task1, conv1_mask_task2, conv1_mask_share], name='conv1_mask_concatenate')
conv1_mask = Lambda(lambda x: K.cast(K.greater(x, 0), dtype='float32'), name='conv1_mask_normalized')(conv1_mask)
nn = Multiply(name='conv1_branched_1')([nn, conv1_mask])
nn = Activation('relu', name='relu_1')(nn)
nn = MaxPooling1D(pool_size=3, strides=3, padding='same', name='maxpool_1')(nn)
nn = BatchNormalization(name='BN_1')(nn)
nn = Conv1D(filters=conv2_dim, kernel_size=11, padding='same', name='conv_2')(nn)
conv2_mask_base = Reshape((int(nn.shape[1]),), name='conv2_mask_base')(RepeatVector(int(nn.shape[1]))(custom_mask_in))
conv2_mask_share = RepeatVector(int(conv2_dim * percent_share[1]), name='conv2_share')(conv2_mask_base)
conv2_mask_share = Permute((2, 1), name='conv2_mask_share_reordered')(conv2_mask_share)
conv2_mask_task1 = RepeatVector(int(conv2_dim * round(1.0 - percent_share[1], 10)/2.0), name='conv2_task1')(conv2_mask_base)
conv2_mask_task1_masking = Lambda(lambda x: K.cast(K.equal(x, 1), dtype='float32'), name='conv2_task1_converted')(conv2_mask_task1)
conv2_mask_task1 = Multiply(name='conv2_mask_task1')([conv2_mask_task1, conv2_mask_task1_masking])
conv2_mask_task1 = Permute((2, 1), name='conv2_mask_task1_reordered')(conv2_mask_task1)
conv2_mask_task2 = RepeatVector(int(conv2_dim * round(1.0 - percent_share[1], 10)/2.0), name='conv2_task2')(conv2_mask_base)
conv2_mask_task2_masking = Lambda(lambda x: K.cast(K.equal(x, 2), dtype='float32'), name='conv2_task2_converted')(conv2_mask_task2)
conv2_mask_task2 = Multiply(name='conv2_mask_task2')([conv2_mask_task2, conv2_mask_task2_masking])
conv2_mask_task2 = Permute((2, 1), name='conv2_mask_task2_reordered')(conv2_mask_task2)
conv2_mask = tf.keras.layers.concatenate([conv2_mask_task1, conv2_mask_task2, conv2_mask_share], name='conv2_mask_concatenate')
conv2_mask = Lambda(lambda x: K.cast(K.greater(x, 0), dtype='float32'), name='conv2_mask_normalized')(conv2_mask)
nn = Multiply(name='conv2_branched_1')([nn, conv2_mask])
nn = Activation('relu', name='relu_2')(nn)
nn = MaxPooling1D(pool_size=4, strides=4, padding='same', name='maxpool_2')(nn)
nn = BatchNormalization(name='BN_2')(nn)
nn = Conv1D(filters=conv3_dim, kernel_size=7, padding='same', name='conv_3')(nn)
conv3_mask_base = Reshape((int(nn.shape[1]),), name='conv3_mask_base')(RepeatVector(int(nn.shape[1]))(custom_mask_in))
conv3_mask_share = RepeatVector(int(conv3_dim * percent_share[2]), name='conv3_share')(conv3_mask_base)
conv3_mask_share = Permute((2, 1), name='conv3_mask_share_reordered')(conv3_mask_share)
conv3_mask_task1 = RepeatVector(int(conv3_dim * round(1.0 - percent_share[2], 10)/2.0), name='conv3_task1')(conv3_mask_base)
conv3_mask_task1_masking = Lambda(lambda x: K.cast(K.equal(x, 1), dtype='float32'), name='conv3_task1_converted')(conv3_mask_task1)
conv3_mask_task1 = Multiply(name='conv3_mask_task1')([conv3_mask_task1, conv3_mask_task1_masking])
conv3_mask_task1 = Permute((2, 1), name='conv3_mask_task1_reordered')(conv3_mask_task1)
conv3_mask_task2 = RepeatVector(int(conv3_dim * round(1.0 - percent_share[2], 10)/2.0), name='conv3_task2')(conv3_mask_base)
conv3_mask_task2_masking = Lambda(lambda x: K.cast(K.equal(x, 2), dtype='float32'), name='conv3_task2_converted')(conv3_mask_task2)
conv3_mask_task2 = Multiply(name='conv3_mask_task2')([conv3_mask_task2, conv3_mask_task2_masking])
conv3_mask_task2 = Permute((2, 1), name='conv3_mask_task2_reordered')(conv3_mask_task2)
conv3_mask = tf.keras.layers.concatenate([conv3_mask_task1, conv3_mask_task2, conv3_mask_share], name='conv3_mask_concatenate')
conv3_mask = Lambda(lambda x: K.cast(K.greater(x, 0), dtype='float32'), name='conv3_mask_normalized')(conv3_mask)
nn = Multiply(name='conv3_branched_1')([nn, conv3_mask])
nn = Activation('relu', name='relu_3')(nn)
nn = MaxPooling1D(pool_size=4, strides=4, padding='same', name='maxpool_3')(nn)
nn = BatchNormalization(name='BN_3')(nn)
nn = Flatten(name='flatten')(nn)
# Split into task1 + task2 + share
fc1_mask_share = Reshape((int(fc1_dim * percent_share[3]),), name='fc1_share')(RepeatVector(int(fc1_dim * 0.65))(custom_mask_in))
fc1_mask_task1 = Reshape((int(fc1_dim * round(1.0 - percent_share[3], 10)/2.0),), name='fc1_task1')(RepeatVector(int(fc1_dim * 0.175))(custom_mask_in))
fc1_mask_task1 = Masking(mask_value=2, name='fc1_mask_task1')(fc1_mask_task1)
fc1_mask_task2 = Reshape((int(fc1_dim * round(1.0 - percent_share[3], 10)/2.0),), name='fc1_task2')(RepeatVector(int(fc1_dim * 0.175))(custom_mask_in))
fc1_mask_task2 = Masking(mask_value=1, name='fc1_mask_task2')(fc1_mask_task2)
fc1_mask = tf.keras.layers.concatenate([fc1_mask_task1, fc1_mask_task2, fc1_mask_share], name='fc1_mask_concatenate')
fc1_mask = Lambda(lambda x: K.cast(K.greater(x, 0), dtype='float32'), name='fc1_mask_normalized')(fc1_mask)
nn = Dense(fc1_dim, name='dense_1')(nn)
nn = Multiply(name='fc1_branched_1')([nn, fc1_mask])
nn = BatchNormalization(name='BN_4')(nn)
nn = Activation('relu', name='relu_4')(nn)
nn = Dropout(0.3, name='dropout_1')(nn)
# Split into task1 + task2 + share
fc2_mask_share = Reshape((int(fc2_dim * percent_share[4]),), name='fc2_share')(RepeatVector(int(fc2_dim * 0.6))(custom_mask_in))
fc2_mask_task1 = Reshape((int(fc2_dim * round(1.0 - percent_share[4], 10)/2.0),), name='fc2_task1')(RepeatVector(int(fc2_dim * 0.2))(custom_mask_in))
fc2_mask_task1 = Masking(mask_value=2, name='fc2_mask_task1')(fc2_mask_task1)
fc2_mask_task2 = Reshape((int(fc2_dim * round(1.0 - percent_share[4], 10)/2.0),), name='fc2_task2')(RepeatVector(int(fc2_dim * 0.2))(custom_mask_in))
fc2_mask_task2 = Masking(mask_value=1, name='fc2_mask_task2')(fc2_mask_task2)
fc2_mask = tf.keras.layers.concatenate([fc2_mask_task1, fc2_mask_task2, fc2_mask_share], name='fc2_mask_concatenate')
fc2_mask = Lambda(lambda x: K.cast(K.greater(x, 0), dtype='float32'), name='fc2_mask_normalized')(fc2_mask)
nn = Dense(fc2_dim, name='dense_2')(nn)
nn = Multiply(name='fc2_branched_1')([nn, fc2_mask])
nn = BatchNormalization(name='BN_5')(nn)
nn = Activation('relu', name='relu_5')(nn)
nn = Dropout(0.3, name='dropout_2')(nn)
## tower head by task:
labels_mask_1 = Input(shape=([num_classes_1]), dtype='float32', name='labels_mask_1')
mask_1 = Masking(mask_value=0.0, name='mask_1')(labels_mask_1)
labels_mask_2 = Input(shape=([num_classes_2]), dtype='float32', name='labels_mask_2')
mask_2 = Masking(mask_value=0.0, name='mask_2')(labels_mask_2)
result_1_ori = Dense(num_classes_1, name='dense_out_1_ori')(nn)
result_2_ori = Dense(num_classes_2, name='dense_out_2_ori')(nn)
result_1 = Multiply(name='dense_out_1')([result_1_ori, mask_1])
result_2 = Multiply(name='dense_out_2')([result_2_ori, mask_2])
model = tf.keras.models.Model(inputs=[inputs, labels_mask_1, labels_mask_2, custom_mask_in],
outputs=[result_1, result_2])
loss_combined = {
'dense_out_1': losses.combine_loss_by_sample(0.01),
'dense_out_2': losses.combine_loss_by_sample(0.0052)
}
lossWeights = {'dense_out_1': ratio, 'dense_out_2': 1.0}
metric = {
'dense_out_1': ['mse', losses.pearson_loss, losses.r2_score],
'dense_out_2': ['mse', losses.pearson_loss, losses.r2_score]
}
opt = tf.keras.optimizers.Adam(lr=learning_rate, beta_1=0.9, beta_2=0.999, epsilon=1e-8, decay=0, amsgrad=False)
model.compile(optimizer=opt,
loss=loss_combined,
loss_weights=lossWeights,
metrics=metric)
return model
def create_model(input_size,
num_classes_1,
num_classes_2,
batch_size,
combining_weight,
weight_flag,
ratio=1.0,
percent_share=None):
from tensorflow.keras.layers import Input, Conv1D, Dense, MaxPooling1D, Flatten, Dropout, Activation, BatchNormalization
from tensorflow.keras.layers import Concatenate, Masking, Multiply, RepeatVector, Reshape, Permute, Lambda
inputs = Input(shape=(input_size, 4), name='inputs')
nn = Conv1D(filters=300, kernel_size=19, padding='same',
name='conv_1')(inputs)
nn = Activation('relu', name='relu_1')(nn)
nn = MaxPooling1D(pool_size=3, strides=3, padding='same',
name='maxpool_1')(nn)
nn = BatchNormalization(name='BN_1')(nn)
nn = Conv1D(filters=200, kernel_size=11, padding='same', name='conv_2')(nn)
nn = Activation('relu', name='relu_2')(nn)
nn = MaxPooling1D(pool_size=4, strides=4, padding='same',
name='maxpool_2')(nn)
nn = BatchNormalization(name='BN_2')(nn)
nn = Conv1D(filters=200, kernel_size=7, padding='same', name='conv_3')(nn)
nn = Activation('relu', name='relu_3')(nn)
nn = MaxPooling1D(pool_size=4, strides=4, padding='same',
name='maxpool_3')(nn)
nn = BatchNormalization(name='BN_3')(nn)
nn = Flatten(name='flatten')(nn)
nn = Dense(1000, name='dense_1')(nn)
nn = BatchNormalization(name='BN_4')(nn)
nn = Activation('relu', name='relu_4')(nn)
nn = Dropout(0.3, name='dropout_1')(nn)
nn = Dense(1000, name='dense_2')(nn)
nn = BatchNormalization(name='BN_5')(nn)
nn = Activation('relu', name='relu_5')(nn)
nn = Dropout(0.3, name='dropout_2')(nn)
## two head version:
labels_mask_1 = Input(shape=([num_classes_1]),
dtype='float32',
name='labels_mask_1')
mask_1 = Masking(mask_value=0.0, name='mask_1')(labels_mask_1)
labels_mask_2 = Input(shape=([num_classes_2]),
dtype='float32',
name='labels_mask_2')
mask_2 = Masking(mask_value=0.0, name='mask_2')(labels_mask_2)
result_1_ori = Dense(num_classes_1, name='dense_out_1_ori')(nn)
result_2_ori = Dense(num_classes_2, name='dense_out_2_ori')(nn)
result_1 = Multiply(name='dense_out_1')([result_1_ori, mask_1])
result_2 = Multiply(name='dense_out_2')([result_2_ori, mask_2])
model = tf.keras.models.Model(
inputs=[inputs, labels_mask_1, labels_mask_2],
outputs=[result_1, result_2])
loss_combined = {
'dense_out_1': losses.combine_loss_by_sample(0.01),
'dense_out_2': losses.combine_loss_by_sample(0.0052)
}
lossWeights = {'dense_out_1': ratio, 'dense_out_2': 1.0}
metric = {
'dense_out_1': ['mse', losses.pearson_loss, losses.r2_score],
'dense_out_2': ['mse', losses.pearson_loss, losses.r2_score]
}
opt = tf.keras.optimizers.Adam(lr=learning_rate,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-8,
decay=0,
amsgrad=False) # decay=1e-6,
model.compile(optimizer=opt,
loss=loss_combined,
loss_weights=lossWeights,
metrics=metric)
return model
model.outputs = [model.layers[-1].output]
from tensorflow.keras.layers import Dense
model.add(Dense(num_classes))
if TL_manner == 'frozen':
for layer in model.layers[:-1]:
print(layer.name)
layer.trainable = False
model.summary()
opt = tensorflow.keras.optimizers.Adam(lr=learning_rate,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-8,
decay=0,
amsgrad=False)
model.compile(loss=losses.combine_loss_by_sample(combining_weight),
optimizer=opt,
metrics=['mse', losses.pearson_loss, losses.r2_score])
checkpoint_path_weights = output_directory + 'cp.ckpt'
## Create checkpoint callback
cp_callback = tensorflow.keras.callbacks.ModelCheckpoint(
checkpoint_path_weights,
save_weights_only=True,
save_best_only=True,
verbose=1)
tensorboard = tensorflow.keras.callbacks.TensorBoard(log_dir=os.path.join(
output_directory, 'logs', '{}'.format(time())),
histogram_freq=1,