def sequence_only_cnn(max_residues, padding):
amino_inputs = Input(shape=(23, max_residues + 2 + 2 * padding,
1)) # 22 amino acids plus null/beginning/end
amino_model = Conv2D(400, (23, 5),
kernel_regularizer=l2(.0),
activation='relu')(amino_inputs)
amino_model = Dropout(0.3)(amino_model)
amino_model = Conv2D(200, (1, 9),
kernel_regularizer=l2(.0),
activation='relu')(amino_model)
amino_model = Dropout(0.3)(amino_model)
amino_model = Conv2D(100, (1, 17),
kernel_regularizer=l2(.0),
activation='relu')(amino_model)
amino_model = Dropout(0.3)(amino_model)
model = Flatten()(amino_model)
model = Dense(80, activation='elu', kernel_regularizer=l2(.0))(model)
model = Dropout(0.3)(model)
model = Dense(40, activation='elu', kernel_regularizer=l2(.0))(model)
model = Dense(1, activation='linear', kernel_regularizer=l2(.0))(model)
model = Model(inputs=amino_inputs, outputs=model)
model.compile(optimizer='adam', loss='mse')
th_model = KerasRegressionTwoDimensional(
model=model,
model_author="Jed",
model_description=
'Sequence CNN regressor: 400x5->200x9->100x17->80->40->1',
batch_size=128,
epochs=25)
return th_model
def sequence_only_cnn_classification(max_residues,
padding,
assign_class_weights=False):
amino_inputs = Input(shape=(23, max_residues + 2 + 2 * padding,
1)) # 22 amino acids plus null/beginning/end
amino_model = Conv2D(400, (23, 5),
kernel_regularizer=l2(.0),
activation='relu')(amino_inputs)
amino_model = Dropout(0.3)(amino_model)
amino_model = Conv2D(200, (1, 9),
kernel_regularizer=l2(.0),
activation='relu')(amino_model)
amino_model = Dropout(0.3)(amino_model)
amino_model = Conv2D(100, (1, 17),
kernel_regularizer=l2(.0),
activation='relu')(amino_model)
amino_model = Dropout(0.3)(amino_model)
model = Flatten()(amino_model)
model = Dense(80, activation='elu', kernel_regularizer=l2(.0))(model)
model = Dropout(0.3)(model)
model = Dense(40, activation='elu', kernel_regularizer=l2(.0))(model)
model = Dense(1, activation='sigmoid', kernel_regularizer=l2(.0))(model)
model = Model(inputs=amino_inputs, outputs=model)
model.compile(optimizer='adam',
loss='binary_crossentropy',
metrics=['binary_crossentropy'])
th_model = KerasClassificationTwoDimensional(
model=model,
model_author="Jed and Hamed",
model_description='Sequence CNN classifier v1, assign_class_weights = {}'
.format(assign_class_weights),
batch_size=128,
epochs=25,
assign_class_weights=assign_class_weights)
return th_model
def get_3d_cnn_model_image_only():
inputA = Input(shape=(4, 108, 192, 3))
model = Convolution3D(filters=32,
kernel_size=(3, 3, 3),
activation='relu',
padding='same',
data_format='channels_last')(inputA)
model = MaxPooling3D(pool_size=(2, 2, 2))(model)
model = Flatten()(model)
model = Dense(1, activation='sigmoid')(model)
model = Model(inputs=inputA, outputs=model)
model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy'])
print_summary(model, line_length=120)
return model