#
with open('%s/fold_%s_model.json' % (modeldir, k_count),
'r') as json_file:
loaded_model_json = json_file.read()
loaded_model = models.model_from_json(
loaded_model_json) # keras.models.model_from_yaml(yaml_string)
loaded_model.load_weights(filepath=filepth)
# loaded_model.compile(optimizer='rmsprop', # SGD,adam,rmsprop
# loss='mse',
# metrics=['mae',pearson_r]) # mae平均绝对误差（mean absolute error） accuracy)
#
# Test model
#
# pearson_coeff, std, mae = test_report_reg(model, x_test, ddg_test)
pearson_coeff, std, mae = test_report_reg(loaded_model, x_test,
ddg_test)
print('\n----------Predict:'
'\npearson_coeff: %s, std: %s, mae: %s' %
(pearson_coeff, std, mae))
score_dict['pearson_coeff'].append(pearson_coeff)
score_dict['std'].append(std)
score_dict['mae'].append(mae)
train_score_dict['pearson_coeff'].append(history_dict['pearson_r'][-1])
train_score_dict['std'].append(history_dict['rmse'][-1])
train_score_dict['mae'].append(history_dict['mean_absolute_error'][-1])
es_train_score_dict['pearson_coeff'].append(
history_dict['pearson_r'][-11])
es_train_score_dict['std'].append(history_dict['rmse'][-11])
es_train_score_dict['mae'].append(
# validation_mae = np.amax(result.history['val_mean_absolute_error'])
# print('Best validation mae of epoch:', validation_mae)
# return {'loss': validation_mae, 'status': STATUS_OK}
if __name__ == '__main__':
import sys
neighbor_obj, CUDA_rate = sys.argv[1:]
## config TF
#os.environ['CUDA_VISIBLE_DEVICES'] = '0'
from mCNN.queueGPU import queueGPU
queueGPU()
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
if CUDA_rate != 'full':
config = tf.ConfigProto()
if float(CUDA_rate) < 0.1:
config.gpu_options.allow_growth = True
else:
config.gpu_options.per_process_gpu_memory_fraction = float(
CUDA_rate)
set_session(tf.Session(config=config))
x_train, y_train, ddg_train, x_test, y_test, ddg_test, class_weights_dict, obj, kneighbor = data(
neighbor_obj)
model = Conv2DRegressorIn1(x_train, y_train, ddg_train, x_test, y_test,
ddg_test, class_weights_dict, obj, kneighbor)
pearson_coeff, std = test_report_reg(model, x_test, ddg_test)
print('\n--reg_pearson_coeff_%s: %s'
'\n--reg_std_%s: %s' % (kneighbor, pearson_coeff, kneighbor, std))
def Conv2DRegressorIn1(x_train, y_train, ddg_train, x_test, y_test, ddg_test,
class_weights_dict, obj):
K.clear_session()
summary = True
verbose = 0
# setHyperParams------------------------------------------------------------------------------------------------
batch_size = 64
epochs = {{choice([50, 100, 150, 200, 250])}}
lr = {{loguniform(np.log(1e-4), np.log(1e-2))}}
optimizer = {{choice(['adam', 'sgd', 'rmsprop'])}}
activator = {{choice(['elu', 'relu', 'tanh'])}}
basic_conv2D_layers = {{choice([1, 2])}}
basic_conv2D_filter_num = {{choice([16, 32])}}
loop_dilation2D_layers = {{choice([2, 4, 6])}}
loop_dilation2D_filter_num = {{choice([16, 32, 64])}} #used in the loop
loop_dilation2D_dropout_rate = {{uniform(0.001, 0.35)}}
dilation_lower = 2
dilation_upper = 16
reduce_layers = 5 # conv 3 times: 120 => 60 => 30 => 15
reduce_conv2D_filter_num = {{choice([8, 16,
32])}} #used for reduce dimention
reduce_conv2D_dropout_rate = {{uniform(0.001, 0.25)}}
residual_stride = 2
dense1_num = {{choice([64, 128, 256])}}
dense2_num = {{choice([32, 64])}}
drop_num = {{uniform(0.0001, 0.3)}}
kernel_size = (3, 3)
pool_size = (2, 2)
initializer = 'random_uniform'
padding_style = 'same'
loss_type = 'mse'
metrics = ('mae', )
my_callbacks = [
callbacks.ReduceLROnPlateau(
monitor='val_loss',
factor=0.8,
patience=10,
)
]
if lr > 0:
if optimizer == 'adam':
chosed_optimizer = optimizers.Adam(lr=lr)
elif optimizer == 'sgd':
chosed_optimizer = optimizers.SGD(lr=lr)
elif optimizer == 'rmsprop':
chosed_optimizer = optimizers.RMSprop(lr=lr)
# build --------------------------------------------------------------------------------------------------------
## basic Conv2D
input_layer = Input(shape=x_train.shape[1:])
y = layers.Conv2D(basic_conv2D_filter_num,
kernel_size,
padding=padding_style,
kernel_initializer=initializer,
activation=activator)(input_layer)
y = layers.BatchNormalization(axis=-1)(y)
if basic_conv2D_layers == 2:
y = layers.Conv2D(basic_conv2D_filter_num,
kernel_size,
padding=padding_style,
kernel_initializer=initializer,
activation=activator)(y)
y = layers.BatchNormalization(axis=-1)(y)
## loop with Conv2D with dilation (padding='same')
for _ in range(loop_dilation2D_layers):
y = layers.Conv2D(loop_dilation2D_filter_num,
kernel_size,
padding=padding_style,
dilation_rate=dilation_lower,
kernel_initializer=initializer,
activation=activator)(y)
y = layers.BatchNormalization(axis=-1)(y)
y = layers.Dropout(loop_dilation2D_dropout_rate)(y)
dilation_lower *= 2
if dilation_lower > dilation_upper:
dilation_lower = 2
## Conv2D with dilation (padding='valaid') and residual block to reduce dimention.
for _ in range(reduce_layers):
y = layers.Conv2D(reduce_conv2D_filter_num,
kernel_size,
padding=padding_style,
kernel_initializer=initializer,
activation=activator)(y)
y = layers.BatchNormalization(axis=-1)(y)
y = layers.Dropout(reduce_conv2D_dropout_rate)(y)
y = layers.MaxPooling2D(pool_size, padding=padding_style)(y)
residual = layers.Conv2D(reduce_conv2D_filter_num,
1,
strides=residual_stride,
padding='same')(input_layer)
y = layers.add([y, residual])
residual_stride *= 2
## flat & dense
y = layers.Flatten()(y)
y = layers.Dense(dense1_num, activation=activator)(y)
y = layers.BatchNormalization(axis=-1)(y)
y = layers.Dropout(drop_num)(y)
y = layers.Dense(dense2_num, activation=activator)(y)
y = layers.BatchNormalization(axis=-1)(y)
y = layers.Dropout(drop_num)(y)
output_layer = layers.Dense(1)(y)
model = models.Model(inputs=input_layer, outputs=output_layer)
if summary:
model.summary()
model.compile(
optimizer=chosed_optimizer,
loss=loss_type,
metrics=list(metrics) # accuracy
)
K.set_session(tf.Session(graph=model.output.graph))
init = K.tf.global_variables_initializer()
K.get_session().run(init)
result = model.fit(
x=x_train,
y=ddg_train,
batch_size=batch_size,
epochs=epochs,
verbose=verbose,
callbacks=my_callbacks,
validation_data=(x_test, ddg_test),
shuffle=True,
)
# print('\n----------History:\n%s'%result.history)
if obj == 'test_report_reg':
pearson_coeff, std = test_report_reg(model, x_test, ddg_test)
print('\n----------Predict:\npearson_coeff: %s, std: %s' %
(pearson_coeff, std))
objective = pearson_coeff * 2 + std
return {'loss': -objective, 'status': STATUS_OK}
elif obj == 'val_mae':
validation_mae = np.amax(result.history['val_mean_absolute_error'])
print('Best validation mae of epoch:', validation_mae)
return {'loss': validation_mae, 'status': STATUS_OK}
