model = load_model('..\model\\best\\' + model_name + '.h5',
custom_objects={'mean_squared_error': mean_squared_error
, 'root_mean_squared_error': root_mean_squared_error,
'mean_absolute_error': mean_absolute_error})
#
print(len(model.layers))
training_start = 0
training_num = 464
batch_size = 32
epochs = 2000
#
data_preprocess_method = name_list.data_preprocess_method
all_data, testing_data = file_helper_unformatted.load_sstha_for_conv2d(training_start, training_num)
all_data = data_preprocess.data_preprocess(all_data, 0, data_preprocess_method)
data_x, data_y = data_preprocess.sequence_data(all_data, input_length=time_step,
prediction_month=prediction_month)
if is_retrain:
data_x = file_helper_unformatted.exchange_rows(data_x)
data_y_nino = index_calculation.get_nino34_from_data_y(data_y)
# print(data_y_nino)
sc = np.linspace(0, 11, 12, dtype='int32')
sc = np.tile(sc, int((training_num - training_start) / 12 + 1))
data_sc = sc[:(training_num - training_start - time_step + 1 - prediction_month + 1)]
# all_data, testing_data = file_helper_unformatted.load_sstha_for_conv2d(training_start, all_num)
# for layer in model.layers[:]:
# layer.trainable = False
# for layer in model.layers[6:20]:
# layer.trainable = True
beta_1=0.9,
beta_2=0.999,
amsgrad=False)
model.compile(optimizer=adam,
loss=mean_squared_error,
metrics=[
root_mean_squared_error, mean_absolute_error,
mean_squared_error
])
model.summary()
# to train model
# the data for training is ssta and ha
training_data, testing_data = file_helper_unformatted.load_sstha_for_conv2d(
training_start, training_num)
training_data = data_preprocess.data_preprocess(training_data, 0,
data_preprocess_method)
data_x, data_y = data_preprocess.sequence_data(training_data,
input_length=time_step)
print(data_x.shape, data_y.shape)
# sc = np.linspace(0, 11, 12, dtype='int32')
# sc = np.tile(sc, int((training_num-training_start)/12+1))
# data_sc = sc[:(training_num-training_start)]
# tesorboard = TensorBoard('..\..\model\\tensorboard\\' + model_name)
save_best = ModelCheckpoint('..\..\model\\best\\' + model_name + '.h5',
monitor='val_root_mean_squared_error',
verbose=1,
save_best_only=True,
mode='min',
period=1)
train_hist = model.fit(data_x,
data_sc = np.empty([1, 1], dtype='int32')
data_y = np.empty([1, 20, 27, 2])
if is_nino_output:
nino34_output = []
rmse_sst = []
nino_temp = 0
nino34 = []
# The time step for non-sequence prediction should be 1, but it was originally written as 0.
# So it caused the inconsistency.
for start_month in range(
file_num - prediction_month * directly_month, file_num + month -
prediction_month * directly_month + 1 - time_step, interval):
if not is_sequence:
predict_data[0] = file_helper_unformatted.read_data_sstaha(
start_month)
predict_data = data_preprocess.data_preprocess(predict_data, 0)
else:
for i in range(time_step):
predict_data[0][i] = file_helper_unformatted.read_data_sstaha(
start_month + i + 1)
predict_data[0] = data_preprocess.data_preprocess(
predict_data[0], 0)
if is_retrain:
predict_data = file_helper_unformatted.exchange_rows(predict_data)
if model_type == 'conv':
data_x[0] = predict_data[0]
elif model_type == 'dense':
data_x[0] = np.reshape(predict_data[0], (1, 1080))
for i in range(directly_month):