def test_fwbw_convlstm(data, experiment):
print('|-- Run model testing.')
gpu = Config.GPU
params = Config.set_comet_params_fwbw_convlstm()
data_name = Config.DATA_NAME
if 'Abilene' in data_name:
day_size = Config.ABILENE_DAY_SIZE
else:
day_size = Config.GEANT_DAY_SIZE
if not Config.ALL_DATA:
data = data[0:Config.NUM_DAYS * day_size]
print('|--- Splitting train-test set.')
train_data2d, valid_data2d, test_data2d = prepare_train_valid_test_2d(data=data, day_size=day_size)
print('|--- Normalizing the train set.')
if 'Abilene' in data_name:
print('|--- Remove last 3 days in test data.')
test_data2d = test_data2d[0:-day_size * 3]
_, valid_data_normalized2d, test_data_normalized2d, scalers = data_scalling(train_data2d,
valid_data2d,
test_data2d)
input_shape = (Config.FWBW_CONVLSTM_STEP,
Config.FWBW_CONVLSTM_WIDE, Config.FWBW_CONVLSTM_HIGH, Config.FWBW_CONVLSTM_CHANNEL)
with tf.device('/device:GPU:{}'.format(gpu)):
net = load_trained_models(input_shape, Config.FWBW_CONVLSTM_BEST_CHECKPOINT)
run_test(test_data2d, test_data_normalized2d, valid_data_normalized2d[-Config.FWBW_CONVLSTM_STEP:],
net, params, scalers)
return
def train_fwbw_convlstm(data):
print('|-- Run model training.')
params = Config.set_comet_params_fwbw_convlstm()
gpu = Config.GPU
data_name = Config.DATA_NAME
if 'Abilene' in data_name:
day_size = Config.ABILENE_DAY_SIZE
assert Config.FWBW_CONVLSTM_HIGH == 12
assert Config.FWBW_CONVLSTM_WIDE == 12
else:
day_size = Config.GEANT_DAY_SIZE
assert Config.FWBW_CONVLSTM_HIGH == 23
assert Config.FWBW_CONVLSTM_WIDE == 23
print('|--- Splitting train-test set.')
train_data2d, valid_data2d, test_data2d = prepare_train_valid_test_2d(data=data, day_size=day_size)
print('|--- Normalizing the train set.')
train_data_normalized2d, valid_data_normalized2d, _, scalers = data_scalling(np.copy(train_data2d),
np.copy(valid_data2d),
np.copy(test_data2d))
train_data_normalized = np.reshape(np.copy(train_data_normalized2d), newshape=(train_data_normalized2d.shape[0],
Config.FWBW_CONVLSTM_WIDE,
Config.FWBW_CONVLSTM_HIGH))
valid_data_normalized = np.reshape(np.copy(valid_data_normalized2d), newshape=(valid_data_normalized2d.shape[0],
Config.FWBW_CONVLSTM_WIDE,
Config.FWBW_CONVLSTM_HIGH))
input_shape = (Config.FWBW_CONVLSTM_STEP,
Config.FWBW_CONVLSTM_WIDE, Config.FWBW_CONVLSTM_HIGH, Config.FWBW_CONVLSTM_CHANNEL)
with tf.device('/device:GPU:{}'.format(gpu)):
net = build_model(input_shape)
# --------------------------------------------------------------------------------------------------------------
# --------------------------------------------Training fw model-------------------------------------------------
if os.path.isfile(path=net.checkpoints_path + 'weights-{:02d}.hdf5'.format(Config.FWBW_CONVLSTM_N_EPOCH)):
print('|--- Forward model exist! Load model from epoch: {}'.format(Config.FWBW_CONVLSTM_BEST_CHECKPOINT))
net.load_model_from_check_point(_from_epoch=Config.FWBW_CONVLSTM_BEST_CHECKPOINT)
else:
print('|--- Compile model. Saving path %s --- ' % net.saving_path)
# -------------------------------- Create offline training and validating dataset ------------------------------
print('|--- Create offline train set!')
trainX, trainY_1, trainY_2 = create_offline_fwbw_convlstm_data(train_data_normalized,
input_shape, Config.FWBW_CONVLSTM_MON_RAIO,
train_data_normalized.mean(),
4)
print('|--- Create offline valid set!')
validX, validY_1, validY_2 = create_offline_fwbw_convlstm_data(valid_data_normalized,
input_shape, Config.FWBW_CONVLSTM_MON_RAIO,
train_data_normalized.mean(),
1)
# Load model check point
from_epoch = net.load_model_from_check_point()
if from_epoch > 0:
print('|--- Continue training forward model from epoch %i --- ' % from_epoch)
training_fw_history = net.model.fit(x=trainX,
y={'pred_data': trainY_1, 'corr_data': trainY_2},
batch_size=Config.FWBW_CONVLSTM_BATCH_SIZE,
epochs=Config.FWBW_CONVLSTM_N_EPOCH,
callbacks=net.callbacks_list,
validation_data=(
validX, {'pred_data': validY_1, 'corr_data': validY_2}),
shuffle=True,
initial_epoch=from_epoch,
verbose=2)
else:
print('|--- Training new forward model.')
training_fw_history = net.model.fit(x=trainX,
y={'pred_data': trainY_1, 'corr_data': trainY_2},
batch_size=Config.FWBW_CONVLSTM_BATCH_SIZE,
epochs=Config.FWBW_CONVLSTM_N_EPOCH,
callbacks=net.callbacks_list,
validation_data=(
validX, {'pred_data': validY_1, 'corr_data': validY_2}),
shuffle=True,
verbose=2)
# Plot the training history
if training_fw_history is not None:
net.plot_training_history(training_fw_history)
# --------------------------------------------------------------------------------------------------------------
run_test(valid_data2d, valid_data_normalized2d, train_data_normalized2d[-Config.FWBW_CONVLSTM_STEP:],
net, params, scalers)
return