train_predictions = np.multiply(train_predictions + 1, sMax - sMin) / 2 + sMin
train_predictions[train_predictions < 0.0] = 0.0
dev_y = np.multiply(dev_y + 1, sMax - sMin) / 2 + sMin
dev_predictions = np.multiply(dev_predictions + 1, sMax - sMin) / 2 + sMin
dev_predictions[dev_predictions < 0.0] = 0.0
test_y = np.multiply(test_y + 1, sMax - sMin) / 2 + sMin
test_predictions = np.multiply(test_predictions + 1, sMax - sMin) / 2 + sMin
test_predictions[test_predictions < 0.0] = 0.0
dum_pred_results(
path=model_path + MODEL_NAME + '.csv',
train_y=train_y,
train_predictions=train_predictions,
dev_y=dev_y,
dev_predictions=dev_predictions,
test_y=test_y,
test_predictions=test_predictions,
time_cost=time_cost,
)
plot_rela_pred(train_y,
train_predictions,
fig_savepath=model_path + MODEL_NAME + '-TRAIN-PRED.png')
plot_rela_pred(dev_y,
dev_predictions,
fig_savepath=model_path + MODEL_NAME + "-DEV-PRED.png")
plot_rela_pred(test_y,
test_predictions,
fig_savepath=model_path + MODEL_NAME + "-TEST-PRED.png")
plot_error_distribution(test_predictions, test_y,
model_path + MODEL_NAME + '-ERROR-DSTRI.png')
def my_lstm(path,pattern,HU1,DR1,
HL=1,
HU2=8,
DR2=0.0,
LR=0.007,
EPS=1000,
lev=None,
EARLY_STOPING = True,
MODEL_ID=None,
loss='mean_squared_error',
wavelet=None):
if wavelet == None and MODEL_ID==None:
data_path = path + 'data\\'+pattern+'\\'
model_path = path+'projects\\lstm-models-history\\'+pattern+'\\history\\'
elif wavelet==None and MODEL_ID!=None:
data_path = path + 'data\\'+pattern+'\\'
model_path = path+'projects\\lstm-models-history\\'+pattern+'\\history\\s'+str(MODEL_ID)+'\\'
elif wavelet!=None and MODEL_ID==None:
data_path = path + 'data\\'+wavelet+'-'+str(lev)+'\\'+pattern+'\\'
model_path = path+'projects\\lstm-models-history\\'+wavelet+'-'+str(lev)+'\\'+pattern+'\\history\\'
elif wavelet!=None and MODEL_ID!=None:
data_path = path + 'data\\'+wavelet+'-'+str(lev)+'\\'+pattern+'\\'
model_path = path+'projects\\lstm-models-history\\'+wavelet+'-'+str(lev)+'\\'+pattern+'\\history\\s'+str(MODEL_ID)+'\\'
# 1.Import the sampled normalized data set from disk
if MODEL_ID==None:
train = pd.read_csv(data_path+'minmax_unsample_train.csv')
dev = pd.read_csv(data_path+'minmax_unsample_dev.csv')
test = pd.read_csv(data_path+'minmax_unsample_test.csv')
else:
train = pd.read_csv(data_path+'minmax_unsample_train_s'+str(MODEL_ID)+'.csv')
dev = pd.read_csv(data_path+'minmax_unsample_dev_s'+str(MODEL_ID)+'.csv')
test = pd.read_csv(data_path+'minmax_unsample_test_s'+str(MODEL_ID)+'.csv')
# Split features from labels
train_x = train
train_y = train.pop('Y')
train_y = train_y.values
dev_x = dev
dev_y = dev.pop('Y')
dev_y = dev_y.values
test_x = test
test_y = test.pop('Y')
test_y = test_y.values
# reshape the input features for LSTM
train_x = (train_x.values).reshape(train_x.shape[0],1,train_x.shape[1])
dev_x = (dev_x.values).reshape(dev_x.shape[0],1,dev_x.shape[1])
test_x = (test_x.values).reshape(test_x.shape[0],1,test_x.shape[1])
RE_TRAIN = False
WARM_UP = False
# EARLY_STOPING = True
INITIAL_EPOCH = 6000
# For initialize weights and bias
SEED=1
# set hyper-parameters
# EPS=1000 #epochs number
#########--1--###########
# LR=0.007 #learnin rate 0.0001, 0.0003, 0.0007, 0.001, 0.003, 0.007,0.01, 0.03 0.1
#########--2--############
#HU1 = 32 #hidden units for hidden layer 1: [8,16,24,32]
BS = 512 #batch size
#########--3--###########
# HL = 1 #hidden layers
# HU2 = 16 #hidden units for hidden layer 2
DC=0.000 #decay rate of learning rate
#########--4--###########
#DR1=0.7 #dropout rate for hidden layer 1:[0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9]
# DR2=0.0 #dropout rate for hidden layer 2
# 2.Build LSTM model with keras
# set the hyper-parameters
LEARNING_RATE=LR
EPOCHS = EPS
BATCH_SIZE = BS
if HL==2:
HIDDEN_UNITS = [HU1,HU2]
DROP_RATE = [DR1,DR2]
else:
HIDDEN_UNITS = [HU1]
DROP_RATE = [DR1]
DECAY_RATE = DC
if MODEL_ID==None:
MODEL_NAME = 'LSTM-LR['+str(LEARNING_RATE)+']-HU'+str(HIDDEN_UNITS)+'-EPS['+str(EPOCHS)+']-BS['+str(BATCH_SIZE)+']-DR'+str(DROP_RATE)+'-DC['+str(DECAY_RATE)+']-SEED['+str(SEED)+']'
else:
MODEL_NAME = 'LSTM-S'+str(MODEL_ID)+'-LR['+str(LEARNING_RATE)+']-HU'+str(HIDDEN_UNITS)+'-EPS['+str(EPOCHS)+']-BS['+str(BATCH_SIZE)+']-DR'+str(DROP_RATE)+'-DC['+str(DECAY_RATE)+']-SEED['+str(SEED)+']'
# RESUME_TRAINING = True
def build_model():
if HL==2:
model = keras.Sequential(
[
layers.LSTM(HIDDEN_UNITS[0],activation=tf.nn.relu,return_sequences=True,input_shape=(train_x.shape[1],train_x.shape[2])),
layers.Dropout(DROP_RATE[0], noise_shape=None, seed=None),
layers.LSTM(HIDDEN_UNITS[1],activation=tf.nn.relu,return_sequences=False), # first hidden layer if hasnext hidden layer
layers.Dropout(DROP_RATE[1], noise_shape=None, seed=None),
# layers.LSTM(20,activation=tf.nn.relu,return_sequence=True),
layers.Dense(1)
]
)
else:
model = keras.Sequential(
[
layers.LSTM(HIDDEN_UNITS[0],activation=tf.nn.relu,input_shape=(train_x.shape[1],train_x.shape[2])),
layers.Dropout(DROP_RATE[0], noise_shape=None, seed=None),
# layers.LSTM(HIDDEN_UNITS1,activation=tf.nn.relu,return_sequences=True,input_shape=(train_x.shape[1],train_x.shape[2])), # first hidden layer if hasnext hidden layer
# layers.LSTM(20,activation=tf.nn.relu,return_sequence=True),
layers.Dense(1)
]
)
optimizer = keras.optimizers.Adam(LEARNING_RATE,
decay=DECAY_RATE
)
if loss=='mean_squared_error':
print('Loss Function:mean_square_error')
model.compile(
loss='mean_squared_error',
optimizer=optimizer,
metrics=['mean_absolute_error','mean_squared_error'])
elif loss=='custom_loss':
print('Custom Loss Function')
model.compile(
loss=custom_loss,
optimizer=optimizer,
metrics=['mean_absolute_error','mean_squared_error',custom_loss])
return model
# set model's parameters restore path
cp_path = model_path+MODEL_NAME+'\\'
if not os.path.exists(cp_path):
os.makedirs(cp_path)
checkpoint_path = model_path+MODEL_NAME+'\\cp.h5' #restore only the latest checkpoint after every update
# checkpoint_path = model_path+'cp-{epoch:04d}.ckpt' #restore the checkpoint every period=x epoch
checkpoint_dir = os.path.dirname(checkpoint_path)
print('checkpoint dir:{}'.format(checkpoint_dir))
cp_callback = keras.callbacks.ModelCheckpoint(checkpoint_path,save_best_only=True,mode='min',save_weights_only=True,verbose=1)
model = build_model()
model.summary() #print a simple description for the model
"""
# Evaluate before training or load trained weights and biases
loss, mae, mse = model.evaluate(test_x, test_y, verbose=1)
# Try the model with initial weights and biases
example_batch = train_x[:10]
example_result = model.predict(example_batch)
print(example_result)
"""
# 3.Train the model
# Display training progress by printing a single dot for each completed epoch
class PrintDot(keras.callbacks.Callback):
def on_epoch_end(self, epoch, logs):
if epoch % 100 == 0: print('')
print('.', end='')
files = os.listdir(checkpoint_dir)
from tensorflow.keras.callbacks import ReduceLROnPlateau,EarlyStopping
# reduce_lr = ReduceLROnPlateau(monitor='val_loss', patience=10, mode='auto')
reduce_lr = ReduceLROnPlateau(monitor='val_loss',min_lr=0.00001,factor=0.2, verbose=1,patience=10, mode='min')
early_stopping = EarlyStopping(monitor='val_loss', mode='min',verbose=1,patience=100,restore_best_weights=True)
if MODEL_ID==None:
warm_dir = 'LSTM-LR['+str(LEARNING_RATE)+']-HU'+str(HIDDEN_UNITS)+'-EPS['+str(INITIAL_EPOCH)+']-BS['+str(BATCH_SIZE)+']-DR'+str(DROP_RATE)+'-DC['+str(DECAY_RATE)+']-SEED['+str(SEED)+']'
else:
warm_dir = 'LSTM-S'+str(MODEL_ID)+'-LR['+str(LEARNING_RATE)+']-HU'+str(HIDDEN_UNITS)+'-EPS['+str(INITIAL_EPOCH)+']-BS['+str(BATCH_SIZE)+']-DR'+str(DROP_RATE)+'-DC['+str(DECAY_RATE)+']-SEED['+str(SEED)+']'
print(os.path.exists(model_path+warm_dir))
if RE_TRAIN:
print('retrain the model')
if EARLY_STOPING:
history2 = model.fit(train_x,train_y,epochs=EPOCHS,batch_size=BATCH_SIZE ,validation_data=(dev_x,dev_y),verbose=1,
callbacks=[
cp_callback,
early_stopping,
])
else:
history2 = model.fit(train_x,train_y,epochs=EPOCHS,batch_size=BATCH_SIZE ,validation_data=(dev_x,dev_y),verbose=1,callbacks=[cp_callback])
hist2 = pd.DataFrame(history2.history)
hist2.to_csv(model_path+MODEL_NAME+'-HISTORY-TRAIN-TEST.csv')
hist2['epoch']=history2.epoch
# print(hist.tail())
plot_history(history2,model_path+MODEL_NAME+'-MAE-ERRORS-TRAINTEST.png',model_path+MODEL_NAME+'-MSE-ERRORS-TRAINTEST.png')
elif len(files)==0:
if os.path.exists(model_path+warm_dir) and WARM_UP:
print('WARM UP FROM EPOCH '+str(INITIAL_EPOCH))
warm_path=model_path+warm_dir+'\\cp.ckpt'
model.load_weights(warm_path)
if EARLY_STOPING:
history2 = model.fit(train_x,train_y,initial_epoch=INITIAL_EPOCH,epochs=EPOCHS,batch_size=BATCH_SIZE ,validation_data=(dev_x,dev_y),verbose=1,
callbacks=[
cp_callback,
early_stopping,
])
else:
history2 = model.fit(train_x,train_y,initial_epoch=INITIAL_EPOCH,epochs=EPOCHS,batch_size=BATCH_SIZE ,validation_data=(dev_x,dev_y),verbose=1,callbacks=[cp_callback,])
hist2 = pd.DataFrame(history2.history)
hist2.to_csv(model_path+MODEL_NAME+'-HISTORY-TRAIN-TEST.csv')
hist2['epoch']=history2.epoch
# print(hist.tail())
plot_history(history2,model_path+MODEL_NAME+'-MAE-ERRORS-TRAINTEST.png',model_path+MODEL_NAME+'-MSE-ERRORS-TRAINTEST.png')
else:
print('new train')
if EARLY_STOPING:
history2 = model.fit(train_x,train_y,epochs=EPOCHS,batch_size=BATCH_SIZE ,validation_data=(dev_x,dev_y),verbose=1,callbacks=[
cp_callback,
early_stopping,
])
else:
history2 = model.fit(train_x,train_y,epochs=EPOCHS,batch_size=BATCH_SIZE ,validation_data=(dev_x,dev_y),verbose=1,callbacks=[cp_callback,])
hist2 = pd.DataFrame(history2.history)
hist2.to_csv(model_path+MODEL_NAME+'-HISTORY-TRAIN-TEST.csv')
hist2['epoch']=history2.epoch
# print(hist.tail())
plot_history(history2,model_path+MODEL_NAME+'-MAE-ERRORS-TRAINTEST.png',model_path+MODEL_NAME+'-MSE-ERRORS-TRAINTEST.png')
else:
print('#'*10+'Already Trained')
model.load_weights(checkpoint_path)
model.load_weights(checkpoint_path)
# loss, mae, mse = model.evaluate(test_x, test_y, verbose=1)
"""
# Evaluate after training or load trained weights and biases
loss, mae, mse = model.evaluate(test_x, test_y, verbose=1)
print("Testing set Mean Abs Error: {:5.2f} ".format(mae))
"""
# 4. Predict the model
# load the unsample data
train_predictions = model.predict(train_x).flatten()
dev_predictions = model.predict(dev_x).flatten()
test_predictions = model.predict(test_x).flatten()
# plt.figure()
# plt.plot(train_y,c='b')
# plt.plot(train_predictions,c='r')
# plt.show()
# renormized the predictions and labels
# load the normalized traindev indicators
if MODEL_ID==None:
norm = pd.read_csv(data_path+'norm_id.csv')
sMax = norm['series_max'][norm.shape[0]-1]
sMin = norm['series_min'][norm.shape[0]-1]
else:
norm = pd.read_csv(data_path+'norm_id_s'+str(MODEL_ID)+'.csv')
sMax = norm['series_max'][norm.shape[0]-1]
sMin = norm['series_min'][norm.shape[0]-1]
print('Series min:{}'.format(sMin))
print('Series max:{}'.format(sMax))
train_y = np.multiply(train_y + 1,sMax - sMin) / 2 + sMin
train_predictions = np.multiply(train_predictions + 1,sMax - sMin) / 2 + sMin
dev_y = np.multiply(dev_y + 1,sMax - sMin) / 2 + sMin
dev_predictions = np.multiply(dev_predictions + 1,sMax - sMin) / 2 + sMin
test_y = np.multiply(test_y + 1,sMax - sMin) / 2 + sMin
test_predictions = np.multiply(test_predictions + 1,sMax - sMin) / 2 + sMin
print("pattern.find('multi')={}".format(pattern.find('multi')))
print("pattern.find('one')={}".format(pattern.find('one')))
if pattern.find('one')>=0:
print('decomposition ensemble model!!!!!!!!!!!!!!!!!!!!!!!')
train_predictions[train_predictions<0.0]=0.0
dev_predictions[dev_predictions<0.0]=0.0
test_predictions[test_predictions<0.0]=0.0
elif pattern.find('one')<0 and pattern.find('multi')<0:
print('monoscale model$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$')
train_predictions[train_predictions<0.0]=0.0
dev_predictions[dev_predictions<0.0]=0.0
test_predictions[test_predictions<0.0]=0.0
dum_pred_results(
path = model_path+MODEL_NAME+'.csv',
train_y = train_y,
train_predictions=train_predictions,
dev_y = dev_y,
dev_predictions = dev_predictions,
test_y = test_y,
test_predictions = test_predictions)
plot_rela_pred(train_y,train_predictions,fig_savepath=model_path + MODEL_NAME + '-TRAIN-PRED.png')
plot_rela_pred(dev_y,dev_predictions,fig_savepath=model_path + MODEL_NAME + "-DEV-PRED.png")
plot_rela_pred(test_y,test_predictions,fig_savepath=model_path + MODEL_NAME + "-TEST-PRED.png")
plot_error_distribution(test_predictions,test_y,model_path+MODEL_NAME+'-ERROR-DSTRI.png')
plt.close('all')
tf.keras.backend.clear_session()