Question

我试图通过使用LSTM捕获长期依赖性，方法是每62点创建一个单位脉冲信号。

想法是返回62个时间步长并复制下一个时间步长的值，以便预测脉冲，但是lstm没有这样做...

import sys
import os
import numpy as np
import math
import pandas as pd
from matplotlib import pyplot as plt

from sklearn.preprocessing import MinMaxScaler
from sklearn.model_selection import train_test_split
from sklearn.metrics import mean_squared_error

from keras.models import Sequential, load_model
from keras.layers import LSTM, Dense, Flatten, Dropout
from keras.callbacks import EarlyStopping, ModelCheckpoint
from keras import backend as K
import tensorflow as tf
from tensorflow.python.client import device_lib

K.clear_session() #pulire eventuali sessioni precedenti (cosi i nomi dei layer ripartono da 0)

print(K.tensorflow_backend._get_available_gpus())
print(device_lib.list_local_devices())
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
config = tf.ConfigProto( device_count = {'GPU': 1 , 'CPU': 4} ) 
sess = tf.Session(config=config) 
K.set_session(sess)
# hyper-parametri
params = {
    "batch_size": 20,
    "epochs": 1000,
    "time_steps": 70,
}

OUTPUT_PATH = "/home/..."
TIME_STEPS = params["time_steps"]
BATCH_SIZE = params["batch_size"]

def generate_impulse(dim):
arr = np.zeros(dim)
frequency = 62
for i in range(0, len(arr)):
    if i % frequency == 0:
        arr[i] = 1
return arr
y = generate_impulse(1300)

plt.figure(figsize=(20,5))
plt.plot(y)
plt.title('unit impulse')
plt.ylabel('y')
plt.xlabel('x')
plt.show()

def create_timeseries(arr):
    # Costruzione time series univariata, predict di un single-step. 
    # Prende i primi TIME_STEPS valori come input e calcola il sin del valore TIME_STEPS+1
    dim_0 = len(arr) - TIME_STEPS

    x = np.zeros((dim_0, TIME_STEPS))
    y = np.zeros((dim_0,))

    for i in range(dim_0):
        x[i] = arr[i:TIME_STEPS+i] #TIME_STEPS+i non compreso
        y[i] = arr[TIME_STEPS+i]
        #print(x[i], y[i])
    print("length of time-series i/o",x.shape,y.shape)
    return x, y

x_ts, y_ts = create_timeseries(y)

len_train = int(len(x_ts)*80/100)
len_val = int(len(x_ts)*10/100)
#DATASET DI TRAINING: 80%
x_train = x_ts[0:len_train]
y_train = y_ts[0:len_train]
#DATASET DI VALIDATION: 10%
x_val = x_ts[len_train:len_train+len_val]
y_val = y_ts[len_train:len_train+len_val]
#DATASET DI TEST 10%
x_test = x_ts[len_train+len_val:]
y_test = y_ts[len_train+len_val:]

x_train =x_train.reshape((x_train.shape[0], x_train.shape[1], 1))
x_val =x_val.reshape((x_val.shape[0], x_val.shape[1], 1))
x_test = x_test.reshape(x_test.shape[0], x_test.shape[1], 1)

def create_model():
    model = Sequential()
    model.add(LSTM(1, input_shape=(TIME_STEPS, 1)))

    model.compile(optimizer='adam', loss='mse')
    return model 

es = EarlyStopping(monitor='val_loss', mode='min', verbose=1,
                       patience=50, min_delta=0.0001)
model = create_model()

history = model.fit(x_train, y_train, epochs=params["epochs"], verbose=2, batch_size=BATCH_SIZE, shuffle=False,
                    validation_data=(x_val, y_val), callbacks=[es])

plt.figure()
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('MSE LOSS')
plt.ylabel('Loss')
plt.xlabel('Epochs')
plt.legend(['Train', 'Validation'], loc='upper left')
plt.show()

dataset

y_pred = model.predict(x_test, batch_size=BATCH_SIZE)
y_pred = y_pred.flatten()

error = mean_squared_error(y_test, y_pred)

plt.figure(figsize=(20,5))
plt.plot(y_pred)
plt.plot(y_test)
plt.title('PREDICTION ON TEST SET')
plt.ylabel('sin(x)')
plt.xlabel('x')
plt.legend(['Prediction', 'Real'], loc='upper left')
plt.show()

mse loss

训练集给我相同的结果（它是相同的信号..）。我尝试了其他具有更多神经元的LSTM模型，但是无论如何它还是不起作用。

Answer 1

您可能会考虑训练更多的纪元。我根据您认为的核心思想创建了一个简化的模型和训练集：

from keras.models import Sequential
from keras.layers import LSTM
import numpy as np

TIME_STEPS=10

x_train = np.array([ [ [1],[0],[0],[0],[0],[0],[0],[0],[0],[0] ],
                     [ [0],[1],[0],[0],[0],[0],[0],[0],[0],[0] ],
                     [ [0],[0],[1],[0],[0],[0],[0],[0],[0],[0] ],
                     [ [0],[0],[0],[1],[0],[0],[0],[0],[0],[0] ],
                     [ [0],[0],[0],[0],[1],[0],[0],[0],[0],[0] ],
                     [ [0],[0],[0],[0],[0],[1],[0],[0],[0],[0] ],
                     [ [0],[0],[0],[0],[0],[0],[1],[0],[0],[0] ],
                     [ [0],[0],[0],[0],[0],[0],[0],[1],[0],[0] ],
                     [ [0],[0],[0],[0],[0],[0],[0],[0],[1],[0] ],
                     [ [0],[0],[0],[0],[0],[0],[0],[0],[0],[1] ]])

y_train = np.array([[1],[0],[0],[0],[0],[0],[0],[0],[0],[0]])
print(x_train.shape)
print(y_train.shape)

model = Sequential()
model.add(LSTM(1, input_shape=(TIME_STEPS,1)))
model.compile(optimizer='adam', loss='mse', metrics=['mse'])

model.fit(x_train, y_train, epochs=10000, verbose=0)

训练后，我得到以下预测：

model.predict(x_train)
array([[ 0.9870746 ],
       [ 0.00665453],
       [-0.00303702],
       [ 0.00697759],
       [-0.02432432],
       [-0.00701594],
       [ 0.01387464],
       [ 0.02281112],
       [ 0.00439195],
       [-0.04109564]], dtype=float32)

我不确定它是否可以完全解决您的问题，但可能会为您提供建议的研究方向。我希望这会有所帮助。

LSTM脉冲信号预测

1 个答案: