自定义loss_lstm.zip

# 导入包 from math import sqrt from keras.activations import sigmoid from keras.metrics import accuracy from keras.optimizer_v1 import SGD from numpy import concatenate from matplotlib import pyplot from pandas import read_csv from pandas import DataFrame from pandas import concat from sklearn.preprocessing import MinMaxScaler from sklearn.preprocessing import LabelEncoder from sklearn.metrics import mean_squared_error from keras.models import Sequential from keras.layers import Dense, Activation from keras.layers import LSTM import pandas as pd import numpy as np import openpyxl from sklearn import metrics from sklearn.metrics import r2_score import tensorflow as tf from sklearn.metrics import accuracy_score from sklearn import metrics # import eli5 # from eli5.sklearn import PermutationImportance from keras.layers import Input, Embedding, LSTM, Dense from keras.models import Model from keras import backend as K # 1、数据预处理 # 1.1 定义有监督数据 def series_to_supervised(data, n_in=1, n_out=1, dropnan=True): n_vars = 1 if type(data) is list else data.shape[1] df = DataFrame(data) cols, names = [], [] # i: n_in, n_in-1, ..., 1，为滞后期数 # 分别代表t-n_in, ... ,t-1期 for i in range(n_in, 0, -1): cols.append(df.shift(i)) names += [('var%d(t-%d)' % (j + 1, i)) for j in range(n_vars)] # i: 0, 1, ..., n_out-1，为超前预测的期数 # 分别代表t，t+1， ... ,t+n_out-1期 for i in range(0, n_out): cols.append(df.shift(-i)) if i == 0: names += [('var%d(t)' % (j + 1)) for j in range(n_vars)] else: names += [('var%d(t+%d)' % (j + 1, i)) for j in range(n_vars)] agg = concat(cols, axis=1) agg.columns = names if dropnan: agg.dropna(inplace=True) return agg # 1.2 定义前瞻数，准备数据 def prepare_data(dataset, n_in, n_out, n_vars=15, train_proportion=0.8): # 读取数据集 values = dataset.values # 保证所有数据都是float32类型 values = values.astype('float32') # 变量归一化 scaler = MinMaxScaler(feature_range=(0, 1)) scaled = scaler.fit_transform(values) print('缩放') print(scaled[0:5]) # 将时间序列问题转化为监督学习问题 reframed = series_to_supervised(scaled, n_in, n_out) print('有监督') print(reframed[0:5]) # 取出保留的变量，t之前的有4个变量，t(包括t)之后的只有第一列，也就是预测列 contain_vars = [] for i in range(1, n_in + 1): contain_vars += [('var%d(t-%d)' % (j, i)) for j in range(1, n_vars + 1)] data = reframed[contain_vars + ['var1(t)'] + [('var1(t+%d)' % (j)) for j in range(1, n_out)]] print(data[0:5]) # 修改列名 ''' col_names = ['Y', 'X1', 'X2', 'X3'] contain_vars = [] for i in range(n_vars): contain_vars += [('%s(t-%d)' % (col_names[i], j)) for j in range(1, n_in + 1)] ''' col_names = ['Y', 'X1', 'X2', 'X3', 'X4', 'X5', 'X6', 'X7', 'X8', 'X9', 'X10', 'X11', 'X12', 'X13', 'X14'] contain_vars = [] for i in range(1, n_in + 1): contain_vars += [('%s(t-%d)' % (col_names[j], i)) for j in range(n_vars)] data.columns = contain_vars + ['Y(t)'] + [('Y(t+%d)' % (j)) for j in range(1, n_out)] print('改列名后\n', data[0:5]) # 分隔数据集，分为训练集和测试集 values = data.values n_train = round(data.shape[0] * train_proportion) train = values[:n_train, :] test = values[n_train:, :] # 分隔输入X和输出y train_X, train_y = train[:, :n_in * n_vars], train[:, n_in * n_vars:] test_X, test_y = test[:, :n_in * n_vars], test[:, n_in * n_vars:] print(train_X.shape) print(train_y.shape) print(test_X.shape) print(test_y.shape) # 将输入X改造为LSTM的输入格式，即[samples,timesteps,features] train_X = train_X.reshape((train_X.shape[0], n_in, n_vars)) # 原来是n_vars test_X = test_X.reshape((test_X.shape[0], n_in, n_vars)) print(train_X.shape) return scaler, data, train_X, train_y, test_X, test_y, dataset # 1.4 训练模型 def fit_lstm(data_prepare, n_neurons, n_batch, n_epoch=500, lr=0.0001, repeats=1): train_X = data_prepare[2] train_y = data_prepare[3] test_X = data_prepare[4] test_y = data_prepare[5] scaler = data_prepare[0] model_list = [] for i in range(repeats): # 设计神经网络 model = Sequential() model.add(LSTM(n_neurons, input_shape=(train_X.shape[1], train_X.shape[2]))) # (时间步长，原始变量数) # model.add(LSTM(n_neurons, return_sequences=True, input_shape=(train_X.shape[1], train_X.shape[2])))# (时间步长，原始变量数) # model.add(LSTM(n_neuron_2)) model.add(Dense(train_y.shape[1], activation='sigmoid')) # 全连接层预测的列数 # model.add(Activation(sigmoid)) # loss = tf.keras.losses.BinaryCrossentropy(from_logits=False, label_smoothing=0.0, axis=-1, reduction="auto", name="binary_crossentropy",) optimizer = tf.keras.optimizers.Adam(learning_rate=lr) BCELoss = tf.keras.losses.BinaryCrossentropy() # model.compile(optimizer=optimizer, loss=myloss, metrics=[accuracy]) # 拟合神经网络 losses = [] for epoch in range(n_epoch): index = np.random.choice(len(train_X), n_batch, replace=False) x = train_X[index] y = train_y[index] e = 0.1 '''with tf.GradientTape() as tape: hat_ys = model(x) z = y*(4292 + 2686)-2686 #将y进行反归一化 q = tf.clip_by_value(1/(tf.abs(z)-30), 0.001, 1) #在（-30，30）之间增加权重 mse = tf.reduce_mean(tf.square(y - hat_ys)*q)/2 #增加权重后的mseloss # mse = tf.reduce_mean( # tf.square(y - hat_ys)) / 2 #没有权重的mse # bce = tf.abs(hat_zs-zs) # bce = tf.reduce_mean(bce)#原来是交叉熵公式，但由于张量问题无法运行，该公式要改 zs = tf.cast(y*(4292 + 2686)-2686>-30, dtype=tf.float32) hat_zs = tf.cast(hat_zs,dtype=tf.float32) #print(f'zs:{zs}') #print(f'hat_zs:{hat_zs}') # print(zs.shape(),hat_zs.shape()) celoss = CELoss(zs,hat_zs) ''' hat_ys = model(x) # z = y * (4292 + 2686) - 2686 # 将y进行反归一化 # q = tf.clip_by_value(1/(tf.abs(z)-30), 0.001, 1) #在（-30，30）之间增加权重 # mse = tf.reduce_mean( # tf.square(y - hat_ys)*q)/2 #增加权重后的mseloss mse = tf.reduce_mean( tf.square(y - hat_ys)) / 2 # 没有权重的mse # 定义转换标签函数 def exchange(hat_ys): hat_zs = [] for hat_y in hat_ys: if hat_y > -30 and hat_y < 30: hat_z = 1 else: hat_z = -1 hat_zs.append(hat_z) return hat_zs zs = exchange(y * (4292 + 2686) - 2686) # 设置真实标签 print(zs) hat_ys = tf.reshape(hat_ys, [1, 128])[0] print(hat_ys) hat_zs = exchange(hat_ys * (4292 + 2686) - 2686) # zs = tf.where( # tf.logical_and( # tf.greater(y, (-30 + 2686) / (4292 + 2686)), # tf.less(y, (30 + 2686) / (4292 + 2686)) # ) # , 1, 0) # 设置预测值标签 # print(zs) # hat_zs = tf.where( # tf.logical_and(