[tensorflow 手搓DLSTM] python tensorflow2.9.0

import tensorflow as tf
# 用于模型搭建
from tensorflow.keras.models import Model, Sequential
from tensorflow.keras.layers import SimpleRNN, LSTM, GRU, Dropout, Dense, Activation, Input

# 数据处理的辅助工具
from tensorflow.keras import utils

from tensorflow.keras.callbacks import CSVLogger, ReduceLROnPlateau

class DLSTMcell(tf.keras.Model):
    """
    
    return : x_t, c_t, h_t
    """
    def __init__(self, hiddensize=None):
        super().__init__()
        
        self.f = Dense(hiddensize, use_bias=True)
        self.i = Dense(hiddensize, use_bias=True)
        self.c = Dense(hiddensize, use_bias=True)
        self.o = Dense(hiddensize, use_bias=True)
        self.d = Dense(hiddensize, use_bias=True)
        
        
    def call(self, inputs, dinputs, c, h):

        c = c * tf.nn.sigmoid(self.f(tf.concat([inputs, h], axis=1)))
        c = c + tf.nn.sigmoid(self.i(tf.concat([inputs, h], axis=1))) * \
                tf.nn.tanh(self.c(tf.concat([inputs, h], axis=1)))
        
        o = tf.nn.sigmoid(self.o(tf.concat([inputs, h], axis=1)) + self.d(dinputs))
        
        h = tf.nn.tanh(c) * o
        
        return h, c, h
    
class DLSTM(tf.keras.Model):
    """
    
    return : x_t, c_t, h_t
    """
    def __init__(self, batchsize, timesteps, hiddensize=None, return_sequences=True):
        super().__init__()
        self.ts = timesteps
        self.return_sequences = return_sequences
        self.hiddensize = hiddensize
        self.batchsize = batchsize
        
        self.cell = DLSTMcell(hiddensize)
        
    def call(self, inputs, dinputs):
        outx = []
        c = tf.random.uniform([self.batchsize, self.hiddensize])
        h = tf.random.uniform([self.batchsize, self.hiddensize])
        
        for i in range(self.ts):
            _, c, h = self.cell(inputs[:,i,:], dinputs[:,i,:], c, h)
            outx.append(h)
            
        if self.return_sequences:
            return tf.stack(outx, axis=1)
        else:
            return outx[-1]

batchsize = 32
timesteps = 10
featuresize = 4
units = 40

inputs = tf.random.uniform([batchsize, timesteps, featuresize])
dinputs = tf.random.uniform([batchsize, timesteps, featuresize])
model = DLSTM(batchsize, timesteps, units)

out = model(inputs, dinputs)
out.shape

>>>  TensorShape([32, 10, 40])

def dlstm(batchsize, timesteps, featuresize):
    
    inputs = Input(shape=[timesteps, featuresize])
    dinputs = Input(shape=[timesteps, featuresize])
    
    x = DLSTM(batchsize, timesteps, 40, return_sequences=True)(inputs, dinputs)
    x = DLSTM(batchsize, timesteps, 40, return_sequences=True)(x, dinputs)
    x = DLSTM(batchsize, timesteps, 10, return_sequences=False)(x, dinputs)
    x = Dense(featuresize)(x)
    x = Activation("relu")(x)
    
    model = Model([inputs, dinputs], x, name='DLSTMnet')
    
    return model


batchsize = 32
timesteps = 10
featuresize = 4
units = 40

inputs = tf.random.uniform([batchsize, timesteps, featuresize])
dinputs = tf.random.uniform([batchsize, timesteps, featuresize])
model = dlstm(batchsize, timesteps, featuresize)
model([inputs, dinputs]).shape

>>> TensorShape([32, 4])

def train(train_x, train_dx, train_y, test_x, test_y):
    
    model = dlstm()
    adam = tf.keras.optimizers.Adam(learning_rate=0.01)

    # 模型编译，设置模型的相关参数：优化器，损失函数和评价指标
    model.compile(loss='mse', 
                  optimizer=adam, 
                  metrics=['acc'])
    
    
    # 训练模型时每个epoch的参数[epoch,acc,loss,val_acc,val_loss]保存到csv文件中
    # loss:训练集损失值,accuracy:训练集准确率,val_loss:测试集损失值,val_accruacy:测试集准确率
    log = CSVLogger(f"./log.csv", separator=",", append=False)
    # 当度量停止改进时，降低学习率
    reduce = ReduceLROnPlateau(monitor='val_acc', 
                               factor=0.5, 
                               patience=3, 
                               verbose=1,
                               mode='auto', min_delta=0.001, cooldown=0, min_lr=0.001)

    model.fit([train_x, train_dx], 
              train_y, 
              epochs=5, 
              batch_size=32, 
              verbose=1,
              validation_split=0.1,
              callbacks=[log, reduce])
    
    # 模型评估：与sklearn.score方法对应的是tf.keras.evaluate()方法，
    # 返回的是损失值和在compile模型时要求的指标metrics
    loss, acc = model.evaluate(test_x, test_y, verbose=1)
    print('Loss : {}, Accuracy: {}'.format(loss, acc))
    
    
    # model.save("./model.h5")
    return model

model=train(train_seq, train_label,test_seq,test_label)
model.summary()

git clone https://github.com/pyecharts/pyecharts-assets.git

# 安装并激活插件
cd pyecharts-assets

jupyter nbextension install assets

jupyter nbextension enable assets/main