algorithm_platform/models_tensorflow/model_cnn_lstm_cdq.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
# time: 2024/5/6 13:25
# file: time_series.py
# author: David
# company: shenyang JY
import os.path
from keras.layers import Input, Dense, LSTM, concatenate, Conv1D, Conv2D, MaxPooling1D, BatchNormalization, Flatten, Dropout
from keras.models import Model, load_model
from keras.callbacks import ModelCheckpoint, EarlyStopping, TensorBoard
from keras import optimizers, regularizers
import keras.backend as K
import numpy as np
from sqlalchemy.ext.instrumentation import find_native_user_instrumentation_hook

np.random.seed(42)
from cache.sloss import SouthLoss, NorthEastLoss
import tensorflow as tf
tf.compat.v1.set_random_seed(1234)
from threading import Lock
model_lock = Lock()

def rmse(y_true, y_pred):
    return K.sqrt(K.mean(K.square(y_pred - y_true)))


var_dir = os.path.dirname(os.path.dirname(__file__))


class FMI(object):
    model = None
    train = False

    def __init__(self, log, args, graph, sess):
        self.logger = log
        self.graph = graph
        self.sess = sess
        opt = args.parse_args_and_yaml()
        with self.graph.as_default():
            tf.compat.v1.keras.backend.set_session(self.sess)
            FMI.get_model(opt)

    @staticmethod
    def get_model(opt):
        """
        单例模式+线程锁，防止在异步加载时引发线程安全
        """
        try:
            if FMI.model is None or FMI.train is True:
                with model_lock:
                    FMI.model = FMI.get_keras_model(opt)
                    FMI.model.load_weights(os.path.join(var_dir, 'var', 'fmi.h5'))
        except Exception as e:
            print("加载模型权重失败:{}".format(e.args))

    @staticmethod
    def get_keras_model(opt):
        db_loss = NorthEastLoss(opt)
        south_loss = SouthLoss(opt)
        l1_reg = regularizers.l1(opt.Model['lambda_value_1'])
        l2_reg = regularizers.l2(opt.Model['lambda_value_2'])
        nwp_input = Input(shape=(opt.Model['time_step'], opt.Model['input_size_nwp']), name='nwp')
        env_input = Input(shape=(opt.Model['his_points'], opt.Model['input_size_env']), name='env')

        con1 = Conv1D(filters=64, kernel_size=5, strides=1, padding='valid', activation='relu', kernel_regularizer=l2_reg)(nwp_input)
        nwp = MaxPooling1D(pool_size=5, strides=1, padding='valid', data_format='channels_last')(con1)
        nwp_lstm = LSTM(units=opt.Model['hidden_size'], return_sequences=False, kernel_regularizer=l2_reg)(nwp)

        con2 = Conv1D(filters=64, kernel_size=5, strides=1, padding='valid', activation='relu', kernel_regularizer=l2_reg)(env_input)
        env = MaxPooling1D(pool_size=5, strides=1, padding='valid', data_format='channels_last')(con2)
        env_lstm = LSTM(units=opt.Model['hidden_size'], return_sequences=False, name='env_lstm',kernel_regularizer=l2_reg)(env)
        tiao = Dense(4, name='d4', kernel_regularizer=l1_reg)(env_lstm)

        if opt.Model['fusion']:
            fusion = concatenate([nwp_lstm, tiao])
        else:
            fusion = nwp_lstm

        output = Dense(opt.Model['output_size'], name='cdq_output')(fusion)

        model = Model([env_input, nwp_input], output)
        adam = optimizers.Adam(learning_rate=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-7, amsgrad=True)
        model.compile(loss=south_loss, optimizer=adam)
        return model

    def train_init(self, opt):
        tf.compat.v1.keras.backend.set_session(self.sess)
        model = FMI.get_keras_model(opt)
        try:
            if opt.Model['add_train'] and opt.authentication['repair'] != "null":
                # 进行加强训练，支持修模
                model.load_weights(os.path.join(var_dir, 'var', 'fmi.h5'))
                self.logger.info("已加载加强训练基础模型")
        except Exception as e:
            self.logger.info("加强训练加载模型权重失败:{}".format(e.args))
        model.summary()
        return model

    def training(self, opt, train_and_valid_data):
        model = self.train_init(opt)
        train_X, train_Y, valid_X, valid_Y = train_and_valid_data
        print("----------", np.array(train_X[0]).shape)
        print("++++++++++", np.array(train_X[1]).shape)
        # weight_lstm_1, bias_lstm_1 = model.get_layer('d1').get_weights()
        # print("weight_lstm_1 = ", weight_lstm_1)
        # print("bias_lstm_1 = ", bias_lstm_1)

        check_point = ModelCheckpoint(filepath='./var/' + 'fmi.h5', monitor='val_loss',
                                      save_best_only=True, mode='auto')
        early_stop = EarlyStopping(monitor='val_loss', patience=opt.Model['patience'], mode='auto')
        # tbCallBack = TensorBoard(log_dir='../figure',
        #                          histogram_freq=0,
        #                          write_graph=True,
        #                          write_images=True)
        history = model.fit(train_X, train_Y, batch_size=opt.Model['batch_size'], epochs=opt.Model['epoch'], verbose=2,
                            validation_data=(valid_X, valid_Y), callbacks=[check_point, early_stop])
        loss = np.round(history.history['loss'], decimals=5)
        val_loss = np.round(history.history['val_loss'], decimals=5)
        self.logger.info("-----模型训练经过{}轮迭代-----".format(len(loss)))
        self.logger.info("训练集损失函数为：{}".format(loss))
        self.logger.info("验证集损失函数为：{}".format(val_loss))
        self.logger.info("训练结束，原模型地址：{}".format(id(FMI.model)))
        with self.graph.as_default():
            tf.compat.v1.keras.backend.set_session(self.sess)
            FMI.train = True
            FMI.get_model(opt)
            FMI.train = False
        self.logger.info("保护线程，加载模型，地址：{}".format(id(FMI.model)))

    def predict(self, test_X, batch_size=1):
        with self.graph.as_default():
            with self.sess.as_default():
                result = FMI.model.predict(test_X, batch_size=batch_size)
        self.logger.info("执行预测方法")
        return result

    def train_custom(self, train_X, train_Y, model, opt):
        epochs = opt.Model['epoch']
        batch_size = opt.Model['batch_size']
        num_batches = len(train_X) // batch_size    # 取整

        optimizer = tf.keras.optimizers.Adam(learning_rate=opt.Model[""])
        for epoch in range(epochs):
            for batch_index in range(epochs):
                start = batch_index * batch_size
                end = start + batch_size
                x_batch, y_batch = train_X[start: end], train_Y[start: end]

                with tf.GradientTape() as tape:
                    res = model(x_batch)
                    loss = rmse(y_batch, res)

                gradients = tape.gradient(loss, model.trainable_variables)