import tensorflow as tf
from tensorflow import keras
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from datetime import datetime
from matplotlib.font_manager import FontProperties
font = FontProperties(fname='C:/Windows/Fonts/STXINGKA.ttf')

# 生成数据
def gen_datas():
    data = pd.read_csv("housing.csv")
    data = np.array(data)
    data1 = pd.read_csv("001.csv")
    data1 = np.array(data1)
    print(data1)
    inputs = data[:, 0:data.shape[1] - 1]
    outputs = data[:, data.shape[1] - 1]
    # inputs = np.linspace(-1, 1, 250, dtype=np.float32)[:, np.newaxis]
    # noise = np.random.normal(0, 0.05, inputs.shape).astype(np.float32)
    # outputs = np.square(inputs) - 0.5*inputs + noise
    return data1, inputs, outputs

def complie_model(model):
    # 神经网络参数配置
    model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.01), loss='mse', metrics=["mae", "mse"])

def create_model():
    # 创建新神经网络
    model = tf.keras.Sequential([tf.keras.layers.Dense(10, activation=tf.nn.relu, input_shape=(3,), name='layer1'), tf.keras.layers.Dense(1, name="outputs")])
    complie_model(model)
    return model

def display_nn_structure(model, nn_structure_path):
    # 展示神经网络结构
    model.summary()
    keras.utils.plot_model(model, nn_structure_path, show_shapes=True)

def callback_only_params(model_path):
    # 保存模型回调参数
    ckpt_callback = tf.keras.callbacks.ModelCheckpoint(filepath=model_path, verbose=1, save_best_only=True, save_freq='epoch')
    return ckpt_callback

def tb_callback(model_path):
    tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=model_path, histogram_freq=1)
    return tensorboard_callback

def plot_history(history):
    # 绘制损失数据
    hist = pd.DataFrame(history.history)
    hist['epoch'] = history.epoch
    # 打开绘图区
    plt.figure()
    plt.xlabel("训练次数", fontproperties=font)
    plt.ylabel("损失值", fontproperties=font)
    plt.plot(hist["epoch"], hist["mse"], label='loss')
    plt.legend(prop=font)
    # plt.savefig
    plt.show()

def train_model(model, inputs, outputs, model_path, log_path):
    # 训练神经网络
    # 回调函数
    ckpt_callback = callback_only_params(model_path)
    tensorboard_callback = tb_callback(log_path)
    # 保存参数
    model.save_weights(model_path.format(epoch=0))
    history = model.fit(inputs, outputs, epochs=10, callbacks=[ckpt_callback, tensorboard_callback], verbose=0)
    plot_history(history)

def load_model(model, model_path):
    latest = tf.train.latest_checkpoint(model_path)
    print("latest:{}".format(latest))
    # model.load_weights(latest)

def prediction(model, model_path, inputs):
    load_model(model, model_path)
    pres = model.predict(inputs)
    # print("prediction:{}".format(pres))
    return pres

def plot_prediction(model, model_path, inputs, outputs):
    pres = prediction(model, model_path, inputs)
    # plt.plot(outputs, s=10, c="r", marker="*", label="实际值")
    plt.plot(outputs, label="实际值")
    plt.plot(pres, label="预测结果")
    plt.xlabel("输入数据", fontproperties=font)
    plt.ylabel("预测值", fontproperties=font)
    plt.legend(prop=font)
    # plt.savefig()
    plt.show()
    print(pres)


if __name__ == "__main__":
    stamp = datetime.now().strftime("%Y%m%d-%H:%M:%S")
    model_path = "D:/Machine-Learning-Model/line-fit-high"  # + stamp
    log_path = "D:/Machine-Learning-Model/line-fit-high"  # + stamp
    data1, inputs, outputs = gen_datas()
    model = create_model()
    display_nn_structure(model, "D:/Machine-Learning-Model/001.png")
    train_model(model, inputs, outputs, model_path, log_path)
    model_path = "D:/Machine-Learning-Model/high"
    plot_prediction(model, model_path, data1, outputs)