import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import torch
from sklearn import preprocessing

data = pd.read_excel(r'C:\Users\user\PycharmProjects\pytorch2\入模数据\杭州数据.xlsx',index_col='dtdate')
print(data.columns)


y = np.array(data['售电量'])  # 制作标签，用于比对训练结果
x = data.drop(columns=['售电量','city_name'])  # 在特征数据集中去掉label
# df.drop(label, axis=0)
# label：要删除的列或者行，如果要删除多个，传入列表
# axis:轴的方向，0为行，1为列，默认为0
fea_train = np.array(x)  # 转换为ndarray格式

# 数据标准化操作：(x-均值μ) / 标准差σ ，使数据关于原点对称，提升训练效率
input_features = preprocessing.StandardScaler().fit_transform(fea_train)  # fit：求出均值和标准差 transform：求解

# y归一化 防止梯度爆炸
y = (y - np.min(y))/(np.max(y) - np.min(y))
print(y)

# 设定神经网络的输入参数、隐藏层神经元、输出参数的个数
input_size = input_features.shape[1]  # 设定输入特征个数
# np.shape[1]
# 0为行，1为列，默认为0
# 在此表格中因为每行为各样本的值，每列为不同的特征分类，所以此处0表示样本数，1表示特征数
hidden_size = 64  # 设定隐藏层包含64个神经元
output_size = 1  # 设定输出特征个数为1
batch_size = 32  # 每一批迭代的特征数量

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")  # 选择使用GPU训练

my_nn = torch.nn.Sequential(
    torch.nn.Linear(input_size, hidden_size).to(device),  # 输入层 → 第一层
    torch.nn.ReLU().to(device),
    torch.nn.Linear(hidden_size, hidden_size).to(device),  # 第一层 → 第二层
    torch.nn.ReLU().to(device),
    torch.nn.Linear(hidden_size, hidden_size).to(device),  # 第二层 → 第三层
    torch.nn.ReLU().to(device),
    torch.nn.Linear(hidden_size, hidden_size).to(device),  # 第三层 → 第四层
    torch.nn.ReLU().to(device),
    torch.nn.Linear(hidden_size, output_size).to(device)  # 第四层 → 输出层
).to(device)
cost = torch.nn.MSELoss().to(device)
optimizer = torch.optim.Adam(my_nn.parameters(), lr=0.001)

# 训练网络
losses = []
for i in range(300):
    batch_loss = []
    # 采用MINI-Batch的方法进行训练
    for start in range(0, len(input_features), batch_size):
        end = start + batch_size if start + batch_size < len(input_features) else len(input_features)
        x_train = torch.tensor(input_features[start:end], dtype=torch.float32, requires_grad=True).to(device)
        y_train = torch.tensor(y[start:end], dtype=torch.float32, requires_grad=True).to(device)
        prediction = my_nn(x_train)
        loss = cost(y_train, prediction)
        optimizer.zero_grad()
        loss.backward(retain_graph=True)
        optimizer.step()
        batch_loss.append(loss.data.cpu().numpy())

    if i % 10 == 0:
        losses.append(np.mean(batch_loss))
        print(losses)
        print(i, np.mean(batch_loss))

# 保存模型
# torch.save(my_nn, 'BP.pt')

# 绘制图像
dev_x = [i * 10 for i in range(20)]
plt.xlabel('step count')
plt.ylabel('loss')
plt.xlim((0, 200))
plt.ylim((0, 1000))
plt.plot(dev_x, losses)
plt.show()