Graph Wavenet：入门图神经网络训练的demo

1. 数据的准备

DCRNN的github项目有完整的数据

Graph Wavenet 提供的数据缺少  '/data/sensor_graph/adj_mx.pkl'

2. engine.train 与engine.eval 的区别

 区别仅在于model.train()与model.eval()

具体区别见https://blog.csdn.net/qq_43118572/article/details/118615095

class trainer():
    def __init__(self, scaler, in_dim, seq_length, num_nodes, nhid, dropout, lrate, wdecay,
                 device, supports, gcn_bool, addaptadj, aptinit):
        self.model = gwnet(device, num_nodes, dropout, supports=supports, gcn_bool=gcn_bool, aptinit=aptinit, in_dim=in_dim,
                           out_dim=seq_length, residual_channels=nhid, skip_channels=nhid * 8, end_channels=nhid *16)
        self.model.to(device)
        self.optimizer = optim.Adam(self.model.parameters(), lr=lrate, weight_decay=wdecay)
        self.loss = util.masked_mae
        self.scaler = scaler
        self.clip = 5
    
    def train(self, input, real_val):
        self.model.train()
        self.optimizer.zero_grad()
        input = nn.functional.pad(input, (1,0,0,0))
        output = self.model(input)
        output = output.transpose(1,3)
        real = torch.unsqueeze(real_val, dim=1)
        predict = self.scaler.inverse_transform(output)

        loss = self.loss(predict, real, 0.0)
        loss.backward()
        if self.clip is not None:
            # 梯度剪切，规定了最大不能超过的max_norm.
            torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.clip)
        self.optimizer.step()
        mape = util.masked_mape(predict, real, 0.0).item()
        rmse = util.masked_rmse(predict, real, 0.0).item()
        return loss.item(), mape, rmse

    def eval(self, input, real_val):
        self.model.eval()
        input = nn.functional.pad(input, (1, 0, 0, 0))
        with torch.no_grad():
            output = self.model(input)
        output = output.transpose(1, 3)

        real = torch.unsqueeze(real_val, dim=1)
        predict = self.scaler.inverse_transform(output)
        loss = self.loss(predict, real, 0.0)
        mape = util.masked_mape(predict, real, 0.0).item()
        rmse = util.masked_rmse(predict, real, 0.0).item()
        return loss.item(), mape, rmse

3. masked loss

具体原因查看：https://blog.csdn.net/qq_43118572/article/details/118698359

4. utils用到的函数：

import pickle
import numpy as np
import os
import scipy.sparse as sp
import torch
from scipy.sparse import linalg

class DataLoader(object):
    def __init__(self, xs, ys, batch_size, pad_with_last_sample=True):
        self.batch_size = batch_size
        self.current_ind = 0
        if pad_with_last_sample:
            num_padding = (batch_size - (len(xs) % batch_size)) % batch_size
            x_padding = np.repeat(xs[-1:], num_padding, axis=0)
            y_padding = np.repeat(ys[-1:], num_padding, axis=0)
            xs = np.concatenate([xs, x_padding], axis=0)
            ys = np.concatenate([ys, y_padding], axis=0)

        self.size = len(xs)
        self.num_batch = int(self.size // self.batch_size)
        self.xs = xs
        self.ys = ys

    def shuffle(self):
        permutation = np.random.permutation(self.size)
        xs, ys = self.xs[permutation], self.ys[permutation]
        self.xs = xs
        self.ys = ys

    # 产生批数据的生成器
    def get_iterator(self):
        self.current_ind = 0

        def _wrapper():
            while self.current_ind < self.num_batch:
                start_ind = self.batch_size * self.current_ind
                end_ind = min(self.size, self.batch_size * (self.current_ind + 1))
                x_i = self.xs[start_ind:end_ind, ...]
                y_i = self.ys[start_ind:end_ind, ...]
                yield (x_i, y_i)

        return _wrapper()

class StandardScaler():
    def __init__(self, mean, std):
        self.mean = mean
        self.std = std

    # 批正则化数据
    def transform(self, data):
        return (data - self.mean) / self.std

    # 批正则化数据逆运算
    def inverse_transform(self, data):
        return data*self.std + self.mean

# [A * D^(-1/2)]^T * D^(-1/2) = D^(-1/2) * A * D^(-1/2)
def sym_adj(adj):
    """Symmetrically normalize adjacency matrix."""
    adj = sp.coo_matrix(adj)  # 转换为coordinate形式的压缩邻接矩阵
    rowsum = np.array(adj.sum(1))
    d_inv_sqrt = np.power(rowsum, - 0.5).flatten()  # 将n*1的矩阵转换为1个向量
    d_inv_sqrt[np.isinf(d_inv_sqrt)] = 0.
    d_mat_inv_sqrt = sp.diags(d_inv_sqrt) # ndarray类型
    # toarray returns an ndarray; todense returns a matrix. If you want a matrix, use todense otherwise, use toarray
    return adj.dot(d_mat_inv_sqrt).transpose().dot(d_mat_inv_sqrt).astype(np.float32).todense()

# D^(-1/2) * A
def asym_adj(adj):
    adj = sp.coo_matrix(adj)
    rowsum = np.array(adj.sum(1))
    d_inv = np.power(rowsum, -1).flatten()
    d_inv[np.isinf(d_inv)] = 0.
    d_mat = sp.diags(d_inv)
    return d_mat.dot(adj).astype(np.float32).todense()

def calculate_normalized_laplacian(adj):
    # L = D^-1/2 (D-A) D^-1/2 = I - D^-1/2 A D^-1/2
    # D = diag(A 1)

    # 由于下面用到的是与稀疏的单位阵相减，故需要tocoo()
    D_tilde = sym_adj(adj).tocoo()
    normarlized_laplacian = sp.identity(np.size(adj)[0]) - D_tilde
    return normarlized_laplacian

def calculate_scaled_laplacian(adj_mx, lambda_max=2, undirected=True):
    if undirected:
        adj_mx = np.maximum.reduce([adj_mx, adj_mx.T])
    L = calculate_normalized_laplacian(adj_mx)
    if lambda_max is None:
        # ‘LM’ : Largest (in magnitude) eigenvalues.
        # 返回1个绝对值最大的特征值与特征向量
        lambda_max, _ = linalg.eigsh(L, 1, which='LM')
        lambda_max = lambda_max[0]
    # 转换为稀疏矩阵
    L = sp.csc_matrix(L)
    M, _ = L.shape  # 原始矩阵的行数
    I = sp.identity(M, format='csr', dtype=L.dtype)
    L = (2 / lambda_max * L) - I
    return L.astype(np.float32).todense()

def load_pickle(pickle_file):
    try:
        with open(pickle_file, 'rb') as f:
            pickle_data = pickle.load(f)
    except UnicodeDecodeError as e:
        with open(pickle_file, 'rb') as f:
            pickle_data = pickle.load(f, encoding='latin1')
    except Exception as e:
        print('Unable to load data', pickle_file, ':', e)
        raise
    return pickle_data

def load_adj(pkl_filename, adjtype):
    sensor_ids, sensor_id_to_ind, adj_mx = load_pickle(pkl_filename)
    if adjtype == "scalap":
        adj = [calculate_scaled_laplacian(adj_mx)]
    elif adjtype == "normlap":
        adj = [calculate_normalized_laplacian(adj_mx).astype(np.float32).todense()]
    elif adjtype == "symnadj":
        adj = [sym_adj(adj_mx)]
    elif adjtype == "transition":
        adj = [asym_adj(adj_mx)]
    elif adjtype == "doubletransition":
        adj = [asym_adj(adj_mx), asym_adj(np.transpose(adj_mx))]
    elif adjtype == "identity":
        adj = [np.diag(np.ones(adj_mx.shape[0])).astype(np.float32)]
    else:
        error = 0
        assert error, "adj type not defined"
    return sensor_ids, sensor_id_to_ind, adj

# 返回train_loader、val_loader、test_loader及归一化所需的均值与方差参数
def load_dataset(dataset_dir, batch_size, valid_batch_size=None, test_batch_size=None):
    data = {}
    for category in ['train', 'val', 'test']:
        cat_data = np.load(os.path.join(dataset_dir, category+'.npz'))
        data['x_' + category] = cat_data['x']
        data['y_' + category] = cat_data['y']

    scaler = StandardScaler(mean=data['x_train'][..., 0].mean(), std=data['x_train'][..., 0].std())
    # Data format
    for category in ['train', 'val', 'test']:
        data['x_' + category][..., 0] = scaler.transform(data['x_' + category][..., 0])
    data['train_loader'] = DataLoader(data['x_train'], data['y_train'], batch_size)
    data['val_loader'] = DataLoader(data['x_val'], data['y_val'], valid_batch_size)
    data['test_loader'] = DataLoader(data['x_test'], data['y_test'], test_batch_size)
    data['scaler'] = scaler
    return data

def masked_mse(preds, labels, null_val=np.nan):
    if np.isnan(null_val):
        mask = ~torch.isnan(labels)
    else:
        mask = (labels != null_val)
    mask = mask.float()
    mask /= torch.mean((mask))
    mask = torch.where(torch.isnan(mask), torch.zeros_like(mask), mask)
    loss = (preds - labels) ** 2
    loss = loss * mask
    loss = torch.where(torch.isnan(loss), torch.zeros_like(loss), loss)
    return torch.mean(loss)

def masked_rmse(preds, labels, null_val=np.nan):
    return torch.sqrt(masked_mse(preds=preds, labels=labels, null_val=null_val))

def masked_mae(preds, labels, null_val=np.nan):
    if np.isnan(null_val):
        mask = ~torch.isnan(labels)
    else:
        mask = (labels != null_val)
    mask = mask.float()
    mask /= torch.mean((mask))
    mask = torch.where(torch.isnan(mask), torch.zeros_like(mask), mask)
    loss = torch.abs(preds-labels)
    loss *= mask
    loss = torch.where(torch.isnan(loss), torch.zeros_like(loss), loss)
    return torch.mean(loss)

def masked_mape(preds, labels, null_val=np.nan):
    if np.isnan(null_val):
        mask = ~torch.isnan(labels)
    else:
        mask = (labels != null_val)
    mask = mask.float()
    mask /= torch.mean((mask))
    mask = torch.where(torch.isnan(mask), torch.zeros_like(mask), mask)
    loss = torch.abs(preds-labels)/labels
    loss *= mask
    loss = torch.where(torch.isnan(loss), torch.zeros_like(loss), loss)
    return torch.mean(loss)

def metric(pred, real):
    # torch.tensor.item(): Returns the value of this tensor as a standard Python number.
    mae = masked_mse(pred, real, 0.0).item()
    mape = masked_mape(pred, real, 0.0).item()
    rmse = masked_rmse(pred, real, 0.0).item()
    return mae, mape, rmse