无法成功地创建火炬训练模型。

Question

我想用基于本站的optuna神经网络进行回归分析。我想创建一个模型，两个一维数据作为输入，一个一维数据作为批量学习的输出。

x是培训数据，y是教师数据。

class Model(nn.Module):
    # コンストラクタ(インスタンス生成時の初期化)
    def __init__(self,trial, mid_units1, mid_units2):
        super(Model, self).__init__()
        self.linear1 = nn.Linear(2, mid_units1)
        self.bn1 = nn.BatchNorm1d(mid_units1)
        self.linear2 = nn.Linear(mid_units1, mid_units2)
        self.bn2 = nn.BatchNorm1d(mid_units2)
        self.linear3 = nn.Linear(mid_units2, 1)
        self.activation = trial_activation(trial)

    def forward(self, x):
        x = self.linear1(x)
        x = self.bn1(x)
        x = self.activation(x)
        x = self.linear2(x)

device = "cuda" if torch.cuda.is_available() else "cpu"

EPOCH = 100
x = torch.from_numpy(a[0].astype(np.float32)).to(device)
y = torch.from_numpy(a[1].astype(np.float32)).to(device)

def train_epoch(model, optimizer, criterion):
    model.train()
    optimizer.zero_grad()    # 勾配情報を0に初期化
    y_pred = model(x)                                               # 予測
    loss = criterion(y_pred.reshape(y.shape), y)          # 損失を計算(shapeを揃える)
    loss.backward()                                                       # 勾配の計算
    optimizer.step()                                                      # 勾配の更新
    return loss.item()

def trial_activation(trial):
    activation_names = ['ReLU','logsigmoid']
    activation_name = trial.suggest_categorical('activation', activation_names)
    if activation_name == activation_names[0]:
        activation = F.relu
    else:
        activation = F.logsigmoid
    return activation

def objective(trial):
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    
    # 中間層のユニット数の試行
    mid_units1 = int(trial.suggest_discrete_uniform("mid_units1", 1024*2,1024*4, 64*2))
    mid_units2 = int(trial.suggest_discrete_uniform("mid_units2", 1024, 1024*2, 64*2))

    net = Model(trial, mid_units1, mid_units2).to(device)

    criterion = nn.MSELoss() 
    # 最適化手法の試行
    optimizer = trial_optimizer(trial, net)
    train_loss = 0
    for epoch in range(EPOCH):
        train_loss = train_epoch(net, optimizer, criterion, device)
    torch.save(net.state_dict(), str(trial.number) + "new1.pth")
    return train_loss

strage_name = "a.sql"
study_name = 'a'

study = optuna.create_study(
    study_name = study_name,
    storage='sqlite:///'  + strage_name, 
    load_if_exists=True,
    direction='minimize')
TRIAL_SIZE = 100

study.optimize(objective, n_trials=TRIAL_SIZE)

错误信息

---> 28     loss = criterion(y_pred.reshape(y.shape), y)          # 損失を計算(shapeを揃える)
     29     loss.backward()                                                       # 勾配の計算
     30     optimizer.step()                                                      # 勾配の更新

AttributeError: 'NoneType' object has no attribute 'reshape'

由于上面的错误，我检查了y_pred的值，发现它是None。

    model.train()
    optimizer.zero_grad()

我认为这两句话可能是错的，但我不知道如何解决这个问题。

Answer 1

使用PyTorch时，当调用y_pred = model(x)时，将调用在Model类中定义的forward函数。

因此，y_pred将得到forward函数的结果，在您的例子中，它什么也不返回，这就是为什么您得到一个None值。您可以如下所示更改forward函数：

    def forward(self, x):
        x = self.linear1(x)
        x = self.bn1(x)
        x = self.activation(x)
        x = self.linear2(x)
        return x