Question

鉴于职能范围为[0,xmax]的减少,如果是参数,我想用Py Torch书写一个方案,从一些双面价值的培训数据中学习l的价值。我拥有的培训数据是(x_i,g(x_i)), i=1,2,...,n,在 f(x)>时,g(x)=1;t 和其他ise 页: 1 这里的t是固定门槛值。我界定了这一类别。

class Function(torch.nn.Module):
    def __init__(self, l_init): # l is the parameter, its initial value is l_init
        self.l = torch.nn.Parameter(torch.tensor(l_init, dtype=float))
    def forward(self, x):
        return f(x)

我还有一个变量train_data,其中train_data[0]为x_i s(一个浮动阵列)和train_data***为g(x_i) s(a 0/1阵列)。即时发送<条码>功能模块:

f = Function()

将损失功能界定为

loss_func = lambda x, y: (x - y) ** 2

对于每一项培训,我对<代码>pred = Function(train_data[0])进行了计算,我不得不在不失去梯度功能的情况下计算损失。但是,如果我写了字的话,我会写一下。

loss = loss_func((pred > t).float(), train_data[0])

梯度功能将丧失。我应该怎样利用这一双轨培训数据来学习升值?

Answer 1

https://github.com/ZouJiu1/CNN_numpy/blob/master/net/loss.py#L36-L66

(pred > t) can not be derivated or loss the gradient information. (pred > t) should be used when you predict input. it should not be used when you train.

您可以通过使用BCEloss和BEWith式行车来做到这一点。无损损失

你的屋顶界定了这种损失功能,因此,培训——数据应当是一种热形式。

BCEloss

https://pytorch.org/docs/stable/generated/torch.nn.BCELoss.html?highlight=bce#torch.nn.BCELoss

loss_func = torch.nn.BCEloss()
train_datas = torch.zeros(2)
train_datas[train_data[1]] = 1
softmax = torch.nn.Softmax(pred, dim = -1)
pred = softmax(pred)
loss = loss_func(pred, train_datas)

当你预测投入时,使用

softmax = torch.nn.Softmax(pred, dim = -1)
pred = softmax(pred)
for i in range(len(pred)):
    if pred[i] > 0.6:
        ...do something...
        break

BCEWithLogitsLoss

https://pytorch.org/docs/stable/生成/torch.nn.BCEWithlogitsLoss.html?highlight=bce#torch.nn.

loss_func = torch.nn.BCEWithLogitsLoss()
train_datas = torch.zeros(2)
train_datas[train_data[1]] = 1
loss = loss_func(pred, train_datas)

当你预测投入时,使用

softmax = torch.nn.Softmax(pred, dim = -1)
pred = softmax(pred)
for i in range(len(pred)):
    if pred[i] > 0.6:
        ...do something...
        break

besides those, you can use crossentropyloss in multi classes classify too. https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html?highlight=cross#torch.nn.CrossEntropyLoss

BCEloss

BCEWithLogitsLoss

友情链接