pytorch 找不到有关“梯度计算所需的变量之一已被原地操作修改”的原地操作

vuktfyat  于 2023-11-19  发布在  其他
关注(0)|答案(1)|浏览(107)
dataset = Data(params)
detector = Detector(params)
optimizer = torch.optim.Adam(detector.parameters(), lr=params['learning_rate'])

Loss = []  
criterion = torch.nn.MSELoss()
for epoch in range(params['maxEpoch']):
    y, h_a, h_b, plus, hTy, hTh = dataset.generate()
    x_ = detector(hTy, hTh)

    loss = 0.0
    optimizer.zero_grad()
    for _ in range(1, params['DetNet_layer']):
        loss += criterion(x_[:,:,_], torch.from_numpy(plus).to(torch.double)) * math.log(_)

    loss.backward(retain_graph=True)
    optimizer.step()

    Loss.append(loss.item())

字符串
上面是我的main.py文件,错误发生在loss.backward()中。类Detetor是:

def forward(self, HTy, HTH):
    HTy_torch = torch.from_numpy(HTy).unsqueeze(1)
    HTH_torch = torch.from_numpy(HTH).unsqueeze(1)

    x_torch = torch.from_numpy(np.zeros((self.batch_size, 1, self.L)))
    v_torch = torch.from_numpy(np.zeros((self.batch_size, 1, self.L)))

    for i in range(1, self.L):
        x_tmp, v_tmp = self.layers[i](HTy_torch, HTH_torch, x_torch[:, :, i-1], v_torch[:, :, i-1])
        x_torch[:, :, i] = x_tmp
        v_torch[:, :, i] = v_tmp

    return x_torch


滑动操作[:,:,i]是否进行原地操作?
如果是,我该如何修改它?
我在函数loss.backword中尝试了'retain_graph=True'。

uklbhaso

uklbhaso1#

这可能是loss +=的用法。

losses = []
for _ in range(1, params['DetNet_layer']):
    losses.append(criterion(x_[:,:,_], torch.from_numpy(plus).to(torch.double)) * math.log(_))
loss = torch.cat(losses).sum()

字符串
另外,你可能想在那里使用浮点数而不是双精度。

相关问题