嗨，我在运行时在loss.backward（）中出现此错误。我在Internet上搜索，并且在就地操作中使用和编辑张量时确保了该错误。我没有找到错误，我也不明白是什么张量给了这个错误这里是感兴趣的2类的代码：
class PolicyNetwork():
def __init__(self, n_state, n_action, n_hidden1=128, n_hidden2 = 128, lr=0.00001):
self.model = ActorCriticModel(n_state, n_action, n_hidden1, n_hidden2)
self.optimizer = torch.optim.Adam(self.model.parameters(), lr)
def predict(self, s):
return self.model(s)
def get_action(self, s):
distXZ, distO, state_value = self.predict(s)
action1 = distXZ.sample()
action2 = distO.sample()
log_prob1 = distXZ.log_prob(action1).view(-1,1)
log_prob2 = distO.log_prob(action2).view(-1,1)
return action1, action2, log_prob1, log_prob2, state_value
def update(self, log_prob1, log_prob2,value, reward):
#for log_prob1, log_prob2, value, reward in zip(log_probs1, log_probs2, state_values, rewards):
advantage = reward - value
policy_loss = ((-log_prob1 * advantage) + (-log_prob2 * advantage)).mean()
value_loss = F.smooth_l1_loss(value, reward)
loss = policy_loss + value_loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
return loss, reward, policy_loss, value_loss
这里是调用更新的火车类，在这里我称为loss.backward（）：
def train_regression(model, estimator, train_loader, test_loader,exp_name='train_regressor', epochs=200):
# device
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
# definiamo un dizionario contenente i loader
loader = {'train': train_loader, 'test': test_loader}
for e in range(epochs):
#Iteriamo in due modalità, training e testing
for mode in ['train', 'test']:
model.train() if mode == 'train' else model.eval()
with torch.set_grad_enabled(mode == 'train'): # abilitiamo i gradienti solo in training
for i, batch in enumerate(loader[mode]):
state_image = batch['image'].to(device)
state = batch['start'][0]
action1, action2, log_prob1, log_prob2, state_value = estimator.get_action(state_image)
rewards = []
for a1, a2 in zip(action1, action2):
rewards.append(loader[mode].dataset.get_reward(state.item(), a1[0], a1[1], a2[0], a2[1]))
reward= torch.Tensor(rewards).view(-1,1).cuda()
if mode == 'train':
loss_to_plot, policy_to_plot, reward_to_plot, value_to_plot = estimator.update(log_prob1, log_prob2, state_value, reward)
torch.save(model.state_dict(), '%s.pth' % exp_name)
return model
很抱歉，如果我输入了大量代码，但是我没有发现错误（我的操作已修改张量）我也打印了4张量，它们是16X1
tensor([[-1.6174],
[-1.7020],
[-1.4557],
[-1.8915],
[-5.1338],
[-4.6408],
[-2.1750],
[-1.7954],
[-2.6198],
[-2.5692],
[-2.4872],
[-2.4471],
[-2.6823],
[-1.6389],
[-2.3344],
[-2.3923]], device='cuda:0', grad_fn=<ViewBackward>)
tensor([[-3.1944],
[-2.3279],
[-1.6919],
[-2.1784],
[-2.9512],
[-3.1380],
[-2.6455],
[-1.8039],
[-2.1176],
[-3.2955],
[-2.2406],
[-1.8591],
[-4.2133],
[-3.0645],
[-5.1651],
[-1.8645]], device='cuda:0', grad_fn=<ViewBackward>)
tensor([[-0.0178],
[-0.0075],
[-0.0216],
[-0.0035],
[-0.0750],
[-0.0397],
[-0.0263],
[-0.0412],
[-0.0118],
[-0.0376],
[-0.0511],
[-0.0626],
[ 0.0066],
[-0.0344],
[-0.0438],
[-0.0028]], device='cuda:0', grad_fn=<ViewBackward>)
tensor([[0.5000],
[0.2500],
[0.7500],
[0.5000],
[0.0000],
[0.5000],
[0.0000],
[0.2500],
[0.5000],
[0.2500],
[0.5000],
[0.2500],
[0.5000],
[0.0000],
[0.2500],
[0.5000]], device='cuda:0')
这里是班级演员评论家，模特：
class ActorCriticModel(nn.Module):
def __init__(self, n_input, n_output, n_hidden1, n_hidden2):
super(ActorCriticModel, self).__init__()
#ResNet50 per features sharing
self.model = resnet50(pretrained=True)
#Cambiamo il Fully Conncted con un lineare con input di un MLP
#self.model.fc = nn.Linear(self.model.fc.in_features, n_input)
out = self.model.fc.in_features
self.model.fc = nn.Identity()
#"Spezzettiamo" l' hiidden layer in due (devo dividere il numero delle features in due?
self.fcN = nn.Linear(out, n_hidden1)
self.fcV = nn.Linear(out, n_hidden2)
#self.muX, self.muZ, self.sigma, self.muO1, self.muO2, self.sigmaO = nn.Linear(n_hidden1, n_output)
#Il primo hidden layer ci ritorna 6 parametri per due distribuzioni Normali
#6 linear con un solo output per ogni parametro
self.muX = nn.Linear(n_hidden1, 1)
self.muZ = nn.Linear(n_hidden1, 1)
self.sigma = nn.Linear(n_hidden1, 1)
self.muO1 = nn.Linear(n_hidden1, 1)
self.muO2 = nn.Linear(n_hidden1, 1)
self.sigmaO = nn.Linear(n_hidden1, 1)
#Il secondo hidden layer ci ritorna uno scalare, value
self.value = nn.Linear(n_hidden2, 1)
#istanziamo due distribuzioni Normali, relative al movimento....
self.distributionXZ = torch.distributions.MultivariateNormal
# e all'orientamento
self.distributionO = torch.distributions.MultivariateNormal
def forward(self, x):
#Passiamo l'immagine sulla ResNet50
#x = x.cuda()
x = self.model(x)
#Attivazioni -> ReLU
x1 = F.relu(self.fcN(x))
x2 = F.relu(self.fcV(x))
#WLS linux terminal on wondows
#Normale per movimento
muX = 2 * torch.tanh(self.muX(x1))
muZ = 5 * torch.sigmoid(self.muZ(x1))
sigma = F.softplus(self.sigma(x1)) + 1e-5
#Normale per orientamento
muO1 = torch.tanh(self.muO1(x1))
muO2 = torch.tanh(self.muO2(x1))
mod = torch.sqrt(muO1**2+muO2**2)
mod[mod<1e-5]=1
muO1=muO1/mod
muO2=muO2/mod
sigmaO = F.softplus(self.sigmaO(x1))
#Creiamo le distribuzioni relativa al movimento...
cov = torch.eye(2).view(1,2,2).cuda()*sigma.view(-1,1,1)
distXZ = self.distributionXZ(torch.cat([muX,muZ],1), cov)
#relativa all'orientamento
covO = torch.eye(2).view(1,2,2).cuda()*sigmaO.view(-1,1,1)
distO = self.distributionO(torch.cat([muO1,muO2],1), covO)
#Tensore, 16x1
value = self.value(x2)
value = value.view(-1,1)
#print(value)
return distXZ, distO, value
问题出在您的网络中。不在前进中。
.forward中的某处您有类似
h = foo(x)
h = bar(h)
或实际的镶嵌物，例如一种“ _”方法（例如unsqueeze_()与unsqueeze()）。