Hi. I have the problem here.
Now, I am trying to get predict interval with using Monte Calro dropout.
However, I got following errors when I call loss.backward() after the running of following code
I do not kown why, any reply will be appreciated.
thanks.
class Net(nn.Module):
def __init__(self, x1, x2, p, m, mode=0):
super().__init__()
input_dim = p*m
hidden_dim = 128
self.dropout = nn.Dropout(p=0.1)
if mode == 1:
self.embed = Encoder1(x1, x2, p)
elif mode == 2:
self.embed = Encoder2(x1, x2, p)
else:
self.embed = Encoder(x1, x2, p)
self.share_block = nn.Sequential(
・・・・・
)
self.head = nn.Sequential(
nn.BatchNorm1d(hidden_dim),
nn.utils.weight_norm(nn.Linear(hidden_dim, 1))
)
def forward(self, x):
x = self.share_block(x)
_out = self.head(x)
_out = self.dropout(_out)
return _out
def predict(self, x1, x2, alpha=0.95, samples=100):
x1 = x1.to(device)
x2 = x2.to(device)
x = self.embed(x1, x2)
z_c = norm.ppf(alpha)
samples_ = [self.forward(x).detach()
for _ in range(samples)]
pred_sample = torch.cat(samples_, dim=1)
pred_mean = torch.mean(pred_sample, dim=1)
pred_std = z_c * torch.std(pred_sample, dim=1)
return pred_mean, pred_std
def train_and_test_gpu(model, num_epochs, dataset):
train_dataset = utils.data.Subset(dataset, train_idx)
train_loader = utils.data.DataLoader(
train_dataset, batch_size=64, worker_init_fn=seed_worker, generator=g, shuffle=True)
test_dataset = utils.data.Subset(dataset, test_idx)
test_loader = utils.data.DataLoader(
test_dataset, batch_size=64, shuffle=False)
optimizer = optim.Adam(model.parameters(), lr=1e-6)
criterion = nn.MSELoss()
scheduler = optim.lr_scheduler.OneCycleLR(
optimizer,
max_lr=1e-3,
steps_per_epoch=len(train_loader),
epochs=num_epochs,
)
model.train()
for epoch in tqdm(range(num_epochs)):
for x1, x2, y in train_loader:
x1 = x1.to(device)
x2 = x2.to(device)
y = torch.Tensor(y).to(device)
optimizer.zero_grad()
y_pred, _ = model.predict(x1, x2)
loss = criterion(y_pred, y)
loss.backward() <ーーーーーー error occured here
optimizer.step()
scheduler.step()
model.eval()
with torch.no_grad():
y_true_list = []
y_pred_list = []
for x1, x2, y_true in test_loader:
x1 = torch.Tensor(x1).to(device)
x2 = torch.Tensor(x2).to(device)
y_pred, _ = model.predict(x1, x2)
y_pred_list.extend(y_pred.detach().flatten().cpu().tolist())
y_true_list.extend(y_true.detach().cpu().tolist())
return y_true_list, y_pred_list