Why I am getting this error:" RuntimeError: Trying to backward through the graph a second time

Question

First the setup:

I have the following simple model architecture:

class Lin2Prop(Module):

def __init__(self):
    super().__init__()
    # Property A
    self.al1 = Linear(159, 32)  
    self.al2 = Linear(32, 2)
    # Property B
    #self.bl1 = Linear(159, 32)  
    #self.bl2 = Linear(32, 1)
    
def forward(self, x):
    #Property A prediction layers
    a_l1 = self.al1(x)
    a_l1 = LeakyReLU(negative_slope=0.2)(a_l1)
    a_l2 = self.al2(a_l1)
    #Property B prediction layers
    b_l1 = self.bl1(x)
    b_l1 = LeakyReLU(negative_slope=0.2)(b_l1)
    b_l2 = self.b2(b_l1)
    return a_l1, b_l2

And this is how I train the model:

model = Lin2Prop()
model = model.to('cuda')
loss_fn = MSELoss()
optimizer = Adam(model.parameters(), lr=0.01)
all_losses = []

for epoch in range(1,50):
    total_loss = 0
    for x, y in tqdm(train_1_dl):
        print(x.shape, y.shape)
        x = x.to('cuda')
        y = y.to('cuda')
        model.train()
        optimizer.zero_grad()
        a_pred, b_pred = model(x)
        a_loss = loss_fn(a_pred, y[:,:0])
        b_loss = loss_fn(b_pred, y[:,:1])
        loss = a_loss + b_loss
        total_loss += loss.item()
        loss.backward(retain_graph=True)
        optimizer.step()
    all_losses.append(loss.detach().cpu().numpy())

I thought that this is a simple network but I am getting this error:

---------------------------------------------------------------------------
---> 21         loss.backward(retain_graph=True)
     22         optimizer.step()
     23         all_losses.append(loss.detach().cpu().numpy())

~/anaconda3/envs/alpha-ren/lib/python3.7/site-packages/torch/_tensor.py inbackward(self, gradient, retain_graph, create_graph, inputs)
     487             )
     488         torch.autograd.backward(
 --> 489             self, gradient, retain_graph, create_graph, inputs=inputs
     490         )
     491 

 ~/anaconda3/envs/alpha-ren/lib/python3.7/site-packages/torch/autograd/__init__.py in  backward(tensors, grad_tensors, retain_graph, create_graph, grad_variables, inputs)
     197     Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
     198         tensors, grad_tensors_, retain_graph, create_graph, inputs,
 --> 199         allow_unreachable=True, accumulate_grad=True)  # Calls into the C++ engine to run the backward pass
     200 
     201 def grad(

     RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors after they have already been freed). Saved intermediate values of the graph are freed when you call .backward() or autograd.grad(). Specify retain_graph=True if you need to backward through the graph a second time or if you need to access saved tensors after calling backward.

I am very aware of topics like these:

But I don’t fully understand how those answers apply to my case. In my case, I think I have a simple network. Please help me understand what I am obviously doing wrong. I should mention that I am still getting the error if I simplify the model to the following:

class Lin2Prop(Module):

    def __init__(self):
        super().__init__()
        # Prediction layers
        self.a1 = Linear(159, 32) 
        self.a2 = Linear(32, 2)
    
    def forward(self, x):
        #Prediction layers
        al1 = self.a1(x)
        al1 = LeakyReLU(negative_slope=0.2)(al1)
        al2 = self.ht12_linear_2(al1)
        return al2

.... and later in the training loop...
    preds = model(x)
    a_loss = loss_fn(preds[:,:0], y[:,:0])
    b_loss = loss_fn(preds[:,:1], y[:,:1])
    total_loss = a_loss + b_loss

Asked By: Thornhale

||

Source

Answer 1

It seems you’re trying to backward your graph multiple times. The error message is indicating that the saved intermediate values of the graph have been freed when you call backward or grad for the first time, and it’s not possible to backward the graph a second time.

The reason you’re encountering this error is because you’re using the same variable loss to store the value for different epochs in the training loop. Every time you’re updating loss, you’re also freeing up the intermediate values of the previous loss.

To resolve the issue, you should create a new variable to store the value of loss in each iteration.

Here’s how you can modify your code:

for epoch in range(1,50):
total_loss = 0
for x, y in tqdm(train_1_dl):
    print(x.shape, y.shape)
    x = x.to('cuda')
    y = y.to('cuda')
    model.train()
    optimizer.zero_grad()
    a_pred, b_pred = model(x)
    a_loss = loss_fn(a_pred, y[:,:0])
    b_loss = loss_fn(b_pred, y[:,:1])
    curr_loss = a_loss + b_loss
    total_loss += curr_loss.item()
    curr_loss.backward(retain_graph=True)
    optimizer.step()
all_losses.append(total_loss)

This should resolve your error, and you should be able to successfully backward the graph.

Answered By: outgaze

Why I am getting this error:" RuntimeError: Trying to backward through the graph a second time

Question:

Answers: