RuntimeError: volatile can only be set on leaf variables

autograd
1

output = model(*input)
output.volatile = False

for i in range(BATCH_SIZE):
    if nextState_batch[i][0][0].data[0] == 0:
        nextStateValues[i] = 0
    else:
        nextStateValues[i] = output[b[i]: b[i+1]].max(0)[0].data
expectedStateActionValues = (nextStateValues * GAMMA) + reward_batch.data.view(reward_batch.size(0), -1)

loss = F.smooth_l1_loss(stateActionValues, Variable(expectedStateActionValues))

optimizer.zero_grad()
loss.backward()
for param in model.parameters():
    if param.grad is not None:
        param.grad.data.clamp_(-0.5, 0.5)
optimizer.step()

This is the code snippet which is giving me the error, basically I am implementing a dqn and have followed the official dqn implementation given on the pytorch website.

Can someone please help resolve the runtime error.

2

Basically, you can only set variable.volatile = True on a variable that you just created. See this post for further explanation.
In short, you can not set volatile to True on a variable that is the result of an operation. If you want to set volatile to True, you can do it by calling output.detach_() if you don’t need to train model.

1 Like