Loss backward() core dump MWE

1

This code is modified from the DQN tutorial (which runs on my machine). However, this code does not. It core dumps in the backward method. Am I doing something dumb?

Thank you, in advance, for any assistance.

class DQN(nn.Module):

    def __init__(self):
        super(DQN, self).__init__()
        self.conv1 = nn.Conv2d(3, 16, kernel_size=5, stride=2)
        self.bn1 = nn.BatchNorm2d(16)
        self.conv2 = nn.Conv2d(16, 32, kernel_size=5, stride=2)
        self.bn2 = nn.BatchNorm2d(32)
        self.conv3 = nn.Conv2d(32, 32, kernel_size=5, stride=2)
        self.bn3 = nn.BatchNorm2d(32)
        self.head = nn.Linear(192, 3)

    def forward(self, x):
        x = F.relu(self.bn1(self.conv1(x)))
        x = F.relu(self.bn2(self.conv2(x)))
        x = F.relu(self.bn3(self.conv3(x)))
        return self.head(x.view(x.size(0), -1))

model = DQN()
optimizer = optim.RMSprop(model.parameters())

action_batch = Variable(LongTensor([[0],[1],[2],[0],[1],[2],[0],[1],[2],[0]]))
next_states = Variable(torch.rand(10,3,26,70),volatile=True)
state_batch = Variable(torch.rand(10,3,26,70))
state_action_values = model(state_batch)
next_state_values = model(next_states)
next_state_values.volatile = False
expected_state_action_values = (next_state_values * 0.999) + 1.0
loss = F.smooth_l1_loss(state_action_values, expected_state_action_values)
optimizer.zero_grad(); model.zero_grad();
loss.backward()
2

I removed one of the conv layers and it works. I assume my core dump must have something to do with some of the reported bugs.

class DQN(nn.Module):

    def __init__(self):
        super(DQN, self).__init__()
        self.conv1 = nn.Conv2d(3, 16, kernel_size=5, stride=2)
        self.bn1 = nn.BatchNorm2d(16)
        self.conv2 = nn.Conv2d(16, 32, kernel_size=5, stride=2)
        self.bn2 = nn.BatchNorm2d(32)
        self.head = nn.Linear(1920, 3)

    def forward(self, x):
        x = F.relu(self.bn1(self.conv1(x)))
        x = F.relu(self.bn2(self.conv2(x)))
        return self.head(x.view(x.size(0), -1))

model = DQN()
optimizer = optim.RMSprop(model.parameters())

action_batch = Variable(LongTensor([[0],[1],[2],[0],[1],[2],[0],[1],[2],[0]]))
next_states = Variable(torch.rand(10,3,26,70),volatile=True)
state_batch = Variable(torch.rand(10,3,26,70))
state_action_values = model(state_batch)
next_state_values = model(next_states)
next_state_values.volatile = False
expected_state_action_values = (next_state_values * 0.999) + 1.0
loss = F.smooth_l1_loss(state_action_values, expected_state_action_values)
optimizer.zero_grad(); model.zero_grad();
loss.backward()
3

where is the coredump stopped? can you look at it in gdb?
also are you on pytorch 0.1.12? can you see if master fixes it? we fixed some rare segfaults recently