Hello everyone, I’ve found myself having the following problem:
when calling
trainloader = DataLoader(train_subset, batch_size=batch_size, shuffle = True, num_workers=8, worker_init_fn=np.random.seed(0), pin_memory=True)
the following pops up:
RuntimeError: Caught RuntimeError in DataLoader worker process 0.
Original Traceback (most recent call last):
File "/home/iferfoglia/mambaforge/envs/conceptSTL/lib/python3.10/site-packages/torch/utils/data/_utils/worker.py", line 302, in _worker_loop
data = fetcher.fetch(index)
File "/home/iferfoglia/mambaforge/envs/conceptSTL/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 52, in fetch
return self.collate_fn(data)
File "/home/iferfoglia/mambaforge/envs/conceptSTL/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py", line 175, in default_collate
return [default_collate(samples) for samples in transposed] # Backwards compatibility.
File "/home/iferfoglia/mambaforge/envs/conceptSTL/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py", line 175, in <listcomp>
return [default_collate(samples) for samples in transposed] # Backwards compatibility.
File "/home/iferfoglia/mambaforge/envs/conceptSTL/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py", line 141, in default_collate
return torch.stack(batch, 0, out=out)
RuntimeError: stack(): functions with out=... arguments don't support automatic differentiation, but one of the arguments requires grad.
If I don’t define the number of workers and it defaults to 0, everything works.
The training loop:
for epoch in range(epochs):
model.train()
for batch, labels in trainloader:
batch, labels = batch.to(device), labels.to(device)
y_preds = model(batch)
loss = criterion(y_preds, labels.float())
# instead of optimizer.zero_grad()
for param in model.parameters():
param.grad = None
loss.backward()
optimizer.step()
and the model:
class DCR_classifier(nn.Module):
def __init__(self, input_shape):
super().__init__()
# Define your MLP layers for classification
self.fc1 = nn.Linear(input_shape, 64)
self.fc2 = nn.Linear(64, 128)
self.fc3 = nn.Linear(128, 1)
def forward(self, x, traj_emb):
x_new = x.view(x.size(0), -1) # (batch, nvar, points) -> (batch, nvar x points)
combined_features = torch.cat((traj_emb, x_new), dim=1) # (batch, kemb_size + nvar x points)
output = self.fc1(combined_features)
output = F.relu(output)
output = self.fc2(output)
output = F.relu(output)
output = self.fc3(output)
return output
class DCR(nn.Module):
def __init__(self, nvar, points, device):
super().__init__()
self.concepts = load_phis_dataset('concepts', nvar)
self.kemb = get_kernel_embedding(self.concepts, nvar).to(device) # (concepts, kemb_size)
_ = self.kemb.requires_grad_()
self.classifier = DCR_classifier(self.kemb.shape[1] + (nvar*points))
self.nvar = nvar
def forward(self, x):
# concept truth degrees
rhos = get_robustness(x, self.concepts, time = False) # (trajectories, concepts)
_ = rhos.requires_grad_()
# embed trajectories in kernel space
traj_emb = torch.matmul(rhos, self.kemb) # (trajectories, kemb_size)
_ = traj_emb.requires_grad_()
output = self.classifier(x, traj_emb)
return output.squeeze(1)
Because I’m working with custom objects, I need to call the requires_grad(), otherwise it doesn’t work.
I can’t disclose all the called functions, but they shouldn’t be the problem.
Thank you!