wannabeog / mas-pytorch Goto Github PK

Could you tell me how to apply Local MAS?

hi, is the classification_head right in utils/model_class?

class classification_head(nn.Module):
"""

Each task has a seperate classification head which houses the features that
are specific to that particular task. These features are unshared across tasks
as described in section 5.1 of the paper

"""

def __init__(self, in_features, out_features):
	super(classification_head, self).__init__()
	self.fc = nn.Linear(in_features, out_features)

def forward(self, x):
	return x

When executing main.py, I only find this

The training process on the 0 tasks is completed
Testing the model now
and it crash
WHY?

Hello,
I have a question regarding the freeze layers here. I notice that the frozen layers are set to require gradient, in contrast, the not frozen layers don't require gradient. I'm not sure if I understand correctly, but this looks like we actually train frozen layers and ignore the not frozen layers, which is a little unnatural to me. Besides, we still initialize omega for not frozen layers here. Do I miss something? Thanks in advance for your help!

In the file model_train.py, in the training loop, I see below code in L175-179:

output = model.tmodel(input_data)
_, preds = torch.max(output, 1)
running_corrects += torch.sum(preds == labels.data)
loss = model_criterion(output, labels)

What I understood is model.tmodel is the feature extractor, but yet we are comparing the outputs with target labels. Shouldn't we compare it with the last layer (classifier) output?