gorilla-lab-scut/multiclassda

TPAMI2020 "Unsupervised Multi-Class Domain Adaptation: Theory, Algorithms, and Practice"

multiclassda's Issues

the version of the Torch:: there occurs a bug when I computed the gradient twice

Hi, I am sorry to bother you. there occurs a bug when I computed the gradient twice such that
self.optimizer_classifier.zero_grad()
loss_summary_classifier.backward(retain_graph=True)
self.optimizer_classifier.step()
self.optimizer_feature_extractor.zero_grad()
loss_summary_feature_extractor.backward()
self.optimizer_feature_extractor.step()

File "./MultiClassDA-master/solver/SymmNetsV2SC_solver.py", line 284
loss_summary_feature_extractor.backward()
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [2048, 62]], which is output 0 of TBackward, is at version 2; expected version 1 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient.

How did you solve the problem? Is there something wrong with my version of the Torch?

Not reproducible result

Hello.
I am trying to reproduce your result from paper for Office-31 dataset.
You do not provide the correct configuration for Office-31 for the best result in the paper (in repo for SymmNet only one strange version with 11 class were found). So I follow the default parameters you have provided with 400 epoch and 31 classes for Office-31.
I've got the following results
W -> A accuracy: 73.9 instead the best 77.0
A -> D accuracy: 88.15 instead the best 95.6
So I can not get the same result as you even for SymmNets-V2. Can you provide the appropriate parameters to get the same score?
Also in code you mentioned your best model
from solver.SymmNetsV2SC_solver import SymmNetsV2SolverSC as Solver
But there is not code for solver.SymmNetsV2SC_solver. Are you going to release it ?