chrisallenming / mixup_for_uda Goto Github PK

Hi, thanks for sharing the code.
I have gone through the code in Mixup_for_UDA/DM-ADA/trainer_mixed.py, but I can not find the related code about the loss described in the eq. (21).
Can you explain which line of code is related to eq. (21)?

Nice work! Could you please explain the mixup ratio clipping operation used in the code? Why do you clip it to [0.,0.2] and [0.8,1.0]? Is there any heuristics in choosing the alpha or mixup ratio?

Hello and thank you ver much for sharing this code with us. I am working on a wear detection of parts of different geometries and materials ( I have made 22000 labeled photos). For this I have already written a net which has a high accuracy when all different types (geometry and material) are known, but as soon as I test on a dataset which has a minimal changed geometry or material, the performance jumps from 90+ to 20%. Do you think your code could close the gap?
Furthermore, I have read that the code is only partially available (pseudo) and therefore a simple implementation cannot be done. I am also relatively new to programming, so I am looking for an example to follow. If applicable, do you have a reference that could make this easy for me?

Thanks for your Help

Hi,

Thank you for sharing the code. Interesting work!

I tried to reproduce the results for svhn -> mnist but the result obtained is lowered than source only. Could you please provide the data processing part for these two datasets? Currently I replaced the dataloaders by reading in the .mat files in original datasets. Not sure if this contributes to the performance issue.

Hi, thanks for sharing the code.
However, I'm getting an error

Traceback (most recent call last):
  File "main.py", line 139, in <module>
    main()
  File "main.py", line 131, in main
    DM_ADA_trainer.train()
  File "/home/usr/code/Test/Mixup/trainer_mixed.py", line 290, in train
    errF.backward()
  File "/home/usr/.conda/envs/emad/lib/python3.7/site-packages/torch/tensor.py", line 185, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph)
  File "/home/usr/.conda/envs/emad/lib/python3.7/site-packages/torch/autograd/__init__.py", line 127, in backward
    allow_unreachable=True)  # allow_unreachable flag
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [16, 1, 1, 4]], which is output 0 of UnsqueezeBackward0, is at version 2; expected version 1 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

can you suggest the reason of this error?

I want to use other object detection data sets for this experiment, can I do this, such as Pascal VOC,thanks!

Hi. The value of 'dlabel_src' should be 0, and the value of 'dlabel_tgt 'should be 1?

Hi

Did you also run on Office-31 datasets? What's should be the directory structure for that dataset?