loyoyang / decota Goto Github PK

Are there any examples of using this library with a custom non-image dataset for either classification or regression tasks? For example (classification):

# Source train labeled
Xs = pd.DataFrame({
  'Feature1': [1,2,3],
  'Feature2': [10,20,30],
  'Feature3': [100,200,300],
})
ys = [0,1,0]

# Target train labeled
Xt = pd.DataFrame({
  'Feature1': [1,2],
  'Feature2': [10,20],
  'Feature3': [100,200],
})
yt = [0,1]

# Target unlabeled train
Xt_unlabeled = pd.DataFrame({
  'Feature1': [1,2,3,4,5,6],
  'Feature2': [10,20,30,40,50,60],
  'Feature3': [100,200,300,400,500,600],
})

# Target test
Xt = pd.DataFrame({
  'Feature1': [1,20],
  'Feature2': [10,200],
  'Feature3': [100,2000],
})
yt = [0,1]

Thanks for sharing the code of this work. Recently I am trying to run the code with default setting, and find before starting, the code will load the pretrained models under pretrained_models folder in default by this code. When I try to take the original ImageNet pretraining weight of torch, I found the results on Real-Clipart 3 shot varies obviously.

When training with provided pretrained model, the results ended up with test acc 77.4 after 20000 steps of training, while when training from an ImageNet pretraining model, the results can only achieve 70.1.

I noticed that in the implementation detail of the paper, the two networks are first pretrained with labeled data, but did not find detailed information about the pretraining process, can you provide any suggustion on the setting of pretraining (e.g. starting model, training steps, learning rate), thanks~