ahmetius / lp-deepssl Goto Github PK

Could you please give the instructions on the processing of mini-imagenet dataset?

Hi! Nice work!

I wanted to reproduce CIFAR100 + your method + mean teacher, 4000 labelled samples. I just run your code and I am getting this in the stage 2:

• Prec@1 56.080 Prec@5 84.920
  Evaluating the EMA model:
• Prec@1 67.520 Prec@5 91.080

However I don't know how to interpret these in the context of the paper. Could you help me please?

Hello,

I have a question because the mini-imagenet file uploaded to github is no longer downloaded.

I would be really grateful if you could tell me how you configured train/test of 50000/10000 about mini-imagenet dataset.

If there is a file, I would be very grateful if you could give me a link.

I look forward to your kind response. Thank you once again for your valuable time.

In this paper, are the training set and the test set put together to construct a graph?

What the parameters of MeanTeacher such as "consistency_type, consistency and lr" do you set on mini-imagenet? Appreciated to be told.

Hello, first thanks for the great paper, and i have a question on the point that has not described in the paper.

according to `line 309 of lp/db_semisuper.py', the one-hot label vector is normalized by the class population, and it's very new implementation detail to me which is not described in the paper.

Please can you give me any evidence for this?

Thank you.
Sincerely.

Hi,

The cosine similarity is used to construct the affinity matrix, but the cosine similarity can be smaller than 0. Is it ok to have negative value in the affinity matrix?

I tried the CIFAR-100 dataset using the parameters specified in the paper. LR =0.02 , B_L = 31 B=128 etc but the code performs very poorly in trainstage2. For CIFAR-100 should the parameter B_L be 31?