vikasverma1077 / ict Goto Github PK

Hi,
Thank you for your impressive work.
When I try to run the code with CIFAR-10 I am getting this error:

line 244, in load_data_subset
    train_sampler, valid_sampler, unlabelled_sampler = get_sampler(train_data.train_labels, labels_per_class, valid_labels_per_class)
AttributeError: 'CIFAR10' object has no attribute 'train_labels'

Any idea how to solve the problem?
Thanks,
Ido

Would you please give me the results of using the same network structure (e.g. with dropout and Gaussian Noise) as in the previous papers if you‘ve run the experiments??? Thank you very much!!!

Hi authors,
The idea of ICT is really interesting, and I want to reproduce the results in the paper. However, the result is really bad when I change the --num_labeled parameter to 100 and 200. For example, if I set num_labeled to 100 and keep other hyper-parameters unchanged, test error on the model with best validation error is 68.83999633789062.
Can you please tell me the hyper-parameters used in CIFAR-10 with 1000 and 2000 labels?
Thank you.

A more semantically coherent interpolation such as McCann's Interpolation (https://twitter.com/gabrielpeyre/status/1006053480455426049) based on optimal transport might significantly generalize this work.

Hey folks - really interested to read the paper behind this code, but I guess it's not out on arxiv yet (or at least I can't find it). Hope you don't mind me opening this issue as a way to be reminded when the paper is out and the link is in the readme.

When you train the model on very few labeled data, like 250 samples, do you use cross validation? Do you split 250 samples, for example, into 220 training and 30 validation? How do you validate?

edit:
Oh I see. You validated on outside labeled data.
So for my case, maybe I can only split a small set out...

Hi there - Thank you for posting this well organized repo! I might have missed it, but could you share the hyperparameter settings (max consistency coefficient and ramp-up schedule, beta distribution parameter, EMA rate, L2 coefficient, batch size) that were used for the two moons dataset? (Figure 1. in the paper)

Thanks!