jlindsey15/feedbackandlocalplasticity

feedbackandlocalplasticity's Issues

data in meta training and Output dimensionality

Hello,

In MetaLearningClassification class, is {x_traj[0:K], y_traj[0:K]} the training data and {x_rand[0]} the test input for the meta learning phase? I was not sure what the names "traj" and "rand" stand for here. Also, is there any reason why your example, x_rand[0] includes 5 instances of 5 characters in order (meaning x_rand[0, :5] is 5 instances of the same character, then x_rand[0, 5:10] is for the next character, and so on) but x_traj[k] contains randomly selected instances of these characters?

I also was wondering why the output dimension of the model is 1000? I was expecting that to be the same as the number of the classes in Omniglot dataset, i.e. 963.

Thanks!

Sampling training data

Hello,

In your paper, it is stated that

"On Omniglot, the meta-training dataset consists of the first 963 character classes, and the meta-testing dataset consists of the the remaining 660 classes. "

in the code, d_traj_iterators

FeedbackAndLocalPlasticity/run_training_classification.py

Line 259 in 2f2df41

d_traj_iterators.append(sampler.sample_task([t]))

and d_rand_iterator

FeedbackAndLocalPlasticity/run_training_classification.py

Line 261 in 2f2df41

d_rand_iterator = sampler.get_complete_iterator()

are both sampled from tasks [0-962]. Based on the paper, I was expecting that the former is sampled from [0-962], while the latter is sampled from [962-1622]. Is there something that I'm missing or the code was simply changed later?

I'm also having hard time understanding what is happening in the sample_training_data function. Would appreciate if you can shed some light on it.