Specifying the class frequencies

[dtafler]

Hi,

thank you for the inspired paper and for uploading your code.

Maybe I am overlooking it, but I cannot find the files specifying the class frequencies for the balanced softmax loss. The path to the file is specified in the config:

https://github.com/jiawei-ren/BalancedMetaSoftmax-Classification/blob/34a61e432881816c2da14d577d6ed63501288f5f/config/CIFAR100_LT/balanced_softmax_imba200.yaml#L6C32-L6C32

and used in the balanced softmax loss function to get the class frequencies:

BalancedMetaSoftmax-Classification/loss/BalancedSoftmaxLoss.py

Lines 26 to 29 in 34a61e4

	def __init__(self, freq_path):
	super(BalancedSoftmax, self).__init__()
	with open(freq_path, 'r') as fd:
	freq = json.load(fd)

Am I correct in assuming that the file simply contains the class frequencies and, once converted to a torch tensor, that the indices of that tensor correspond to the class labels?

[jiawei-ren]

Yes, that's right.