Generation of asymmetric noise for cifar10 might have problems?

[he-zh]

Hi, thanks for your inspiring work.

In ELR/data_loader/cifar10.py, line 17 in this page, after self.train_labels[idx]=5, the indices of class '5' is enlarged because some labels in class '3' are switched to class '5'. While in the following iterations where i=5, line 4 will select the changed train_labels which equal '5'.
Then a larger set of examples with label '5' will be selected than expected, which results in that more examples being assigned with label '3' after executing the command in line 20.

I assume that if we set the percent = 0.5, after the label flipping step, we will get 12.5% examples with label '3' and 7.5% examples with label '5'. I am wondering if my understanding is right? Or maybe I just misunderstand your code. I would appreciate it very much if you could spare time to respond to my concerns.

1.     def asymmetric_noise(self):
2.         self.train_labels_gt = self.train_labels.copy()
3.         for i in range(self.num_classes):
4.             indices = np.where(self.train_labels == i)[0]
5.             np.random.shuffle(indices)
6.             for j, idx in enumerate(indices):
7.                 if j < self.cfg_trainer['percent'] * len(indices):
8.                     self.noise_indx.append(idx)
9.                     # truck -> automobile
10.                     if i == 9:
11.                         self.train_labels[idx] = 1
12.                     # bird -> airplane
13.                     elif i == 2:
14.                         self.train_labels[idx] = 0
15.                     # cat -> dog
16.                     elif i == 3:
17.                         self.train_labels[idx] = 5
18.                     # dog -> cat
19.                     elif i == 5:
20.                         self.train_labels[idx] = 3
21.                     # deer -> horse
22.                     elif i == 4:
23.                         self.train_labels[idx] = 7

[shengliu66]

Yeah, the definition is following previous work such as https://arxiv.org/abs/1803.11364.