Dear authors,
Thanks for your interesting work.
Since we usually train neural network many epochs, and after each epoch's training, we may test the neural network and get a test accuracy. Usually, the highest test accuracy may occur in some epoch of the intermediate training process, not necessarily the last epoch.

In your paper, is the accuracy reported in tables 1, 2, and 3 from the best epoch or the last epoch?

Hi, I want to apply SCEloss to other visual tasks, such as semantic segmentation.How should I consider tweaking the alpha and beta of SCEloss? And,I can't find the parameter A of the paper in the code.

Hi,
thanks for sharing your implementation. I have two questions about it:

1. Does it also work on tabular data?
2. Is it possible to identify the noisy instances (return the noisy IDs or the clean set)?

Thanks!

Hi @YisenWang @xingjunm
I happened to find that you are authors of both SL and D2L.
It is common that results are different if they are from different authors because of implementation and training details.
However, since you are authors of both SL and D2L, I am wondering why the reported results on CIFAR-100 with the same network ResNet44 are quite different.

CIFAR-100 (Results From D2L ICML 2018)
Noise Rate cross-entropy forward backward boot-hard boot-soft D2L
0% 68.20±0.2 68.54±0.3 68.48±0.3 68.31±0.2 67.89±0.2 68.60±0.3
20% 52.88±0.2 60.25±0.2 58.74±0.3 58.49±0.4 57.32±0.3 62.20±0.4
40% 42.85±0.2 51.27±0.3 45.42±0.2 44.41±0.1 41.87±0.1 52.01±0.3
60% 30.09±0.2 41.22±0.3 34.49±0.2 36.65±0.3 32.29±0.1 42.27±0.2

CIFAR-100 (Results From SL ICCV 2019)
Noise rate 0.0 0.2 0.4 0.6
CE 64.34 ± 0.37 59.26 ± 0.39 50.82 ± 0.19 25.39 ± 0.09
LSR 63.68 ± 0.54 58.83 ± 0.40 50.05 ± 0.31 24.68 ± 0.43
Bootstrap 63.26 ± 0.39 57.91 ± 0.42 48.17 ± 0.18 12.27 ± 0.11
Forward 63.99 ± 0.52 59.75 ± 0.34 53.13 ± 0.28 24.70 ± 0.26
D2L 64.60 ± 0.31 59.20 ± 0.43 52.01 ± 0.37 35.27 ± 0.28
GCE 64.43 ± 0.20 59.06 ± 0.27 53.25 ± 0.65 36.16 ± 0.74
SL 66.75 ± 0.04 60.01 ± 0.19 53.69 ± 0.07 41.47 ± 0.04

Results From SL ICCV 2019 are generally lower than Results From D2L IMCL 2018.

I would like to compare with your reported results. I am wondering which one to compare against?
Thanks.

To my understanding, you only want to clip "y_pred_1" and "y_true_2" for calculating \ell_rce, and expect to use "y_pred_2" and "y_true_1" to record the previous value to calculate \ell_ce. However, Python handles variables as references. In other words, when you clip "y_pred_1" and "y_true_2", the variables "y_pred_2" and "y_true_1" are also changed.

Is symmetric_cross_entropy suitable for binary classification issue?
If so, someone can give some hints to modify the symmetric_cross_entropy function?

Why the settings of alpha\beta in the task of cifar10 and cifar100 are vastly differently?
For cifar10, the setting is α = 0.1 β = 1.0.
For cifar100, the setting is α = 6.0 β = 0.1.

import os

import numpy as np

np.random.seed(123)

NUM_CLASSES = {'mnist': 10, 'svhn': 10, 'cifar-10': 10, 'cifar-100': 100}

noise_ratio=40
n = noise_ratio/100.0

nb_subclasses = 5

def build_for_cifar100(size, noise):
""" random flip between two random classes.
"""
assert(noise >= 0.) and (noise <= 1.)

P = np.eye(size)
cls1, cls2 = np.random.choice(range(size), size=2, replace=False)
P[cls1, cls2] = noise
P[cls2, cls1] = noise
P[cls1, cls1] = 1.0 - noise
P[cls2, cls2] = 1.0 - noise

#assert_array_almost_equal(P.sum(axis=1), 1, 1)
return P

P = build_for_cifar100(nb_subclasses, n)

print(np.matrix(P))

[[1. 0. 0. 0. 0. ]
[0. 0.6 0. 0.4 0. ]
[0. 0. 1. 0. 0. ]
[0. 0.4 0. 0.6 0. ]
[0. 0. 0. 0. 1. ]]

I do not understand this noise-transition matrix.

Thanks.

Hi what are optimal values for alpha beta in acse my noise rate is around 0.2 and type asymmetric

By replacing the categorical cross entropy with binary cross entropy and a sigmoid activation can this be extended to be used for a multilabel case?

On the first attempt with this even with a very small beta value it doesn't seem to converge fast.
Is this expected behavior?