Hi,

I see the files below are 0 bytes.

CIFAR-100_human_ordered.npy CIFAR-10_human_ordered.npy

I noticed this after running below from the instructions in the readme:

import numpy as np
noise_file = np.load('./data/CIFAR-100_human_ordered.npy', allow_pickle=True)
clean_label = noise_file.item().get('clean_label')
noise_label = noise_file.item().get('noise_label')
# The noisy label matches with following tensorflow dataloader
train_ds, test_ds = tfds.load('cifar100', split=['train','test'], as_supervised=True, batch_size = -1)
train_images, train_labels = tfds.as_numpy(train_ds) 
# You may want to replace train_labels by CIFAR-N noisy label sets

UnpicklingError: Failed to interpret file './data/CIFAR-100_human_ordered.npy' as a pickle

Thanks

Hello,

Can you share the individual labeling information of each labeler for CIFAR-100? (e.g., noise_file['random_label2'])

It will be great if they are provided as CIFAR-10-N.

Thank you.

Hello,

I'm trying to reproduce the performance of "CE" in the paper.

In "worst" label noise, the paper reports the test accuracy of 77.69 (in CIFAR-10-N).

However, when I run the provided code on my machine, the last test epoch accuracy is only 67.89 and it seems overfitting occurs to the training noisy labels.

Did you use validation set for evaluation? Or could you point out if I'm missing something?

Also, there's a discrepancy on the learning rate scheduling between the paper and code.

Learning rate decay is applied in 60th epoch in the code, but the paper says that it is applied in 50th epoch.

Could you check about it?

Thank you.

Hi,

I would like to load the side information and associate it with the correct sample in Tensorflow. What order do the side_info_cifar10N.csv and side_info_cifar100N.csv follow, the one of the PyTorch or the Tensorflow files?

And if they don't come in the Tensorflow order, should I load them like this?

import numpy as np
import pandas as pd

noise_file = np.load('./data/CIFAR-10_human_ordered.npy', allow_pickle=True)
random_label1 = noise_file.item().get('random_label1')

train_ds, test_ds = tfds.load('cifar10', split=['train','test'], as_supervised=True, batch_size = -1)
train_images, train_labels = tfds.as_numpy(train_ds) 

side_info_df = pd.read_csv('side_info_cifar10N.csv')
worker1_id = df['Worker1-id'].to_numpy()

# Reorder side information with correct order
image_order = np.load('image_order_c10.npy')
worker1_id_ordered = worker1_id[image_order // 10]

# Now, the indexing of all arrays matches correctly
first_example = (train_images[0], train_labels[0], worker1_id_ordered[0])

Thank you very much!

Dear game initiator, I noticed that the official code divides the verification set. Do we also need to divide the verification set? Can we train on the complete training set and then test on the test set？