噪声标签学习下的一个任务是：训练集上存在开集噪声和闭集噪声；然后在测试集上对闭集样本进行分类。
训练集中被加入的开集样本，会被均匀得打上闭集样本的标签充当开集噪声；而闭集噪声的设置与一般的噪声标签学习一致，分为对称噪声：随机将闭集样本的标签替换为其他类别；和非对称噪声：将闭集样本的标签替换为特定的类别。
论文实验中，常用cifar数据集模拟这类任务。目前已知有两类方法：
第一类基于cifar100，将100个类的一部分，通常是20个类作为开集样本，将它们标签替换了前80个类作为开集噪声；然后对于后续80个类，选择部分样本设置为对称/非对称闭集噪声。CVPR2022的PNP: Robust Learning From Noisy Labels by Probabilistic Noise Prediction提供的代码中，使用了这种方法。但是，如果要考虑非对称噪声，在cifar10上就很难实现，cifar10的类的顺序不像cifar100那样有规律，不好设置闭集噪声。
第二类方法适用cifar10和cifar100，保持原始数据集的样本数不变，使用额外的数据集（通常是imagenet32、places365）代替部分样本作为开集噪声，对于剩下的非开集噪声样本再设置闭集噪声。ECCV2022的Embedding contrastive unsupervised features to cluster in-and out-of-distribution noise in corrupted image datasets提供的代码使用了这种方式。
places365可以使用torchvision.datasets.Places365下载，由于训练集较大，通常是用它的验证集作为辅助数据集。
imagenet32是imagnet的32x32版本，同样是1k类，但是类的具体含义的顺序与imagenet不同，imagenet32类的具体含义可见这里。image32下载地址在对应论文A downsampled variant of imagenet as an alternative to the cifar datasets提供的链接。
接下来是用第二种方法，辅助数据集使用imagenet32，基于cifar构造含开集闭集噪声的训练集。
设计imagenet32数据集
import os
import pickle
import numpy as np
from PIL import Image
from torch.utils.data import Dataset
_train_list = ['train_data_batch_1',
'train_data_batch_2',
'train_data_batch_3',
'train_data_batch_4',
'train_data_batch_5',
'train_data_batch_6',
'train_data_batch_7',
'train_data_batch_8',
'train_data_batch_9',
'train_data_batch_10']
_val_list = ['val_data']
def get_dataset(transform_train, transform_test):
# prepare datasets
# Train set
train = Imagenet32(train=True, transform=transform_train) # Load all 1000 classes in memory
# Test set
test = Imagenet32(train=False, transform=transform_test) # Load all 1000 test classes in memory
return train, test
class Imagenet32(Dataset):
def __init__(self, root='~/data/imagenet32', train=True, transform=None):
if root[0] == '~':
root = os.path.expanduser(root)
self.transform = transform
size = 32
# Now load the picked numpy arrays
if train:
data, labels = [], []
for f in _train_list:
file = os.path.join(root, f)
with open(file, 'rb') as fo:
entry = pickle.load(fo, encoding='latin1')
data.append(entry['data'])
labels += entry['labels']
data = np.concatenate(data)
else:
f = _val_list[0]
file = os.path.join(root, f)
with open(file, 'rb') as fo:
entry = pickle.load(fo, encoding='latin1')
data = entry['data']
labels = entry['labels']
data = data.reshape((-1, 3, size, size))
self.data = data.transpose((0, 2, 3, 1)) # Convert to HWC
labels = np.array(labels) - 1
self.labels = labels.tolist()
def __getitem__(self, index):
img, target = self.data[index], self.labels[index]
img = Image.fromarray(img)
if self.transform is not None:
img = self.transform(img)
return img, target, index
def __len__(self):
return len(self.data)
目录结构：
imagenet32
├─ train_data_batch_1
├─ train_data_batch_10
├─ train_data_batch_2
├─ train_data_batch_3
├─ train_data_batch_4
├─ train_data_batch_5
├─ train_data_batch_6
├─ train_data_batch_7
├─ train_data_batch_8
├─ train_data_batch_9
└─ val_data
设计cifar数据集
import torchvision
import numpy as np
from dataset.imagenet32 import Imagenet32
class CIFAR10(torchvision.datasets.CIFAR10):
def __init__(self, root='~/data', train=True, transform=None,
r_ood=0.2, r_id=0.2, seed=0, corruption='imagenet', ):
nb_classes = 10
self.nb_classes = nb_classes
super().__init__(root, train=train, transform=transform)
if train is False:
return
np.random.seed(seed)
if r_ood > 0.:
ids_ood = [i for i in range(len(self.targets)) if np.random.random() < r_ood]
if corruption == 'imagenet':
imagenet32 = Imagenet32(root='~/data/imagenet32', train=True)
img_ood = imagenet32.data[np.random.permutation(range(len(imagenet32)))[:len(ids_ood)]]
else:
raise ValueError(f'Unknown corruption: {corruption}')
self.ids_ood = ids_ood
self.data[ids_ood] = img_ood
if r_id > 0.:
ids_not_ood = [i for i in range(len(self.targets)) if i not in ids_ood]
ids_id = [i for i in ids_not_ood if np.random.random() < (r_id / (1 - r_ood))]
for i, t in enumerate(self.targets):
if i in ids_id:
self.targets[i] = int(np.random.random() * nb_classes)
self.ids_id = ids_id
class CIFAR100(torchvision.datasets.CIFAR100):
def __init__(self, root='~/data', train=True, transform=None,
r_ood=0.2, r_id=0.2, seed=0, corruption='imagenet', ):
nb_classes = 100
self.nb_classes = nb_classes
super().__init__(root, train=train, transform=transform)
if train is False:
return
np.random.seed(seed)
if r_ood > 0.:
ids_ood = [i for i in range(len(self.targets)) if np.random.random() < r_ood]
if corruption == 'imagenet':
imagenet32 = Imagenet32(root='~/data/imagenet32', train=True)
img_ood = imagenet32.data[np.random.permutation(range(len(imagenet32)))[:len(ids_ood)]]
else:
raise ValueError(f'Unknown corruption: {corruption}')
self.ids_ood = ids_ood
self.data[ids_ood] = img_ood
if r_id > 0.:
ids_not_ood = [i for i in range(len(self.targets)) if i not in ids_ood]
ids_id = [i for i in ids_not_ood if np.random.random() < (r_id / (1 - r_ood))]
for i, t in enumerate(self.targets):
if i in ids_id:
self.targets[i] = int(np.random.random() * nb_classes)
self.ids_id = ids_id
查看统计结果
import pandas as pd
import altair as alt
from dataset.cifar import CIFAR10, CIFAR100
# Initialize CIFAR10 dataset
cifar10 = CIFAR10()
cifar100 = CIFAR100()
def statistics_samples(dataset):
ids_ood = dataset.ids_ood
ids_id = dataset.ids_id
# Collect statistics
statistics = []
for i in range(dataset.nb_classes):
statistics.append({
'class': i,
'id': 0,
'ood': 0,
'clear': 0
})
for i, t in enumerate(dataset.targets):
if i in ids_ood:
statistics[t]['ood'] += 1
elif i in ids_id:
statistics[t]['id'] += 1
else:
statistics[t]['clear'] += 1
df = pd.DataFrame(statistics)
# Melt the DataFrame for Altair
df_melt = df.melt(id_vars='class', var_name='type', value_name='count')
# Create the bar chart
chart = alt.Chart(df_melt).mark_bar().encode(
x=alt.X('class:O', title='Classes'),
y=alt.Y('count:Q', title='Sample Count'),
color='type:N'
)
return chart
chart1 = statistics_samples(cifar10)
chart2 = statistics_samples(cifar100)
chart1 = chart1.properties(
title='cifar10',
width=100, # Adjust width to fit both charts side by side
height=400
)
chart2 = chart2.properties(
title='cifar100',
width=800,
height=400
)
combined_chart = alt.hconcat(chart1, chart2).configure_axis(
labelFontSize=12,
titleFontSize=14
).configure_legend(
titleFontSize=14,
labelFontSize=12
)
combined_chart
运行环境
# Name Version Build Channel
altair 5.3.0 pypi_0 pypi
pytorch 2.3.1 py3.12_cuda12.1_cudnn8_0 pytorch
pandas 2.2.2 pypi_0 pypi