Spaces:
Configuration error
Configuration error
File size: 8,028 Bytes
72fc481 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 |
import sys
import numpy as np
from PIL import Image
import torchvision
from torch.utils.data.dataset import Subset
from sklearn.metrics.pairwise import cosine_similarity, euclidean_distances
import torch
import torch.nn.functional as F
import random
import json
import os
def get_cifar10(root, cfg_trainer, train=True,
transform_train=None, transform_val=None,
download=False, noise_file = ''):
base_dataset = torchvision.datasets.CIFAR10(root, train=train, download=download)
if train:
train_idxs, val_idxs = train_val_split(base_dataset.targets)
train_dataset = CIFAR10_train(root, cfg_trainer, train_idxs, train=True, transform=transform_train)
val_dataset = CIFAR10_val(root, cfg_trainer, val_idxs, train=train, transform=transform_val)
if cfg_trainer['asym']:
train_dataset.asymmetric_noise()
val_dataset.asymmetric_noise()
else:
train_dataset.symmetric_noise()
val_dataset.symmetric_noise()
print(f"Train: {len(train_idxs)} Val: {len(val_idxs)}") # Train: 45000 Val: 5000
else:
train_dataset = []
val_dataset = CIFAR10_val(root, cfg_trainer, None, train=train, transform=transform_val)
print(f"Test: {len(val_dataset)}")
return train_dataset, val_dataset
def train_val_split(base_dataset: torchvision.datasets.CIFAR10):
num_classes = 10
base_dataset = np.array(base_dataset)
train_n = int(len(base_dataset) * 0.9 / num_classes)
train_idxs = []
val_idxs = []
for i in range(num_classes):
idxs = np.where(base_dataset == i)[0]
np.random.shuffle(idxs)
train_idxs.extend(idxs[:train_n])
val_idxs.extend(idxs[train_n:])
np.random.shuffle(train_idxs)
np.random.shuffle(val_idxs)
return train_idxs, val_idxs
class CIFAR10_train(torchvision.datasets.CIFAR10):
def __init__(self, root, cfg_trainer, indexs, train=True,
transform=None, target_transform=None,
download=False):
super(CIFAR10_train, self).__init__(root, train=train,
transform=transform, target_transform=target_transform,
download=download)
self.num_classes = 10
self.cfg_trainer = cfg_trainer
self.train_data = self.data[indexs]#self.train_data[indexs]
self.train_labels = np.array(self.targets)[indexs]#np.array(self.train_labels)[indexs]
self.indexs = indexs
self.prediction = np.zeros((len(self.train_data), self.num_classes, self.num_classes), dtype=np.float32)
self.noise_indx = []
#self.all_refs_encoded = torch.zeros(self.num_classes,self.num_ref,1024, dtype=np.float32)
def symmetric_noise(self):
self.train_labels_gt = self.train_labels.copy()
#np.random.seed(seed=888)
indices = np.random.permutation(len(self.train_data))
for i, idx in enumerate(indices):
if i < self.cfg_trainer['percent'] * len(self.train_data):
self.noise_indx.append(idx)
self.train_labels[idx] = np.random.randint(self.num_classes, dtype=np.int32)
def asymmetric_noise(self):
self.train_labels_gt = self.train_labels.copy()
for i in range(self.num_classes):
indices = np.where(self.train_labels == i)[0]
np.random.shuffle(indices)
for j, idx in enumerate(indices):
if j < self.cfg_trainer['percent'] * len(indices):
self.noise_indx.append(idx)
# truck -> automobile
if i == 9:
self.train_labels[idx] = 1
# bird -> airplane
elif i == 2:
self.train_labels[idx] = 0
# cat -> dog
elif i == 3:
self.train_labels[idx] = 5
# dog -> cat
elif i == 5:
self.train_labels[idx] = 3
# deer -> horse
elif i == 4:
self.train_labels[idx] = 7
def __getitem__(self, index):
"""
Args:
index (int): Index
Returns:
tuple: (image, target) where target is index of the target class.
"""
img, target, target_gt = self.train_data[index], self.train_labels[index], self.train_labels_gt[index]
# doing this so that it is consistent with all other datasets
# to return a PIL Image
img = Image.fromarray(img)
if self.transform is not None:
img = self.transform(img)
if self.target_transform is not None:
target = self.target_transform(target)
return img,target, index, target_gt
def __len__(self):
return len(self.train_data)
class CIFAR10_val(torchvision.datasets.CIFAR10):
def __init__(self, root, cfg_trainer, indexs, train=True,
transform=None, target_transform=None,
download=False):
super(CIFAR10_val, self).__init__(root, train=train,
transform=transform, target_transform=target_transform,
download=download)
# self.train_data = self.data[indexs]
# self.train_labels = np.array(self.targets)[indexs]
self.num_classes = 10
self.cfg_trainer = cfg_trainer
if train:
self.train_data = self.data[indexs]
self.train_labels = np.array(self.targets)[indexs]
else:
self.train_data = self.data
self.train_labels = np.array(self.targets)
self.train_labels_gt = self.train_labels.copy()
def symmetric_noise(self):
indices = np.random.permutation(len(self.train_data))
for i, idx in enumerate(indices):
if i < self.cfg_trainer['percent'] * len(self.train_data):
self.train_labels[idx] = np.random.randint(self.num_classes, dtype=np.int32)
def asymmetric_noise(self):
for i in range(self.num_classes):
indices = np.where(self.train_labels == i)[0]
np.random.shuffle(indices)
for j, idx in enumerate(indices):
if j < self.cfg_trainer['percent'] * len(indices):
# truck -> automobile
if i == 9:
self.train_labels[idx] = 1
# bird -> airplane
elif i == 2:
self.train_labels[idx] = 0
# cat -> dog
elif i == 3:
self.train_labels[idx] = 5
# dog -> cat
elif i == 5:
self.train_labels[idx] = 3
# deer -> horse
elif i == 4:
self.train_labels[idx] = 7
def __len__(self):
return len(self.train_data)
def __getitem__(self, index):
"""
Args:
index (int): Index
Returns:
tuple: (image, target) where target is index of the target class.
"""
img, target, target_gt = self.train_data[index], self.train_labels[index], self.train_labels_gt[index]
# doing this so that it is consistent with all other datasets
# to return a PIL Image
img = Image.fromarray(img)
if self.transform is not None:
img = self.transform(img)
if self.target_transform is not None:
target = self.target_transform(target)
return img, target, index, target_gt
|