dataloader_cifar.py

import json
import os
import pickle
import random

import _pickle as cPickle
import numpy as np
import torch
import torchvision.transforms as transforms
from PIL import Image
from torch.utils.data import DataLoader, Dataset

from autoaugment import CIFAR10Policy
from randaugment import *


def unpickle(file):
    with open(file, "rb") as fo:
        return cPickle.load(fo, encoding="latin1")


transform_none_10_compose = transforms.Compose(
    [
        transforms.ToTensor(),
        transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
    ]
)


transform_none_100_compose = transforms.Compose(
    [
        transforms.ToTensor(),
        transforms.Normalize((0.507, 0.487, 0.441), (0.267, 0.256, 0.276)),
    ]
)


transform_weak_10_compose = transforms.Compose(
    [
        transforms.RandomCrop(32, padding=4),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
    ]
)


transform_weak_100_compose = transforms.Compose(
    [
        transforms.RandomCrop(32, padding=4),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        transforms.Normalize((0.507, 0.487, 0.441), (0.267, 0.256, 0.276)),
    ]
)


transform_strong_10_compose = transforms.Compose(
    [
        transforms.RandomCrop(32, padding=4),
        transforms.RandomHorizontalFlip(),
        CIFAR10Policy(),
        transforms.ToTensor(),
        transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
    ]
)


transform_strong_100_compose = transforms.Compose(
    [
        transforms.RandomCrop(32, padding=4),
        transforms.RandomHorizontalFlip(),
        CIFAR10Policy(),
        transforms.ToTensor(),
        transforms.Normalize((0.507, 0.487, 0.441), (0.267, 0.256, 0.276)),
    ]
)

transform_strong_randaugment_10_compose = transforms.Compose(
    [
        transforms.RandomCrop(32, padding=4),
        transforms.RandomHorizontalFlip(),
        # RandAugment(1, 6),
        transforms.ToTensor(),
        transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
    ]
)

transform_strong_randaugment_100_compose = transforms.Compose(
    [
        transforms.RandomCrop(32, padding=4),
        transforms.RandomHorizontalFlip(),
        # RandAugment(1, 6),
        transforms.ToTensor(),
        transforms.Normalize((0.507, 0.487, 0.441), (0.267, 0.256, 0.276)),
    ]
)


class cifar_dataset(Dataset):
    def __init__(
        self,
        dataset,
        r,
        noise_mode,
        root_dir,
        transform,
        mode,
        noise_file="",
        preaug_file="",
        pred=[],
        probability=[],
    ):
        self.r = r
        self.transform = transform
        self.mode = mode
        self.preaug_file = preaug_file
        self.transition = {
            0: 0,
            2: 0,
            4: 7,
            7: 7,
            1: 1,
            9: 1,
            3: 5,
            5: 3,
            6: 6,
            8: 8,
        }  # class transition for asymmetric noise

        if self.mode == "test":
            if dataset == "cifar10":
                test_dic = unpickle("%s/test_batch" % root_dir)
                self.test_data = test_dic["data"]
                self.test_data = self.test_data.reshape((10000, 3, 32, 32))
                self.test_data = self.test_data.transpose((0, 2, 3, 1))
                self.test_label = test_dic["labels"]
            elif dataset == "cifar100":
                test_dic = unpickle("%s/test" % root_dir)
                self.test_data = test_dic["data"]
                self.test_data = self.test_data.reshape((10000, 3, 32, 32))
                self.test_data = self.test_data.transpose((0, 2, 3, 1))
                self.test_label = test_dic["fine_labels"]
        else:
            train_data = []
            train_label = []
            if dataset == "cifar10":
                for n in range(1, 6):
                    dpath = "%s/data_batch_%d" % (root_dir, n)
                    data_dic = unpickle(dpath)
                    train_data.append(data_dic["data"])
                    train_label = train_label + data_dic["labels"]
                train_data = np.concatenate(train_data)
            elif dataset == "cifar100":
                train_dic = unpickle("%s/train" % root_dir)
                train_data = train_dic["data"]
                train_label = train_dic["fine_labels"]
            train_data = train_data.reshape((50000, 3, 32, 32))
            train_data = train_data.transpose((0, 2, 3, 1))

            self.train_label = train_label

            if os.path.exists(noise_file):
                noise_label = json.load(open(noise_file, "r"))
            else:  # inject noise
                noise_label = []
                idx = list(range(50000))
                random.shuffle(idx)
                num_noise = int(self.r * 50000)
                noise_idx = idx[:num_noise]
                for i in range(50000):
                    if i in noise_idx:
                        if noise_mode == "sym":
                            if dataset == "cifar10":
                                noiselabel = random.randint(0, 9)
                            elif dataset == "cifar100":
                                noiselabel = random.randint(0, 99)
                            noise_label.append(noiselabel)
                        elif noise_mode == "asym":
                            noiselabel = self.transition[train_label[i]]
                            noise_label.append(noiselabel)
                    else:
                        noise_label.append(train_label[i])
                print(f"saving noisy labels to {noise_file}...")
                json.dump(noise_label, open(noise_file, "w"), indent=4, sort_keys=True)

            if self.preaug_file != "":
                all_augmented = torch.load(self.preaug_file)
                train_data = np.concatenate(
                    (
                        train_data,
                        np.array(all_augmented["samples"], dtype=np.uint8).transpose(
                            (0, 2, 3, 1)
                        ),
                    )
                )
                noise_label = np.concatenate(
                    (
                        noise_label,
                        np.array(all_augmented["labels"]),
                    )
                )

            if self.mode == "all":
                self.train_data = train_data
                self.noise_label = noise_label
            else:
                if self.mode == "labeled":
                    pred_idx = pred.nonzero()[0]
                    self.probability = [probability[i] for i in pred_idx]

                elif self.mode == "unlabeled":
                    pred_idx = (1 - pred).nonzero()[0]

                self.train_data = train_data[pred_idx]

                self.noise_label = [noise_label[i] for i in pred_idx]
                print("%s data has a size of %d" % (self.mode, len(self.train_data)))

    def __getitem__(self, index):
        if self.mode == "labeled":
            img, target, prob = (
                self.train_data[index],
                self.noise_label[index],
                self.probability[index],
            )
            img = Image.fromarray(img)
            img1 = self.transform[0](img)
            img2 = self.transform[1](img)
            if self.transform[2] == None:
                img3 = img1
                img4 = img2
            else:
                img3 = self.transform[2](img)
                img4 = self.transform[3](img)
            return img1, img2, img3, img4, target, prob
        elif self.mode == "unlabeled":
            img = self.train_data[index]
            img = Image.fromarray(img)
            img1 = self.transform[0](img)
            img2 = self.transform[1](img)
            if self.transform[2] == None:
                img3 = img1
                img4 = img2
            else:
                img3 = self.transform[2](img)
                img4 = self.transform[3](img)
            return img1, img2, img3, img4
        elif self.mode == "all":
            img, target = self.train_data[index], self.noise_label[index]
            img = Image.fromarray(img)
            img = self.transform(img)
            return img, target, index
        elif self.mode == "test":
            img, target = self.test_data[index], self.test_label[index]
            img = Image.fromarray(img)
            img = self.transform(img)
            return img, target

    def __len__(self):
        if self.mode != "test":
            return len(self.train_data)
        else:
            return len(self.test_data)


class cifar_dataloader:
    # workaround for windows because
    # python can't pickle lambdas :(

    def prob_transform_100(self, x):
        if random.random() < self.warmup_aug_prob:
            return transform_strong_100_compose(x)
        else:
            return transform_weak_100_compose(x)

    def prob_transform_10(self, x):
        if random.random() < self.warmup_aug_prob:
            return transform_strong_10_compose(x)
        else:
            return transform_weak_10_compose(x)

    def transform_strong_100(self, x):
        return transform_strong_100_compose(x)

    def transform_strong_10(self, x):
        return transform_strong_10_compose(x)

    def transform_weak_100(self, x):
        return transform_weak_100_compose(x)

    def transform_weak_10(self, x):
        return transform_weak_10_compose(x)

    def transform_strong_randaugment_10(self, x):
        return transform_strong_randaugment_10_compose(x)

    def transform_strong_randaugment_100(self, x):
        return transform_strong_randaugment_100_compose(x)

    def transform_none_10(self, x):
        return transform_none_10_compose(x)

    def transform_none_100(self, x):
        return transform_none_100_compose(x)

    def __init__(
        self,
        dataset,
        r,
        noise_mode,
        batch_size,
        warmup_batch_size,
        num_workers,
        root_dir,
        noise_file="",
        preaug_file="",
        augmentation_strategy={},
    ):
        self.dataset = dataset
        self.r = r
        self.noise_mode = noise_mode
        self.batch_size = batch_size
        self.warmup_batch_size = warmup_batch_size
        self.num_workers = num_workers
        self.root_dir = root_dir
        self.noise_file = noise_file
        self.preaug_file = preaug_file
        self.warmup_aug_prob = augmentation_strategy.warmup_aug_probability
        if "randaugment_params" in augmentation_strategy:
            p = augmentation_strategy["randaugment_params"]
            a = RandAugment(p["n"], p["m"])
            transform_strong_randaugment_10_compose.transforms.insert(2, a)
            transform_strong_randaugment_100_compose.transforms.insert(2, a)
        self.transforms = {
            "warmup": self.__getattribute__(augmentation_strategy.warmup_transform),
            "unlabeled": [None for i in range(4)],
            "labeled": [None for i in range(4)],
            "test": None,
        }
        # workaround so it works on both windows and linux
        for i in range(len(augmentation_strategy.unlabeled_transforms)):
            self.transforms["unlabeled"][i] = self.__getattribute__(
                augmentation_strategy.unlabeled_transforms[i]
            )
        for i in range(len(augmentation_strategy.labeled_transforms)):
            self.transforms["labeled"][i] = self.__getattribute__(
                augmentation_strategy.labeled_transforms[i]
            )
        if self.dataset == "cifar10":
            self.transforms["test"] = transforms.Compose(
                [
                    transforms.ToTensor(),
                    transforms.Normalize(
                        (0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)
                    ),
                ]
            )
        elif self.dataset == "cifar100":
            self.transforms["test"] = transforms.Compose(
                [
                    transforms.ToTensor(),
                    transforms.Normalize((0.507, 0.487, 0.441), (0.267, 0.256, 0.276)),
                ]
            )
        if augmentation_strategy.preaugment and not os.path.exists(self.preaug_file):
            print(f"Preaugmenting and saving to {self.preaug_file}...")
            test_dataset = cifar_dataset(
                dataset=self.dataset,
                noise_mode=self.noise_mode,
                noise_file=self.noise_file,
                r=self.r,
                root_dir=self.root_dir,
                transform=self.__getattribute__(
                    augmentation_strategy.preaugment["transform"]
                ),
                mode="all",
            )
            test_loader = DataLoader(
                dataset=test_dataset,
                batch_size=self.batch_size,
                shuffle=False,
                num_workers=self.num_workers,
            )
            all_augmented = {"samples": [], "labels": []}
            for i in range(augmentation_strategy.preaugment["ratio"] - 1):
                for img, target, index in test_loader:
                    for j in range(len(img)):
                        all_augmented["samples"].append(img[j].numpy())
                        all_augmented["labels"].append(target[j])
            torch.save(all_augmented, self.preaug_file)

    def run(self, mode, pred=[], prob=[]):
        if mode == "warmup":
            all_dataset = cifar_dataset(
                dataset=self.dataset,
                noise_mode=self.noise_mode,
                r=self.r,
                root_dir=self.root_dir,
                transform=self.transforms["warmup"],
                mode="all",
                noise_file=self.noise_file,
                preaug_file=self.preaug_file,
            )
            trainloader = DataLoader(
                dataset=all_dataset,
                batch_size=self.warmup_batch_size,
                shuffle=True,
                num_workers=self.num_workers,
            )
            return trainloader

        elif mode == "train":
            labeled_dataset = cifar_dataset(
                dataset=self.dataset,
                noise_mode=self.noise_mode,
                r=self.r,
                root_dir=self.root_dir,
                transform=self.transforms["labeled"],
                mode="labeled",
                noise_file=self.noise_file,
                pred=pred,
                probability=prob,
                preaug_file=self.preaug_file,
            )
            labeled_trainloader = DataLoader(
                dataset=labeled_dataset,
                batch_size=self.batch_size,
                shuffle=True,
                num_workers=self.num_workers,
            )

            unlabeled_dataset = cifar_dataset(
                dataset=self.dataset,
                noise_mode=self.noise_mode,
                r=self.r,
                root_dir=self.root_dir,
                transform=self.transforms["unlabeled"],
                mode="unlabeled",
                noise_file=self.noise_file,
                pred=pred,
                preaug_file=self.preaug_file,
            )
            unlabeled_trainloader = DataLoader(
                dataset=unlabeled_dataset,
                batch_size=self.batch_size,
                shuffle=True,
                num_workers=self.num_workers,
            )
            return labeled_trainloader, unlabeled_trainloader

        elif mode == "test":
            test_dataset = cifar_dataset(
                dataset=self.dataset,
                noise_mode=self.noise_mode,
                r=self.r,
                root_dir=self.root_dir,
                transform=self.transforms["test"],
                mode="test",
            )
            test_loader = DataLoader(
                dataset=test_dataset,
                batch_size=self.batch_size,
                shuffle=False,
                num_workers=self.num_workers,
            )
            return test_loader

        elif mode == "eval_train":
            eval_dataset = cifar_dataset(
                dataset=self.dataset,
                noise_mode=self.noise_mode,
                r=self.r,
                root_dir=self.root_dir,
                transform=self.transforms["test"],
                # ^- this is a small mistake in our implementation!
                # augmentations for eval_train should be weak, not none.
                # although this has a neglibible effect on performance,
                # we feel it is important to note for future readers of this code.
                # see: https://github.com/KentoNishi/Augmentation-for-LNL/issues/4
                mode="all",
                noise_file=self.noise_file,
                preaug_file=self.preaug_file,
            )
            eval_loader = DataLoader(
                dataset=eval_dataset,
                batch_size=self.batch_size,
                shuffle=False,
                num_workers=self.num_workers,
            )
            return eval_loader