自动数据增强

下载Notebook下载样例代码查看源文件

MindSpore除了可以让用户自定义数据增强的使用,还提供了一种自动数据增强方式,可以基于特定策略自动对图像进行数据增强处理。

下面分为基于概率基于回调参数两种不同的自动数据增强方式进行介绍。

基于概率的数据增强

MindSpore提供了一系列基于概率的自动数据增强API,用户可以对各种数据增强操作进行随机选择与组合,使数据增强更加灵活。

RandomApply操作

RandomApply操作接收一个数据增强操作列表,以一定的概率顺序执行列表中各数据增强操作,默认概率为0.5,否则都不执行。

在下面的代码示例中,通过调用RandomApply接口来以0.5的概率来顺序执行RandomCropRandomColorAdjust操作。

[1]:
import mindspore.dataset.vision as vision
from mindspore.dataset.transforms import RandomApply

transforms_list = [vision.RandomCrop(512), vision.RandomColorAdjust()]
rand_apply = RandomApply(transforms_list)

RandomChoice

RandomChoice操作接收一个数据增强操作列表transforms,从中随机选择一个数据增强操作执行。

在下面的代码示例中,通过调用RandomChoice操作等概率地在CenterCropRandomCrop中选择一个操作执行。

[2]:
import mindspore.dataset.vision as vision
from mindspore.dataset.transforms import RandomChoice

transforms_list = [vision.CenterCrop(512), vision.RandomCrop(512)]
rand_choice = RandomChoice(transforms_list)

RandomSelectSubpolicy

RandomSelectSubpolicy操作接收一个预置策略列表,包含一系列子策略组合,每一子策略由若干个顺序执行的数据增强操作及其执行概率组成。

对各图像先等概率随机选择一种子策略,再依照子策略中的概率顺序执行各个操作。

在下面的代码示例中,预置了两条子策略:

  • 子策略1中包含RandomRotationRandomVerticalFlip两个操作,概率分别为0.5、1.0。

  • 子策略2中包含RandomRotationRandomColorAdjust两个操作,概率分别为1.0和0.2。

[3]:
import mindspore.dataset.vision as vision
from mindspore.dataset.vision import RandomSelectSubpolicy

policy_list = [
    # policy 1: (transforms, probability)
    [(vision.RandomRotation((45, 45)), 0.5),
     (vision.RandomVerticalFlip(), 1.0)],
    # policy 2: (transforms, probability)
    [(vision.RandomRotation((90, 90)), 1.0),
     (vision.RandomColorAdjust(), 0.2)]
]

policy = RandomSelectSubpolicy(policy_list)

基于回调参数的数据增强

MindSpore的sync_wait接口支持按训练数据的batch或epoch粒度,在训练过程中动态调整数据增强策略,用户可以设定阻塞条件来触发特定的数据增强操作。

sync_wait将阻塞整个数据处理pipeline,直到sync_update触发用户预先定义的callback函数,两者需配合使用,对应说明如下:

  • sync_wait(condition_name, num_batch=1, callback=None):为数据集添加一个阻塞条件condition_name,当sync_update调用时执行指定的callback函数。

  • sync_update(condition_name, num_batch=None, data=None):用于释放对应condition_name的阻塞,并对data触发指定的callback函数。

下面将演示基于回调参数的自动数据增强的用法。

  1. 用户预先定义Augment类,其中preprocess为自定义的数据增强函数,update为更新数据增强策略的回调函数。

[4]:
import numpy as np

class Augment:
    def __init__(self):
        self.ep_num = 0
        self.step_num = 0

    def preprocess(self, input_):
        return np.array((input_ + self.step_num ** self.ep_num - 1),)

    def update(self, data):
        self.ep_num = data['ep_num']
        self.step_num = data['step_num']
  1. 数据处理pipeline先回调自定义的增强策略更新函数update,然后在map操作中按更新后的策略来执行preprocess中定义的数据增强操作。

[5]:
import mindspore.dataset as ds

arr = list(range(1, 4))
dataset = ds.NumpySlicesDataset(arr, shuffle=False)

aug = Augment()
dataset = dataset.sync_wait(condition_name="policy", callback=aug.update)
dataset = dataset.map(operations=[aug.preprocess])
  1. 在每个step中调用sync_update进行数据增强策略的更新。

[6]:
epochs = 5
itr = dataset.create_tuple_iterator(num_epochs=epochs)

step_num = 0
for ep_num in range(epochs):
    for data in itr:
        print("epcoh: {}, step:{}, data :{}".format(ep_num, step_num, data))
        step_num += 1
        dataset.sync_update(condition_name="policy",
                            data={'ep_num': ep_num, 'step_num': step_num})
epcoh: 0, step:0, data :[Tensor(shape=[], dtype=Int64, value= 1)]
epcoh: 0, step:1, data :[Tensor(shape=[], dtype=Int64, value= 2)]
epcoh: 0, step:2, data :[Tensor(shape=[], dtype=Int64, value= 3)]
epcoh: 1, step:3, data :[Tensor(shape=[], dtype=Int64, value= 1)]
epcoh: 1, step:4, data :[Tensor(shape=[], dtype=Int64, value= 5)]
epcoh: 1, step:5, data :[Tensor(shape=[], dtype=Int64, value= 7)]
epcoh: 2, step:6, data :[Tensor(shape=[], dtype=Int64, value= 6)]
epcoh: 2, step:7, data :[Tensor(shape=[], dtype=Int64, value= 50)]
epcoh: 2, step:8, data :[Tensor(shape=[], dtype=Int64, value= 66)]
epcoh: 3, step:9, data :[Tensor(shape=[], dtype=Int64, value= 81)]
epcoh: 3, step:10, data :[Tensor(shape=[], dtype=Int64, value= 1001)]
epcoh: 3, step:11, data :[Tensor(shape=[], dtype=Int64, value= 1333)]
epcoh: 4, step:12, data :[Tensor(shape=[], dtype=Int64, value= 1728)]
epcoh: 4, step:13, data :[Tensor(shape=[], dtype=Int64, value= 28562)]
epcoh: 4, step:14, data :[Tensor(shape=[], dtype=Int64, value= 38418)]

ImageNet自动数据增强

下面以ImageNet数据集上实现AutoAugment作为示例。

针对ImageNet数据集的数据增强策略包含25条子策略,每条子策略中包含两种变换,针对一个batch中的每张图像随机挑选一个子策略的组合,以预定的概率来决定是否执行子策略中的每种变换。

用户可以使用MindSpore中mindspore.dataset.vision模块的RandomSelectSubpolicy接口来实现AutoAugment,在ImageNet分类训练中标准的数据增强方式分以下几个步骤:

  • RandomCropDecodeResize:随机裁剪后进行解码。

  • RandomHorizontalFlip:水平方向上随机翻转。

  • Normalize:归一化。

  • HWC2CHW:图片通道变化。

  1. 定义MindSpore操作到AutoAugment操作的映射:

[1]:
import mindspore.dataset.vision as vision
import mindspore.dataset.transforms as transforms

# define Auto Augmentation operations
PARAMETER_MAX = 10

def float_parameter(level, maxval):
    return float(level) * maxval /  PARAMETER_MAX

def int_parameter(level, maxval):
    return int(level * maxval / PARAMETER_MAX)

def shear_x(level):
    transforms_list = []
    v = float_parameter(level, 0.3)

    transforms_list.append(vision.RandomAffine(degrees=0, shear=(-v, -v)))
    transforms_list.append(vision.RandomAffine(degrees=0, shear=(v, v)))
    return transforms.RandomChoice(transforms_list)

def shear_y(level):
    transforms_list = []
    v = float_parameter(level, 0.3)

    transforms_list.append(vision.RandomAffine(degrees=0, shear=(0, 0, -v, -v)))
    transforms_list.append(vision.RandomAffine(degrees=0, shear=(0, 0, v, v)))
    return transforms.RandomChoice(transforms_list)

def translate_x(level):
    transforms_list = []
    v = float_parameter(level, 150 / 331)

    transforms_list.append(vision.RandomAffine(degrees=0, translate=(-v, -v)))
    transforms_list.append(vision.RandomAffine(degrees=0, translate=(v, v)))
    return transforms.RandomChoice(transforms_list)

def translate_y(level):
    transforms_list = []
    v = float_parameter(level, 150 / 331)

    transforms_list.append(vision.RandomAffine(degrees=0, translate=(0, 0, -v, -v)))
    transforms_list.append(vision.RandomAffine(degrees=0, translate=(0, 0, v, v)))
    return transforms.RandomChoice(transforms_list)

def color_impl(level):
    v = float_parameter(level, 1.8) + 0.1
    return vision.RandomColor(degrees=(v, v))

def rotate_impl(level):
    transforms_list = []
    v = int_parameter(level, 30)

    transforms_list.append(vision.RandomRotation(degrees=(-v, -v)))
    transforms_list.append(vision.RandomRotation(degrees=(v, v)))
    return transforms.RandomChoice(transforms_list)

def solarize_impl(level):
    level = int_parameter(level, 256)
    v = 256 - level
    return vision.RandomSolarize(threshold=(0, v))

def posterize_impl(level):
    level = int_parameter(level, 4)
    v = 4 - level
    return vision.RandomPosterize(bits=(v, v))

def contrast_impl(level):
    v = float_parameter(level, 1.8) + 0.1
    return vision.RandomColorAdjust(contrast=(v, v))

def autocontrast_impl(level):
    return vision.AutoContrast()

def sharpness_impl(level):
    v = float_parameter(level, 1.8) + 0.1
    return vision.RandomSharpness(degrees=(v, v))

def brightness_impl(level):
    v = float_parameter(level, 1.8) + 0.1
    return vision.RandomColorAdjust(brightness=(v, v))
  1. 定义ImageNet数据集的AutoAugment策略:

[ ]:
# define the Auto Augmentation policy
imagenet_policy = [
    [(posterize_impl(8), 0.4), (rotate_impl(9), 0.6)],
    [(solarize_impl(5), 0.6), (autocontrast_impl(5), 0.6)],
    [(vision.Equalize(), 0.8), (vision.Equalize(), 0.6)],
    [(posterize_impl(7), 0.6), (posterize_impl(6), 0.6)],

    [(vision.Equalize(), 0.4), (solarize_impl(4), 0.2)],
    [(vision.Equalize(), 0.4), (rotate_impl(8), 0.8)],
    [(solarize_impl(3), 0.6), (vision.Equalize(), 0.6)],
    [(posterize_impl(5), 0.8), (vision.Equalize(), 1.0)],
    [(rotate_impl(3), 0.2), (solarize_impl(8), 0.6)],
    [(vision.Equalize(), 0.6), (posterize_impl(6), 0.4)],

    [(rotate_impl(8), 0.8), (color_impl(0), 0.4)],
    [(rotate_impl(9), 0.4), (vision.Equalize(), 0.6)],
    [(vision.Equalize(), 0.0), (vision.Equalize(), 0.8)],
    [(vision.Invert(), 0.6), (vision.Equalize(), 1.0)],
    [(color_impl(4), 0.6), (contrast_impl(8), 1.0)],

    [(rotate_impl(8), 0.8), (color_impl(2), 1.0)],
    [(color_impl(8), 0.8), (solarize_impl(7), 0.8)],
    [(sharpness_impl(7), 0.4), (vision.Invert(), 0.6)],
    [(shear_x(5), 0.6), (vision.Equalize(), 1.0)],
    [(color_impl(0), 0.4), (vision.Equalize(), 0.6)],

    [(vision.Equalize(), 0.4), (solarize_impl(4), 0.2)],
    [(solarize_impl(5), 0.6), (autocontrast_impl(5), 0.6)],
    [(vision.Invert(), 0.6), (vision.Equalize(), 1.0)],
    [(color_impl(4), 0.6), (contrast_impl(8), 1.0)],
    [(vision.Equalize(), 0.8), (vision.Equalize(), 0.6)],
]
  1. RandomCropDecodeResize操作后插入AutoAugment变换。

[ ]:
import mindspore.dataset as ds
import mindspore as ms


def create_dataset(dataset_path, train, repeat_num=1,
                   batch_size=32, shuffle=False, num_samples=5):
    # create a train or eval imagenet2012 dataset for ResNet-50
    dataset = ds.ImageFolderDataset(dataset_path, num_parallel_workers=8,
                                    shuffle=shuffle, decode=True)

    image_size = 224

    # define map operations
    if train:
        trans = RandomSelectSubpolicy(imagenet_policy)
    else:
        trans = [vision.Resize(256),
                 vision.CenterCrop(image_size)]
    type_cast_op = transforms.TypeCast(ms.int32)

    # map images and labes
    dataset = dataset.map(operations=[vision.Resize(256), vision.CenterCrop(image_size)], input_columns="image")
    dataset = dataset.map(operations=trans, input_columns="image")
    dataset = dataset.map(operations=type_cast_op, input_columns="label")

    # apply the batch and repeat operation
    dataset = dataset.batch(batch_size, drop_remainder=True)
    dataset = dataset.repeat(repeat_num)

    return dataset
  1. 验证自动数据增强效果:

[ ]:
import matplotlib.pyplot as plt
from download import download

# Define the path to image folder directory.
url = "https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/datasets/ImageNetSimilar.tar.gz"
download(url, "./", kind="tar.gz", replace=True)
dataset = create_dataset(dataset_path="ImageNetSimilar",
                         train=True,
                         batch_size=5,
                         shuffle=False)

epochs = 5
columns = 5
rows = 5
fig = plt.figure(figsize=(8, 8))
itr = dataset.create_dict_iterator(num_epochs=epochs)

for ep_num in range(epochs):
    step_num = 0
    for data in itr:
        for index in range(rows):
            fig.add_subplot(rows, columns, step_num * rows + index + 1)
            plt.imshow(data['image'].asnumpy()[index])
        step_num += 1
plt.show()

为了更好地演示效果,此处只加载5张图片,且读取时不进行shuffle操作,自动数据增强时也不进行NormalizeHWC2CHW操作。

augment

运行结果可以看到,batch中每张图像的增强效果,垂直方向表示1个batch的5张图像,水平方向表示5个batch。

参考文献

[1] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, Quoc V. Le AutoAugment: Learning Augmentation Policies from Data.