自动数据增强
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。
下面分为基于概率和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RandomCrop
和RandomColorAdjust
操作。
[1]:
import mindspore.dataset.vision.c_transforms as c_vision
from mindspore.dataset.transforms.c_transforms import RandomApply
transforms_list = [c_vision.RandomCrop(512), c_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CenterCrop
和RandomCrop
中选择一个操作执行。
[2]:
import mindspore.dataset.vision.c_transforms as c_vision
from mindspore.dataset.transforms.c_transforms import RandomChoice
transforms_list = [c_vision.CenterCrop(512), c_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
RandomRotation
、RandomVerticalFlip
两个操作,概率分别为0.5、1.0。子策略2中包含
RandomRotation
和RandomColorAdjust
两个操作,概率分别为1.0和0.2。
[3]:
import mindspore.dataset.vision.c_transforms as c_vision
from mindspore.dataset.vision.c_transforms import RandomSelectSubpolicy
policy_list = [
# policy 1: (transforms, probability)
[(c_vision.RandomRotation((45, 45)), 0.5),
(c_vision.RandomVerticalFlip(), 1.0)],
# policy 2: (transforms, probability)
[(c_vision.RandomRotation((90, 90)), 1.0),
(c_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。
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']
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])
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c_transforms
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。
定义MindSpore算子到AutoAugment算子的映射:
[1]:
import mindspore.dataset.vision.c_transforms as c_vision
import mindspore.dataset.transforms.c_transforms as c_transforms
# define Auto Augmentation operators
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(c_vision.RandomAffine(degrees=0, shear=(-v, -v)))
transforms_list.append(c_vision.RandomAffine(degrees=0, shear=(v, v)))
return c_transforms.RandomChoice(transforms_list)
def shear_y(level):
transforms_list = []
v = float_parameter(level, 0.3)
transforms_list.append(c_vision.RandomAffine(degrees=0, shear=(0, 0, -v, -v)))
transforms_list.append(c_vision.RandomAffine(degrees=0, shear=(0, 0, v, v)))
return c_transforms.RandomChoice(transforms_list)
def translate_x(level):
transforms_list = []
v = float_parameter(level, 150 / 331)
transforms_list.append(c_vision.RandomAffine(degrees=0, translate=(-v, -v)))
transforms_list.append(c_vision.RandomAffine(degrees=0, translate=(v, v)))
return c_transforms.RandomChoice(transforms_list)
def translate_y(level):
transforms_list = []
v = float_parameter(level, 150 / 331)
transforms_list.append(c_vision.RandomAffine(degrees=0, translate=(0, 0, -v, -v)))
transforms_list.append(c_vision.RandomAffine(degrees=0, translate=(0, 0, v, v)))
return c_transforms.RandomChoice(transforms_list)
def color_impl(level):
v = float_parameter(level, 1.8) + 0.1
return c_vision.RandomColor(degrees=(v, v))
def rotate_impl(level):
transforms_list = []
v = int_parameter(level, 30)
transforms_list.append(c_vision.RandomRotation(degrees=(-v, -v)))
transforms_list.append(c_vision.RandomRotation(degrees=(v, v)))
return c_transforms.RandomChoice(transforms_list)
def solarize_impl(level):
level = int_parameter(level, 256)
v = 256 - level
return c_vision.RandomSolarize(threshold=(0, v))
def posterize_impl(level):
level = int_parameter(level, 4)
v = 4 - level
return c_vision.RandomPosterize(bits=(v, v))
def contrast_impl(level):
v = float_parameter(level, 1.8) + 0.1
return c_vision.RandomColorAdjust(contrast=(v, v))
def autocontrast_impl(level):
return c_vision.AutoContrast()
def sharpness_impl(level):
v = float_parameter(level, 1.8) + 0.1
return c_vision.RandomSharpness(degrees=(v, v))
def brightness_impl(level):
v = float_parameter(level, 1.8) + 0.1
return c_vision.RandomColorAdjust(brightness=(v, v))
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)],
[(c_vision.Equalize(), 0.8), (c_vision.Equalize(), 0.6)],
[(posterize_impl(7), 0.6), (posterize_impl(6), 0.6)],
[(c_vision.Equalize(), 0.4), (solarize_impl(4), 0.2)],
[(c_vision.Equalize(), 0.4), (rotate_impl(8), 0.8)],
[(solarize_impl(3), 0.6), (c_vision.Equalize(), 0.6)],
[(posterize_impl(5), 0.8), (c_vision.Equalize(), 1.0)],
[(rotate_impl(3), 0.2), (solarize_impl(8), 0.6)],
[(c_vision.Equalize(), 0.6), (posterize_impl(6), 0.4)],
[(rotate_impl(8), 0.8), (color_impl(0), 0.4)],
[(rotate_impl(9), 0.4), (c_vision.Equalize(), 0.6)],
[(c_vision.Equalize(), 0.0), (c_vision.Equalize(), 0.8)],
[(c_vision.Invert(), 0.6), (c_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), (c_vision.Invert(), 0.6)],
[(shear_x(5), 0.6), (c_vision.Equalize(), 1.0)],
[(color_impl(0), 0.4), (c_vision.Equalize(), 0.6)],
[(c_vision.Equalize(), 0.4), (solarize_impl(4), 0.2)],
[(solarize_impl(5), 0.6), (autocontrast_impl(5), 0.6)],
[(c_vision.Invert(), 0.6), (c_vision.Equalize(), 1.0)],
[(color_impl(4), 0.6), (contrast_impl(8), 1.0)],
[(c_vision.Equalize(), 0.8), (c_vision.Equalize(), 0.6)],
]
在
RandomCropDecodeResize
操作后插入AutoAugment变换。
[ ]:
import mindspore.dataset as ds
from mindspore import dtype as mstype
def create_dataset(dataset_path, train, repeat_num=1,
batch_size=32, shuffle=True, num_samples=5):
# create a train or eval imagenet2012 dataset for ResNet-50
dataset = ds.ImageFolderDataset(dataset_path, num_parallel_workers=8,
shuffle=shuffle, num_samples=num_samples)
image_size = 224
mean = [0.485 * 255, 0.456 * 255, 0.406 * 255]
std = [0.229 * 255, 0.224 * 255, 0.225 * 255]
# define map operations
if train:
trans = imagenet_policy
else:
trans = [c_vision.Decode(),
c_vision.Resize(256),
c_vision.CenterCrop(image_size),
c_vision.Normalize(mean=mean, std=std),
c_vision.HWC2CHW()]
type_cast_op = c_transforms.TypeCast(mstype.int32)
# map images and labes
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
验证自动数据增强效果:
[ ]:
import matplotlib.pyplot as plt
# Define the path to image folder directory.
DATA_DIR = "/path/to/image_folder_directory"
dataset = create_dataset(dataset_path=DATA_DIR,
train=True,
batch_size=5,
shuffle=False,
num_samples=5)
epochs = 5
columns = 5
rows = 5
step_num = 0
fig = plt.figure(figsize=(8, 8))
itr = dataset.create_dict_iterator()
for ep_num in range(epochs):
for data in itr:
step_num += 1
for index in range(rows):
fig.add_subplot(rows, columns, ep_num * rows + index + 1)
plt.imshow(data['image'].asnumpy()[index])
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Normalize
和HWC2CHW
操作。
运行结果可以看到,batch中每张图像的增强效果,水平方向表示1个batch的5张图像,垂直方向表示5个batch。