{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# 自动数据增强\n", "\n", "[![下载Notebook](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.7/resource/_static/logo_notebook.png)](https://obs.dualstack.cn-north-4.myhuaweicloud.com/mindspore-website/notebook/r1.7/tutorials/experts/zh_cn/dataset/mindspore_augment.ipynb) [![下载样例代码](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.7/resource/_static/logo_download_code.png)](https://obs.dualstack.cn-north-4.myhuaweicloud.com/mindspore-website/notebook/r1.7/tutorials/experts/zh_cn/dataset/mindspore_augment.py) [![查看源文件](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.7/resource/_static/logo_source.png)](https://gitee.com/mindspore/docs/blob/r1.7/tutorials/experts/source_zh_cn/dataset/augment.ipynb)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "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。\n", "\n", "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。" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 基于概率的数据增强\n", "\n", "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。\n", "\n", "### RandomApply操作\n", "\n", "`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。\n", "\n", "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。" ] }, { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [], "source": [ "import mindspore.dataset.vision.c_transforms as c_vision\n", "from mindspore.dataset.transforms.c_transforms import RandomApply\n", "\n", "transforms_list = [c_vision.RandomCrop(512), c_vision.RandomColorAdjust()]\n", "rand_apply = RandomApply(transforms_list)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### RandomChoice\n", "\n", "`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。\n", "\n", "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。" ] }, { "cell_type": "code", "execution_count": 2, "metadata": {}, "outputs": [], "source": [ "import mindspore.dataset.vision.c_transforms as c_vision\n", "from mindspore.dataset.transforms.c_transforms import RandomChoice\n", "\n", "transforms_list = [c_vision.CenterCrop(512), c_vision.RandomCrop(512)]\n", "rand_choice = RandomChoice(transforms_list)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### RandomSelectSubpolicy\n", "\n", "`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。\n", "\n", "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。\n", "\n", "在下面的代码示例中，预置了两条子策略：\n", "\n", "- 子策略1中包含`RandomRotation`、`RandomVerticalFlip`两个操作，概率分别为0.5、1.0。\n", "\n", "- 子策略2中包含`RandomRotation`和`RandomColorAdjust`两个操作，概率分别为1.0和0.2。" ] }, { "cell_type": "code", "execution_count": 3, "metadata": {}, "outputs": [], "source": [ "import mindspore.dataset.vision.c_transforms as c_vision\n", "from mindspore.dataset.vision.c_transforms import RandomSelectSubpolicy\n", "\n", "policy_list = [\n", " # policy 1: (transforms, probability)\n", " [(c_vision.RandomRotation((45, 45)), 0.5),\n", " (c_vision.RandomVerticalFlip(), 1.0)],\n", " # policy 2: (transforms, probability)\n", " [(c_vision.RandomRotation((90, 90)), 1.0),\n", " (c_vision.RandomColorAdjust(), 0.2)]\n", "]\n", "\n", "policy = RandomSelectSubpolicy(policy_list)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 基于回调参数的数据增强\n", "\n", "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。\n", "\n", "`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：\n", "\n", "- 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。\n", "\n", "- 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。\n", "\n", "下面将演示基于回调参数的自动数据增强的用法。\n", "\n", "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。" ] }, { "cell_type": "code", "execution_count": 4, "metadata": {}, "outputs": [], "source": [ "import numpy as np\n", "\n", "class Augment:\n", " def __init__(self):\n", " self.ep_num = 0\n", " self.step_num = 0\n", "\n", " def preprocess(self, input_):\n", " return np.array((input_ + self.step_num ** self.ep_num - 1),)\n", "\n", " def update(self, data):\n", " self.ep_num = data['ep_num']\n", " self.step_num = data['step_num']" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "2. 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。" ] }, { "cell_type": "code", "execution_count": 5, "metadata": {}, "outputs": [], "source": [ "import mindspore.dataset as ds\n", "\n", "arr = list(range(1, 4))\n", "dataset = ds.NumpySlicesDataset(arr, shuffle=False)\n", "\n", "aug = Augment()\n", "dataset = dataset.sync_wait(condition_name=\"policy\", callback=aug.update)\n", "dataset = dataset.map(operations=[aug.preprocess])" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "3. 在每个step中调用`sync_update`进行数据增强策略的更新。" ] }, { "cell_type": "code", "execution_count": 6, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "epcoh: 0, step:0, data :[Tensor(shape=[], dtype=Int64, value= 1)]\n", "epcoh: 0, step:1, data :[Tensor(shape=[], dtype=Int64, value= 2)]\n", "epcoh: 0, step:2, data :[Tensor(shape=[], dtype=Int64, value= 3)]\n", "epcoh: 1, step:3, data :[Tensor(shape=[], dtype=Int64, value= 1)]\n", "epcoh: 1, step:4, data :[Tensor(shape=[], dtype=Int64, value= 5)]\n", "epcoh: 1, step:5, data :[Tensor(shape=[], dtype=Int64, value= 7)]\n", "epcoh: 2, step:6, data :[Tensor(shape=[], dtype=Int64, value= 6)]\n", "epcoh: 2, step:7, data :[Tensor(shape=[], dtype=Int64, value= 50)]\n", "epcoh: 2, step:8, data :[Tensor(shape=[], dtype=Int64, value= 66)]\n", "epcoh: 3, step:9, data :[Tensor(shape=[], dtype=Int64, value= 81)]\n", "epcoh: 3, step:10, data :[Tensor(shape=[], dtype=Int64, value= 1001)]\n", "epcoh: 3, step:11, data :[Tensor(shape=[], dtype=Int64, value= 1333)]\n", "epcoh: 4, step:12, data :[Tensor(shape=[], dtype=Int64, value= 1728)]\n", "epcoh: 4, step:13, data :[Tensor(shape=[], dtype=Int64, value= 28562)]\n", "epcoh: 4, step:14, data :[Tensor(shape=[], dtype=Int64, value= 38418)]\n" ] } ], "source": [ "epochs = 5\n", "itr = dataset.create_tuple_iterator(num_epochs=epochs)\n", "\n", "step_num = 0\n", "for ep_num in range(epochs):\n", " for data in itr:\n", " print(\"epcoh: {}, step:{}, data :{}\".format(ep_num, step_num, data))\n", " step_num += 1\n", " dataset.sync_update(condition_name=\"policy\",\n", " data={'ep_num': ep_num, 'step_num': step_num})" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## ImageNet自动数据增强\n", "\n", "下面以ImageNet数据集上实现AutoAugment作为示例。\n", "\n", "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。\n", "\n", "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：\n", "\n", "- `RandomCropDecodeResize`：随机裁剪后进行解码。\n", "- `RandomHorizontalFlip`：水平方向上随机翻转。\n", "- `Normalize`：归一化。\n", "- `HWC2CHW`：图片通道变化。" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "1. 定义MindSpore算子到AutoAugment算子的映射：" ] }, { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [], "source": [ "import mindspore.dataset.vision.c_transforms as c_vision\n", "import mindspore.dataset.transforms.c_transforms as c_transforms\n", "\n", "# define Auto Augmentation operators\n", "PARAMETER_MAX = 10\n", "\n", "def float_parameter(level, maxval):\n", " return float(level) * maxval / PARAMETER_MAX\n", "\n", "def int_parameter(level, maxval):\n", " return int(level * maxval / PARAMETER_MAX)\n", "\n", "def shear_x(level):\n", " transforms_list = []\n", " v = float_parameter(level, 0.3)\n", "\n", " transforms_list.append(c_vision.RandomAffine(degrees=0, shear=(-v, -v)))\n", " transforms_list.append(c_vision.RandomAffine(degrees=0, shear=(v, v)))\n", " return c_transforms.RandomChoice(transforms_list)\n", "\n", "def shear_y(level):\n", " transforms_list = []\n", " v = float_parameter(level, 0.3)\n", "\n", " transforms_list.append(c_vision.RandomAffine(degrees=0, shear=(0, 0, -v, -v)))\n", " transforms_list.append(c_vision.RandomAffine(degrees=0, shear=(0, 0, v, v)))\n", " return c_transforms.RandomChoice(transforms_list)\n", "\n", "def translate_x(level):\n", " transforms_list = []\n", " v = float_parameter(level, 150 / 331)\n", "\n", " transforms_list.append(c_vision.RandomAffine(degrees=0, translate=(-v, -v)))\n", " transforms_list.append(c_vision.RandomAffine(degrees=0, translate=(v, v)))\n", " return c_transforms.RandomChoice(transforms_list)\n", "\n", "def translate_y(level):\n", " transforms_list = []\n", " v = float_parameter(level, 150 / 331)\n", "\n", " transforms_list.append(c_vision.RandomAffine(degrees=0, translate=(0, 0, -v, -v)))\n", " transforms_list.append(c_vision.RandomAffine(degrees=0, translate=(0, 0, v, v)))\n", " return c_transforms.RandomChoice(transforms_list)\n", "\n", "def color_impl(level):\n", " v = float_parameter(level, 1.8) + 0.1\n", " return c_vision.RandomColor(degrees=(v, v))\n", "\n", "def rotate_impl(level):\n", " transforms_list = []\n", " v = int_parameter(level, 30)\n", "\n", " transforms_list.append(c_vision.RandomRotation(degrees=(-v, -v)))\n", " transforms_list.append(c_vision.RandomRotation(degrees=(v, v)))\n", " return c_transforms.RandomChoice(transforms_list)\n", "\n", "def solarize_impl(level):\n", " level = int_parameter(level, 256)\n", " v = 256 - level\n", " return c_vision.RandomSolarize(threshold=(0, v))\n", "\n", "def posterize_impl(level):\n", " level = int_parameter(level, 4)\n", " v = 4 - level\n", " return c_vision.RandomPosterize(bits=(v, v))\n", "\n", "def contrast_impl(level):\n", " v = float_parameter(level, 1.8) + 0.1\n", " return c_vision.RandomColorAdjust(contrast=(v, v))\n", "\n", "def autocontrast_impl(level):\n", " return c_vision.AutoContrast()\n", "\n", "def sharpness_impl(level):\n", " v = float_parameter(level, 1.8) + 0.1\n", " return c_vision.RandomSharpness(degrees=(v, v))\n", "\n", "def brightness_impl(level):\n", " v = float_parameter(level, 1.8) + 0.1\n", " return c_vision.RandomColorAdjust(brightness=(v, v))" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "2. 定义ImageNet数据集的AutoAugment策略：" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# define the Auto Augmentation policy\n", "imagenet_policy = [\n", " [(posterize_impl(8), 0.4), (rotate_impl(9), 0.6)],\n", " [(solarize_impl(5), 0.6), (autocontrast_impl(5), 0.6)],\n", " [(c_vision.Equalize(), 0.8), (c_vision.Equalize(), 0.6)],\n", " [(posterize_impl(7), 0.6), (posterize_impl(6), 0.6)],\n", "\n", " [(c_vision.Equalize(), 0.4), (solarize_impl(4), 0.2)],\n", " [(c_vision.Equalize(), 0.4), (rotate_impl(8), 0.8)],\n", " [(solarize_impl(3), 0.6), (c_vision.Equalize(), 0.6)],\n", " [(posterize_impl(5), 0.8), (c_vision.Equalize(), 1.0)],\n", " [(rotate_impl(3), 0.2), (solarize_impl(8), 0.6)],\n", " [(c_vision.Equalize(), 0.6), (posterize_impl(6), 0.4)],\n", "\n", " [(rotate_impl(8), 0.8), (color_impl(0), 0.4)],\n", " [(rotate_impl(9), 0.4), (c_vision.Equalize(), 0.6)],\n", " [(c_vision.Equalize(), 0.0), (c_vision.Equalize(), 0.8)],\n", " [(c_vision.Invert(), 0.6), (c_vision.Equalize(), 1.0)],\n", " [(color_impl(4), 0.6), (contrast_impl(8), 1.0)],\n", "\n", " [(rotate_impl(8), 0.8), (color_impl(2), 1.0)],\n", " [(color_impl(8), 0.8), (solarize_impl(7), 0.8)],\n", " [(sharpness_impl(7), 0.4), (c_vision.Invert(), 0.6)],\n", " [(shear_x(5), 0.6), (c_vision.Equalize(), 1.0)],\n", " [(color_impl(0), 0.4), (c_vision.Equalize(), 0.6)],\n", "\n", " [(c_vision.Equalize(), 0.4), (solarize_impl(4), 0.2)],\n", " [(solarize_impl(5), 0.6), (autocontrast_impl(5), 0.6)],\n", " [(c_vision.Invert(), 0.6), (c_vision.Equalize(), 1.0)],\n", " [(color_impl(4), 0.6), (contrast_impl(8), 1.0)],\n", " [(c_vision.Equalize(), 0.8), (c_vision.Equalize(), 0.6)],\n", "]" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "3. 在`RandomCropDecodeResize`操作后插入AutoAugment变换。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import mindspore.dataset as ds\n", "from mindspore import dtype as mstype\n", "\n", "\n", "def create_dataset(dataset_path, train, repeat_num=1,\n", " batch_size=32, shuffle=True, num_samples=5):\n", " # create a train or eval imagenet2012 dataset for ResNet-50\n", " dataset = ds.ImageFolderDataset(dataset_path, num_parallel_workers=8,\n", " shuffle=shuffle, num_samples=num_samples)\n", "\n", " image_size = 224\n", " mean = [0.485 * 255, 0.456 * 255, 0.406 * 255]\n", " std = [0.229 * 255, 0.224 * 255, 0.225 * 255]\n", "\n", " # define map operations\n", " if train:\n", " trans = imagenet_policy\n", " else:\n", " trans = [c_vision.Decode(),\n", " c_vision.Resize(256),\n", " c_vision.CenterCrop(image_size),\n", " c_vision.Normalize(mean=mean, std=std),\n", " c_vision.HWC2CHW()]\n", " type_cast_op = c_transforms.TypeCast(mstype.int32)\n", "\n", " # map images and labes\n", " dataset = dataset.map(operations=trans, input_columns=\"image\")\n", " dataset = dataset.map(operations=type_cast_op, input_columns=\"label\")\n", "\n", " # apply the batch and repeat operation\n", " dataset = dataset.batch(batch_size, drop_remainder=True)\n", " dataset = dataset.repeat(repeat_num)\n", "\n", " return dataset\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "4. 验证自动数据增强效果:" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import matplotlib.pyplot as plt\n", "\n", "# Define the path to image folder directory.\n", "DATA_DIR = \"/path/to/image_folder_directory\"\n", "dataset = create_dataset(dataset_path=DATA_DIR,\n", " train=True,\n", " batch_size=5,\n", " shuffle=False,\n", " num_samples=5)\n", "\n", "epochs = 5\n", "columns = 5\n", "rows = 5\n", "step_num = 0\n", "fig = plt.figure(figsize=(8, 8))\n", "itr = dataset.create_dict_iterator()\n", "\n", "for ep_num in range(epochs):\n", " for data in itr:\n", " step_num += 1\n", " for index in range(rows):\n", " fig.add_subplot(rows, columns, ep_num * rows + index + 1)\n", " plt.imshow(data['image'].asnumpy()[index])\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "> 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。\n", "\n", "![augment](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.7/tutorials/experts/source_zh_cn/dataset/images/auto_augmentation.png)\n", "\n", "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。\n", "\n", "## 参考文献\n", "\n", "[1] [AutoAugment: Learning Augmentation Policies from Data](https://arxiv.org/abs/1805.09501).\n" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.7.3" } }, "nbformat": 4, "nbformat_minor": 4 }