{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# 1.Dataloader与Dataset\n", "\n", "Q:如何用torch.utils.data.DataLoader构建可迭代的数据装载器？\n", "- 这是个类：`torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=False, sampler=None, batch_sampler=None, num_workers=0, collate_fn=None, pin_memory=False, drop_last=False, timeout=0, worker_init_fn=None, multiprocessing_context=None, generator=None)`\n", "- dataset: Dataset类，决定数据从哪读取及如何读取\n", "- batchsize：批大小\n", "- num_works：是否多进程读取数据\n", "- shuffle：每个epoch是否乱序\n", "- drop_last：当样本数不能被batchsize整除时，是否舍弃最后一批数据\n", "\n", "Q:torch.utils.data.Dataset的功能是什么？\n", "- Dataset抽象类，所有自定义的Dataset需要继承它，并且复写__getitem__()\n", "- getitem:接收一个索引，返回一个样本\n", "- 要如何读取样本，是用户需要自定义的\n", "- ![](http://anki190912.xuexihaike.com/20200920111241.png?imageView2/2/h/150)\n", "\n", "Q:DataLoader数据读取三问：读哪些，从哪读，怎么读\n", "- 1.读哪些？Sampler输出的Index\n", "- 2.从哪读？Dataset中的data_dir\n", "- 3.怎么读？Dataset的getitem\n", "\n", "Q:DataLoader数据读取流程是怎样？\n", "- ![](http://anki190912.xuexihaike.com/20200920115859.png)\n", "- 1.在for循环中使用DataLoader\n", "- 2.根据是否采用多进程还是单进程的DataLoaderIter\n", "- 3.使用Sampler获得index索引\n", "- 4.拿到index索引然后给DatasetFetcher\n", "- 5.在这里调用Dataset\n", "- 6.Dataset根据给定的索引在getitem中在硬盘里读取实际的图像和和标签\n", "- 7.读取了一个batchsize大小的图像后，通过collate_fn进行整理成一个Batch Data形式，然后输入到模型中训练" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "metadata": {}, "source": [ "# 2.transforms与normalize\n", "\n", "Q:torchvision（计算机视觉工具包）中有哪些模块以及相应的作用？\n", "- torchvision.transforms：常用的图像预处理方法\n", "- torchvision.datasets：常用数据集的dataset实现，MNIST,CIFAR-10,ImageNet等\n", "- torchvision.model：常用的模型预训练，AlexNet,VGG,ResNet,GoogLeNet等\n", "\n", "Q:pytorch的transform机制是在哪一步做？\n", "- ![](http://anki190912.xuexihaike.com/20200920150244.png)\n", "\n", "Q:如何逐channel地对图像进行标准化？\n", "- `transforms.Normalize(mean, std, inplace=False)`\n", "- output = (input - mean) / std\n", "- mean：各通道的均值\n", "- std：各通道的标准差\n", "- inplace：是否原地操作\n", "\n", "# 3.transforms图像增强（一）\n", "Q:如何从图像中心裁剪图片？\n", "- `transforms.CenterCrop`\n", "- size：所需裁剪图片尺寸\n", "\n", "Q:如何从图片中随机裁剪出尺寸为size的图片？\n", "- `transforms.RandomCrop(size, padding=None, pad_if_needed=False, fill=0, padding_mode='constant')`\n", "- size：所需裁剪图片尺寸\n", "- padding：设置填充大小\n", " - 当为a时，上下左右均填充a个像素\n", " - 当为(a,b)时，上下填充b个像素，左右填充a个像素\n", " - 党委(a,b,c,d)时，左，上，右，下分别填充a,b,c,d\n", "- pad_if_need：若图像小于设定size，则填充\n", "- padding_mode：填充模式，有4种模式\n", " - constand：像素值由fill设定\n", " - edge：像素值由图像边缘像素决定\n", " - reflect：镜像填充，最后一个像素不镜像,e.g.：[1,2,3,4]->[3,2,1,2,3,4,3,2]\n", " - symmetric：镜像填充，最后一个像素镜像，e.g.：[1,2,3,4]->[2,1,1,2,3,4,4,3]\n", "- fill：constant时，设置填充的像素值，(R,G,B) or (Gray)\n", "\n", "Q:如何随机大小、长宽比裁剪图片？\n", "- `RandomResizedCrop(size, scale=(0.08, 1.0), ratio=(3/4, 4/3), interpolation)`\n", "- size：所需裁剪图片尺寸\n", "- scale：随机裁剪面积比例，默认(0.08, 1)\n", "- ratio：随机长宽比，默认(3/4, 4/3)\n", "- interpolation：插值方法\n", " - PIL.Image.NEAREST\n", " - PIL.Image.BILINEAR\n", " - PIL.Image.BICUBIC\n", "\n", "Q:FiveCrop和TenCrop的功能是什么？\n", "- `transforms.FiveCrop(size)`\n", "- `transforms.TenCrop(size, vertival_flip=False)`\n", "- 功能：在图像的上下左右以及中心裁剪出尺寸为size的5张图片，TenCrop对这5张图片进行水平或者垂直镜像获得10张图片\n", "- size：所需裁剪图片尺寸\n", "- vertical_flip：是否垂直翻转\n", "\n", "Q:如何按照概率水平（左右）或垂直（上下）翻转图片？\n", "- `RandomHorizontalFlip(p=0.5)`\n", "- `RandomVerticalFlip(p=0.5)`\n", "- p：翻转概率\n", "\n", "Q:如何随机旋转图片？\n", "- `RandomRotation(degress, resample=False, expand=False, center=None)`\n", "- degree：旋转角度\n", " - 当为a时，在(-a, a)之间选择旋转角度\n", " - 当为(a, b)时，在(a,b)之间选择旋转角度\n", "- resample：重采样方法\n", "- expand：是否扩大图片，以保持原图信息\n", "- center：旋转点设置，默认中心选择\n", "\n", "# 4.transforms图像增强(二)\n", "Q:如何对图像边缘进行填充？\n", "- `transforms.Pad(padding, fill=0, padding_mode='constant')`\n", "- padding：设置填充大小\n", " - 当为a时，上下左右均填充a个像素\n", " - 当为(a,b)时，上下填充b个像素，左右填充a个像素\n", " - 党委(a,b,c,d)时，左，上，右，下分别填充a,b,c,d\n", "- padding_mode：填充模式，有4种模式\n", " - constand：像素值由fill设定\n", " - edge：像素值由图像边缘像素决定\n", " - reflect：镜像填充，最后一个像素不镜像,e.g.：[1,2,3,4]->[3,2,1,2,3,4,3,2]\n", " - symmetric：镜像填充，最后一个像素镜像，e.g.：[1,2,3,4]->[2,1,1,2,3,4,4,3]\n", "- fill：constant时，设置填充的像素值，(R,G,B) or (Gray)\n", "\n", "Q:如何调整亮度、对比度、 饱和度和色相？\n", "- `transforms.ColorJitter(brightness=0, contrast=0, saturation=0, hue=0)`\n", "- brightness：亮度调整因子\n", " - 当为a时，从[max(0, 1-a), 1+a]中随机选择\n", " - 当为(a,b)时，从[a,b]中\n", "- contrast：对比度参数，同brightness\n", "- saturation：饱和度参数，同brightness\n", "- hue：色相参数\n", " - 当为a时，从[-a, a]中选择参数，注：0<=a<=0.5\n", " - 当为(a,b)时，从[a,b]中选择参数，注：-0.5 <= a <= b <= 0.5\n", "\n", "Q:如何对把彩色图像转换为灰度图像？\n", "- `Grayscale(num_output_channels)`\n", "- `RandomGrayscale(num_output_channels, p=0.1)`\n", "- num_output_channels：输出通道数，只能设1或3\n", "- p：概率值，图像被转换为灰度图的概率\n", "\n", "Q:如何对图像进行仿射变换？\n", "- `RandomAffile(degrees, translate=None, scale=None, shear=None, resample=False, fillcolor=0)`\n", "- 仿射变换是二维的线性变换，由五种基本原子变换构成，分别是选择、平移、缩放、错切和翻转\n", "- degrees：旋转角度设置\n", "- translate：平移区间设置，如(a,b)，a设置宽(width)，b设置高(height)，图像在宽维度平移的区间为-img_width * a < dx < img_width * a\n", "- scale：缩放比例（以面积为单位）\n", "- fill_color：填充颜色设置\n", "- shear：错切角度设置，有水平错切和垂直错切\n", " - 若为a，则仅在x轴错切，错切角度为(-a, a)之间\n", " - 若为(a, b)，则a设置x轴角度，b设置y的角度\n", " - 若为(a, b, c, d)，则a, b设置x轴角度，c, d设置y轴角度\n", "- resample：重采样方式，有NEAREST、BILINEAR、BICUBIC\n", "\n", "Q:如何对图像进行随机遮挡？\n", "- `RandomErasing(p=0.5, scale=(0.02, 0.33), ratio=(0.3, 3.3), value=0, inplace=Flase)`\n", "- p：概率值，执行该操作的概率\n", "- scale：遮挡区域的面积\n", "- ratio：遮挡区域长宽比\n", "- value：设置遮挡区域的像素值，(R,G,B) or (Gray) or 任意字符串\n", "- 参考文献：《Random Erasing Data Augmentation》\n", "- 注意：这个接收的是Tensor，所以在上一步需要执行`transforms.ToTensor()`\n", "\n", "Q:如何自定义lambda方法？\n", "- `transforms.Lambda(lambd)`\n", "- lambd：lambda匿名函数\n", "- lambda [arg1 [, arg2, ..., argn]]:expression\n", "- ```python\n", "transforms.TenCrop(200, vertical_flip=True)\n", "transforms.Lambda(lambda crops: torch.stack([transforms.Totensor()(crop) for crop in crops]))\n", "```\n", "\n", "Q:如何从一系列transforms方法中随机挑选一个？\n", "- `transforms.RandomChoice([transforms1, transforms2, transforms3])`\n", "\n", "Q:如何依概率执行一组transforms操作？\n", "- `transforms.RandomApply([transforms1, transforms2, transforms3], p=0.5)`\n", "\n", "Q:如何对一组transforms操作打乱顺序？\n", "- `transforms.RandomOrder([transforms1, transforms2, transforms3])`\n", "\n", "Q:用户如何自定义transforms方法？\n", "- 调用方法：\n", "- ```python\n", "class Compose(object):\n", " def __call__(self, img):\n", " for t in self.transforms:\n", " img = t(img)\n", " return img\n", "```\n", "- 要素：\n", " - 1.仅接收一个参数，返回一个参数\n", " - 2.注意上下游的输出与输入\n", "- 通过类实现多参数传入：\n", "- ```python\n", "class YourTransforms(object):\n", " def __init__(self, ...):\n", " ...\n", " def __call__(self, img):\n", " ...\n", " return img\n", "```" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] } ], "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": "3.6.12" } }, "nbformat": 4, "nbformat_minor": 4 }