|
一、transforms的介绍
在pytorch中，图像的预处理过程中常常需要对图片的格式、尺寸等做一系列的变化，这就需要借助transforms。
__all__ = ["Compose", "ToTensor", "PILToTensor", "ConvertImageDtype", "ToPILImage", "Normalize", "Resize", "Scale",
"CenterCrop", "Pad", "Lambda", "RandomApply", "RandomChoice", "RandomOrder", "RandomCrop",
"RandomHorizontalFlip", "RandomVerticalFlip", "RandomResizedCrop", "RandomSizedCrop", "FiveCrop", "TenCrop",
"LinearTransformation", "ColorJitter", "RandomRotation", "RandomAffine", "Grayscale", "RandomGrayscale",
"RandomPerspective", "RandomErasing", "GaussianBlur", "InterpolationMode", "RandomInvert", "RandomPosterize",
"RandomSolarize", "RandomAdjustSharpness", "RandomAutocontrast", "RandomEqualize"]
这是官方文档里所有的transforms下的操作，以下就根据官方文档和我自己的理解介绍几个常用的transforms方法。
1、transforms.ToTensor()
ToTensor就是将 "PIL Image" 或者 "numpy.ndarray" 格式转换为tensor格式，tensor格式的数据可以直接作为网络的输入。
from PIL import Image
from torchvision import transforms
img = Image.open("../**.png")
trans_totensor = transforms.ToTensor()
img_tensor = trans_totensor(img)
运行上述范例代码可以很清楚的看到transforms.ToTensor前后的变化。
#原始的图片信息
<PIL.PngImagePlugin.PngImageFile image mode=RGB size=477x362 at 0x1E58CEB8FD0>
#transforms.ToTensor后的图片信息
tensor([[[0.2510, 0.2431, 0.2392, ..., 0.2392, 0.2275, 0.2314],
[0.2471, 0.2392, 0.2353, ..., 0.2471, 0.2627, 0.2627],
[0.2549, 0.2549, 0.2549, ..., 0.2549, 0.2471, 0.2471],
...,
[0.2784, 0.2902, 0.2784, ..., 0.9843, 0.9961, 1.0000],
[0.3020, 0.3059, 0.2824, ..., 0.9608, 0.9843, 0.9647],
[0.3176, 0.2902, 0.2863, ..., 1.0000, 0.9686, 0.9608]],
[[0.2510, 0.2431, 0.2392, ..., 0.2392, 0.2275, 0.2314],
[0.2471, 0.2392, 0.2353, ..., 0.2471, 0.2627, 0.2627],
[0.2549, 0.2549, 0.2549, ..., 0.2549, 0.2471, 0.2471],
...,
[0.2784, 0.2902, 0.2784, ..., 0.9843, 0.9961, 1.0000],
[0.3020, 0.3059, 0.2824, ..., 0.9608, 0.9843, 0.9647],
[0.3176, 0.2902, 0.2863, ..., 1.0000, 0.9686, 0.9608]],
[[0.2510, 0.2431, 0.2392, ..., 0.2392, 0.2275, 0.2314],
[0.2471, 0.2392, 0.2353, ..., 0.2471, 0.2627, 0.2627],
[0.2549, 0.2549, 0.2549, ..., 0.2549, 0.2471, 0.2471],
...,
[0.2784, 0.2902, 0.2784, ..., 0.9843, 0.9961, 1.0000],
[0.3020, 0.3059, 0.2824, ..., 0.9608, 0.9843, 0.9647],
[0.3176, 0.2902, 0.2863, ..., 1.0000, 0.9686, 0.9608]]])
2、transforms.Normalize
Normalize就是使用均值和标准差对张量进行归一化处理，其输出通道计算方法为：
output[channel] = (input[channel] - mean[channel]) / std[channel]
?举例说明，这里设置均值和标准差都为0.5，
trans_norm = transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
img_norm = trans_norm(img_tensor)
那么output=(input-0.5)/0.5=2*input-1，根据这个公式计算出对应的输出，结果为：
tensor([[[-0.4980, -0.5137, -0.5216, ..., -0.5216, -0.5451, -0.5373],
[-0.5059, -0.5216, -0.5294, ..., -0.5059, -0.4745, -0.4745],
[-0.4902, -0.4902, -0.4902, ..., -0.4902, -0.5059, -0.5059],
...,
[-0.4431, -0.4196, -0.4431, ..., 0.9686, 0.9922, 1.0000],
[-0.3961, -0.3882, -0.4353, ..., 0.9216, 0.9686, 0.9294],
[-0.3647, -0.4196, -0.4275, ..., 1.0000, 0.9373, 0.9216]],
[[-0.4980, -0.5137, -0.5216, ..., -0.5216, -0.5451, -0.5373],
[-0.5059, -0.5216, -0.5294, ..., -0.5059, -0.4745, -0.4745],
[-0.4902, -0.4902, -0.4902, ..., -0.4902, -0.5059, -0.5059],
...,
[-0.4431, -0.4196, -0.4431, ..., 0.9686, 0.9922, 1.0000],
[-0.3961, -0.3882, -0.4353, ..., 0.9216, 0.9686, 0.9294],
[-0.3647, -0.4196, -0.4275, ..., 1.0000, 0.9373, 0.9216]],
[[-0.4980, -0.5137, -0.5216, ..., -0.5216, -0.5451, -0.5373],
[-0.5059, -0.5216, -0.5294, ..., -0.5059, -0.4745, -0.4745],
[-0.4902, -0.4902, -0.4902, ..., -0.4902, -0.5059, -0.5059],
...,
[-0.4431, -0.4196, -0.4431, ..., 0.9686, 0.9922, 1.0000],
[-0.3961, -0.3882, -0.4353, ..., 0.9216, 0.9686, 0.9294],
[-0.3647, -0.4196, -0.4275, ..., 1.0000, 0.9373, 0.9216]]])
这里可以稍做验证，取第一个数0.2510，计算“（0.2510-0.5）/0.5=－0.4980”，发现与结果吻合，其他同理，可自行挑选验证。
3、transforms.Resize
Resize就是对图像的尺寸进行变换，这种变换并不会改变图像的格式。
trans_resize = transforms.Resize((224, 224))
img_resize = trans_resize(img)
输出结果为：
<PIL.Image.Image image mode=RGB size=224x224 at 0x2129684C130>
4、transforms.Compose
Compose的功能就是将多个transforms组合起来，这里举例来说明，还是对图片的尺寸进行变换，然后转换为tensor形式
trans_resize_2 = transforms.Resize((224, 224))
trans_compose = transforms.Compose([trans_resize_2, trans_totensor])
img_resize_2 = trans_compose(img)
输出结果为：
tensor([[[0.2471, 0.2588, 0.2510, ..., 0.2392, 0.2431, 0.2431],
[0.2510, 0.2471, 0.2471, ..., 0.2471, 0.2471, 0.2471],
[0.2510, 0.2510, 0.2510, ..., 0.2510, 0.2392, 0.2510],
...,
[0.2824, 0.2902, 0.3020, ..., 0.3255, 0.7608, 0.9882],
[0.2863, 0.2784, 0.2784, ..., 0.5686, 0.9412, 0.9922],
[0.3020, 0.2863, 0.2863, ..., 0.8471, 0.9804, 0.9725]],
[[0.2471, 0.2588, 0.2510, ..., 0.2392, 0.2431, 0.2431],
[0.2510, 0.2471, 0.2471, ..., 0.2471, 0.2471, 0.2471],
[0.2510, 0.2510, 0.2510, ..., 0.2510, 0.2392, 0.2510],
...,
[0.2824, 0.2902, 0.3020, ..., 0.3255, 0.7608, 0.9882],
[0.2863, 0.2784, 0.2784, ..., 0.5686, 0.9412, 0.9922],
[0.3020, 0.2863, 0.2863, ..., 0.8471, 0.9804, 0.9725]],
[[0.2471, 0.2588, 0.2510, ..., 0.2392, 0.2431, 0.2431],
[0.2510, 0.2471, 0.2471, ..., 0.2471, 0.2471, 0.2471],
[0.2510, 0.2510, 0.2510, ..., 0.2510, 0.2392, 0.2510],
...,
[0.2824, 0.2902, 0.3020, ..., 0.3255, 0.7608, 0.9882],
[0.2863, 0.2784, 0.2784, ..., 0.5686, 0.9412, 0.9922],
[0.3020, 0.2863, 0.2863, ..., 0.8471, 0.9804, 0.9725]]])
5、transforms.RandomCrop
RandomCrop是对图像进行随机裁剪，这里可以用来进行数据增强等，具体实现代码如范例程序。
trans_random = transforms.RandomCrop(224)
trans_compose_2 = transforms.Compose([trans_random, trans_totensor])
for i in range(10):
img_crop = trans_compose_2(img)
以上就是一些常见的transforms操作，为了更加直观的看出图像预处里过程，可以使用tensorboard。
以上就是我关于tansforms中一些常见操作的介绍，写的不是很规范，欢迎大家批评与讨论，我们一起进步！
|