Paddle：Normalize算子与pytorch的区别

今天看了pytorch、paddle、发现归一化参数设置是不同，所以深入看看到底是怎么使用的。

pytorch：

torchvision.transforms.Normalize(mean, std, inplace=False)

Normalize：是直接减去mean，再除以std
totensor：先除以255,再转成CHW

示例

'''
transforms.ToTensor的作用是将一个PIL Image格式的图片或者是取值范围为[ 0 , 255 ],
形状为[ H × W × C ],numpy.ndarray的数组转换为取值范围为[ 0.0 , 1.0 ]，
形状为[ C × H × W ] 的tensor格式图片。
'''

import cv2
import numpy as np
from PIL import Image
import torchvision.transforms as transforms


img_trans = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize([0.5,0.5,0.5], [0.5,0.5,0.5])])


img_bgr = cv2.imread("1".jpg)
img_rgb = cv2.cvtColor(img_bgr,cv2.COLOR_BGR2RGB)

img = Image.fromarray(img.astype(np.uint8))
img = img_trans(img)

albumentations：

Normalize：源码里面会自动乘上255。所以mean，std参数要自己除以255

ToTensorV2：是调整成CHW

示例

import cv2
import numpy as np
import albumentations as A
from albumentations.core.transforms_interface import ImageOnlyTransform
from albumentations.pytorch import ToTensorV2


img_trans = A.Compose([
    A.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
    ToTensorV2(),
], p=1.0)


img_bgr = cv2.imread("1".jpg)
img_rgb = cv2.cvtColor(img_bgr,cv2.COLOR_BGR2RGB)

img_t = img_trans(image=img_rgb)['image']

paddle：

paddle.vision.transforms.normalize(img, mean, std, data_format='CHW', to_rgb=False)

源码解释： 

示例代码：

import cv2
import numpy as np
from PIL import Image
from paddle.vision.transforms import functional as F


img_trans = T.Compose([
    T.Normalize(
mean = [127.5, 127.5, 127.5], 
std = [127.5, 127.5, 127.5], 
data_format='HWC', 
to_rgb=True)])


img_bgr = cv2.imread("1".jpg)

img = Image.fromarray(img_bgr)
img = img_trans(img)

numpy版本的也记录一下：

def pre_processing(img_roi):
    img = cv2.cvtColor(img_roi,cv2.COLOR_BGR2RGB)
    img = cv2.resize(img_roi,(96,96))
    img = np.float32(img) / 255.0
    
    # norm
    means = np.array([0.5, 0.5, 0.5])
    stds  = np.array([0.5, 0.5, 0.5])
    img  -= means
    img  /= stds
    
    # channel first
    img = np.ascontiguousarray(np.transpose(img, (2, 0, 1)))
    # add batch axis
    img = img[np.newaxis, ...]

    return img