transform 等效转换(参考源码)

下述转换因为原来的transform封装和一些情况的考虑，耗时大概24ms，手动的简化版本是1.2ms.

 71         self.predict_transform = transforms.Compose([
 72             transforms.Resize((256, 256)),
 73             transforms.ToTensor(),
 74             transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
 75         ])

 95     def my_predict_transform(self, Xt):
 96         #Xt = Xt.transpose(2, 0, 1)
 97         Xt = np.array(Xt, dtype=np.float32)
 98         Xt = Xt.transpose(2, 0, 1)
 99         print("Xt.shape:", Xt.shape)
100         #Xt = cv2.resize(Xt, (256, 256)) #already256,last step
101         Xt = Xt/255
102         print("Xt[0]:", Xt[0].shape)
103         print(Xt[0])
104         Xt -= 0.5
105         Xt = Xt/0.5
106 
107         return torch.from_numpy(Xt)

经验证，结果一致