explain.py 文件源码

def preprocess_image(img):
    means=[0.485, 0.456, 0.406]
    stds=[0.229, 0.224, 0.225]

    preprocessed_img = img.copy()[: , :, ::-1]
    for i in range(3):
        preprocessed_img[:, :, i] = preprocessed_img[:, :, i] - means[i]
        preprocessed_img[:, :, i] = preprocessed_img[:, :, i] / stds[i]
    preprocessed_img = \
        np.ascontiguousarray(np.transpose(preprocessed_img, (2, 0, 1)))

    if use_cuda:
        preprocessed_img_tensor = torch.from_numpy(preprocessed_img).cuda()
    else:
        preprocessed_img_tensor = torch.from_numpy(preprocessed_img)

    preprocessed_img_tensor.unsqueeze_(0)
    return Variable(preprocessed_img_tensor, requires_grad = False)