Question about the accuracy of the result about dsn

[codenie]

The accuracy I get is about 76%. I am trying to get higher accyracy.

python 3.6.12
pytorch 1.4.0
torchvision 0.5.0

To run the code with python3 and pytorch 1.x, I modified some codes. I find something interesting:

1. The transformation of the image dataset matters:

# mnist transformation 1
# I got very low accuracy with the hyper-parameter not modified.
# img_transform_source = transforms.Compose([
#     transforms.Resize(image_size),
#     transforms.ToTensor(),
#     transforms.Normalize(mean=(0.1307,), std=(0.3081,))
# ])

# mnist transformation 2
# accuracy about 76%
img_transform_source = transforms.Compose([
    transforms.Resize(image_size),
    transforms.ToTensor(),
    transforms.Normalize(mean=(0.5,), std=(0.5,))
])

# mnist-m transformation
img_transform_target = transforms.Compose([
    transforms.Resize(image_size),
    transforms.ToTensor(),
    transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])

2. the implementation of SI-MSE:
   The author utilizes loss_recon1 + loss_recon2 in the train.py. Actually, loss_recon1 - loss_recon2 is right. The author finds + is better than -. I find the same result.

If anyone who gets higher accuracy than 76%, can you tell me your implemantation details? Many thanks!