Can't Reproduce CIFAR10 Test Image Results

[bri-davis]

Testing Issue

I am unable to reproduce the cifar10 results seen in the paper while using the pre-trained weights. Any insight to what I might be doing wrong?

Setup

1. I downloaded cifar10 data and placed it in ./dataset/cifar10
2. I downloaded the pre-trained cifar10 weights and placed them in ./checkpoints/cifar10.
3. I converted some cifar10 images to grayscale and put them in ./checkpoints/test
4. I run the test script as follows: python test.py --dataset cifar10 --checkpoints-path ./checkpoints --test-input ./checkpoints/test --test-output ./checkpoints/output --gpu-ids 0

Results

The output is blotchy and not what I expected. Here are some example output images produced from the setup above:

[karoly-hars]

@bri-davis I have the same issue.

[yunhao1996]

@bri-davis I have the same issue.

me too!!

[VladVin]

The same problem with Places365 model. Is there any special preprocessing? I tried grayscale images in ranges [0, 1] and [0, 255]. Both give nearly the same output. I'm guessing, there should be some preprocessing, but no information in the repo neither in the arcticle.

P.S. Test on an image not from the dataset:

[Chansi]

Same issue with the Places365 model