test with val.py about imgcomp-cvpr

[fab-jul]

Please follow the README. You need to download the checkpoints and extract them to ckpts. Then, the possible values for MODEL_ID are given in the README. Let me know if this works for you.

[Jillian2017]

@fab-jul Sorry for the late reply, maybe some version is not the same in my environment, I will try again. Recently I have tried to use your ms_ssim.py to train my network, but the value of ms_ssim was quite high(about 0.99) even at the very begining as I trained the network was trained from scratch. The ms_ssim' value is not reasonable, while the mean square loss is quite low. I wonder have you ever tested the ms_ssim.py before, I did not change any thing to the ms_ssim.py. Need your help, thanks a lot.

[fab-jul]

How did you try training, with which parameters?

We have thoroughly tested ms_ssim.py before using it to train.

[Jillian2017]

Er, I found that in my situation the image's range value was set incorrectly, I noticed that there was a layer of tf.clip_by_value in your decoder network, so I adapted it in my network, and it worked.
Another problem appeared! I test an image with my trained model, and print the ms-ssim value as the way of training process, here I can get the ms-ssim value calculating in your way. After testing, I can get the reconstructed image, then I calculate the ms-ssim value with the original image using a python version ms-ssim , but these 2 values are different, e.g. the former one is 0.186 and the latter one is 0.255, so I'm very confused.

[fab-jul]

Hard to say what the problem is. One thing you have to be careful about is the dtype, it should be uint8 for these kind of comparisons. Also double check the ranges of the values.

[Jillian2017]

Okay, thanks for your reply, I will check.