Worse translation results after updating to pytorch 0.4.0 about munit HOT 6 OPEN

@HsinYingLee and @milankolkata I think the code is now working properly in pytorch 0.4 now.

We spent sometime in the past few days playing with this summer2winter_yosemite256. We found that when enabling explicit cycle consistency loss, the model converges faster and better. The new config file can be found in configs/summer2winter_yosemite256_folder.yaml .

from munit.

@HsinYingLee I also have the similar problem, how many images do you have in your training set(trainA and trainB)?

from munit.

@milankolkata According to the the recent commit, the degraded issue is due to setting <track_running_stats=True> for instance normalization. I haven't tried the updated code yet but I believe the problem should be fixed.

from munit.

@HsinYingLee Thanks for mentioning it. I am trying the new code. How many images did you use when training the model?

from munit.

@milankolkata In the commit 972e42, the custom layernorm only supports one image per batch. With the new commit 4c21350, it supports multiple images per batch. However, the time required for each iteration would increase about 4 times when you use a batch size greater than 1. (This is due to the change of the way pytorch implements view function in 0.4. For training with batch size greater than 1, please roll back to pytorch 0.3 and use munit_pytorch0.3.

BTW, I am still confirming if the performance is the same for pytorch 0.3 and 0.4.

from munit.

@mingyuliutw Thanks! I will try munit_pytorch0.3 and hope it could speed the training up. Also, expect the code supports the multi-GPUs in the future.

from munit.