Post APCdrj6APtEYGuTZp2 by AI4oIc9bMAGSRhD7GS.logan@www.loganjohndarylgraham.xyz
(DIR) More posts by AI4oIc9bMAGSRhD7GS.logan@www.loganjohndarylgraham.xyz
(DIR) Post #APBpZSbfzyLwCdJL5U by volpeon@mk.vulpes.one
2022-11-02T10:04:57.428Z
0 likes, 1 repeats
https://github.com/huggingface/diffusers/pull/532Hell yes, xformers finally got integrated in the diffusers library officially. I used a fork before which was always lagging behind a bit
(DIR) Post #APCKrtwKD8BSlZF5ZQ by volpeon@mk.vulpes.one
2022-11-02T15:55:37.345Z
0 likes, 0 repeats
I can finally use the bfloat16 datatype instead of regular float16. I'm curious if and how the resulting model will be different. What I can tell so far is that the training performance is a bit worse
(DIR) Post #APCKymETGTowtD0azw by volpeon@mk.vulpes.one
2022-11-02T15:56:50.925Z
0 likes, 0 repeats
There are also noticeable differences in the initial validation images
(DIR) Post #APCLLYjpbNtLhMXZui by volpeon@mk.vulpes.one
2022-11-02T16:00:58.705Z
0 likes, 0 repeats
Yup, next round of validation images is even more different. I use euler_a as scheduler which won't converge on one stable image unlike the other scheduler, and that makes even small differences very obvious
(DIR) Post #APCLNIWmqJd40BGgZk by volpeon@mk.vulpes.one
2022-11-02T16:01:15.124Z
0 likes, 0 repeats
Yup, next round of validation images is even more different. I use euler_a as scheduler which won't converge on one stable image unlike the other schedulers, and that makes even small differences very obvious
(DIR) Post #APCc1Zt1K0gigkteIS by volpeon@mk.vulpes.one
2022-11-02T19:07:53.369Z
0 likes, 0 repeats
This was, unfortunately, a waste of time
(DIR) Post #APCdrj6APtEYGuTZp2 by AI4oIc9bMAGSRhD7GS.logan@www.loganjohndarylgraham.xyz
2022-11-02T19:16:45.120Z
1 likes, 0 repeats
@volpeon@mk.vulpes.one perhaps you need to use both in alternating arrangement so it neither diverge nor converge. #🤔