Artsiom

@ium121, I can advice you to dig into the exemplarsvm/internal/esvm_update_svm.m (this file contains the line with this error) and try to find out why libSVM not learning. Try to look...

@taivanbat Email of the first author, if you need it [email protected]

Hi, Thanks for the interest! 15k steps is indeed to few. As stated in [launch_multistep_training.sh](https://github.com/CompVis/content-style-disentangled-ST/blob/dev_subm_1/launch_multistep_training.sh): ``` ... declare -a IMAGE_SIZES=(128 256 384 512 640 768) declare -a TOTAL_STEPS=(50000 75000 100000...

@ZilinIB any results after the last round (with 1M iteration) ended?

@CrossLee1 you mean 150rk step from the very beginning? @dimakot55 could you pelase take a look?

Guys, try changing hyperparameters a bit if it doesn't work. @dimakot55 is busy with NeurIPS submission right now and will be able to answer you after the deadline.

**Solution:** Do the following changes in File "/content/consistency_models/cm/unet.py", line 359, in __init__ ``` - self.inner_attn = FlashAttention( - attention_dropout=attention_dropout, **factory_kwargs - ) + self.inner_attn = FlashAttention(attention_dropout=attention_dropout) ```

Since I'm running on V100, I also had to disable flash-attention (apparently it only works on A100) ``` index 3fe5184..d9f7c2f 100644 --- a/cm/unet.py +++ b/cm/unet.py @@ -270,7 +270,7 @@ class...

Still doesn't work for me. This is what i get for CD on Imagenet 64, the similar result I get with EDM 

@kebinC did you solve the issue? What is the highest result you get?