Terry Zhang
Terry Zhang
The default dump method `pickle.dumps` has inconsistency in python3.0-3.7 and python3.8+ because the default value for `protocol` changes. Protocol version 3 is the default protocol in Python 3.0–3.7. Protocol version...
> > https://huggingface.co/LanguageBind/Open-Sora-Plan-v1.0.0/tree/main > > 请问这个怎样能成功下载呢？使用网站提供的方式总是无法下载，lfs无法拉取总是报443错误，请问有无直接的模型链接呢？ 用hf-mirror.com
I notice that word ### will result in drawContour in training_masks. Perhaps you can try comment this line: https://github.com/whai362/pan_pp_stable/blob/d52a93a6c6ca4f2ef8b31aa3c6d6bad06daae2b5/dataset/ic15.py#L364
After training, each token of the encoder's output is very close to a certain vector in the codebook. I've tried adding the VQ to refine the encoder's output and find...
It is implemented in `CaptionEmbedder`. Have a look at the method `token_drop`
pip install colossalai==0.4.0
If you strictly follow the installation instructions in readme, you shouldn't have the issue about mmengine. It's been written in requirements.txt. I suggest you rebuild the environment following the docs...
I used hostfile but the program got stuck. --host works well for me. --hostfile, stuck there forever.
don't worry. pad_off is always 0. therefore this function is never called.
尤其是细节比较丰富的视频，比如花海、草丛这种，网格非常明显。VAE还有比较大的优化空间。