Zhipei Xu comments

Results 9 comments of


                                            Zhipei Xu

[Usage] `resume_from_checkpoint` fails when finetuning in the lora settings

> I encountered this error while resuming the checkpoint of Lora training. I found that this is basically due to the old version of Transformers that LLaVA is using. Please...

[Usage] Extremely Slow Inference Speed

I have the same problem as you, how did you solve it?

About training dataset 'Fantastic Reality'

@CauchyComplete Hello! I just sent you an email to apply for the Fantastic Reality Dataset. I really need it, thank you very much!

Llava-video slowfast mode

I have the same question, has anyone found the answer?

Serious misalignment in LLaVA implementation

+1， looking forward to the author’s improvement @hiyouga

> 你在llava中使用了“--mm_projector_lr 2e-5”，在llama-factory里没有这个设置，因此mm_projector应该也使用了lr 2e-4。这会导致训练不同。感谢您的回复，那请问我在llama-factory中直接把学习率设置为2e-5，是否会有更好的效果呢？另外，使用llama-factory微调其他MLLM如Qwen，LLAVA-Next等是否有相似的问题呢？

与LLaVA官方代码训练结果性能相差较大

> > > 你在llava中使用了“--mm_projector_lr 2e-5”，在llama-factory里没有这个设置，因此mm_projector应该也使用了lr 2e-4。这会导致训练不同。 > > > > > > 感谢您的回复，那请问我在llama-factory中直接把学习率设置为2e-5，是否会有更好的效果呢？另外，使用llama-factory微调其他MLLM如Qwen，LLAVA-Next等是否有相似的问题呢？ > > 我没试过。估计不行。如果采用类似LLaVA的visual-encoder → mm_projector → LLM的architecture，应该都会存在类似的问题。mm_projector一般在训练中被要求变动更小。好的，非常感谢~

Code Bug for llava/model/llava_arch.py

I have the same question, has anyone found the answer?

关于处理后的数据集获取

作者您好，我们对您的工作非常感兴趣，想获取一下处理后的FaceForensics++, Celeb-DF和DFDC数据集来使用，非常感谢！这是我的工作邮箱：[[email protected]](mailto:[email protected])