Da Mu

Results 8 comments of Da Mu

looking forward to the publication of pretrained model:)

> Sorry, now the code can run on GPU, but the computer memory occupation rate increases gradually with the code running until it crashes. I can't solve this problem. Is...

> > Because it has been a long time, I can't remember clearly. You need to download the GPU version of TensorFlow and the corresponding cudatoolkit version, and the model...

> Sorry, now the code can run on GPU, but the computer memory occupation rate increases gradually with the code running until it crashes. I can't solve this problem. Is...

I have another question: For example, sound events A and B are detected in a time period, and at the same time, two sound event coordinates X and Y are...

> [@xuyifan-0731](https://github.com/xuyifan-0731) 感谢分享,相较于next token loss的训练,直接使用分类头在您的任务上效果提升了吗,如果有提升,大概提升了多少 他源代码有点问题(可能是Qwen2.5_VL库更新的缘故),我基于他的代码上改了下。在猫狗分类任务上效果很明显,第0个step:测试集准确率3%。第200个step:测试集准确率100%

> > > [@xuyifan-0731](https://github.com/xuyifan-0731) 感谢分享,相较于next token loss的训练,直接使用分类头在您的任务上效果提升了吗,如果有提升,大概提升了多少 > > > > > > 他源代码有点问题(可能是Qwen2.5_VL库更新的缘故),我基于他的代码上改了下。在猫狗分类任务上效果很明显,第0个step:测试集准确率3%。第200个step:测试集准确率100% > > 请问是否有开源实现吗 改下模型代码就可以啦 环境: ``` torch 2.1.2+cu121 transformers 4.53.3 ``` 代码: ``` import os os.environ["TOKENIZERS_PARALLELISM"]...

> > > > > [@xuyifan-0731](https://github.com/xuyifan-0731) 感谢分享,相较于next token loss的训练,直接使用分类头在您的任务上效果提升了吗,如果有提升,大概提升了多少 > > > > > > > > > > > > 他源代码有点问题(可能是Qwen2.5_VL库更新的缘故),我基于他的代码上改了下。在猫狗分类任务上效果很明显,第0个step:测试集准确率3%。第200个step:测试集准确率100% > > > > > > > >...