Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

MVbench的测试情况 #9

Open
ruishuzhao opened this issue Apr 26, 2024 · 6 comments
Open

MVbench的测试情况 #9

ruishuzhao opened this issue Apr 26, 2024 · 6 comments

Comments

@ruishuzhao
Copy link

作者好,我在本地复现了模型的训练过程。采用videochat2相同的训练集,并且修改了你所提到的两个数据集(videochat1 videochatgpt)的标注内容。
采用4个epoch,在mvbench上的性能大概是51.2%(开源模型本地复现性能54.85%)
基于存在差异较大,
请问,在训练过程中有什么需要注意的事项吗?

@farewellthree
Copy link
Collaborator

你好,可能是两个方面

  1. 两个epoch应该足够了,多了性能会变差。
  2. 确认用的是qa的config吗,训qa时是不用videochat1的数据的

@ruishuzhao
Copy link
Author

你好,可能是两个方面

  1. 两个epoch应该足够了,多了性能会变差。
  2. 确认用的是qa的config吗,训qa时是不用videochat1的数据的

您好,感谢您的回答。
按照您的建议,我使用qa的config配置,然后使用的数据集是qa里面对应的数据集。
因为qa里面的数据集较少,我这里训练出来的模型,在MVbench上的性能更低了。

请问,在训练qa的过程中,都具体用到了哪些训练集呢?

@farewellthree
Copy link
Collaborator

训练qa的数据集都在config/instructblipbase_stllm_qa.yaml里。确定每个数据集都可以正常load吗。如果还是解决不掉可以把训练log邮我看一下

@ruishuzhao
Copy link
Author

训练qa的数据集都在config/instructblipbase_stllm_qa.yaml里。确定每个数据集都可以正常load吗。如果还是解决不掉可以把训练log邮我看一下

感谢邮件回复。
对于config/instructblipbase_stllm_qa.yaml这里配置里面的: caption_videochatgpt: num_frames: 16。
通过videochat2中caption中没有对应的videochatgpt。
这个数据集对应的是conversation_videochatgpt 这个数据集吗?

videochat2的数据集配置如下:https://github.com/OpenGVLab/Ask-Anything/blob/main/video_chat2/configs/instruction_data.py
另外,我这边最后的loss降至0.3+左右。

@farewellthree
Copy link
Collaborator

是的。conversation_videochatgpt和caption_videochatgpt是同一个数据。loss最后在0.3-0.5是正常的。另外在128的batch下epoch数是多少呢

@ruishuzhao
Copy link
Author

是的。conversation_videochatgpt和caption_videochatgpt是同一个数据。loss最后在0.3-0.5是正常的。另外在128的batch下epoch数是多少呢

感谢回复~
在epoch为2,128batch的情况下,step数量为6146。总视频标注条数接近39.4w。
因为这些数据集在下载的过程中,有些视频是丢失的。我这边根据丢失视频情况,对的anno标注数据进行了少量的清洗。
请问你那边step的数量大概是多少呢?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants