search

RenShuhuai-Andy / TimeChat

[CVPR 2024] TimeChat: A Time-sensitive Multimodal Large Language Model for Long Video Understanding
https://arxiv.org/abs/2312.02051
BSD 3-Clause "New" or "Revised" License
267 stars 23 forks source link

Question about batch size #23

Closed gyxxyg closed 4 months ago

gyxxyg commented 4 months ago

Hello, I attempted to replicate the results from the paper using the specified settings. According to the paper, the experiments were conducted on a single server with 8 V100 GPUs, and the total batch size was 32. Consequently, the batch size for each GPU should be 4. However, when I used this value, the training consistently failed.

Could you please provide the training scripts that were used in the paper? I would greatly appreciate your assistance.

RenShuhuai-Andy commented 4 months ago

Hi, what does "The Training Consistently Failed" mean?

Do you mean gpu out-of-memory? If so, please refer to https://github.com/RenShuhuai-Andy/TimeChat/issues/10#issuecomment-1959292668