search

happyharrycn / actionformer_release

Code release for ActionFormer (ECCV 2022)
MIT License
446 stars 78 forks source link

Clarification on mAP Score Variations Across Machines/GPUs #142

Open krkrkrrk opened 1 week ago

krkrkrrk commented 1 week ago

Hello.

In your paper, you mention that there are minor variations in the average mAP scores across machines/GPUs and computing environments. Could you provide more details on the specific causes of these variations? For example, are they related to hardware differences, numerical precision, or software implementation inconsistencies? Additionally, are there any reliable sources or references that discuss such variations in detail?

tzzcl commented 1 week ago

Hello, I think there may exist following causes:

  1. Hardware differences: for different types of GPUs like A100/H100, even with same operation (conv, etc), the results will not be the same.
  2. Software differences. Different PyTorch/CUDA/CUDNN version may also provides different results (but in a reasonable variation).

For more references, you can refer to reproducibility section in PyTorch.