Shen001
Shen001
here is a problem,maybe not very about your perfect tools. my pose comes from a tool called "ContextCapture",but a photo's pose is a rotation data includes "omega/phi/kappa" and a center...
### What is the feature? 我使用了https://mmengine.readthedocs.io/zh-cn/latest/common_usage/large_model_training.html 里面介绍的几种方法在2台3090(无NVLINK)上进行了训练测试,发现使用torchrun 不用任何策略,比直接python调用快20%-30%,使用deepspeed或者colossalai都慢了2倍甚至以上,我使用mmengine0.10.3;请问这种情况是普遍情况吗?我没有使用fsdp,因为我使用的是1.13版本的torch~ ### Any other context? _No response_
i have no gcp ,maybe just get velocity by pixel?