using data parallelization with SageMaker JumpStart

I am trying to train Faster-RCNN model available on SageMaker JumpStart and wonder if it is possible to use Data Parallelization feature to finish the job faster as the size of training data is big? I set the environment variable "LAUNCH_SM_DDP_ENV_NAME" to True inside estimator.JumpStartEstimator class and increased the number of instances to 10 (as example). What happens is that it just launches 10 training jobs running in parallel but it does not finish faster (in fact in finishes the same time as with 1 instance). Any hint is appreciated!!

主题

机器学习和人工智能

标签

计算机视觉 Amazon SageMaker 亚马逊 SageMaker JumpStart

语言

English

alex

已提问 5 个月前2081 查看次数

1 回答

最新
投票最多
评论最多

这些答案有用吗？为正确答案投票，以帮助社区从您的知识中受益。

While I am not sure on the exact Model you are using, I suggest taking a look at training script that JumpStart is using and see if there is any implementation of DDP

Marc

已回答 5 个月前

alex
5 个月前
As per the documentation the only fine-tunable PyTorch Object Detection model on SageMaker JumpStart is "pytorch-od1-fasterrcnn-resnet50-fpn". and I checked its training script and it does not seem to have DDP implemented. So I assume one cannot benefit from the DDP strategy with this model on JS. I assume I will have to implement it by myself by updating the trasnfer_learning.py (docker image entry point).

using data parallelization with SageMaker JumpStart

相关内容