- 最新
- 投票最多
- 评论最多
Hello,
The question is quite broad since there will be multiple factors to consider.
Here is the pricing range provided. You can calculate the same and find out if using 2ml.m5.2xlarge with count as 2 OR ml.m5.4xlarge would be better for your use-case.
Pricing- https://aws.amazon.com/sagemaker/pricing/
You can refer to the training best practice #6 and find out what is the specification and which instance would you choose from Compute/Memory Optimized image or the Standard one.
Here is a 3rd party link provided that would could be helpful, however it is not verified by AWS and is based on the research that I found on the internet based on what to consider while choosing a training instance.
Link- https://datachef.co/blog/how-to-choose-the-best-training-instance-on-sagemaker/
Hello, to help on the experimentation, you can create a pipeline and run multiple training jobs simultaneously with different instance configurations https://github.com/aws/amazon-sagemaker-examples/tree/main/sagemaker-pipeline-multi-model. Hope this helps!
相关内容
- AWS 官方已更新 1 年前
- AWS 官方已更新 2 年前
- AWS 官方已更新 2 年前
- AWS 官方已更新 2 年前
Thank you for your responce.
I tried optimization while performing a training job I used ml.m5.12xlarge which clearly underused the machine capacity. (Log : https://drive.google.com/file/d/1dhV8PQdLELho8o2dVWOhyatmBWNOLE8l/view?usp=sharing) When I used ml.c5.4xlarge this instance failed with the following error.
Error :
numpy.core._exceptions.MemoryError: Unable to allocate 5.19 GiB for an array with shape (83, 8398169) and data type float64
So for this particular job, how would you choose the instance?
Crunching ~6 GB data with data shape (83, 8398169) in CSV I'm not time bound No GPU is required as I'm performing SKLearn RFRegression