CODEWITHSUNDEEP
amazon-sagemaker
Shark Deng
Shark Deng

Reputation: 1068

Why I feel AWS SageMaker GPU not faster?

I spent 5 days configuring AWS SageMaker. ( contacting to increase GPU notebook limit costs 2 days, downloading and unzip all data cost 2 days, detours such as uploading data to S3 costs 1 day).

When I finally can run code on the SageMaker GPU notebook instance, I found the speed is not faster than my local CPU PC. enter image description here

I used the following code to check GPU. Not sure what's going wrong. Does this happen to you pals? And my credits only have $200 left (Original $750). enter image description here

Upvotes: 2

Views: 739

Answers (1)

Shark Deng
Shark Deng

Reputation: 1068

The problem should lie in video loading. When I train a Resnet model on an image dataset, the speed is pretty fast!! Thank you AWS !! enter image description here

Regarding video loading strategy, I made this kernel for the comparison of different methods.

https://www.kaggle.com/muerbingsha/deepfake-face-detector-evaluate

It turns out imageio_ffmpeg is the fastest way.

Upvotes: 2

Related Questions