Rysia is a declartive benchmarking framework for deep learning systems. Users specify deep learning workloads in blueprint files without any coding required. Compilation into platform specific execution is done automatically. Benchmarking experiments can be executed locally or in an AWS cloud environment.
To install CPU-restricted packages, run
python setup_cpu.py install
To install GPU-accelerated packages, run
python setup_gpu.py install
Experiment worklaods are specified fully declarative in blueprint files which are implemented as Python modules.
Examples can be found in the blueprints
folder. An overview over available parameters can be found in
the blueprint_example.py
file.
To execute benchmarking workloads locally:
- Install the package
- Specify a blueprint file and set the parameter
aws
to False - Run
rysia blueprint.py
To run benchmarking workloads within the AWS cloud environment:
- Install the awscli and configure a valid AWS account
- Build and push the package to an Amazon ECR repository by
specifying the required parameters in
build_and_push.py
and then runningpython build_and_push.py
- Specify job definitions that are linked to the ECR container images in the AWS Batch web terminal
- Specify a blueprint file and set the parameter
aws
to True - Run
rysia blueprint.py