Get a Docker Image

The easiest to start is by using one of provided Docker images. They contain the core package, all published primitives, and their dependencies. Include CUDA, Tensorflow and other dependencies to support using GPUs.

We will pull the latest Docker image available:

$ docker pull

Beware, Docker images are not small, but around 10 GB each.


See this HOWTO to learn more about Docker images available, how to build them and how to extend them.


While Docker images contain all primitives and their dependencies, some primitives require also additional static files provided to them at runtime (e.g., pretrained model weights). See static files HOWTO for more information.

Get Datasets

We provide many datasets in an uniform structure in a git repository using git LFS to store large files. You can clone the whole repository, but that can take time and disk space. So let’s clone just one dataset:

$ GIT_LFS_SKIP_SMUDGE=1 git clone --recursive
$ git -C datasets lfs pull -I training_datasets/seed_datasets_archive/185_baseball

Despite the dataset itself being small this will still take around 4 GB of your disk space.

Alongside datasets there are also problem descriptions available, one for each dataset.


Datasets are in D3M dataset format. Similarly problems. You can learn more about them in this repository.


You can use and load datasets and problems from other sources and in other formats, too. See loading datasets and problems HOWTO for more information.

Get and Run a Pipeline

We have a metalearning database containing millions of pipeline runs and associated documents. You can use MARVIN to explore it, but for now we will just fetch one existing pipeline from it:

$ wget -O pipeline.json

There are also other ways to obtain a pipeline: you can use one of compatible AutoML systems or you could write a pipeline by hand.

The core package provides a command line interface (CLI) to many parts of what it provides. One of them is also a reference runtime to run pipelines. We can use core package installed inside the Docker image to run the pipeline on the dataset. Furthermore, we can use one of standard data preparation pipelines and the scoring pipeline to evaluate the pipeline with 5 fold cross validation:

$ docker run --rm -v "$(pwd):/data" \
  python3 -m d3m runtime evaluate --pipeline /data/pipeline.json \
  --problem /data/datasets/training_datasets/seed_datasets_archive/185_baseball/185_baseball_problem/problemDoc.json \
  --input /data/datasets/training_datasets/seed_datasets_archive/185_baseball/185_baseball_dataset/datasetDoc.json \
  --data-pipeline /src/d3m/d3m/contrib/pipelines/c8ed65df-aa68-4ee0-bbb5-c5f76a40bcf8.yml \
  --data-param number_of_folds 5 --data-param shuffle true --data-param stratified true

You should see some logging output and at the end:


Metric to use (and the target column) is specified in the problem description. Observe that every time you run this command you get exactly the same results. D3M works hard to provide full reproducibility.


To really achieve full reproducibility we would have to instruct you to use a Docker image at a fixed version and a dataset at a fixed git commit hash. Read more about reproducibility.

Running this is interesting, but to develop using technologies available, you should first install the core package and basic primitives locally.