ML in a Box

This repo houses a script used to turn any Paperspace Ubuntu 20.04 based Linux instance into a fully functional machine learning environment for interactive development. Users are free to modify and run against their own Paperspace instance, be it Linux headless or Linux desktop. The only requirement is a working Nvidia GPU driver pre-installed. A pre-built template exists in the Paperspace eco-system as a Public Template based on a base Ubuntu 20.04 image.

Who is this for?

We assume a generic advanced data science user who probably wants GPU access, but not any particular specialized subfield of data science such as computer vision or natural language processing. Such users can build upon this base to create their own stack, or we can create other VMs for subfields, similar to what can be done with Gradient containers.

Software included

Category	Software	Version	Install Method	Why / Notes
GPU	NVidia Driver	470.63.01	pre-installed	Enable Nvidia GPUs. Latest version as of VM creation date
	CUDA	11.4	pre-installed	Nvidia A100 GPUs require CUDA 11+ to work, so 10.x is not suitable
	cuDNN	8.2.4.15-1+cuda11.4	Ubuntu repo	Nvidia GPU deep learning library
	CUDA toolkit	10.1.243-3	apt-get	Needed for nvcc command for cuDNN
Infra	Docker Engine CE	20.10.8, build 3967b7d	Ubuntu repo	Docker Enginer community edition
	NVidia Docker	2.6.0-1	Ubuntu repo	Enable NVidia GPU in Docker containers
Python	Python	3.8.10	Done in above	Most widely used programming language for data science. Version 3.8.10 is already installed by one of the above steps, and is compatible with other software and their versions installed here.
	pip3	20.0.2-5ubuntu1.6	apt-get	Enable easy installation of 1000s of other data science, etc., packages. Is a version 21.2 but apt-get install for Python 3.8 gives 20.0.2, which is OK.
	NumPy	1.21.2	pip3	Handle arrays, matrices, etc., in Python
	Pandas	1.3.3	pip3	De facto standard for data science data exploration/preparation in Python
	Matplotlib	3.4.3	pip3	Widely used plotting library in Python for data science, e.g., scikit-learn plotting requires it
	JupyterLab	3.1.12	pip3	De facto standard for data science using Jupyter notebooks
Machine learning	H2O-3	3.34	pip3	Enables in one place a wide range of ML algorithms outside deep learning at considerably higher performance than scikit-learn: gradient boosted trees, random forest, support vector machine, k-means clustering, generalized linear model (includes logistic regression), isolation forest, etc. Plus auto-ml, model ensembling, and other features.
	Scikit-learn	0.24.2	pip3	Widely used ML library for data science, generally for smaller data or models
	TensorFlow	2.5.0	pip3	Most widely used deep learning library, alongside PyTorch. Note 2.5.0 as 2.6 unclear if works with CUDA 11.4, NVidia, etc.
	NVidia RAPIDS	21.08	conda	GPU acceleration for common ML algorithms
	(PyTorch)	(1.9.0)	source	Not installed because doesn't support CUDA 11.4 yet. Most widely used deep learning library, alongside TensorFlow

Licenses

Software	License	Source
CUDA	NVidia EULA	https://docs.nvidia.com/cuda/eula/index.html
cuDNN	NVidia EULA	https://docs.nvidia.com/deeplearning/cudnn/sla/index.html
Docker Engine	Apache 2.0	https://github.com/moby/moby/blob/master/LICENSE
H2O-3	Apache 2.0	https://github.com/h2oai/h2o-3/blob/master/LICENSE
JupyterLab	New BSD	https://github.com/jupyterlab/jupyterlab/blob/master/LICENSE
Matplotlib	PSF-based	https://matplotlib.org/stable/users/license.html
Numpy	New BSD	https://numpy.org/doc/stable/license.html
NVidia Docker	Apache 2.0	https://github.com/NVIDIA/nvidia-docker/blob/master/LICENSE
NVidia Driver	NVidia EULA	https://www.nvidia.com/en-us/drivers/nvidia-license/
NVidia RAPIDS	"Open source"	https://developer.nvidia.com/rapids
Pandas	New BSD	https://github.com/pandas-dev/pandas/blob/master/LICENSE
Pip3	MIT	https://github.com/pypa/pip/blob/main/LICENSE.txt
PyTorch	New BSD	https://github.com/pytorch/pytorch/blob/master/LICENSE
Python	PSF	https://en.wikipedia.org/wiki/Python_(programming_language)
Scikit-learn	New BSD	https://github.com/scikit-learn/scikit-learn/blob/main/COPYING
TensorFlow	Apache 2.0	https://github.com/tensorflow/tensorflow/blob/master/LICENSE

Information about license types:

Apache 2.0: https://opensource.org/licenses/Apache-2.0
MIT: https://opensource.org/licenses/MIT
New BSD: https://opensource.org/licenses/BSD-3-Clause
PSF = Python Software Foundation: https://en.wikipedia.org/wiki/Python_Software_Foundation_License

Open source software can be used for commercial purposes: https://opensource.org/docs/osd#fields-of-endeavor .

Software not included

Other software considered but not included.

The potential data science stack is far larger than any one person will use, for example, the Anaconda Python distribution for data science has over 7500 optional packages, so we don't attempt to cover everything here.

Some generic categories of software not included:

• Non-data-science software
• Commercial software
• Software not licensed to be used on an available VM template
• Software only used in particular specialized data science subfields (although we assume our users probably want a GPU)

Category	Software	Why Not
Apache	Kafka, Parquet
Classifiers	libsvm, XGBoost	H2O contains SVM and GBM, save on installs
Collections	ELKI, GNU Octave, Weka, Mahout
Connectors	S3, Academic Torrents, Google Drive
Dashboarding	panel, dash, voila, streamlit
Databases	MySQL, Hive, PostgreSQL, Prometheus, Neo4j, MongoDB, Cassandra, Redis	No particular infra to connect to databases
Deep Learning	Caffe, Caffe2, Theano, Keras, PaddlePaddle, Chainer, Torch, MXNet	PyTorch and TensorFlow are dominant, rest niche
Deployment	Dash, TFServing, R Shiny, Flask	Use Gradient Deployments
Distributed.	Horovod, OpenMPI	Use Gradient distributed
Distributions	Anaconda	Includes over 250 packages with many licenses
Feature store	Feast
IDEs	PyCharm, Spyder, RStudio
Image proc	OpenCV, Pillow, scikit-image
Interpretability	LIME/SHAP, Fairlearn, AI Fairness 360, InterpretML
Languages	R, SQL, Java, Julia, C++, JavaScript, Python2, Scala	Python is dominant for data science
Monitoring	Grafana
NLP	HuggingFace, NLTK, GenSim, spaCy
Notebooks	Jupyter, Zeppelin	JupyterLab includes Jupyter notebook
Orchestrators	Kubernetes	Use Gradient cluster
Partners	fast.ai	Could add if we want partner functionality
Pipelines	AirFlow, MLFlow, Intake, Kubeflow
Python libraries	SciPy, statsmodels, pymc3, geopandas, Geopy, LIBSVM	Too many to attempt to cover
PyTorch extensions	Lightning
R packages	ggplot, tidyverse	Could add R if customer demand
Recommenders	TFRS, scikit-surprise
Scalable	Dask, Numba, Spark 1 or 2, Koalas, Hadoop
TensorFlow	TF 1.15, Datasets, Recommenders, TensorBoard, TensorRT	Could add TensorFlow 1.x if customer demand. Requires separate tensorflow-gpu for GPU support.
Viz	Bokeh, Plotly, Holoviz (Datashader), Seaborn, Google FACETS, Excalidraw, GraphViz, ggplot2, d3.js