ArcGIS Pro, Server and the ArcGIS API for Python all include tools to use AI and Deep Learning to solve geospatial problems, such as feature extraction, pixel classification, and feature categorization. This installer includes a broad collection of components, such as PyTorch, TensorFlow, Fast.ai and scikit-learn, for performing deep learning and machine learning tasks, a total collection of 95 packages. These packages can be used with the new Deep Learning Training tools, by using the arcgis.learn module within the ArcGIS API for Python, and directly imported into your own scripts and tools. Most of the tools in this collection will work on any machine, but common deep learning workflows require a recent NVIDIA graphics processing unit (GPU), and problem sizes are bound by available GPU memory, see the requirements section.
This installer adds all the included packages to the default arcgispro-py3 environment that Pro and Server both ship with, and no additional environments are necessary to create in order to get started using the tools. If you do create custom environments, these packages will also be included so you can use the same tools in your own custom environments as well. If you cannot use the installer, you can install the libraries manually using these instructions.
For an example of the kinds of workflows this installer and ArcGIS enables, see the AI & Deep Learning in the UC 2020 Plenary video
- Deep Learning Libraries Installer for ArcGIS Pro 2.6
- Deep Learning Libraries Installer for ArcGIS Server 10.8.1
- Deep Learning Libraries Installer for ArcGIS Server Linux 10.8.1
Once you've installed the deep learning libraries, you can use the Deep Learning Tools to train geospatial deep learning models. You can also find out more about the capabilities of the arcgis.learn module which provides specialized access to many geospatial models beyond those directly available as Geoprocessing tools. Finally, you can add any of the above libraries to your own workflows, by importing the packages listed below.
- Learn ArcGIS lesson on using Deep Learning to Access Palm Tree Health
- Join us in October 2020 for the Spatial Data Science MOOC
A collection of recent User Conference 2020 Technical Workshops on Deep Learning:
- Deep Dive into Deep Learning
- Deep Learning for Geographers
- Using Deep Learning with Imagery in ArcGIS
Most of the packages included in the Deep Learning Libraries installer will work out of the box on any machine configuration. For example, PyTorch optionally can take advantage of a GPU, but will fall back to running its calculations on the CPU if a GPU is not available. However, GPU computation is significantly faster, and some packages such as TensorFlow in this distribution only will work with a supported GPU. CUDA, or Compute Unified Device Architecture, is a general purpose computing platform for GPUs, a requirement for current GPU backed deep learning tools.
| GPU requirement | Supported |
|---|---|
| GPU Type | NVIDIA with CUDA Compute Capability* 3.5 or greater |
| Dedicated graphics memory † | minimum: 2GB recommended: 4GB or more, depending on the size of models trained |
* NVIDIA provides a list of CUDA enabled products and their compute capability.
† GPU memory, unlike system memory, cannot be accessed 'virtually'. If a model training consumes more GPU memory than you have available, it will fail. GPU memory is also shared across all uses of the machine, so open Pro projects with maps and other applications can limit the available memory for use with these tools.
| Library Name | Version | Description |
|---|---|---|
| absl-py | 0.9.0 | Abseil Python Common Libraries |
| appdirs | 1.4.3 | A small Python module for determining appropriate platform-specific directories |
| ase | 3.19.1 | Set of tools for atomistic simulations |
| astor | 0.8.0 | Read, rewrite, and write Python ASTs nicely |
| beautifulsoup4 | 4.9.0 | Python library designed for screen-scraping |
| cachetools | 3.1.1 | Extensible memoizing collections and decorators |
| catalogue | 1.0.0 | Super lightweight function registries for your library |
| click | 7.1.2 | Python composable command line interface toolkit |
| cloudpickle | 1.4.1 | Extended pickling support for Python objects |
| cudatoolkit | 10.1.243 | NVIDIA CUDA toolkit |
| cudnn | 7.6.5 | NVIDIA's cuDNN deep neural network acceleration library |
| cymem | 2.0.2 | Manage calls to calloc/free through Cython |
| cython-blis | 0.4.1 | Fast matrix-multiplication as a self-contained Python library |
| cytoolz | 0.10.1 | Cython implementation of Toolz. High performance functional utilities |
| dask-core | 2.17.2 | Parallel Python with task scheduling |
| dataclasses | 0.6 | A backport of the dataclasses module for Python 3.6 |
| fastai | 1.0.60 | fastai makes deep learning with PyTorch faster, more accurate, and easier |
| fastprogress | 0.2.3 | A fast and simple progress bar for Jupyter Notebook and console |
| gast | 0.2.2 | Python AST that abstracts the underlying Python version |
| google-auth | 1.14.1 | Google authentication library for Python |
| google-auth-oauthlib | 0.4.1 | Google Authentication Library, oauthlib integration with google-auth |
| google-pasta | 0.2.0 | pasta is an AST-based Python refactoring library |
| googledrivedownloader | 0.4 | Minimal class to download shared files from Google Drive |
| graphviz | 2.38 | Open Source graph visualization software |
| grpcio | 1.27.2 | HTTP/2-based RPC framework |
| imageio | 2.8.0 | A Python library for reading and writing image data |
| isodate | 0.6.0 | An ISO 8601 date/time/duration parser and formatter |
| joblib | 0.15.1 | Lightweight pipelining: using Python functions as pipeline jobs |
| jpeg | 9b | Read and write jpeg COM, EXIF, and IPTC medata |
| keepalive | 0.5 | An HTTP handler for urllib that supports HTTP 1.1 and keepalive |
| keras-applications | 1.0.8 | Applications module of the Keras deep learning library |
| keras-base | 2.3.1 | Deep Learning Library for Theano and TensorFlow, base package |
| keras-gpu | 2.3.1 | Deep Learning Library for Theano and TensorFlow |
| keras-preprocessing | 1.1.0 | Data preprocessing and data augmentation module of the Keras deep learning library |
| laspy | 1.7.0 | A Python library for reading, modifying and creating LAS files |
| libprotobuf | 3.11.4 | Protocol Buffers - Google's data interchange format. C++ Libraries |
| libtiff | 4.0.10 | Support for the Tag Image File Format (TIFF) |
| llvmlite | 0.32.1 | A lightweight LLVM python binding for writing JIT compilers |
| markdown | 3.1.1 | Python implementation of Markdown |
| murmurhash | 1.0.2 | Cython bindings for MurmurHash2 |
| ninja | 1.9.0 | A small build system with a focus on speed |
| numba | 0.49.1 | NumPy aware dynamic Python compiler using LLVM |
| nvidia-ml-py3 | 7.352.0 | Python bindings to the NVIDIA Management Library |
| olefile | 0.46 | Parse, read and write Microsoft OLE2 files |
| onnx | 1.7.0 | Open Neural Network Exchange library |
| onnx-tf | 1.5.0 | Experimental Tensorflow Backend for ONNX |
| opencv | 4.0.1.24 | Computer vision and machine learning software library |
| opt_einsum | 3.1.0 | Optimizing einsum functions in NumPy, Tensorflow, Dask, and more with contraction order optimization |
| pillow-simd | 7.1.2 | Python image manipulation, SIMD optimized for SSE4 compatible CPUs |
| plac | 1.1.0 | The smartest command line arguments parser in the world |
| plotly | 4.5.4 | An interactive JavaScript-based visualization library for Python |
| pooch | 1.0.0 | A friend to fetch your Python library's sample data files |
| preshed | 3.0.2 | Cython Hash Table for Pre-Hashed Keys |
| protobuf | 3.11.4 | Protocol Buffers - Google's data interchange format |
| pyasn1 | 0.4.8 | ASN.1 types and codecs |
| pyasn1-modules | 0.2.7 | A collection of ASN.1-based protocols modules |
| pytorch | 1.4.0 | PyTorch is an optimized tensor library for deep learning using GPUs and CPUs |
| pywavelets | 1.1.1 | Discrete Wavelet Transforms in Python |
| pyyaml | 5.3.1 | YAML parser and emitter for Python |
| rdflib | 5.0.0 | Library for working with RDF, a simple yet powerful language for representing information |
| retrying | 1.3.3 | Simplify the task of adding retry behavior to just about anything |
| rsa | 4.0 | Pure-Python RSA implementation |
| scikit-image | 0.17.2 | Image processing routines for SciPy |
| scikit-learn | 0.22.1 | A set of python modules for machine learning and data mining |
| soupsieve | 2.0.1 | A modern CSS selector implementation for BeautifulSoup |
| spacy | 2.2.4 | Industrial-strength Natural Language Processing |
| sparqlwrapper | 1.8.5 | SPARQL Endpoint interface to Python for use with rdflib |
| srsly | 1.0.2 | Modern high-performance serialization utilities for Python |
| tensorboard | 2.2.1 | TensorBoard lets you watch Tensors Flow |
| tensorboard-plugin-wit | 1.6.0 | What-If Tool TensorBoard plugin |
| tensorboardx | 1.7 | Tensorboard for PyTorch |
| tensorflow | 2.1.0 | TensorFlow is a machine learning library |
| tensorflow-addons | 0.9.1 | Useful extra functionality for TensorFlow 2 |
| tensorflow-base | 2.1.0 | TensorFlow is a machine learning library, base package |
| tensorflow-estimator | 2.1.0 | TensorFlow Estimator |
| tensorflow-gpu | 2.1.0 | Metapackage for selecting a TensorFlow variant |
| termcolor | 1.1.0 | ANSI Color formatting for output in terminal |
| thinc | 7.4.0 | Learn super-sparse multi-class models |
| tifffile | 0.15.1 | Read and write image data from and to TIFF files |
| toolz | 0.10.0 | A functional standard library for Python |
| torch-cluster | 1.5.4 | Extension library of highly optimized graph cluster algorithms for use in PyTorch |
| torch-geometric | 1.5.0 | Geometric deep learning extension library for PyTorch |
| torch-scatter | 2.0.4 | Extension library of highly optimized sparse update (scatter and segment) operations |
| torch-sparse | 0.6.1 | Extension library of optimized sparse matrix operations with autograd support |
| torch-spline-conv | 1.2.0 | PyTorch implementation of the spline-based convolution operator of SplineCNN |
| torchvision | 0.5.0 | Image and video datasets and models for torch deep learning |
| tqdm | 4.46.0 | A Fast, Extensible Progress Meter |
| transforms3d | 0.3.1 | Functions for 3D coordinate transformations |
| typeguard | 2.7.0 | Runtime type checker for Python |
| wasabi | 0.6.0 | A lightweight console printing and formatting toolkit |
| werkzeug | 0.16.1 | The comprehensive WSGI web application library |
| wrapt | 1.12.1 | Module for decorators, wrappers and monkey patching |
| xz | 5.2.5 | Data compression software with high compression ratio |
| yaml | 0.1.7 | A C library for parsing and emitting YAML |
| zstd | 1.3.7 | Zstandard - Fast real-time compression algorithm |
| _tflow_select | 2.1.0 | Metapackage for selecting GPU or CPU version of TensorFlow |
- Though this package distributes the GPU based versions of packages, CPU versions can still be installed and used on any machine Pro supports. To install TensorFlow for the CPU, from the Python backstage you can install the
tensorflow-mklpackage to get a CPU only version. - This installer adds packages to the default
arcgispro-py3environment. Any subsequent clones of that environment will also include this full collection of packages. This collection of packages is validated and tested against the version of Pro is installed alongside, and upgrades of Pro will also require reinstallation of the deep learning libraries. Note that when you upgrade the software to a new release, you'll need to uninstall the Deep Learning Libraries installation as well as Pro or Server, and reinstall the new version of this package for that release. - This installer is only available for ArcGIS Pro 2.6, and ArcGIS Server 10.8.1 -- for earlier releases, you'll need to follow the documentation on installing the packages through the Python backstage or Python command prompt.
- If you want these packages for a specific environment only, you can install the
deep-learning-essentialspackage which has the same list of dependencies as a standalone conda metapackage.