What's new in version 1.8.5

New

Guides

• Deep Learning with ArcGIS
  • Utilize Multipe GPUs to Train Model
  • Deep learning on time series data

Samples

• GIS Analysts and Data Scientists
  • Streams extraction using ConnectNet
  • Streams extraction using Multi-Task Road Extractor
  • Forecasting monthly rainfall in California using Deep Learning Time Series techniques
  • Landcover mapping using hyperspectral imagery and deep learning

Module Changes

arcgis.gis

• User
  • Adds generate_direct_access_url() method for upload of large files to datastores

arcgis.raster

• Adds new submodule
  • arcgis.raster.utils - functions to interact with ArcGIS Online raster stores
    • generate_direct_access_url()
    • upload_imagery_to_agol_userstore()
• ImageryLayer
  • Adds process_as_multidimensional parameter to following functions:
    • compute_stats_and_histograms()
    • compute_histograms()
    • identify()

arcgis.raster.analytics

• Adds new functions:
• compute_change_raster
• analyze_changes_using_landtrendr
• zonal_statistics_as_table
• Adds note to tiles_only parameter documentation to clarify its use and avaiability to create dynamic imagery in all relevant methods

arcgis.geoanalytics

• Adds ellapse_time property to GPJob returned when tools are run asynchronously

arcgis.learn

• New Table of Contents heading:
  • Replace Tabular and Feature Models with Feature, Tabular and Timeseries Models
• New Table of Contents entry and models:
  • Object Tracking Models
    • SiamMask
    • ObjectTracker
    • Track
• Adds cell size ranges and output units to trained models (Esri Model Definitions)
• Adds monitor parameter to fit() methods of models for use with early_stopping and checkpoint
• Data Preparation Methods
  • Adds documentation to path parameter for prepare_data()
  • Adds keyword parameters and documentation to prepare_data():
    • min_points
    • classes_of_interest
    • extra_features
    • remap_classes
    • background_classcode
    • Adds ability to use a folder or list of paths (multi-folder support) in path parameter for prepare_data()
  • Adds working_dir parameter to data prepartion methods which sets a default path as a prefix for saving trained models and checkpoints
• Unstructured Text Models
  • arcgis.learn.text module
  • TextClassifier
    • adds support for working_dir parameter
• 3D Models
  • PointCNN
    • Adds keyword parameters when using output of prepare_data()
      • min_points
      • classes_of_interest
      • extra_features
      • remap_classes
      • background_classcode
    • Adds ability to remap classes
    • Adds documentation to describe precision, f1, and recall metrics for fit() method
• Feature, Tabular and Timeseries Models
  • Adds ability to publish non-spatial dataframes as Table items with the predict() methods
  • MLModel
    • Adds load() method
    • Adds cabability to save() method to allow saving model without data input
  • TimeSeriesModel
    • Enhances vizualization to show separate graph for each dependent variable when using show_batch() on TabularData object
    • Adds warning message when using predict() if openpyxl dependency not installed
• Pixel Classification Models
• Adds ConnectNet Model
• Adds dice_loss_average keyword parameter to initialization options
• Adds inference and exports support to:
  • ChangeDetector
• Adds threshholding functionality to:
  • BDCNEdgeDetector
  • HEDEdgeDectector
• MultiTaskRoadExtractor
  • Adds support for multispectral data
  • Adds support for using multiple folders for training model
• Image Translations Models
  • Adds inference and export support to:
    • Pix2Pix
    • CycleGAN
  • Adds additional checks for whether multispectral images were used in initialization for:
    • Pix2Pix
    • CycleGAN
• Model Management
• Adds train_model() method

Fixes

Samples

• GIS Analysts and Data Scientists
  • Updates for Try it live Site:
    • Automate Road Surface Investigation Using Deep Learning
    • Reconstructing 3D buildings from Aerial LiDAR with Deep Learning
    • Finding a New Home
    • Automate Building Footprint Extraction using Deep learning
    • Analyzing growth factors of rental properties in New York City

Module Changes

arcgis.raster

• ImageryLayer
• Fixes table structure in documentation for:
  • compute_pixel_location()
  • slices()
  • statistics()
  • get_histograms()
  • mosaic_by()
• Fixes clipping input error when rendering tiles only Sentinel-2 imagery

arcgis.raster.analytics

• Improves warning documentation if Raster Function Template (RFT) initialization fails
• Fixes issue where create_image_collection() created blank output if input_rasters referred to Sentinel-2 or Pleiades data

arcgis.raster.functions

• Fixes table structure in documentation for:
  • rasterize_features()
• Fixes incorrect band_indexes parameter order in nbr documentation

arcgis.network.analysis

• Fixes Token Required errors when using tools with GIS object created using api_key:
  • get_travel_modes()
  • get_tool_info()

arcgis.learn

• Fixes to various functions and model methods for ArcGIS Pro Notebooks
  • show_batch()
  • lr_find()
  • model.show_results()
• Fixes WARNING 003054: Input Raster does not support PIXEL_SPACE (raw image space). Running in MAP_SPACE warning message when running the Classify Pixels Using Deep Learning tool in ArcGIS Pro
• Data Preparation Methods
  • Fixes expected scalar type Float Exception with prepare_data() when using pytorch 1.7.0
• Pixel Classification Models
  • Improves per_class_metrics() results on models initialized from unbalanced datasets
  • Fixes fit() method in segmentation models returning NaN values for certain attributes: UNetClassifier, PSPNetClassifier, DeepLab
  • MultiTaskRoadExtractor
  • Fixes issue for show_results() plotting images incorrectly
  • Fixes issue causing save() only to work after calling fit()
  • Fixes load() so it only needs the model name
  • Fixes errors with fit() when using monitor functionality
  • Fixes parameter table for show_results() documentation
• Unstructured Text Models
  • arcgis.learn.text module
    • Adds capability to use HuggingFace pretrained models for:
      • EntityRecognizer
      • TextClassifier
    • Adds transformer backbone options for:
      • EntityRecognizer
        • Funnel
        • LayoutLM
      • TextClassifier
        • Funnel
• Object Classification Models
  • Feature Classifier
    • Fixes error when using an input object resulting from prepare_data(dataset_type='ImageNet')
    • Fixes issues with plot_confusion_matrix()
      • when using with an object created with backend=tensorflow
      • illegible results when dataset has a large number of classes
      • when run on object created from the from_model() method and with a data argument
    • Fixes issue with lr_find():
      • with object created with mixup=True
      • with MultiLabeled_Tiles dataset type object and mixup=True
    • Fixes issue with using predict() on a trained model
• Feature, Tabular and Timeseries Models
  • MLModel
    • Fixes inaccurate predict() results because training transformations were not incorporating training statistics
  • TimeSeriesModel
    • Fixes AttributeError: '<object-name>' object has no attribute 'inverse_transform' when using the score() method
• 3D Models
  • PointCNN
    • Fixes bug with predict_las() returning class codes from the model instead of the actual class codes in specific scenarios
    • Fixes error message raised with from_model() if library dependencies are missing
    • Fixes show_results() and show_batch() when run against data object created with a remap_classes argument in prepare_data()
• Object Detection Models
  • SingleShotDetector
    • Fixes out of memory error when using show_results() on model with multispectral data
    • Fixes error when loading and executing on CPU device
  • RetinaNet
  • Fixes issue where FasterRCNN and MaskRCNN returned no results with multispectral data
    • Fixes error when loading and executing on CPU device
• Image Translation Models
• Pix2Pix
  • Fixes Key Error when calling save() on object created with mulitspectral data
  • Fixes issue with save() not having an FID metric
  • Fixes issue with compute_metrics() returning too many significant digits
• CycleGAN
  • compute_metrics()
    • Fixes issue with returning FID metrics without respective names
    • Fixes issue with values returning with too many significant digits
• SuperResolution
  • Fixes issue with compute_metrics() returning PSNR and SSIM metrics without respective names
  • Fixes issue with prepare_data() failing for dataset_type='superres' or when using superres data for data argument
• ImageCaptioner
  • Fixes AttributError about missing temporary folder with lr_find()