Train Using AutoDL (GeoAI)

Available with Image Analyst license.

Summary

Trains a deep learning model by building training pipelines and automating much of the training process. This includes data augmentation, model selection, hyperparameter tuning, and batch size deduction. Its outputs include performance metrics of the best model on the training data, as well as the trained deep learning model package (.dlpk file) that can be used as input for the Extract Features Using AI Models tool to predict on a new imagery.

Learn more about how AutoDL works

Usage

• You must install the proper deep learning framework for Python in AllSource.

  Learn how to install deep learning frameworks for ArcGIS

• If you will be training models in a disconnected environment, see Additional Installation for Disconnected Environment for more information.

• The time it takes for the tool to produce the trained model depends on the following:

  • The amount of data provided during training
  • The AutoDL Mode parameter value
  • The Total Time Limit (Hours) parameter value

  By default, the timer for all modes is set at 2 hours. The Basic mode will train the selected networks on the default backbone within the given time. The Advanced mode will divide the total time into two halves, perform the model evaluation in the first half, and finally pick the top two performing models for evaluating on other backbones. If the amount of data being trained is large, all the selected models may not be evaluated within 2 hours. In such cases, the best performing model determined within 2 hours will be considered the optimum model. You can then either use this model or rerun the tool with a higher Total Time Limit (Hours) parameter value.

• This tool can also be used to fine-tune an existing trained model. For example, an existing model that has been trained for cars can be fine-tuned to train a model that identifies trucks.

• To run this tool, a GPU-equipped machine is required. If you have more than one GPU, use the GPU ID environment.

• The input training data for this tool must include the images and labels folders that are generated from the Export Training Data For Deep Learning tool.

• Potential use cases for the tool include training object detection and pixel classification models for extracting features such as building footprint, pools, solar panels, land cover classification, and so on.

• For information about requirements for running this tool and issues you may encounter, see Deep Learning frequently asked questions.

Parameters

Dialog

Label	Explanation	Data Type
Input Training Data	The folders containing the image chips, labels, and statistics required to train the model. This is the output from the Export Training Data For Deep Learning tool. The metadata format of the exported data must be Classified_Tiles, PASCAL_VOC_rectangles, or KITTI_rectangles.	Folder
Output Model	The output trained model that will be saved as a deep learning package (.dlpk file).	File
Pretrained Model (Optional)	A pretrained model that will be used to fine-tune the new model. The input is an Esri model definition file (.emd) or a deep learning package file (.dlpk). A pretrained model with similar classes can be fine-tuned to fit the new model. The pretrained model must have been trained with the same model type and backbone model that will be used to train the new model.	File
Total Time Limit (Hours) (Optional)	The total time limit in hours it will take for AutoDL model training. The default is 2 hours.	Double
AutoDL Mode (Optional)	Specifies the AutoDL mode that will be used and how intensive the AutoDL search will be. Basic—The basic mode will be used. This mode is used to train all selected networks without hyperparameter tuning. Advanced— The advanced mode will be used. This mode is used to perform hyperparameter tuning on the top two performing models.	String
Neural Networks (Optional)	Specifies the architectures that will be used to train the model. By default, all the networks will be used. SingleShotDetector—The SingleShotDetector architecture will be used to train the model. SingleShotDetector is used for object detection. RetinaNet—The RetinaNet architecture will be used to train the model. RetinaNet is used for object detection. FasterRCNN—The FasterRCNN architecture will be used to train the model. FasterRCNN is used for object detection. YOLOv3—The YOLOv3 architecture will be used to train the model. YOLOv3 is used for object detection. HRNet—The HRNet architecture will be used to train the model. HRNet is used for pixel classification. ATSS—The ATSS architecture will be used to train the model. ATSS is used for object detection. CARAFE—The CARAFE architecture will be used to train the model. CARAFE is used for object detection. CascadeRCNN—The CascadeRCNN architecture will be used to train the model. CascadeRCNN is used for object detection. CascadeRPN—The CascadeRPN architecture will be used to train the model. CascadeRPN is used for object detection. DCN—The DCN architecture will be used to train the model. DCN is used for object detection. DeepLab—The DeepLab architecture will be used to train the model. DeepLab is used for pixel classification. UnetClassifier—The UnetClassifier architecture will be used to train the model. UnetClassifier is used for pixel classification. PSPNetClassifier—The PSPNetClassifier architecture will be used to train the model. PSPNetClassifier is used for pixel classification. ANN—The ANN architecture will be used to train the model. ANN is used for pixel classification. APCNet—The APCNet architecture will be used to train the model. APCNet is used for pixel classification. CCNet—The CCNet architecture will be used to train the model. CCNet is used for pixel classification. CGNet—The CGNet architecture will be used to train the model. CGNet is used for pixel classification.	String
Save Evaluated Models (Optional)	Specifies whether all evaluated models will be saved. Checked—All evaluated models will be saved. Unchecked—Only the best performing model will be saved. This is the default.	Boolean

Derived Output

Label	Explanation	Data Type
Output Model File	The output model file.	File

Python

arcpy.geoai.TrainUsingAutoDL(in_data, out_model, {pretrained_model}, {total_time_limit}, {autodl_mode}, {networks}, {save_evaluated_models})

Name	Explanation	Data Type
in_data	The folders containing the image chips, labels, and statistics required to train the model. This is the output from the Export Training Data For Deep Learning tool. The metadata format of the exported data must be Classified_Tiles, PASCAL_VOC_rectangles, or KITTI_rectangles.	Folder
out_model	The output trained model that will be saved as a deep learning package (.dlpk file).	File
pretrained_model (Optional)	A pretrained model that will be used to fine-tune the new model. The input is an Esri model definition file (.emd) or a deep learning package file (.dlpk). A pretrained model with similar classes can be fine-tuned to fit the new model. The pretrained model must have been trained with the same model type and backbone model that will be used to train the new model.	File
total_time_limit (Optional)	The total time limit in hours it will take for AutoDL model training. The default is 2 hours.	Double
autodl_mode (Optional)	Specifies the AutoDL mode that will be used and how intensive the AutoDL search will be. BASIC—The basic mode will be used. This mode is used to train all selected networks without hyperparameter tuning. ADVANCED— The advanced mode will be used. This mode is used to perform hyperparameter tuning on the top two performing models.	String
networks [networks,...] (Optional)	Specifies the architectures that will be used to train the model. SingleShotDetector—The SingleShotDetector architecture will be used to train the model. SingleShotDetector is used for object detection. RetinaNet—The RetinaNet architecture will be used to train the model. RetinaNet is used for object detection. FasterRCNN—The FasterRCNN architecture will be used to train the model. FasterRCNN is used for object detection. YOLOv3—The YOLOv3 architecture will be used to train the model. YOLOv3 is used for object detection. HRNet—The HRNet architecture will be used to train the model. HRNet is used for pixel classification. ATSS—The ATSS architecture will be used to train the model. ATSS is used for object detection. CARAFE—The CARAFE architecture will be used to train the model. CARAFE is used for object detection. CascadeRCNN—The CascadeRCNN architecture will be used to train the model. CascadeRCNN is used for object detection. CascadeRPN—The CascadeRPN architecture will be used to train the model. CascadeRPN is used for object detection. DCN—The DCN architecture will be used to train the model. DCN is used for object detection. DeepLab—The DeepLab architecture will be used to train the model. DeepLab is used for pixel classification. UnetClassifier—The UnetClassifier architecture will be used to train the model. UnetClassifier is used for pixel classification. PSPNetClassifier—The PSPNetClassifier architecture will be used to train the model. PSPNetClassifier is used for pixel classification. ANN—The ANN architecture will be used to train the model. ANN is used for pixel classification. APCNet—The APCNet architecture will be used to train the model. APCNet is used for pixel classification. CCNet—The CCNet architecture will be used to train the model. CCNet is used for pixel classification. CGNet—The CGNet architecture will be used to train the model. CGNet is used for pixel classification. By default, all the networks will be used.	String
save_evaluated_models (Optional)	Specifies whether all evaluated models will be saved. SAVE_ALL_MODELS— All evaluated models will be saved. SAVE_BEST_MODEL—Only the best performing model will be saved. This is the default.	Boolean

Derived Output

Name	Explanation	Data Type
output_model_file	The output model file.	File

Code sample

TrainUsingAutoDL (Python window)

This example shows how to use the TrainUsingAutoDL function.

# Name: TrainUsingAutoDL.py
# Description: Train a deep learning model on imagery data with
# automatic hyperparameter selection.
  
# Import system modules
import arcpy
import os

# Set local variables

datapath = "path_to_training_data" 
out_path = "path_to_trained_model"

out_model = os.path.join(out_path, "mymodel")

# Run Train Using AutoML Model
arcpy.geoai.TrainUsingAutoDL(
    datapath, out_model, None, 2, "BASIC", 
    ["ATSS", "DCN", "FasterRCNN", "RetinaNet", "SingleShotDetector", "YOLOv3"], 
    "SAVE_BEST_MODEL")

Environments

GPU ID