Deep learning arguments

Available with Image Analyst license.

Arguments are one of the many ways to control how deep learning models are trained and used. The first table below lists the supported model arguments for training deep learning models. The second table lists the arguments to control how deep learning models are used for inferencing.

Training arguments

The following arguments are available in the Train Deep Learning Model tool for training deep learning models. These arguments vary, depending on the model architecture. You can change the values of these arguments to train a model.

Model typeArgumentValid values

Change detector (pixel classification)

attention_type

PAM (Pyramid Attention Module) or BAM (Basic Attention Module).

chip_size

Integers between 0 and image size.

monitor

valid_loss, precision, recall, and f1.

ConnectNet (pixel classification)

chip_size

Integers between 0 and image size.

gaussian_thresh

0.0 to 1.0. The default is 0.76.

monitor

valid_loss, accuracy, miou, and dice.

mtl_model

linknet or hourglass.

orient_bin_size

A positive number. The default is 20.

orient_theta

A positive number. The default is 8.

DeepLabv3 (pixel classification)

chip_size

Integers between 0 and image size.

class_balancing

true or false.

focal_loss

true or false.

ignore_classes

Valid class values.

monitor

valid_loss and accuracy.

mixup

true or false.

Image captioner (image translation)

chip_size

Integers between 0 and image size.

The decode_params argument is composed of the following parameters:

  • embed_size
  • hidden_size
  • attention_size
  • teacher_forcing
  • dropout
  • pretrained_emb

The default is {'embed_size':100, 'hidden_size':100, 'attention_size':100, 'teacher_forcing':1, 'dropout':0.1, 'pretrained_emb':False}.

monitor

valid_loss, accuracy, corpus_bleu, and multi_label_fbeta.

MMDetection (object detection)

chip_size

Integers between 0 and image size.

model

atss, carafe, cascade_rcnn, cascade_rpn, dcn, detectors, double_heads, dynamic_rcnn, empirical_attention, fcos, foveabox, fsaf, ghm, hrnet, libra_rcnn, nas_fcos, pafpn, pisa, regnet, reppoints, res2net, sabl, and vfnet.

model_weight

true or false.

MMSegmentation (pixel classification)

chip_size

Integers between 0 and image size.

model

ann, apcnet, ccnet, cgnet, danet, deeplabv3, deeplabv3plus, dmnet , dnlnet, emanet, encnet, fastscnn, fcn, gcnet, hrnet, mobilenet_v2, mobilenet_v3, nonlocal_net, ocrnet, ocrnet_base, pointrend, psanet, pspnet, resnest, sem_fpn, unet, and upernet.

model_weight

true or false.

Multi Task Road Extractor (pixel classification)

chip_size

Integers between 0 and image size.

gaussian_thresh

0.0 to 1.0. The default is 0.76.

monitor

valid_loss, accuracy, miou, and dice.

mtl_model

linknet or hourglass.

orient_bin_size

A positive number. The default is 20.

orient_theta

A positive number. The default is 8.

Pyramid Scene Parsing Network (pixel classification)

chip_size

Integers between 0 and image size.

class_balancing

true or false.

focal_loss

true or false.

ignore_classes

Valid class values.

monitor

valid_loss or accuracy.

mixup

true or false.

pyramid_sizes

[convolution layer 1, convolution layer 2, ... , convolution layer n]

use_net

true or false.

RetinaNet (object detection)

chip_size

Integers between 0 and image size.

monitor

valid_loss or average_precision.

ratios

Ratio value 1, ratio vale 2, ratio value 3.

The default is 0.5,1,2.

scales

[scale value 1, scale value 2, scale value 3]

The default is [1, 0.8, 0.63].

SAMLoRA (pixel classification)

class_balancing

true or false.

ignore_classes

Valid class values.

Single Shot Detector (object detection)

chip_size

Integers between 0 and image size.

grids

Integer values greater than 0.

monitor

valid_loss or average_precision.

ratios

[horizontal value, vertical value]

zooms

The zoom value in which 1.0 is normal zoom.

Super Resolution with SR3 backbone (image translation)

attn_res

Integers greater than 0. The default is 16.

channel_mults 

Integer multiplier sets The default is [1, 2, 4, 4, 8, 8].

dropout

Floating point value. The default is 0.

inner_channel

Integer value greater than 0. The default is 64.

linear_start

Time integer. The default is 1e-02.

linear_end 

Time integer. The default is 1e-06.

n_timestep 

Integer value greater than 0. The default is 1000.

norm_groups 

Integer value greater than 0. The default is 32.

res_blocks

Integer value greater than 0. The default is 3.

schedule 

linear, warmup10, warmup50, const, jsd, or cosine.

The default is linear.

U-Net (pixel classification)

chip_size

Integers between 0 and image size.

class_balancing

true or false.

focal_loss

true or false.

ignore_classes

Valid class values.

monitor

valid_loss or accuracy.

mixup

true or false.

Inferencing arguments

The following arguments are available to control how deep learning models are trained for inferencing. The information from the Model Definition parameter will be used to populate Arguments in the inferencing tools. These arguments vary, depending on the model architecture. ArcGIS pretrained models and custom deep learning models may have additional arguments that the tool supports.

ArgumentInference typeValid values

batch_size

Classify Objects

Classify Pixels

Detect Change

Detect Objects

Integer values greater than 0; it's usually an integer that is a power of 2n.

direction

Classify Pixels

Available options are AtoB and BtoA.

The argument is only available for the CycleGAN architecture.

exclude_pad_detections

Detect Objects

true or false.

The argument is available for SSD, RetinaNet, YOLOv3, DETReg, MMDetection, and Faster RCNN only.

merge_policy

Classify Pixels

Detect Objects

Available options are mean, max, and min.

For the Classify Pixels Using Deep Learning tool, the argument is available for the Multi Task Road Extractor and ConnectNet architectures. If IsEdgeDetection is present in the model's .emd file, BDCN Edge Detector, HED Edge Detector, and MMSegmentation are also available architectures.

For the Detect Objects Using Deep Learning tool, the argument is only available for MaskRCNN.

n_timestep

Classify Pixels

Integer value greater than 0. The default is 200.

The argument is available for Super Resolution with SR3 backbone models.

nms_overlap

Detect Objects

A floating point value of 0.0 to 1.0. The default is 0.1.

output_classified_raster

Detect Objects

The file path and name for the output classified raster.

The argument is only available for MaXDeepLab.

padding

Classify Pixels

Detect Change

Detect Objects

Integer values greater than 0 and less than half the tile size value.

predict_background

Classify Pixels

true or false.

The argument is available for UNET, PSPNET, DeepLab, and MMSegmentation.

return_probability_raster

Classify Pixels

true or false.

If ArcGISLearnVersion is 1.8.4 or later in the model's .emd file, the Multi Task Road Extractor and ConnectNet architectures are available. If ArcGISLearnVersion is 1.8.4 or later and IsEdgeDetection is present in the model's .emd file, the BDCN Edge Detector, HED Edge Detector, and MMSegmentation architectures are also available.

sampling_type

Classify Pixels

ddim and ddpm.

The default is ddim.

The argument is available for Super Resolution with SR3 backbone models.

schedule

Classify Pixels

linear, warmup10, warmup50, const, jsd, or cosine.

The default is set to the same value as the model it was trained on.

The argument is available for Super Resolution with SR3 backbone models.

score_threshold

Classify Objects

0 to 1.0.

test_time_augmentation

Classify Objects

Classify Pixels

true or false.

threshold

Classify Pixels

Detect Objects

0 to 1.0.

For the Classify Pixels Using Deep Learning tool, if ArcGISLearnVersion is 1.8.4 or later in the model's .emd file, the MultiTaskRoadExtractor and ConnectNet architectures are available. If ArcGISLearnVersion is 1.8.4 or later and IsEdgeDetection is present in the model's .emd file, the BDCNEdgeDetector, HEDEdgeDetector, and MMSegmentation architectures are also available.

For the Detect Objects Using Deep Learning tool, the argument is available for all model architectures.

thinning

Classify Pixels

true or false.

If IsEdgeDetection is present in the model's .emd file, BDCNEdgeDetector, HEDEdgeDetector, and MMSegmentation are available architectures.

tile_size

Classify Pixels

Detect Objects

Integer values greater than 0 and less than the size of the image.

For the Classify Pixels Using Deep Learning tool, the argument is only available for the CycleGAN architecture.

For the Detect Objects Using Deep Learning tool, the argument is only available for MaskRCNN.

Related topics