Object Detection YAML References

Object Detection training jobs use the keys listed in the table below.

Dataset

The following parameters can be used to load datasets.

train_split_key

• Type: String
• Default: train
• Choices: train/test/dev
• Description: The name given to the data split that will be used for training.

test_split_key

• Type: String
• Default: dev
• Choices: train/test/dev
• Description: The name given to the data split that will be used for testing.

val_split_key

• Type: String
• Default: dev
• Choices: train/test/dev
• Description: The name given to the data split that will be used for validation.

Train

The following parameters allow you to define the model hyperparameter.

batch_size

• Type: Integer
• Default: 8
• Value Range: 1 <= batch_size
• Description: Number of examples/images in each training batch. The minimum batch size is 1.

epoch

• Type: Integer
• Default: 200
• Value Range: 1 <= epoch <=1000
• Description: The number of epochs to train the model. If you turn on early stopping, the actual training epoch may be smaller.

learning_rate

• Type: Float
• Default: 0.0001
• Value Range: 0 < learning_rate < 1.0
• Description: The learning rate used to update the weights.

validation_run_freq

• Type: Integer
• Default: 1
• Value Range: 1 <= validation_run_freq
• Description: Number of epochs between validation runs.

early_stop

• Type: Object
• Description: Configuration for stopping training when a monitored metric has stopped improving
• Properties:
  • min_delta
    • Type: Float
    • Default: 0.01
    • Description: Minimum change in the monitored quantity to qualify as an improvement.
  • min_epochs
    • Type: Integer
    • Default: 40
    • Description: Number of epochs with no improvement after which training will be stopped.

auto_tuning

• Type: Object
• Description: Configuration for automated hyperparameter tuning.
• Properties:
  • anchors
    • Type: Boolean
    • Default: True
    • Description: Whether to use automated anchor-parameter tuning with train and valid datasets.
  • class_weights
    • Type: Boolean
    • Default: True
    • Description: Whether to use automated class weight tuning with train and valid datasets.
  • class_weights_method
    • Type: Integer
    • Default: 0
    • Choices: 0, 1, 2
    • Description: Which method to use to tune class weights
  • debug
    • Type: Boolean
    • Default: False
    • Description: Whether to turn on the debug mode of auto-tuning process

Model

The following parameters allow you to specify which Object Detection model to use.

avi

• Type: Object
• Description: Specifies the python library from where the model is implemented.\
• Properties:
  • RetinaOD
    • Type: Object
    • Description: The architecture of the RetinaNet model.
    • Properties:
      • backbone
        • Type: String
        • Default: ResNet34
        • Choices:
          • ResNet18
          • ResNet34
          • ResNet50
        • Description: The backbone to use in RetinaNet.
      • backbone_weights
        • Type: String | Null
        • Default: Imagenet
        • Choices:
          • null
          • imagenet
        • Description: Load ImageNet pre-trained weights to the backbone or randomly initialized weights
      • output_depth
        • Type: Integer
        • Value Range: 1 <= output_depth
        • Description: The number of classes. Number of defective class plus OK class.'
      • input_shape
        • Type: array
        • Description: Shape of the images.
        • Items:
          • item_1
            • Default: 400
            • Value Range: 1 <= item_1 <=1500
            • Description: Height of the image
          • item_2
            • Default: 400
            • Value Range: 1 <= item_2 <=1500
            • Description: Width of the image
          • item_3
            • Default: 3
            • Value Range: 1 <= item_3 <=4
            • Description: This is the channels of the image (1 for grayscale, 2 for grayscale + alpha, 3 for RGB, 4 for RGB + alpha).
    • nms_threshold
      • Type: Float
      • Default: 0.1
      • Value Range: 0 <= nms_threshold <= 1
      • Description: Non-maximum suppression threshold.
    • score_threshold
      • Type: float
      • Default: 0.3
      • Value Range: 0 <= score_threshold <= 0.99
      • Multiple of: 0.01
      • Description: Minimum score required to make an object prediction.
    • class_specific_filter
      • Type: Boolean
      • Default: false
      • Description: Whether to perform score-based filtering and NMS filtering per class.
    • anchor_sizes
      • Type: array
      • Description: The anchor_sizes parameter
      • Minimum Items: 1
      • Default: [ 32, 64, 128, 256, 512 ]
      • Items:
        • Type: integer
        • Minimum: 1
    • anchor_sizes
      • Type: array
      • Description: The anchor_strides parameter
      • Minimum Items: 1
      • Default: [ 8, 16, 32, 64, 128 ]
      • Items:
        • Type: integer
        • Minimum: 1
    • anchor_ratios
      • Type: array
      • Description: The anchor_ratios parameter
      • Minimum Items: 1
      • Default: [ 0.25, 0.5, 1, 4 ]
      • Items:
        • Type: Float
        • Minimum: 0.0
    • anchor_scales
      • Type: Array
      • Description: The anchor_scales parameter
      • Minimum Items: 1
      • Default: [ 0.5, 1, 1.5, 2 ]
      • Items:
        • Type: Float
          • Minimum: 0.0

Loss

The following parameters allow you to specify the loss functions to use for the Classification and regression head of the model.

regression

• Type: Object
• Description: Loss function to measure the distance between the predicted and the target box.
• Properties:
  • RetinaNetSmoothL1
  • Type: Object
  • Description: The Smooth L1 loss
    • Properties:
      • sigma
        • Type: Float
        • Default: 3.0
        • Description: Sigma

classification

• Type: Object
• Description: Loss function to measure the misclassification error between the predicted and the target class.
• Properties:
  • RetinaNetFocal
    • Type: Object
    • Description: The Smooth L1 loss
  • Properties:
    • gamma
      • Type: Integer
        • Default: 2
        • Value Range: 1 <= gamma
        • Description: Parameter that gives weight to the hard misclassified examples.

Eval

The following parameters can be used in the evaluation phase. To see the list of supported post-processing transformations, please refer to the Transform YAML References.

postprocessing

• Type: Object
• Description: The configuration for post-processing steps.
• Properties:
  • output_type
    • Type: string
    • Default: object-detection
    • Choices:
      • classification
      • object-detection
    • Description: The kind of output after preprocessing
  • iou_threshold
    • Type: float
    • Default: 0.5
    • Value Range: 0 <= iou_threshold <= 1
    • Description: The IoU threshold to determine True Positive and False Positive.
  • transforms
    • Type: array | null
    • Items:
      • Type: Post-processing transforms
      • Description: The transformations to apply in post-processing steps to both ground truth and prediction.
  • gt_transforms
    • Type: Array | Null
    • Items:
      • Type: Post-processing transforms
      • Description: The transformations to apply in post-processing steps to ground truth only.
  • pred_transforms
    • Type: Array | Null
    • Items:
      • Type: Post-processing transforms
        • Description: The transformations to apply in post-processing steps to prediction only.