tfm.vision.heads.RetinaNetHead

Creates a RetinaNet head.

tfm.vision.heads.RetinaNetHead(
    min_level: int,
    max_level: int,
    num_classes: int,
    num_anchors_per_location: (int | dict[str, int]),
    num_convs: int = 4,
    num_filters: int = 256,
    attribute_heads: Optional[List[Dict[str, Any]]] = None,
    share_classification_heads: bool = False,
    use_separable_conv: bool = False,
    activation: str = 'relu',
    use_sync_bn: bool = False,
    norm_momentum: float = 0.99,
    norm_epsilon: float = 0.001,
    kernel_regularizer: Optional[tf_keras.regularizers.Regularizer] = None,
    bias_regularizer: Optional[tf_keras.regularizers.Regularizer] = None,
    num_params_per_anchor: int = 4,
    share_level_convs: bool = True,
    **kwargs
)

min_level An int number of minimum feature level.
max_level An int number of maximum feature level.
num_classes An int number of classes to predict.
num_anchors_per_location Number of anchors per pixel location. If an int, the same number is used for all levels. If a dict, it specifies the number at each level.
num_convs An int number that represents the number of the intermediate conv layers before the prediction.
num_filters An int number that represents the number of filters of the intermediate conv layers.
attribute_heads If not None, a list that contains a dict for each additional attribute head. Each dict consists of 4 key-value pairs: name, type ('regression' or 'classification'), size (number of predicted values for each instance), and prediction_tower_name (optional, specifies shared prediction towers.)
share_classification_heads A bool that indicates whether sharing weights among the main and attribute classification heads.
use_separable_conv A bool that indicates whether the separable convolution layers is used.
activation A str that indicates which activation is used, e.g. 'relu', 'swish', etc.
use_sync_bn A bool that indicates whether to use synchronized batch normalization across different replicas.
norm_momentum A float of normalization momentum for the moving average.
norm_epsilon A float added to variance to avoid dividing by zero.
kernel_regularizer A tf_keras.regularizers.Regularizer object for Conv2D. Default is None.
bias_regularizer A tf_keras.regularizers.Regularizer object for Conv2D.
num_params_per_anchor Number of parameters required to specify an anchor box. For example, num_params_per_anchor would be 4 for axis-aligned anchor boxes specified by their y-centers, x-centers, heights, and widths.
share_level_convs An optional bool to enable sharing convs across levels for classnet, boxnet, classifier and box regressor. If True, convs will be shared across all levels.
**kwargs Additional keyword arguments to be passed.

Methods

call

View source

call(
    features: Mapping[str, tf.Tensor]
)

Forward pass of the RetinaNet head.

Args
features A dict of tf.Tensor where

  • key: A str of the level of the multilevel features.
  • values: A tf.Tensor, the feature map tensors, whose shape is [batch, height_l, width_l, channels].

Returns
scores A dict of tf.Tensor which includes scores of the predictions.

  • key: A str of the level of the multilevel predictions.
  • values: A tf.Tensor of the box scores predicted from a particular feature level, whose shape is [batch, height_l, width_l, num_classes * num_anchors_per_location].
boxes A dict of tf.Tensor which includes coordinates of the predictions.
  • key: A str of the level of the multilevel predictions.
  • values: A tf.Tensor of the box scores predicted from a particular feature level, whose shape is [batch, height_l, width_l, num_params_per_anchor * num_anchors_per_location].
    • attributes a dict of (attribute_name, attribute_prediction). Each attribute_prediction is a dict of:
  • key: str, the level of the multilevel predictions.
  • values: Tensor, the box scores predicted from a particular feature level, whose shape is [batch, height_l, width_l, attribute_size * num_anchors_per_location]. Can be an empty dictionary if no attribute learning is required.