tfm.vision.decoders.ASPP

Creates an Atrous Spatial Pyramid Pooling (ASPP) layer.

tfm.vision.decoders.ASPP(
    level: int,
    dilation_rates: List[int],
    num_filters: int = 256,
    pool_kernel_size: Optional[int] = None,
    use_sync_bn: bool = False,
    norm_momentum: float = 0.99,
    norm_epsilon: float = 0.001,
    activation: str = 'relu',
    dropout_rate: float = 0.0,
    kernel_initializer: str = 'VarianceScaling',
    kernel_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
    interpolation: str = 'bilinear',
    use_depthwise_convolution: bool = False,
    spp_layer_version: str = 'v1',
    output_tensor: bool = False,
    **kwargs
)

level: An int level to apply ASPP. dilation_rates: A list of dilation rates. num_filters: An int number of output filters in ASPP. pool_kernel_size: A list of [height, width] of pooling kernel size or None. Pooling size is with respect to original image size, it will be scaled down by 2**level. If None, global average pooling is used. use_sync_bn: A bool. If True, use synchronized batch normalization. norm_momentum: A float of normalization momentum for the moving average. norm_epsilon: A float added to variance to avoid dividing by zero. activation: A str activation to be used in ASPP. dropout_rate: A float rate for dropout regularization. kernel_initializer: A str name of kernel_initializer for convolutional layers. kernel_regularizer: A tf.keras.regularizers.Regularizer object for Conv2D. Default is None. interpolation: A str of interpolation method. It should be one of bilinear, nearest, bicubic, area, lanczos3, lanczos5, gaussian, or mitchellcubic. use_depthwise_convolution: If True depthwise separable convolutions will be added to the Atrous spatial pyramid pooling.
spp_layer_version A str of spatial pyramid pooling layer version.
output_tensor Whether to output a single tensor or a dictionary of tensor. Default is false. **kwargs: Additional keyword arguments to be passed.

Methods

call

View source

call(
    inputs: TensorMapUnion
) -> TensorMapUnion

Calls the Atrous Spatial Pyramid Pooling (ASPP) layer on an input.

The output of ASPP will be a dict of {level, tf.Tensor} even if only one level is present, if output_tensor is false. Hence, this will be compatible with the rest of the segmentation model interfaces. If output_tensor is true, a single tensot is output.

Args
inputs A tf.Tensor of shape [batch, height_l, width_l, filter_size] or a dict of tf.Tensor where

  • key: A str of the level of the multilevel feature maps.
  • values: A tf.Tensor of shape [batch, height_l, width_l, filter_size].

Returns
A tf.Tensor of shape [batch, height_l, width_l, filter_size] or a dict of tf.Tensor where

  • key: A str of the level of the multilevel feature maps.
  • values: A tf.Tensor of output of ASPP module.