phygnn.layers.custom_layers

Custom tf layers.

Functions

get_custom_layer_objects()

Get local custom layer classes for Keras deserialization.

Classes

CBAM(*args, **kwargs)	Convolutional Block Attention Module
ExpandDims(*args, **kwargs)	Layer to add an extra dimension to a tensor.
FlattenAxis(*args, **kwargs)	Layer to flatten an axis from a 5D spatiotemporal Tensor into axis-0 observations.
FlexiblePadding(*args, **kwargs)	Class to perform padding on tensors
FunctionalLayer(*args, **kwargs)	Custom layer to implement the tensorflow layer functions (e.g., add, subtract, multiply, maximum, and minimum) with a constant value.
GaussianAveragePooling2D(*args, **kwargs)	Custom layer to implement tensorflow average pooling layer but with a gaussian kernel.
GaussianNoiseAxis(*args, **kwargs)	Layer to apply random noise along a given axis.
LogTransform(*args, **kwargs)	Log transform or inverse transform of data
MaskedSqueezeAndExcitation(*args, **kwargs)	Custom layer for masked squeeze and excitation block for convolutional networks
MultiHeadAttention(*args, **kwargs)	MultiHeadAttention that accepts an additive pre-softmax bias.
PatchDecoder(*args, **kwargs)	Project token features back to the query feature grid.
PatchEncoder(*args, **kwargs)	Project spatial inputs into token features.
PositionEncoder(*args, **kwargs)	Positional encoding layer.
SigLin(*args, **kwargs)	Sigmoid linear unit.
SkipConnection(*args, **kwargs)	Custom layer to implement a skip connection.
SpatialExpansion(*args, **kwargs)	Class to expand the spatial dimensions of tensors with shape: (n_observations, n_spatial_0, n_spatial_1, n_features)
SpatioTemporalExpansion(*args, **kwargs)	Class to expand the spatiotemporal dimensions of tensors with shape: (n_observations, n_spatial_0, n_spatial_1, n_temporal, n_features)
SqueezeAndExcitation(*args, **kwargs)	Custom layer for squeeze and excitation block for convolutional networks
Sup3rAdder(*args, **kwargs)	Layer to add high-resolution data to a sup3r model in the middle of a super resolution forward pass.
Sup3rConcat(*args, **kwargs)	Layer to concatenate a high-resolution feature to a sup3r model in the middle of a super resolution forward pass.
Sup3rConcatObs(*args, **kwargs)	Layer to concatenate sparse data in the middle of a super resolution forward pass.
Sup3rObsModel(*args, **kwargs)	Layer to concatenate sparse data in the middle of a super resolution forward pass, with a learned embedding.
Sup3rTransformerBlock(*args, **kwargs)	Stack of Sup3rTransformerLayer instances with optional window shifting on odd layers in the stack.
Sup3rTransformerLayer(*args, **kwargs)	Transformer layer with shared Q encoding, joint K/V attention, and a shared post-attention pathway.
SwiGLU(*args, **kwargs)	SwiGLU activation function.
TileLayer(*args, **kwargs)	Layer to tile (repeat) data across a given axis.
TransformerLayer(*args, **kwargs)	Custom transformer layer with multi-head attention layer that allows for additive bias pre-softmax.
UnitConversion(*args, **kwargs)	Layer to convert units per feature channel using the linear transform: y = x * scalar + adder
WindowGeometry(batch_size, query_height, ...)	Computed layout for one windowed-attention call.
WindowedMultiHeadAttention(*args, **kwargs)	MultiHeadAttention with overlapping spatial windowing.