from akida.core import Layer, LayerType, ActivationType, LayerParams
[docs]
class BufferTempConv(Layer):
"""This represents the Akida BufferTempConv layer.
To mitigate the memory requirements of a real 3D spatiotemporal convolution,
this spatiotemporal layer can store a FIFO buffer caching the past inputs and
perform a standard 2-D convolution between the buffered inputs and the layer
kernel.
The FIFO has a limited storage capacity (T), determined by its parameter
fifo_size.
Inputs shape must be in the form (X, Y, C). But FIFO has a shape of (X, Y, T, C).
The BufferTempConv convolution operation can be described as follows:
>>> fifo = fifo_op(inputs)
>>> fifo = fifo << input_shift
>>> output = conv2d(fifo, weights)
>>> output = output + (bias << bias_shift) (optional)
>>> output = ReLU(output) (optional)
>>> output = output * output_scale >> output_shift
Where fifo_op pushes a new sample to the FIFO and pop out the oldest one (The storage
occurs on the axis T of the FIFO where FIFO shape is (X, Y, T, C))
Note that output values will be saturated on the range that can be represented with
output_bits.
Args:
filters (int): number of filters.
fifo_size (int): integer value specifying the FIFO capacity.
activation (:obj:`ActivationType`, optional): activation type.
Defaults to `ActivationType.ReLU`.
output_bits (int, optional): output bitwidth. Defaults to 8.
buffer_bits (int, optional): buffer bitwidth. Defaults to 32.
post_op_buffer_bits (int, optional): internal bitwidth for post operations. Defaults to 32.
name (str, optional): name of the layer. Defaults to empty string.
"""
def __init__(self,
filters,
fifo_size,
activation=ActivationType.ReLU,
output_bits=8,
buffer_bits=32,
post_op_buffer_bits=32,
name=""):
try:
params = LayerParams(
LayerType.BufferTempConv, {
"filters": filters,
"fifo_size": fifo_size,
"activation": activation,
"output_bits": output_bits,
"buffer_bits": buffer_bits,
"post_op_buffer_bits": post_op_buffer_bits,
})
# Call parent constructor to initialize C++ bindings
# Note that we invoke directly __init__ instead of using super, as
# specified in pybind documentation
Layer.__init__(self, params, name)
except BaseException:
self = None
raise