def _set_cell_ops(edge, C):
edge.data.set('op', [
ops.Identity(),
ops.Zero(stride=1),
ops.ReLUConvBN(C, C, kernel_size=3),
ops.ReLUConvBN(C, C, kernel_size=1),
ops.AvgPool1x1(kernel_size=3, stride=1),
])
def _set_ops(edge, C, stride):
"""
Replace the 'op' at the edges with the ones defined here.
This function is called by the framework for every edge in
the defined scope.
Args:
current_egde_data (EdgeData): The data that currently sits
at the edge.
C (int): convolutional channels
stride (int): stride for the operation
Returns:
EdgeData: the updated EdgeData object.
"""
edge.data.set('op', [
ops.Identity()
if stride == 1 else FactorizedReduce(C, C, stride, affine=False),
ops.Zero(stride=stride),
ops.MaxPool(3, stride),
ops.AvgPool(3, stride),
ops.SepConv(
C, C, kernel_size=3, stride=stride, padding=1, affine=False),
ops.SepConv(
C, C, kernel_size=5, stride=stride, padding=2, affine=False),
ops.DilConv(C,
C,
kernel_size=3,
stride=stride,
padding=2,
dilation=2,
affine=False),
ops.DilConv(C,
C,
kernel_size=5,
stride=stride,
padding=4,
dilation=2,
affine=False),
])
def _set_cell_ops(edge, C):
edge.data.set('op', [
ops.Identity(),
ops.Zero(stride=1),
])
def _set_cell_ops(current_edge_data, C):
current_edge_data.set('op', [
ops.Identity(),
ops.Zero(stride=1),
])