def __init__(self,
in_channels,
out_channels,
expansion_factor,
stride=1,
bias=False,
depthwise_bias=False,
**kwargs):
super().__init__()
self.stride = stride
hidden_channels = int(round(in_channels * expansion_factor))
if hidden_channels == in_channels:
self.conv1 = ai8x.Empty()
else:
self.conv1 = ai8x.FusedConv2dBNReLU(in_channels,
hidden_channels,
1,
padding=0,
bias=bias,
**kwargs)
if stride == 1:
if depthwise_bias:
self.conv2 = ai8x.FusedDepthwiseConv2dBNReLU(
hidden_channels,
hidden_channels,
3,
padding=1,
stride=stride,
bias=depthwise_bias,
**kwargs)
else:
self.conv2 = ai8x.FusedDepthwiseConv2dReLU(hidden_channels,
hidden_channels,
3,
padding=1,
stride=stride,
bias=depthwise_bias,
**kwargs)
else:
if depthwise_bias:
self.conv2 = ai8x.FusedMaxPoolDepthwiseConv2dBNReLU(
hidden_channels,
hidden_channels,
3,
padding=1,
pool_size=stride,
pool_stride=stride,
bias=depthwise_bias,
**kwargs)
else:
self.conv2 = ai8x.FusedMaxPoolDepthwiseConv2dReLU(
hidden_channels,
hidden_channels,
3,
padding=1,
pool_size=stride,
pool_stride=stride,
bias=depthwise_bias,
**kwargs)
self.conv3 = ai8x.FusedConv2dBN(hidden_channels,
out_channels,
1,
bias=bias,
**kwargs)
if (stride == 1) and (in_channels == out_channels):
self.resid = ai8x.Add()
else:
self.resid = self.NoResidual()
def __init__(
self,
num_classes=100,
num_channels=3,
dimensions=(32, 32), # pylint: disable=unused-argument
bias=False,
**kwargs):
super().__init__()
self.conv1 = ai8x.FusedConv2dReLU(num_channels,
16,
3,
stride=1,
padding=1,
bias=bias,
**kwargs)
self.conv2 = ai8x.FusedConv2dReLU(16,
20,
3,
stride=1,
padding=1,
bias=bias,
**kwargs)
self.conv3 = ai8x.FusedConv2dReLU(20,
20,
3,
stride=1,
padding=1,
bias=bias,
**kwargs)
self.conv4 = ai8x.FusedConv2dReLU(20,
20,
3,
stride=1,
padding=1,
bias=bias,
**kwargs)
self.resid1 = ai8x.Add()
self.conv5 = ai8x.FusedMaxPoolConv2dReLU(20,
20,
3,
pool_size=2,
pool_stride=2,
stride=1,
padding=1,
bias=bias,
**kwargs)
self.conv6 = ai8x.FusedConv2dReLU(20,
20,
3,
stride=1,
padding=1,
bias=bias,
**kwargs)
self.resid2 = ai8x.Add()
self.conv7 = ai8x.FusedConv2dReLU(20,
44,
3,
stride=1,
padding=1,
bias=bias,
**kwargs)
self.conv8 = ai8x.FusedMaxPoolConv2dReLU(44,
48,
3,
pool_size=2,
pool_stride=2,
stride=1,
padding=1,
bias=bias,
**kwargs)
self.conv9 = ai8x.FusedConv2dReLU(48,
48,
3,
stride=1,
padding=1,
bias=bias,
**kwargs)
self.resid3 = ai8x.Add()
self.conv10 = ai8x.FusedMaxPoolConv2dReLU(48,
96,
3,
pool_size=2,
pool_stride=2,
stride=1,
padding=1,
bias=bias,
**kwargs)
self.conv11 = ai8x.FusedMaxPoolConv2dReLU(96,
512,
1,
pool_size=2,
pool_stride=2,
padding=0,
bias=bias,
**kwargs)
self.conv12 = ai8x.FusedConv2dReLU(512,
128,
1,
stride=1,
padding=0,
bias=bias,
**kwargs)
self.conv13 = ai8x.FusedMaxPoolConv2dReLU(128,
128,
3,
pool_size=2,
pool_stride=2,
stride=1,
padding=1,
bias=bias,
**kwargs)
self.conv14 = ai8x.Conv2d(128,
num_classes,
1,
stride=1,
padding=0,
bias=bias,
wide=True,
**kwargs)
def __init__(self,
in_channels,
out_channels,
kernel_size=3,
stride=1,
bias=False,
se_ratio=None,
expand_ratio=1,
fused=False,
**kwargs):
super().__init__()
self.has_se = (se_ratio is not None) and (0 < se_ratio <= 1)
self.in_channels = in_channels
self.out_channels = out_channels
self.stride = stride
self.expand_ratio = expand_ratio
self.fused = fused
# Expansion phase (Inverted Bottleneck)
inp = in_channels # number of input channels
out = in_channels * expand_ratio # number of output channels
if expand_ratio != 1:
if fused is True:
self.expand_conv = ai8x.FusedConv2dBNReLU(
inp,
out,
kernel_size=kernel_size,
padding=1,
batchnorm='Affine',
bias=bias,
eps=1e-03,
momentum=0.01,
**kwargs)
else:
self.expand_conv = ai8x.FusedConv2dBNReLU(inp,
out,
1,
batchnorm='Affine',
bias=bias,
eps=1e-03,
momentum=0.01,
**kwargs)
# Depthwise Convolution phase
if fused is not True:
self.depthwise_conv = ai8x.FusedConv2dBNReLU(
in_channels=out,
out_channels=out,
groups=out, # groups makes it depthwise
padding=1,
kernel_size=kernel_size,
stride=stride,
batchnorm='Affine',
bias=bias,
eps=1e-03,
momentum=0.01,
**kwargs)
# Squeeze and Excitation phase
if self.has_se:
num_squeezed_channels = max(1, int(in_channels * se_ratio))
self.se_reduce = ai8x.FusedConv2dReLU(
in_channels=out,
out_channels=num_squeezed_channels,
kernel_size=1,
stride=1,
bias=bias,
**kwargs)
self.se_expand = ai8x.Conv2d(in_channels=num_squeezed_channels,
out_channels=out,
kernel_size=1,
stride=1,
bias=bias,
**kwargs)
# Output Convolution phase
final_out = out_channels
self.project_conv = ai8x.FusedConv2dBN(in_channels=out,
out_channels=final_out,
kernel_size=1,
batchnorm='Affine',
bias=bias,
eps=1e-03,
momentum=0.01,
**kwargs)
# Skip connection
self.resid = ai8x.Add()