def _make_layers(self, cfg):
layers = []
in_channels = 3
for x in cfg:
if x == 'M':
layers += [
nn.AvgPool2d(kernel_size=2, stride=2),
nn.Dropout2d(0.15)
]
else:
padding = x[1] if isinstance(x, tuple) else 1
out_channels = x[0] if isinstance(x, tuple) else x
layers += [
catSNN.QuantizedConv2d(in_channels,
out_channels,
kernel_size=3,
padding=padding,
bias=self.bias),
nn.BatchNorm2d(out_channels),
catSNN.Clamp(max=self.clamp_max),
nn.Dropout2d(0.15)
]
in_channels = out_channels
layers += [nn.AvgPool2d(kernel_size=1, stride=1)]
return nn.Sequential(*layers)
def _make_layers(self, cfg, quantize_bit=32):
layers = []
in_channels = 3
for x in cfg:
# [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M']
if x == 'M':
#Train the model with dropout
#layers += [nn.AvgPool2d(kernel_size=2, stride=2)]
layers += [
nn.AvgPool2d(kernel_size=2, stride=2),
nn.Dropout2d(0.3)
]
else:
# ReLU()-->Clamp()-->Clamp_q-->fuse bn and dropout
#layers += [catSNN.QuantizedConv2d(in_channels, x, kernel_size=3, padding=1, bias=self.bias, quantize_bit=quantize_bit),
#catSNN.Clamp(max = self.clamp_max)]
#layers += [catSNN.QuantizedConv2d(in_channels, x, kernel_size=3, padding=1, bias=self.bias, quantize_bit=quantize_bit),nn.BatchNorm2d(x),
#catSNN.Clamp(max = self.clamp_max),nn.Dropout2d(0.3)]
layers += [
catSNN.QuantizedConv2d(in_channels,
x,
kernel_size=3,
padding=1,
bias=self.bias,
quantize_bit=quantize_bit),
nn.BatchNorm2d(x),
nn.ReLU(),
nn.Dropout2d(0.2)
]
if self.quantize_factor != -1:
layers += [catSNN.Quantize(self.quantize_factor)]
in_channels = x
layers += [nn.AvgPool2d(kernel_size=1, stride=1)]
return nn.Sequential(*layers)
def _make_layers(self, cfg, quantize_bit=32):
layers = []
in_channels = 3
for x in cfg:
# [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M']
if type(x) is tuple:
layers += [
nn.Conv2d(x[0],
x[1],
kernel_size=2,
stride=2,
bias=self.bias,
groups=x[1]),
catSNN.Clamp(max=self.clamp_max)
]
#layers += [nn.Conv2d(x[0], x[1], kernel_size=2, stride = 2, bias=self.bias, groups = x[1]),catSNN.Clamp(max = self.clamp_max),nn.Dropout2d(0.2)]
else:
layers += [
catSNN.QuantizedConv2d(in_channels,
x,
kernel_size=3,
padding=1,
bias=self.bias,
quantize_bit=quantize_bit),
catSNN.Clamp(max=self.clamp_max)
]
#layers += [catSNN.QuantizedConv2d(in_channels, x, kernel_size=3, padding=1, bias=self.bias, quantize_bit=quantize_bit),catSNN.Clamp(max = self.clamp_max),nn.Dropout2d(0.2)]
if self.quantize_factor != -1:
layers += [catSNN.Quantize(self.quantize_factor)]
in_channels = x
layers += [nn.AvgPool2d(kernel_size=1, stride=1)]
return nn.Sequential(*layers)
def _make_layers(self, cfg, quantize_bit=32):
layers = []
in_channels = 3
for x in cfg:
# [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M']
if x == 'M':
layers += [nn.AvgPool2d(kernel_size=2, stride=2)]
else:
layers += [catSNN.QuantizedConv2d(in_channels, x, kernel_size=3, padding=1, bias=self.bias, quantize_bit=quantize_bit),
catSNN.Clamp(max = self.clamp_max)]#catSNN.Clamp(max = self.clamp_max)
if self.quantize_factor!=-1:
layers += [catSNN.Quantize(self.quantize_factor)]
in_channels = x
layers += [nn.AvgPool2d(kernel_size=1, stride=1)]
return nn.Sequential(*layers)