def __init__(self):
super(LeNet5, self).__init__()
self.num_classes = 10
self.conv1 = layers.ConvVDO(1,
20,
5,
padding=0,
alpha_shape=(1, 1, 1, 1))
self.relu1 = nn.ReLU(True)
self.pool1 = nn.MaxPool2d(2, padding=0)
self.conv2 = layers.ConvVDO(20,
50,
5,
padding=0,
alpha_shape=(1, 1, 1, 1))
self.relu2 = nn.ReLU(True)
self.pool2 = nn.MaxPool2d(2, padding=0)
self.flatten = layers.FlattenLayer(800)
self.dense1 = layers.LinearVDO(800, 500)
self.bn1 = nn.BatchNorm1d(500)
self.relu3 = nn.ReLU()
self.dense2 = layers.LinearVDO(500, 10)
def __init__(self):
super(LeNet5, self).__init__()
self.num_classes = 10
if flags.tanh:
nonlinearity = nn.Tanh
else:
nonlinearity = nn.ReLU
# Conv-BN-Tanh-Pool
if flags.vdo1:
self.conv1 = layers.ConvVDO(1,
20,
5,
alpha_shape=(1, 1),
padding=0,
bias=False)
else:
self.conv1 = nn.Conv2d(1, 20, 5, padding=0, bias=False)
self.bn1 = nn.BatchNorm2d(20, affine=not flags.no_biases)
self.relu1 = nonlinearity()
self.pool1 = nn.MaxPool2d(2, padding=0)
# Conv-BN-Tanh-Pool
if flags.vdo2:
self.conv2 = layers.ConvVDO(20,
50,
5,
alpha_shape=(1, 1),
padding=0,
bias=False)
else:
self.conv2 = nn.Conv2d(20, 50, 5, padding=0, bias=False)
self.bn2 = nn.BatchNorm2d(50, affine=not flags.no_biases)
self.relu2 = nonlinearity()
self.pool2 = nn.MaxPool2d(2, padding=0)
self.flatten = layers.FlattenLayer(800)
if not (flags.vdo1 or flags.vdo2 or flags.vdo3 or flags.vdo4):
self.do1 = nn.Dropout(0.5)
# Dense-BN-Tanh
if flags.vdo3:
self.dense1 = layers.LinearVDO(800,
500,
alpha_shape=(1, 1),
bias=False)
else:
self.dense1 = nn.Linear(800, 500, bias=False)
self.bn3 = nn.BatchNorm1d(500, affine=not flags.no_biases)
self.relu3 = nonlinearity()
# Dense
if flags.vdo4:
self.dense2 = layers.LinearVDO(500,
10,
alpha_shape=(1, 1),
bias=False)
else:
self.dense2 = nn.Linear(500, 10, bias=False)
def __init__(self):
super(LeNet5, self).__init__()
self.num_classes = 10
self.conv1 = nn.Conv2d(1, 20, 5, padding=0)
self.relu1 = nn.ReLU(True)
self.pool1 = nn.MaxPool2d(2, padding=0)
self.conv2 = nn.Conv2d(20, 50, 5, padding=0)
self.relu2 = nn.ReLU(True)
self.pool2 = nn.MaxPool2d(2, padding=0)
self.flatten = layers.FlattenLayer(800)
self.do1 = nn.Dropout(0.5)
self.dense1 = nn.Linear(800, 500)
self.relu3 = nn.ReLU()
self.dense2 = nn.Linear(500, 10)
def __init__(self):
super(LeNet5, self).__init__()
self.num_classes = 10
if flags.tanh:
nonlinearity = nn.Tanh
else:
nonlinearity = nn.ReLU
# Conv-BN-Tanh-Pool
if flags.var1:
self.conv1 = layers.ConvVarianceUnif(1,
int(20 * flags.width),
5,
padding=0,
bias=False)
else:
self.conv1 = nn.Conv2d(1,
int(20 * flags.width),
5,
padding=0,
bias=False)
self.bn1 = nn.BatchNorm2d(int(20 * flags.width),
affine=not flags.no_biases)
self.relu1 = nonlinearity()
self.pool1 = nn.MaxPool2d(2, padding=0)
# Conv-BN-Tanh-Pool
if flags.var2:
self.conv2 = layers.ConvVarianceUnif(int(20 * flags.width),
int(50 * width),
5,
padding=0,
bias=False)
else:
self.conv2 = nn.Conv2d(int(20 * flags.width),
int(50 * flags.width),
5,
padding=0,
bias=False)
self.bn2 = nn.BatchNorm2d(int(50 * flags.width),
affine=not flags.no_biases)
self.relu2 = nonlinearity()
self.pool2 = nn.MaxPool2d(2, padding=0)
self.flatten = layers.FlattenLayer(int(800 * flags.width))
if not (flags.var1 or flags.var2 or flags.var3 or flags.var4):
self.do1 = nn.Dropout(0.5)
# Dense-BN-Tanh
if flags.var3:
self.dense1 = layers.LinearVarianceUnif(int(800 * flags.width),
int(500 * flags.width),
bias=False)
else:
self.dense1 = nn.Linear(int(800 * flags.width),
int(500 * flags.width),
bias=False)
self.bn3 = nn.BatchNorm1d(int(500 * flags.width),
affine=not flags.no_biases)
self.relu3 = nonlinearity()
# Dense
if flags.var4:
self.dense2 = layers.LinearVarianceUnif(int(500 * flags.width),
10,
bias=False)
else:
self.dense2 = nn.Linear(int(500 * flags.width), 10, bias=False)