def __init__(self, n_classes):
super(EnvNet5, self).__init__()
self.model = nn.Sequential(
OrderedDict([
('densenet',
DenseNet(growth_rate=32,
block_config=(6, 12, 24, 16))), #[b, 1024,9,8]
# growth_rate= 32 --> 16 解决当前服务器内存不够
('senet1', SEBottleneck(inplanes=256, planes=256 // 4)),
('senet2', SEBottleneck(inplanes=512, planes=512 // 4)),
# ('senet1', SELayer(channel=136)),
# ('senet2', SELayer(channel=260)),
('global_max_pool2', nn.AdaptiveAvgPool2d(1)),
('flatten', Flatten()),
('fc13',
nn.Linear(in_features=(256 + 512 + 1024),
out_features=n_classes,
bias=True)) #[b, n]
]))
# params initialization
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight,
mode='fan_out',
nonlinearity='relu')
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.normal_(m.weight, 0, 0.01)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
def __init__(self, n_classes):
super(EnvNet3, self).__init__()
self.model = nn.Sequential(
OrderedDict([
('conv1',
EnvReLu(in_channels=1,
out_channels=32,
kernel_size=(1, 64),
stride=(1, 2),
padding=0)), #[b,32, 1, 33294]
('conv2',
EnvReLu(in_channels=32,
out_channels=64,
kernel_size=(1, 16),
stride=(1, 2),
padding=0)), #[b, 64, 1, 16640]
('max_pool2',
nn.MaxPool2d(kernel_size=(1, 64),
stride=(1, 64),
ceil_mode=True)), #[b, 64, 1, 260]
('transpose', Transpose()), #[b, 1, 64, 260]
('densenet',
DenseNet(growth_rate=16, block_config=(6, 12, 24, 16))),
# growth_rate= 32 --> 16 解决当前服务器内存不够
('flatten', Flatten()),
('fc11',
nn.Linear(in_features=516 * 8 * 32,
out_features=1024,
bias=True)),
('relu11', nn.ReLU()),
('dropout11', nn.Dropout()),
('fc13',
nn.Linear(in_features=1024, out_features=n_classes,
bias=True)) #[b, n]
]))
# params initialization
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight,
mode='fan_out',
nonlinearity='relu')
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.normal_(m.weight, 0, 0.01)
nn.init.constant_(m.bias, 0)
def __init__(self, n_classes):
super(EnvNet4, self).__init__()
self.model = nn.Sequential(OrderedDict([
('conv1', EnvReLu(in_channels=1,
out_channels=32,
kernel_size=(1, 64),
stride=(1, 2),
padding=0)), #[b,32, 1, 33294]
('conv2', EnvReLu(in_channels=32,
out_channels=64,
kernel_size=(1, 16),
stride=(1, 2),
padding=0)), #[b, 64, 1, 16640]
('max_pool2', nn.MaxPool2d(kernel_size=(1, 64),
stride=(1, 64),
ceil_mode=True)), #[b, 64, 1, 260]
('transpose', Transpose()), #[b, 1, 64, 260]
('densenet', DenseNet(growth_rate=16, block_config=(6, 12, 24, 16))), #[b, 1024,8,32]
# growth_rate= 32 --> 16 解决当前服务器内存不够
('senet1', SEBottleneck(inplanes=136, planes=136//4)),
('senet2', SEBottleneck(inplanes=260, planes=260//4)),
# ('senet1', SELayer(channel=256)),
# ('senet2', SELayer(channel=512)),
('global avgpool', nn.AdaptiveAvgPool2d(1)),
('flatten', Flatten()),
('fc13', nn.Linear(in_features=(136+260+516), out_features=n_classes, bias=True))#[b, n]
]))
# # params initialization
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.normal_(m.weight, 0, 0.01)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
def __init__(self, n_classes):
super(EnvNet, self).__init__()
self.model = nn.Sequential(
OrderedDict([('conv1', ConvBNReLu(1, 40, (1, 8))),
('conv2', ConvBNReLu(40, 40, (1, 8))),
('max_pool2', nn.MaxPool2d((1, 160), ceil_mode=True)),
('transpose', Transpose()),
('conv3', ConvBNReLu(1, 50, (8, 13))),
('max_pool3', nn.MaxPool2d((3, 3), ceil_mode=True)),
('conv4', ConvBNReLu(50, 50, (1, 5))),
('max_pool4', nn.MaxPool2d((1, 3), ceil_mode=True)),
('flatten', Flatten()),
('fc5',
nn.Linear(in_features=50 * 11 * 14,
out_features=4096,
bias=True)), ('relu5', nn.ReLU()),
('dropout5', nn.Dropout()),
('fc6', nn.Linear(4096, 4096)), ('relu6', nn.ReLU()),
('dropout6', nn.Dropout()),
('fc7', nn.Linear(4096, n_classes))]))
def __init__(self, n_classes):
super(EnvNet2, self).__init__()
self.model = nn.Sequential(
OrderedDict([('conv1',
EnvReLu(in_channels=1,
out_channels=32,
kernel_size=(1, 64),
stride=(1, 2),
padding=0)),
('conv2',
EnvReLu(in_channels=32,
out_channels=64,
kernel_size=(1, 16),
stride=(1, 2),
padding=0)),
('max_pool2',
nn.MaxPool2d(kernel_size=(1, 64),
stride=(1, 64),
ceil_mode=True)),
('transpose', Transpose()),
('conv3',
EnvReLu(in_channels=1,
out_channels=32,
kernel_size=(8, 8),
stride=(1, 1),
padding=0)),
('conv4',
EnvReLu(in_channels=32,
out_channels=32,
kernel_size=(8, 8),
stride=(1, 1),
padding=0)),
('max_pool4',
nn.MaxPool2d(kernel_size=(5, 3),
stride=(5, 3),
ceil_mode=True)),
('conv5',
EnvReLu(in_channels=32,
out_channels=64,
kernel_size=(1, 4),
stride=(1, 1),
padding=0)),
('conv6',
EnvReLu(in_channels=64,
out_channels=64,
kernel_size=(1, 4),
stride=(1, 1),
padding=0)),
('max_pool6',
nn.MaxPool2d(kernel_size=(1, 2),
stride=(1, 2),
ceil_mode=True)),
('conv7',
EnvReLu(in_channels=64,
out_channels=128,
kernel_size=(1, 2),
stride=(1, 1),
padding=0)),
('conv8',
EnvReLu(in_channels=128,
out_channels=128,
kernel_size=(1, 2),
stride=(1, 1),
padding=0)),
('max_pool8',
nn.MaxPool2d(kernel_size=(1, 2),
stride=(1, 2),
ceil_mode=True)),
('conv9',
EnvReLu(in_channels=128,
out_channels=256,
kernel_size=(1, 2),
stride=(1, 1),
padding=0)),
('conv10',
EnvReLu(in_channels=256,
out_channels=256,
kernel_size=(1, 2),
stride=(1, 1),
padding=0)),
('max_pool10',
nn.MaxPool2d(kernel_size=(1, 2),
stride=(1, 2),
ceil_mode=True)),
('flatten', Flatten()),
('fc11',
nn.Linear(in_features=256 * 10 * 8,
out_features=4096,
bias=True)), ('relu11', nn.ReLU()),
('dropout11', nn.Dropout()),
('fc12',
nn.Linear(in_features=4096,
out_features=4096,
bias=True)), ('relu12', nn.ReLU()),
('dropout12', nn.Dropout()),
('fc13',
nn.Linear(in_features=4096,
out_features=n_classes,
bias=True))]))
# params initialization
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight,
mode='fan_out',
nonlinearity='relu')
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.normal_(m.weight, 0, 0.01)
nn.init.constant_(m.bias, 0)
def __init__(self, n_classes):
super(EnvNet2, self).__init__()
self.model = nn.Sequential(
OrderedDict([('conv1',
EnvReLu(in_channels=1,
out_channels=32,
kernel_size=(1, 64),
stride=(1, 2),
padding=0)),
('conv2',
EnvReLu(in_channels=32,
out_channels=64,
kernel_size=(1, 16),
stride=(1, 2),
padding=0)),
('max_pool2',
nn.MaxPool2d(kernel_size=(1, 64),
stride=(1, 64),
ceil_mode=True)),
('transpose', Transpose()),
('conv3',
EnvReLu(in_channels=1,
out_channels=32,
kernel_size=(8, 8),
stride=(1, 1),
padding=0)),
('conv4',
EnvReLu(in_channels=32,
out_channels=32,
kernel_size=(8, 8),
stride=(1, 1),
padding=0)),
('max_pool4',
nn.MaxPool2d(kernel_size=(5, 3),
stride=(5, 3),
ceil_mode=True)),
('conv5',
EnvReLu(in_channels=32,
out_channels=64,
kernel_size=(1, 4),
stride=(1, 1),
padding=0)),
('conv6',
EnvReLu(in_channels=64,
out_channels=64,
kernel_size=(1, 4),
stride=(1, 1),
padding=0)),
('max_pool6',
nn.MaxPool2d(kernel_size=(1, 2),
stride=(1, 2),
ceil_mode=True)),
('conv7',
EnvReLu(in_channels=64,
out_channels=128,
kernel_size=(1, 2),
stride=(1, 1),
padding=0)),
('conv8',
EnvReLu(in_channels=128,
out_channels=128,
kernel_size=(1, 2),
stride=(1, 1),
padding=0)),
('max_pool8',
nn.MaxPool2d(kernel_size=(1, 2),
stride=(1, 2),
ceil_mode=True)),
('conv9',
EnvReLu(in_channels=128,
out_channels=256,
kernel_size=(1, 2),
stride=(1, 1),
padding=0)),
('conv10',
EnvReLu(in_channels=256,
out_channels=256,
kernel_size=(1, 2),
stride=(1, 1),
padding=0)),
('max_pool10',
nn.MaxPool2d(kernel_size=(1, 2),
stride=(1, 2),
ceil_mode=True)),
('flatten', Flatten()),
('fc11',
nn.Linear(in_features=256 * 10 * 8,
out_features=4096,
bias=True)), ('relu11', nn.ReLU()),
('dropout11', nn.Dropout()),
('fc12',
nn.Linear(in_features=4096,
out_features=4096,
bias=True)), ('relu12', nn.ReLU()),
('dropout12', nn.Dropout()),
('fc13',
nn.Linear(in_features=4096,
out_features=n_classes,
bias=True)),
('softmax', nn.Softmax(dim=-1))]))