def __init__(self, input_channel, pool_reduce_channel, conv1_channel,
conv3_reduce_channel, conv3_channel,
conv3_double_reduce_channel, conv3_double_channel_1_1,
conv3_double_channel_1_2, conv3_double_channel_2):
super(Inceptionv4_C, self).__init__()
self.pool = nn.Sequential(
nn.AvgPool2d(kernel_size=3, stride=1, padding=1),
Conv2d(input_channel, pool_reduce_channel, kernel_size=1))
self.conv1 = Conv2d(input_channel, conv1_channel, kernel_size=1)
self.conv3 = nn.Sequential(
Conv2d(input_channel, conv3_reduce_channel, kernel_size=1),
Concat_Separable_Conv2d(conv3_reduce_channel,
conv3_channel,
kernel_size=3,
padding=1))
self.conv3_double = nn.Sequential(
Conv2d(input_channel, conv3_double_reduce_channel, kernel_size=1),
Separable_Conv2d(conv3_double_reduce_channel,
conv3_double_channel_1_1,
conv3_double_channel_1_2,
kernel_size=3,
padding=1),
Concat_Separable_Conv2d(conv3_double_channel_1_2,
conv3_double_channel_2,
kernel_size=3,
padding=1))
def __init__(self, input_channel, conv3_channel,
conv3_double_reduce_channel, conv3_double_channel):
super(Inceptionv4_reduction_B, self).__init__()
self.pool = nn.MaxPool2d(3, 2)
self.conv3 = nn.Sequential(Conv2d(input_channel, conv3_channel, 1),
Conv2d(conv3_channel, conv3_channel, 3, 2))
self.conv3_double = nn.Sequential(
Conv2d(input_channel, conv3_double_reduce_channel, 1),
Separable_Conv2d(conv3_double_reduce_channel,
conv3_double_reduce_channel,
conv3_double_channel,
7,
padding=3),
Conv2d(conv3_double_channel, conv3_double_channel, 3, stride=2))
def __init__(self, input_channel, conv3_channel,
conv3_double_reduce_channel, conv3_double_channel_1,
conv3_double_channel_2):
super(Inceptionv4_reduction_A, self).__init__()
self.pool = nn.MaxPool2d(3, 2)
self.conv3 = Conv2d(input_channel, conv3_channel, 3, 2) # n
self.conv3_double = nn.Sequential(
Conv2d(input_channel, conv3_double_reduce_channel, 1), # k
Conv2d(conv3_double_reduce_channel,
conv3_double_channel_1,
3,
padding=1), # l
Conv2d(conv3_double_channel_1, conv3_double_channel_2, 3,
stride=2) # m
)
def __init__(self):
super(Inceptionv4_stem, self).__init__()
self.stem1 = nn.Sequential(
Conv2d(3, 32, 3, stride=2),
Conv2d(32, 32, 3, stride=1),
Conv2d(32, 64, 3, stride=1, padding=1),
)
self.stem1_1 = nn.MaxPool2d(3, 2)
self.stem1_2 = Conv2d(64, 96, 3, 2)
self.stem2_1 = nn.Sequential(
Conv2d(160, 64, 1),
Conv2d(64, 96, 3),
)
self.stem2_2 = nn.Sequential(
Conv2d(160, 64, 1), Separable_Conv2d(64, 64, 64, 7, padding=3),
Conv2d(64, 96, 3))
self.stem3_1 = nn.MaxPool2d(3, 2)
self.stem3_2 = Conv2d(192, 192, 3, 2)
def __init__(self, num_classes, mode='train'):
super(GoogLeNetv4, self).__init__()
self.num_classes = num_classes
self.mode = mode
self.layers = nn.Sequential(
Inceptionv4_stem(), Inceptionv4_A(384, 96, 96, 64, 96, 64, 96, 96),
Inceptionv4_A(384, 96, 96, 64, 96, 64, 96, 96),
Inceptionv4_A(384, 96, 96, 64, 96, 64, 96, 96),
Inceptionv4_A(384, 96, 96, 64, 96, 64, 96, 96),
Inceptionv4_reduction_A(384, 384, 192, 224, 256),
Inceptionv4_B(1024, 128, 384, 192, 224, 256, 192, 224, 256),
Inceptionv4_B(1024, 128, 384, 192, 224, 256, 192, 224, 256),
Inceptionv4_B(1024, 128, 384, 192, 224, 256, 192, 224, 256),
Inceptionv4_B(1024, 128, 384, 192, 224, 256, 192, 224, 256),
Inceptionv4_B(1024, 128, 384, 192, 224, 256, 192, 224, 256),
Inceptionv4_B(1024, 128, 384, 192, 224, 256, 192, 224, 256),
Inceptionv4_B(1024, 128, 384, 192, 224, 256, 192, 224, 256),
Inceptionv4_reduction_B(1024, 192, 256, 320),
Inceptionv4_C(1536, 256, 256, 384, 256, 384, 448, 512, 256),
Inceptionv4_C(1536, 256, 256, 384, 256, 384, 448, 512, 256),
Inceptionv4_C(1536, 256, 256, 384, 256, 384, 448, 512, 256),
nn.AvgPool2d(8, 1), nn.Dropout2d(0.2, inplace=True),
Conv2d(1536, num_classes, kernel_size=1, output=True), Squeeze())