def __init__(self, phase='train', num_classes=2):
super(SSD_MobileNetV3, self).__init__()
self.phase = phase
self.num_classes = num_classes
self.priors = Variable(priorbox('mobilenetv3'), volatile=True)
self.conv_layer1 = nn.Sequential(
nn.Conv2d(3, 16, 3, 2, 1, bias=False),
nn.BatchNorm2d(16),
HardSwish())
'''
input = 16
kernel, expansion output_channel se nl stride
3 16 16 True 'RE' 2
3 72 24 False 'RE' 2
3 88 24 False 'RE' 1
5 96 40 True 'HS' 2
5 240 40 True 'HS' 1
5 240 40 True 'HS' 1
5 120 48 True 'HS' 1
5 144 48 True 'HS' 1
5 288 96 True 'HS' 2
5 576 96 True 'HS' 1
5 576 96 True 'HS' 1
'''
self.invertedresidual_layer1 = nn.Sequential(
nn.Conv2d(16, 16, 3, 2, 1, groups=16, bias=False),
nn.BatchNorm2d(16),
nn.ReLU6(inplace=True),
SqueezeAndExcite(16, reduction=4),
nn.Conv2d(16, 16, 1, 1, 0, bias=False),
nn.BatchNorm2d(16)
)
self.invertedresidual_layer2 = nn.Sequential(
nn.Conv2d(16, 72, 1, 1, 0, bias=False),
nn.BatchNorm2d(72),
nn.ReLU6(inplace=True),
nn.Conv2d(72, 72, 3, 2, 1, groups=72, bias=False),
nn.BatchNorm2d(72),
nn.ReLU6(inplace=True),
nn.Conv2d(72, 24, 1, 1, 0, bias=False),
nn.BatchNorm2d(24)
)
# Special
self.invertedresidual_layer3 = nn.Sequential(
nn.Conv2d(24, 88, 1, 1, 0, bias=False),
nn.BatchNorm2d(88),
nn.ReLU6(inplace=True),
nn.Conv2d(88, 88, 3, 1, 1, groups=88, bias=False),
nn.BatchNorm2d(88),
nn.Conv2d(88, 24, 1, 1, 0, bias=False),
nn.BatchNorm2d(24)
)
self.invertedresidual_layer4 = nn.Sequential(
nn.Conv2d(24, 96, 1, 1, 0, bias=False),
nn.BatchNorm2d(96),
HardSwish(),
nn.Conv2d(96, 96, 5, 2, 2, groups=96, bias=False),
nn.BatchNorm2d(96),
HardSwish(),
SqueezeAndExcite(96, reduction=4),
nn.Conv2d(96, 40, 1, 1, 0, bias=False),
nn.BatchNorm2d(40)
)
# Special
self.invertedresidual_layer5 = nn.Sequential(
nn.Conv2d(40, 240, 1, 1, 0, bias=False),
nn.BatchNorm2d(240),
HardSwish(),
nn.Conv2d(240, 240, 5, 1, 2, groups=240, bias=False),
nn.BatchNorm2d(240),
HardSwish(),
SqueezeAndExcite(240, reduction=4),
nn.Conv2d(240, 40, 1, 1, 0, bias=False),
nn.BatchNorm2d(40)
)
# Special
self.invertedresidual_layer6 = nn.Sequential(
nn.Conv2d(40, 240, 1, 1, 0, bias=False),
nn.BatchNorm2d(240),
HardSwish(),
nn.Conv2d(240, 240, 5, 1, 2, groups=240, bias=False),
nn.BatchNorm2d(240),
HardSwish(),
SqueezeAndExcite(240, reduction=4),
nn.Conv2d(240, 40, 1, 1, 0, bias=False),
nn.BatchNorm2d(40)
)
self.invertedresidual_layer7 = nn.Sequential(
nn.Conv2d(40, 120, 1, 1, 0, bias=False),
nn.BatchNorm2d(120),
HardSwish(),
nn.Conv2d(120, 120, 5, 1, 2, groups=120, bias=False),
nn.BatchNorm2d(120),
HardSwish(),
SqueezeAndExcite(120, reduction=4),
nn.Conv2d(120, 48, 1, 1, 0, bias=False),
nn.BatchNorm2d(48)
)
# Special
self.invertedresidual_layer8 = nn.Sequential(
nn.Conv2d(48, 144, 1, 1, 0, bias=False),
nn.BatchNorm2d(144),
HardSwish(),
nn.Conv2d(144, 144, 5, 1, 2, groups=144, bias=False),
nn.BatchNorm2d(144),
HardSwish(),
SqueezeAndExcite(144, reduction=4),
nn.Conv2d(144, 48, 1, 1, 0, bias=False),
nn.BatchNorm2d(48)
)
self.regression_layer1 = nn.Sequential(nn.Conv2d(48, 48,
kernel_size=3,
groups=48, stride=1, padding=1),
nn.BatchNorm2d(48),
nn.ReLU(),
nn.Conv2d(48, 6 * 4,
kernel_size=1))
self.confidence_layer1 = nn.Sequential(nn.Conv2d(48, 48,
kernel_size=3,
groups=48, stride=1, padding=1),
nn.BatchNorm2d(48),
nn.ReLU(),
nn.Conv2d(48, 6 * self.num_classes,
kernel_size=1))
self.invertedresidual_layer9 = nn.Sequential(
nn.Conv2d(48, 288, 1, 1, 0, bias=False),
nn.BatchNorm2d(288),
HardSwish(),
nn.Conv2d(288, 288, 5, 2, 2, groups=288, bias=False),
nn.BatchNorm2d(288),
HardSwish(),
SqueezeAndExcite(288, reduction=4),
nn.Conv2d(288, 96, 1, 1, 0, bias=False),
nn.BatchNorm2d(96)
)
# Special
self.invertedresidual_layer10 = nn.Sequential(
nn.Conv2d(96, 576, 1, 1, 0, bias=False),
nn.BatchNorm2d(576),
HardSwish(),
nn.Conv2d(576, 576, 5, 1, 2, groups=576, bias=False),
nn.BatchNorm2d(576),
HardSwish(),
SqueezeAndExcite(576, reduction=4),
nn.Conv2d(576, 96, 1, 1, 0, bias=False),
nn.BatchNorm2d(96)
)
# Special
self.invertedresidual_layer11 = nn.Sequential(
nn.Conv2d(96, 576, 1, 1, 0, bias=False),
nn.BatchNorm2d(576),
HardSwish(),
nn.Conv2d(576, 576, 5, 1, 2, groups=576, bias=False),
nn.BatchNorm2d(576),
HardSwish(),
SqueezeAndExcite(576, reduction=4),
nn.Conv2d(576, 96, 1, 1, 0, bias=False),
nn.BatchNorm2d(96)
)
self.conv_layer2 = nn.Sequential(
nn.Conv2d(96, 576, 1, 1, 0, bias=False),
SqueezeAndExcite(576, reduction=4),
nn.BatchNorm2d(576),
HardSwish()
)
self.conv_layer3 = nn.Sequential(
nn.Conv2d(576, 1280, 1, 1, 0, bias=False),
SqueezeAndExcite(1280, reduction=4),
nn.BatchNorm2d(1280),
HardSwish()
)
self.regression_layer2 = nn.Sequential(nn.Conv2d(1280, 1280,
kernel_size=3,
groups=1280, stride=1, padding=1),
nn.BatchNorm2d(1280),
nn.ReLU(),
nn.Conv2d(1280, 6 * 4,
kernel_size=1))
self.confidence_layer2 = nn.Sequential(nn.Conv2d(1280, 1280,
kernel_size=3,
groups=1280, stride=1, padding=1),
nn.BatchNorm2d(1280),
nn.ReLU(),
nn.Conv2d(1280, 6 * self.num_classes,
kernel_size=1))
self.extra1 = nn.Sequential(nn.Conv2d(1280, 256, 1, 1, 0, bias=False),
nn.BatchNorm2d(256),
nn.ReLU(inplace=True),
nn.Conv2d(256, 256, 3, 2, 1, groups=256, bias=False),
nn.BatchNorm2d(256),
nn.ReLU(inplace=True),
nn.Conv2d(256, 512, 1, 1, 0, bias=False),
nn.BatchNorm2d(512))
self.regression_layer3 = nn.Sequential(nn.Conv2d(512, 512,
kernel_size=3,
groups=512, stride=1, padding=1),
nn.BatchNorm2d(512),
nn.ReLU(),
nn.Conv2d(512, 6 * 4,
kernel_size=1))
self.confidence_layer3 = nn.Sequential(nn.Conv2d(512, 512,
kernel_size=3,
groups=512, stride=1, padding=1),
nn.BatchNorm2d(512),
nn.ReLU(),
nn.Conv2d(512, 6 * self.num_classes,
kernel_size=1))
self.extra2 = nn.Sequential(nn.Conv2d(512, 128, 1, 1, 0, bias=False),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 128, 3, 2, 1, groups=128, bias=False),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 256, 1, 1, 0, bias=False),
nn.BatchNorm2d(256))
self.regression_layer4 = nn.Sequential(nn.Conv2d(256, 256,
kernel_size=3,
groups=256, stride=1, padding=1),
nn.BatchNorm2d(256),
nn.ReLU(),
nn.Conv2d(256, 4 * 4,
kernel_size=1))
self.confidence_layer4 = nn.Sequential(nn.Conv2d(256, 256,
kernel_size=3,
groups=256, stride=1, padding=1),
nn.BatchNorm2d(256),
nn.ReLU(),
nn.Conv2d(256, 4 * self.num_classes,
kernel_size=1))
self.extra3 = nn.Sequential(nn.Conv2d(256, 128, 1, 1, 0, bias=False),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 128, 3, 2, 1, groups=128, bias=False),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 256, 1, 1, 0, bias=False),
nn.BatchNorm2d(256))
self.regression_layer5 = nn.Sequential(nn.Conv2d(256, 256,
kernel_size=3,
groups=256, stride=1, padding=1),
nn.BatchNorm2d(256),
nn.ReLU(),
nn.Conv2d(256, 4 * 4,
kernel_size=1))
self.confidence_layer5 = nn.Sequential(nn.Conv2d(256, 256,
kernel_size=3,
groups=256, stride=1, padding=1),
nn.BatchNorm2d(256),
nn.ReLU(),
nn.Conv2d(256, 4 * self.num_classes,
kernel_size=1))
self.extra4 = nn.Sequential(nn.Conv2d(256, 64, 1, 1, 0, bias=False),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
nn.Conv2d(64, 64, 3, 2, 1, groups=64, bias=False),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
nn.Conv2d(64, 64, 1, 1, 0, bias=False),
nn.BatchNorm2d(64))
self.regression_layer6 = nn.Sequential(nn.Conv2d(64, 64,
kernel_size=3,
groups=64, stride=1, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.Conv2d(64, 4 * 4,
kernel_size=1))
self.confidence_layer6 = nn.Sequential(nn.Conv2d(64, 64,
kernel_size=3,
groups=64, stride=1, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.Conv2d(64, 4 * self.num_classes,
kernel_size=1))
if phase == 'test':
self.softmax = nn.Softmax()
self.inference_layer = InferenceLayer(top_k=200, conf_thresh=0.01, nms_thresh=0.45)
def __init__(self, phase='train', num_classes=21):
super(SSD_VGG, self).__init__()
self.phase = phase
self.num_classes = num_classes
self.priors = Variable(priorbox('vgg'), volatile=True)
self.conv_layer1 = nn.Sequential(
nn.Conv2d(3, 64, kernel_size=3, padding=1), nn.ReLU(inplace=True))
self.conv_layer2 = nn.Sequential(
nn.Conv2d(64, 64, kernel_size=3, padding=1), nn.ReLU(inplace=True))
self.maxpool_layer1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv_layer3 = nn.Sequential(
nn.Conv2d(64, 128, kernel_size=3, padding=1),
nn.ReLU(inplace=True))
self.conv_layer4 = nn.Sequential(
nn.Conv2d(128, 128, kernel_size=3, padding=1),
nn.ReLU(inplace=True))
self.maxpool_layer2 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv_layer5 = nn.Sequential(
nn.Conv2d(128, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True))
self.conv_layer6 = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True))
self.conv_layer7 = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True))
self.maxpool_layer3 = nn.MaxPool2d(kernel_size=2,
stride=2,
ceil_mode=True)
self.conv_layer8 = nn.Sequential(
nn.Conv2d(256, 512, kernel_size=3, padding=1),
nn.ReLU(inplace=True))
self.conv_layer9 = nn.Sequential(
nn.Conv2d(512, 512, kernel_size=3, padding=1),
nn.ReLU(inplace=True))
self.conv_layer10 = nn.Sequential(
nn.Conv2d(512, 512, kernel_size=3, padding=1),
nn.ReLU(inplace=True))
self.l2_norm_scale = L2NormScale(512, 20)
self.maxpool_layer4 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv_layer11 = nn.Sequential(
nn.Conv2d(512, 512, kernel_size=3, padding=1),
nn.ReLU(inplace=True))
self.conv_layer12 = nn.Sequential(
nn.Conv2d(512, 512, kernel_size=3, padding=1),
nn.ReLU(inplace=True))
self.conv_layer13 = nn.Sequential(
nn.Conv2d(512, 512, kernel_size=3, padding=1),
nn.ReLU(inplace=True))
self.maxpool_layer5 = nn.MaxPool2d(kernel_size=3, stride=1, padding=1)
self.conv_layer14 = nn.Sequential(
nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6),
nn.ReLU(inplace=True))
self.conv_layer15 = nn.Sequential(nn.Conv2d(1024, 1024, kernel_size=1),
nn.ReLU(inplace=True))
self.loc0 = nn.Conv2d(512, 4 * 4, kernel_size=3, padding=1)
self.class0 = nn.Conv2d(512, 4 * num_classes, kernel_size=3, padding=1)
self.loc1 = nn.Conv2d(1024, 6 * 4, kernel_size=3, padding=1)
self.class1 = nn.Conv2d(1024,
6 * num_classes,
kernel_size=3,
padding=1)
self.extra2 = nn.Sequential(
nn.Conv2d(1024, 256, kernel_size=1), nn.ReLU(inplace=True),
nn.Conv2d(256, 512, kernel_size=3, stride=2, padding=1),
nn.ReLU(inplace=True))
self.loc2 = nn.Conv2d(512, 6 * 4, kernel_size=3, padding=1)
self.class2 = nn.Conv2d(512, 6 * num_classes, kernel_size=3, padding=1)
self.extra3 = nn.Sequential(
nn.Conv2d(512, 128, kernel_size=1), nn.ReLU(inplace=True),
nn.Conv2d(128, 256, kernel_size=3, stride=2, padding=1),
nn.ReLU(inplace=True))
self.loc3 = nn.Conv2d(256, 6 * 4, kernel_size=3, padding=1)
self.class3 = nn.Conv2d(256, 6 * num_classes, kernel_size=3, padding=1)
self.extra4 = nn.Sequential(nn.Conv2d(256, 128, kernel_size=1),
nn.ReLU(inplace=True),
nn.Conv2d(128, 256, kernel_size=3),
nn.ReLU(inplace=True))
self.loc4 = nn.Conv2d(256, 4 * 4, kernel_size=3, padding=1)
self.class4 = nn.Conv2d(256, 4 * num_classes, kernel_size=3, padding=1)
self.extra5 = nn.Sequential(nn.Conv2d(256, 128, kernel_size=1),
nn.ReLU(inplace=True),
nn.Conv2d(128, 256, kernel_size=3),
nn.ReLU(inplace=True))
self.loc5 = nn.Conv2d(256, 4 * 4, kernel_size=3, padding=1)
self.class5 = nn.Conv2d(256, 4 * num_classes, kernel_size=3, padding=1)
if phase == 'test':
self.softmax = nn.Softmax()
self.inference_layer = InferenceLayer(top_k=200,
conf_thresh=0.01,
nms_thresh=0.45)
def __init__(self,
num_classes,
phase='train',
blocks_args=None,
global_params=None):
super().__init__()
assert isinstance(blocks_args, list), 'blocks_args should be a list'
assert len(blocks_args) > 0, 'block args must be greater than 0'
self._global_params = global_params
self._blocks_args = blocks_args
self.phase = phase
self.num_classes = num_classes
self.priors = Variable(priorbox('efficientnet'), volatile=True)
bn_mom = 1 - self._global_params.batch_norm_momentum
bn_eps = self._global_params.batch_norm_epsilon
in_channels = 3
out_channels = round_filters(32, self._global_params)
self._conv_stem = Conv2dSamePadding(in_channels,
out_channels,
kernel_size=3,
stride=2,
bias=False)
self._bn0 = nn.BatchNorm2d(num_features=out_channels,
momentum=bn_mom,
eps=bn_eps)
self._blocks = nn.ModuleList([])
for block_args in self._blocks_args:
block_args = block_args._replace(
input_filters=round_filters(block_args.input_filters,
self._global_params),
output_filters=round_filters(block_args.output_filters,
self._global_params),
num_repeat=round_repeats(block_args.num_repeat,
self._global_params))
self._blocks.append(MBConvBlock(block_args, self._global_params))
if block_args.num_repeat > 1:
block_args = block_args._replace(
input_filters=block_args.output_filters, stride=1)
for _ in range(block_args.num_repeat - 1):
self._blocks.append(
MBConvBlock(block_args, self._global_params))
in_channels = block_args.output_filters # output of final block
out_channels = round_filters(1280, self._global_params)
self._conv_head = Conv2dSamePadding(in_channels,
out_channels,
kernel_size=1,
bias=False)
self._bn1 = nn.BatchNorm2d(num_features=out_channels,
momentum=bn_mom,
eps=bn_eps)
self.regression_layer1 = nn.Sequential(
nn.Conv2d(160, 160, kernel_size=3, groups=160, stride=1,
padding=1), nn.BatchNorm2d(160), nn.ReLU(),
nn.Conv2d(160, 6 * 4, kernel_size=1))
self.confidence_layer1 = nn.Sequential(
nn.Conv2d(160, 160, kernel_size=3, groups=160, stride=1,
padding=1), nn.BatchNorm2d(160), nn.ReLU(),
nn.Conv2d(160, 6 * self.num_classes, kernel_size=1))
self.regression_layer2 = nn.Sequential(
nn.Conv2d(1792,
1792,
kernel_size=3,
groups=1792,
stride=1,
padding=1), nn.BatchNorm2d(1792), nn.ReLU(),
nn.Conv2d(1792, 6 * 4, kernel_size=1))
self.confidence_layer2 = nn.Sequential(
nn.Conv2d(1792,
1792,
kernel_size=3,
groups=1792,
stride=1,
padding=1), nn.BatchNorm2d(1792), nn.ReLU(),
nn.Conv2d(1792, 6 * self.num_classes, kernel_size=1))
self.regression_layer3 = nn.Sequential(
nn.Conv2d(512, 512, kernel_size=3, groups=512, stride=1,
padding=1), nn.BatchNorm2d(512), nn.ReLU(),
nn.Conv2d(512, 6 * 4, kernel_size=1))
self.confidence_layer3 = nn.Sequential(
nn.Conv2d(512, 512, kernel_size=3, groups=512, stride=1,
padding=1), nn.BatchNorm2d(512), nn.ReLU(),
nn.Conv2d(512, 6 * self.num_classes, kernel_size=1))
self.regression_layer4 = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, groups=256, stride=1,
padding=1), nn.BatchNorm2d(256), nn.ReLU(),
nn.Conv2d(256, 4 * 4, kernel_size=1))
self.confidence_layer4 = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, groups=256, stride=1,
padding=1), nn.BatchNorm2d(256), nn.ReLU(),
nn.Conv2d(256, 4 * self.num_classes, kernel_size=1))
self.regression_layer5 = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, groups=256, stride=1,
padding=1), nn.BatchNorm2d(256), nn.ReLU(),
nn.Conv2d(256, 4 * 4, kernel_size=1))
self.confidence_layer5 = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, groups=256, stride=1,
padding=1), nn.BatchNorm2d(256), nn.ReLU(),
nn.Conv2d(256, 4 * self.num_classes, kernel_size=1))
self.regression_layer6 = nn.Sequential(
nn.Conv2d(64, 64, kernel_size=3, groups=64, stride=1, padding=1),
nn.BatchNorm2d(64), nn.ReLU(), nn.Conv2d(64, 4 * 4, kernel_size=1))
self.confidence_layer6 = nn.Sequential(
nn.Conv2d(64, 64, kernel_size=3, groups=64, stride=1, padding=1),
nn.BatchNorm2d(64), nn.ReLU(),
nn.Conv2d(64, 4 * self.num_classes, kernel_size=1))
self.extra1 = nn.Sequential(
nn.Conv2d(1792, 256, 1, 1, 0, bias=False), nn.BatchNorm2d(256),
nn.ReLU(inplace=True),
nn.Conv2d(256, 256, 3, 2, 1, groups=256, bias=False),
nn.BatchNorm2d(256), nn.ReLU(inplace=True),
nn.Conv2d(256, 512, 1, 1, 0, bias=False), nn.BatchNorm2d(512))
self.extra2 = nn.Sequential(
nn.Conv2d(512, 128, 1, 1, 0, bias=False), nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 128, 3, 2, 1, groups=128, bias=False),
nn.BatchNorm2d(128), nn.ReLU(inplace=True),
nn.Conv2d(128, 256, 1, 1, 0, bias=False), nn.BatchNorm2d(256))
self.extra3 = nn.Sequential(
nn.Conv2d(256, 128, 1, 1, 0, bias=False), nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 128, 3, 2, 1, groups=128, bias=False),
nn.BatchNorm2d(128), nn.ReLU(inplace=True),
nn.Conv2d(128, 256, 1, 1, 0, bias=False), nn.BatchNorm2d(256))
self.extra4 = nn.Sequential(
nn.Conv2d(256, 64, 1, 1, 0, bias=False), nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
nn.Conv2d(64, 64, 3, 2, 1, groups=64, bias=False),
nn.BatchNorm2d(64), nn.ReLU(inplace=True),
nn.Conv2d(64, 64, 1, 1, 0, bias=False), nn.BatchNorm2d(64))
if self.phase == 'test':
self.softmax = nn.Softmax()
self.inference_layer = InferenceLayer(top_k=200,
conf_thresh=0.01,
nms_thresh=0.45)
def __init__(self, phase='train', num_classes=2):
super(SSD_MobileNetV2, self).__init__()
self.phase = phase
self.num_classes = num_classes
self.priors = Variable(priorbox('mobilenetv2'), volatile=True)
self.conv_layer1 = nn.Sequential(nn.Conv2d(3, 32, 3, 2, 1, bias=False),
nn.BatchNorm2d(32),
nn.ReLU(inplace=True))
self.invertedresidual_layer1 = nn.Sequential(
nn.Conv2d(32, 32, 3, 1, 1, groups=32, bias=False),
nn.BatchNorm2d(32), nn.ReLU(inplace=True),
nn.Conv2d(32, 16, 1, 1, 0, bias=False), nn.BatchNorm2d(16))
self.invertedresidual_layer2 = nn.Sequential(
nn.Conv2d(16, 96, 1, 1, 0, bias=False), nn.BatchNorm2d(96),
nn.ReLU(inplace=True),
nn.Conv2d(96, 96, 3, 2, 1, groups=96, bias=False),
nn.BatchNorm2d(96), nn.ReLU(inplace=True),
nn.Conv2d(96, 24, 1, 1, 0, bias=False), nn.BatchNorm2d(24))
self.invertedresidual_layer3 = nn.Sequential(
nn.Conv2d(24, 144, 1, 1, 0, bias=False), nn.BatchNorm2d(144),
nn.ReLU(inplace=True),
nn.Conv2d(144, 144, 3, 1, 1, groups=144, bias=False),
nn.BatchNorm2d(144), nn.ReLU(inplace=True),
nn.Conv2d(144, 24, 1, 1, 0, bias=False), nn.BatchNorm2d(24))
self.invertedresidual_layer4 = nn.Sequential(
nn.Conv2d(24, 144, 1, 1, 0, bias=False), nn.BatchNorm2d(144),
nn.ReLU(inplace=True),
nn.Conv2d(144, 144, 3, 2, 1, groups=144, bias=False),
nn.BatchNorm2d(144), nn.ReLU(inplace=True),
nn.Conv2d(144, 32, 1, 1, 0, bias=False), nn.BatchNorm2d(32))
self.invertedresidual_layer5 = nn.Sequential(
nn.Conv2d(32, 192, 1, 1, 0, bias=False), nn.BatchNorm2d(192),
nn.ReLU(inplace=True),
nn.Conv2d(192, 192, 3, 1, 1, groups=192, bias=False),
nn.BatchNorm2d(192), nn.ReLU(inplace=True),
nn.Conv2d(192, 32, 1, 1, 0, bias=False), nn.BatchNorm2d(32))
self.invertedresidual_layer6 = nn.Sequential(
nn.Conv2d(32, 192, 1, 1, 0, bias=False), nn.BatchNorm2d(192),
nn.ReLU(inplace=True),
nn.Conv2d(192, 192, 3, 1, 1, groups=192, bias=False),
nn.BatchNorm2d(192), nn.ReLU(inplace=True),
nn.Conv2d(192, 32, 1, 1, 0, bias=False), nn.BatchNorm2d(32))
self.invertedresidual_layer7 = nn.Sequential(
nn.Conv2d(32, 192, 1, 1, 0, bias=False), nn.BatchNorm2d(192),
nn.ReLU(inplace=True),
nn.Conv2d(192, 192, 3, 2, 1, groups=192, bias=False),
nn.BatchNorm2d(192), nn.ReLU(inplace=True),
nn.Conv2d(192, 64, 1, 1, 0, bias=False), nn.BatchNorm2d(64))
self.invertedresidual_layer8 = nn.Sequential(
nn.Conv2d(64, 384, 1, 1, 0, bias=False), nn.BatchNorm2d(384),
nn.ReLU(inplace=True),
nn.Conv2d(384, 384, 3, 1, 1, groups=384, bias=False),
nn.BatchNorm2d(384), nn.ReLU(inplace=True),
nn.Conv2d(384, 64, 1, 1, 0, bias=False), nn.BatchNorm2d(64))
self.invertedresidual_layer9 = nn.Sequential(
nn.Conv2d(64, 384, 1, 1, 0, bias=False), nn.BatchNorm2d(384),
nn.ReLU(inplace=True),
nn.Conv2d(384, 384, 3, 1, 1, groups=384, bias=False),
nn.BatchNorm2d(384), nn.ReLU(inplace=True),
nn.Conv2d(384, 64, 1, 1, 0, bias=False), nn.BatchNorm2d(64))
self.invertedresidual_layer10 = nn.Sequential(
nn.Conv2d(64, 384, 1, 1, 0, bias=False), nn.BatchNorm2d(384),
nn.ReLU(inplace=True),
nn.Conv2d(384, 384, 3, 1, 1, groups=384, bias=False),
nn.BatchNorm2d(384), nn.ReLU(inplace=True),
nn.Conv2d(384, 64, 1, 1, 0, bias=False), nn.BatchNorm2d(64))
self.invertedresidual_layer11 = nn.Sequential(
nn.Conv2d(64, 384, 1, 1, 0, bias=False), nn.BatchNorm2d(384),
nn.ReLU(inplace=True),
nn.Conv2d(384, 384, 3, 1, 1, groups=384, bias=False),
nn.BatchNorm2d(384), nn.ReLU(inplace=True),
nn.Conv2d(384, 96, 1, 1, 0, bias=False), nn.BatchNorm2d(96))
self.invertedresidual_layer12 = nn.Sequential(
nn.Conv2d(96, 576, 1, 1, 0, bias=False), nn.BatchNorm2d(576),
nn.ReLU(inplace=True),
nn.Conv2d(576, 576, 3, 1, 1, groups=576, bias=False),
nn.BatchNorm2d(576), nn.ReLU(inplace=True),
nn.Conv2d(576, 96, 1, 1, 0, bias=False), nn.BatchNorm2d(96))
self.invertedresidual_layer13 = nn.Sequential(
nn.Conv2d(96, 576, 1, 1, 0, bias=False), nn.BatchNorm2d(576),
nn.ReLU(inplace=True),
nn.Conv2d(576, 576, 3, 1, 1, groups=576, bias=False),
nn.BatchNorm2d(576), nn.ReLU(inplace=True),
nn.Conv2d(576, 96, 1, 1, 0, bias=False), nn.BatchNorm2d(96))
self.regression_layer1 = nn.Sequential(
nn.Conv2d(96, 96, kernel_size=3, groups=96, stride=1, padding=1),
nn.BatchNorm2d(96), nn.ReLU(), nn.Conv2d(96, 6 * 4, kernel_size=1))
self.confidence_layer1 = nn.Sequential(
nn.Conv2d(96, 96, kernel_size=3, groups=96, stride=1, padding=1),
nn.BatchNorm2d(96), nn.ReLU(),
nn.Conv2d(96, 6 * self.num_classes, kernel_size=1))
self.invertedresidual_layer14 = nn.Sequential(
nn.Conv2d(96, 576, 1, 1, 0, bias=False), nn.BatchNorm2d(576),
nn.ReLU(inplace=True),
nn.Conv2d(576, 576, 3, 2, 1, groups=576, bias=False),
nn.BatchNorm2d(576), nn.ReLU(inplace=True),
nn.Conv2d(576, 160, 1, 1, 0, bias=False), nn.BatchNorm2d(160))
self.invertedresidual_layer15 = nn.Sequential(
nn.Conv2d(160, 960, 1, 1, 0, bias=False), nn.BatchNorm2d(960),
nn.ReLU(inplace=True),
nn.Conv2d(960, 960, 3, 1, 1, groups=960, bias=False),
nn.BatchNorm2d(960), nn.ReLU(inplace=True),
nn.Conv2d(960, 160, 1, 1, 0, bias=False), nn.BatchNorm2d(160))
self.invertedresidual_layer16 = nn.Sequential(
nn.Conv2d(160, 960, 1, 1, 0, bias=False), nn.BatchNorm2d(960),
nn.ReLU(inplace=True),
nn.Conv2d(960, 960, 3, 1, 1, groups=960, bias=False),
nn.BatchNorm2d(960), nn.ReLU(inplace=True),
nn.Conv2d(960, 160, 1, 1, 0, bias=False), nn.BatchNorm2d(160))
self.invertedresidual_layer17 = nn.Sequential(
nn.Conv2d(160, 960, 1, 1, 0, bias=False), nn.BatchNorm2d(960),
nn.ReLU(inplace=True),
nn.Conv2d(960, 960, 3, 1, 1, groups=960, bias=False),
nn.BatchNorm2d(960), nn.ReLU(inplace=True),
nn.Conv2d(960, 320, 1, 1, 0, bias=False), nn.BatchNorm2d(320))
self.conv_layer2 = nn.Sequential(
nn.Conv2d(320, 1280, 1, 1, 0, bias=False), nn.BatchNorm2d(1280),
nn.ReLU(inplace=True))
self.regression_layer2 = nn.Sequential(
nn.Conv2d(1280,
1280,
kernel_size=3,
groups=1280,
stride=1,
padding=1), nn.BatchNorm2d(1280), nn.ReLU(),
nn.Conv2d(1280, 6 * 4, kernel_size=1))
self.confidence_layer2 = nn.Sequential(
nn.Conv2d(1280,
1280,
kernel_size=3,
groups=1280,
stride=1,
padding=1), nn.BatchNorm2d(1280), nn.ReLU(),
nn.Conv2d(1280, 6 * self.num_classes, kernel_size=1))
self.extra1 = nn.Sequential(
nn.Conv2d(1280, 256, 1, 1, 0, bias=False), nn.BatchNorm2d(256),
nn.ReLU(inplace=True),
nn.Conv2d(256, 256, 3, 2, 1, groups=256, bias=False),
nn.BatchNorm2d(256), nn.ReLU(inplace=True),
nn.Conv2d(256, 512, 1, 1, 0, bias=False), nn.BatchNorm2d(512))
self.regression_layer3 = nn.Sequential(
nn.Conv2d(512, 512, kernel_size=3, groups=512, stride=1,
padding=1), nn.BatchNorm2d(512), nn.ReLU(),
nn.Conv2d(512, 6 * 4, kernel_size=1))
self.confidence_layer3 = nn.Sequential(
nn.Conv2d(512, 512, kernel_size=3, groups=512, stride=1,
padding=1), nn.BatchNorm2d(512), nn.ReLU(),
nn.Conv2d(512, 6 * self.num_classes, kernel_size=1))
self.extra2 = nn.Sequential(
nn.Conv2d(512, 128, 1, 1, 0, bias=False), nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 128, 3, 2, 1, groups=128, bias=False),
nn.BatchNorm2d(128), nn.ReLU(inplace=True),
nn.Conv2d(128, 256, 1, 1, 0, bias=False), nn.BatchNorm2d(256))
self.regression_layer4 = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, groups=256, stride=1,
padding=1), nn.BatchNorm2d(256), nn.ReLU(),
nn.Conv2d(256, 4 * 4, kernel_size=1))
self.confidence_layer4 = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, groups=256, stride=1,
padding=1), nn.BatchNorm2d(256), nn.ReLU(),
nn.Conv2d(256, 4 * self.num_classes, kernel_size=1))
self.extra3 = nn.Sequential(
nn.Conv2d(256, 128, 1, 1, 0, bias=False), nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 128, 3, 2, 1, groups=128, bias=False),
nn.BatchNorm2d(128), nn.ReLU(inplace=True),
nn.Conv2d(128, 256, 1, 1, 0, bias=False), nn.BatchNorm2d(256))
self.regression_layer5 = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, groups=256, stride=1,
padding=1), nn.BatchNorm2d(256), nn.ReLU(),
nn.Conv2d(256, 4 * 4, kernel_size=1))
self.confidence_layer5 = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, groups=256, stride=1,
padding=1), nn.BatchNorm2d(256), nn.ReLU(),
nn.Conv2d(256, 4 * self.num_classes, kernel_size=1))
self.extra4 = nn.Sequential(
nn.Conv2d(256, 64, 1, 1, 0, bias=False), nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
nn.Conv2d(64, 64, 3, 2, 1, groups=64, bias=False),
nn.BatchNorm2d(64), nn.ReLU(inplace=True),
nn.Conv2d(64, 64, 1, 1, 0, bias=False), nn.BatchNorm2d(64))
self.regression_layer6 = nn.Sequential(
nn.Conv2d(64, 64, kernel_size=3, groups=64, stride=1, padding=1),
nn.BatchNorm2d(64), nn.ReLU(), nn.Conv2d(64, 4 * 4, kernel_size=1))
self.confidence_layer6 = nn.Sequential(
nn.Conv2d(64, 64, kernel_size=3, groups=64, stride=1, padding=1),
nn.BatchNorm2d(64), nn.ReLU(),
nn.Conv2d(64, 4 * self.num_classes, kernel_size=1))
if phase == 'test':
self.softmax = nn.Softmax()
self.inference_layer = InferenceLayer(top_k=200,
conf_thresh=0.01,
nms_thresh=0.45)