def __init__(self, args, num_classes):
"""
Initializes Deeplabv3_plus
Args:
args (argparse.ArgumentParser): command line arguments
num_classes (int): specifies the output channels of the last convolution
"""
super(Deeplabv3_plus, self).__init__()
self.args = args
if args.model == DEEPLAB:
self.backbone = ResNet101(args.output_stride, args.init_type)
if self.args.refine_network:
self.refine_backbone = ResNet101(args.output_stride, args.init_type)
elif args.model == DEEPLAB_50:
self.backbone = ResNet50(args.output_stride, args.init_type)
if self.args.refine_network:
self.refine_backbone = ResNet50(args.output_stride, args.init_type)
else:
raise NotImplementedError()
self.aspp = ASPP(args.output_stride, self.args.resize)
self.decoder = Decoder(num_classes)
self.upsampling = tf.keras.layers.UpSampling2D((4,4), interpolation='bilinear')
if self.args.refine_network:
self.refine_aspp = ASPP(args.output_stride, self.args.resize)
self.refine_decoder = Decoder(num_classes)
def __init__(self):
super(SSD, self).__init__()
self.num_classes = NUM_CLASSES
self.anchor_ratios = ASPECT_RATIOS
self.backbone = ResNet50()
self.learnable_factor = self.add_weight(
shape=(1, 1, 1, 512),
dtype=tf.float32,
initializer=tf.keras.initializers.Ones(),
trainable=True,
name='fckr')
# self.conv1 = tf.keras.layers.Conv2D(filters=1024, kernel_size=(1, 1), strides=1, padding="same")
self.conv2_1 = tf.keras.layers.Conv2D(filters=256,
kernel_size=(1, 1),
strides=1,
padding="same")
self.conv2_2 = tf.keras.layers.Conv2D(filters=512,
kernel_size=(3, 3),
strides=2,
padding="same")
self.conv3_1 = tf.keras.layers.Conv2D(filters=128,
kernel_size=(1, 1),
strides=1,
padding="same")
self.conv3_2 = tf.keras.layers.Conv2D(filters=256,
kernel_size=(3, 3),
strides=2,
padding="same")
self.conv4_1 = tf.keras.layers.Conv2D(filters=128,
kernel_size=(1, 1),
strides=1,
padding="same")
self.conv4_2 = tf.keras.layers.Conv2D(filters=256,
kernel_size=(3, 3),
strides=2,
padding="same")
self.pool = tf.keras.layers.GlobalAveragePooling2D()
self.predict_1 = self._predict_layer(k=self._get_k(i=0))
self.predict_2 = self._predict_layer(k=self._get_k(i=1))
self.predict_3 = self._predict_layer(k=self._get_k(i=2))
self.predict_4 = self._predict_layer(k=self._get_k(i=3))
self.predict_5 = self._predict_layer(k=self._get_k(i=4))
self.predict_6 = self._predict_layer(k=self._get_k(i=5))
def __init__(self):
super(SSD, self).__init__()
self.num_classes = NUM_CLASSES + 1
self.anchor_ratios = ASPECT_RATIOS
self.backbone = ResNet50()
self.conv1 = tf.keras.layers.Conv2D(filters=1024,
kernel_size=(1, 1),
strides=1,
padding="same")
self.conv2_1 = tf.keras.layers.Conv2D(filters=256,
kernel_size=(1, 1),
strides=1,
padding="same")
self.conv2_2 = tf.keras.layers.Conv2D(filters=512,
kernel_size=(3, 3),
strides=2,
padding="same")
self.conv3_1 = tf.keras.layers.Conv2D(filters=128,
kernel_size=(1, 1),
strides=1,
padding="same")
self.conv3_2 = tf.keras.layers.Conv2D(filters=256,
kernel_size=(3, 3),
strides=2,
padding="same")
self.conv4_1 = tf.keras.layers.Conv2D(filters=128,
kernel_size=(1, 1),
strides=1,
padding="same")
self.conv4_2 = tf.keras.layers.Conv2D(filters=256,
kernel_size=(3, 3),
strides=2,
padding="same")
self.pool = tf.keras.layers.GlobalAveragePooling2D()
self.predict_1 = self._predict_layer(k=self._get_k(i=0))
self.predict_2 = self._predict_layer(k=self._get_k(i=1))
self.predict_3 = self._predict_layer(k=self._get_k(i=2))
self.predict_4 = self._predict_layer(k=self._get_k(i=3))
self.predict_5 = self._predict_layer(k=self._get_k(i=4))
self.predict_6 = self._predict_layer(k=self._get_k(i=5))
def __init__(self, args, num_classes=21, norm_layer=nn.BatchNorm2d, input_channels=3):
super(DeepLabv3_plus, self).__init__()
self.args = args
if args.model == DEEPLAB:
self.backbone = ResNet101(args.output_stride, norm_layer=norm_layer, pretrained=args.pretrained, input_channels=input_channels)
self.aspp_inplanes = 2048
self.decoder_inplanes = 256
if self.args.refine_network:
self.refine_backbone = ResNet101(args.output_stride, norm_layer=norm_layer, pretrained=args.pretrained, input_channels=input_channels + num_classes)
elif args.model == DEEPLAB_50:
self.backbone = ResNet50(args.output_stride, norm_layer=norm_layer, pretrained=args.pretrained)
self.aspp_inplanes = 2048
self.decoder_inplanes = 256
if self.args.refine_network:
self.refine_backbone = ResNet50(args.output_stride, norm_layer=norm_layer, pretrained=args.pretrained, input_channels=input_channels + num_classes)
elif args.model == DEEPLAB_34:
self.backbone = ResNet34(args.output_stride, norm_layer=norm_layer, pretrained=args.pretrained)
self.aspp_inplanes = 2048
self.decoder_inplanes = 256
if self.args.refine_network:
self.refine_backbone = ResNet34(args.output_stride, norm_layer=norm_layer, pretrained=args.pretrained, input_channels=input_channels + num_classes)
elif args.model == DEEPLAB_18:
self.backbone = ResNet18(args.output_stride, norm_layer=norm_layer, pretrained=args.pretrained)
self.aspp_inplanes = 2048
self.decoder_inplanes = 256
if self.args.refine_network:
self.refine_backbone = ResNet18(args.output_stride, norm_layer=norm_layer, pretrained=args.pretrained, input_channels=input_channels + num_classes)
elif args.model == DEEPLAB_MOBILENET:
self.backbone = MobileNet_v2(pretrained=args.pretrained, first_layer_input_channels=input_channels)
self.aspp_inplanes = 320
self.decoder_inplanes = 24
if self.args.refine_network:
self.refine_backbone = MobileNet_v2(pretrained=args.pretrained, first_layer_input_channels=input_channels + num_classes)
elif args.model == DEEPLAB_MOBILENET_DILATION:
self.backbone = MobileNet_v2_dilation(pretrained=args.pretrained, first_layer_input_channels=input_channels)
self.aspp_inplanes = 320
if self.args.refine_network:
self.refine_backbone = MobileNet_v2_dilation(pretrained=args.pretrained, first_layer_input_channels=input_channels + num_classes)
self.decoder_inplanes = 24
else:
raise NotImplementedError
if self.args.use_aspp:
self.aspp = ASPP(args.output_stride, norm_layer=norm_layer, inplanes=self.aspp_inplanes)
aspp_outplanes = 256 if self.args.use_aspp else self.aspp_inplanes
self.decoder = Decoder(num_classes, norm_layer=norm_layer, inplanes=self.decoder_inplanes, aspp_outplanes=aspp_outplanes)
if self.args.learned_upsampling:
self.learned_upsampling = nn.Sequential(nn.ConvTranspose2d(num_classes, num_classes, kernel_size=4, stride=2, padding=1),
nn.ConvTranspose2d(num_classes, num_classes, kernel_size=4, stride=2, padding=1))
if self.args.refine_network:
if self.args.use_aspp:
self.refine_aspp = ASPP(args.output_stride, norm_layer=norm_layer, inplanes=self.aspp_inplanes)
self.refine_decoder = Decoder(num_classes, norm_layer=norm_layer, inplanes=self.decoder_inplanes, aspp_outplanes=aspp_outplanes)
if self.args.learned_upsampling:
self.refine_learned_upsampling = nn.Sequential(nn.ConvTranspose2d(num_classes, num_classes, kernel_size=4, stride=2, padding=1),
nn.ConvTranspose2d(num_classes, num_classes, kernel_size=4, stride=2, padding=1))