def call(self, inputs, training=None, mask=None):
x = compose(self.conv1, self.normalize1, self.activity1)(inputs)
x = compose(self.conv2, self.normalize2, self.activity2)(x)
x = compose(self.conv3, self.normalize3)(x)
x = layers.add([x, inputs])
x = self.activity3(x)
return x
def call(self, inputs, training=None, mask=None):
x = compose(self.conv1, self.normalize1, self.activity1)(inputs)
x = compose(self.conv2, self.normalize2, self.activity2)(x)
x = compose(self.conv3, self.normalize3)(x)
y = compose(self.conv4, self.normalize4)(inputs)
z = layers.add([x, y])
z = self.activity4(z)
return z
def YOLOv4(input_layer, NUM_CLASS):
route_1, route_2, route_3 = cspdarknet53(input_layer)
conv = YoloConv(filters=256, kernel_size=1)(route_3)
conv = up_sample(conv)
route_2 = YoloConv(filters=256, kernel_size=1)(route_2)
# conv = tf.concat([route_2, conv], axis=-1)
# conv = compose(YoloConv(filters=256, kernel_size=1),
# YoloConv(filters=512, kernel_size=3),
# YoloConv(filters=256, kernel_size=1),
# YoloConv(filters=512, kernel_size=3),
# YoloConv(filters=256, kernel_size=1))(conv)
conv = YoloNeck(route_2, conv, filters=256)
route_tmp = conv
route_1 = YoloConv(filters=128, kernel_size=1)(route_1)
conv = YoloConv(filters=128, kernel_size=1)(conv)
conv = up_sample(conv)
conv = YoloNeck(route_1, conv, filters=128)
conv_sbbox = compose(
YoloConv(filters=256, kernel_size=3),
YoloConv(filters=3 * (NUM_CLASS + 5),
kernel_size=1,
activate=False,
bn=False))(conv)
conv = YoloConv(filters=256, kernel_size=3, down_sample=True)(conv)
conv = YoloNeck(conv, route_tmp, filters=256)
conv_mbbox = compose(
YoloConv(filters=512, kernel_size=3),
YoloConv(filters=3 * (NUM_CLASS + 5),
kernel_size=1,
activate=False,
bn=False))(conv)
conv = YoloConv(filters=512, kernel_size=3, down_sample=True)(conv)
conv = YoloNeck(conv, route_3, filters=512)
conv_lbbox = compose(
YoloConv(filters=1024, kernel_size=3),
YoloConv(filters=3 * (NUM_CLASS + 5),
kernel_size=1,
activate=False,
bn=False))(conv)
return [conv_sbbox, conv_mbbox, conv_lbbox]
def call(self, inputs, **kwargs):
output = compose(*self.compose_list)(inputs)
if self.activate:
if self.activate_type == "leaky":
output = tf.nn.leaky_relu(output, alpha=0.1)
elif self.activate_type == "mish":
output = mish(output)
return output
def YoloNeck(conv, route_tmp, filters):
conv = tf.concat([conv, route_tmp], axis=-1)
conv = compose(YoloConv(filters=filters, kernel_size=1),
YoloConv(filters=filters * 2, kernel_size=3),
YoloConv(filters=filters, kernel_size=1),
YoloConv(filters=filters * 2, kernel_size=3),
YoloConv(filters=filters, kernel_size=1))(conv)
return conv
def call(self, inputs, **kwargs):
out0 = self.layer0(inputs)
out1 = self.block1(out0)
out2 = self.block2(out1)
out3 = self.block3(out2)
out4 = self.block4(out3)
out5 = self.block5(out4)
x = self.make_list[0][0](out5)
y1 = compose(*self.make_list[0][1:])(x)
x = compose(*self.conv_samp[0])(x)
x = self.concat[0]([x, out4])
x = self.make_list[1][0](x)
y2 = compose(*self.make_list[1][1:])(x)
x = compose(*self.conv_samp[1])(x)
x = self.concat[1]([x, out3])
x = self.make_list[2][0](x)
y3 = compose(*self.make_list[2][1:])(x)
return y1, y2, y3
def call(self, inputs, **kwargs):
residual_output = inputs + compose(*self.compose_list)(inputs)
return residual_output
def call(self, inputs, **kwargs):
return compose(*self.conv_list)(inputs)
def call(self, inputs, **kwargs):
output = compose(*self.compose_list)(inputs)
return output
def call(self, inputs, **kwargs):
x = compose(self.zero, self.conv)(inputs)
for conv1, conv2, add in self.block_layers:
y = compose(conv1, conv2)(x)
x = add([x, y])
return x
