def yolo3lite_spp_shufflenetv2_body(inputs, num_anchors, num_classes):
'''Create YOLO_v3 Lite SPP ShuffleNetV2 model CNN body in keras.'''
shufflenetv2 = ShuffleNetV2(input_tensor=inputs, weights=None, include_top=False)
# input: 416 x 416 x 3
# 1x1conv5_out: 13 x 13 x 1024
# stage4/block1/relu_1x1conv_1: 26 x 26 x 464
# stage3/block1/relu_1x1conv_1: 52 x 52 x 232
f1 = shufflenetv2.get_layer('1x1conv5_out').output
# f1 :13 x 13 x 1024
#x, y1 = make_depthwise_separable_last_layers(f1, 464, num_anchors * (num_classes + 5), block_id_str='14')
x, y1 = make_spp_depthwise_separable_last_layers(f1, 464, num_anchors * (num_classes + 5), block_id_str='17')
x = compose(
DarknetConv2D_BN_Leaky(232, (1,1)),
UpSampling2D(2))(x)
f2 = shufflenetv2.get_layer('stage4/block1/relu_1x1conv_1').output
# f2: 26 x 26 x 464
x = Concatenate()([x,f2])
x, y2 = make_depthwise_separable_last_layers(x, 232, num_anchors * (num_classes + 5), block_id_str='18')
x = compose(
DarknetConv2D_BN_Leaky(116, (1,1)),
UpSampling2D(2))(x)
f3 = shufflenetv2.get_layer('stage3/block1/relu_1x1conv_1').output
# f3 : 52 x 52 x 232
x = Concatenate()([x, f3])
x, y3 = make_depthwise_separable_last_layers(x, 116, num_anchors * (num_classes + 5), block_id_str='19')
return Model(inputs = inputs, outputs=[y1,y2,y3])
def yolo3lite_xception_body(inputs, num_anchors, num_classes):
'''Create YOLO_v3 Lite Xception model CNN body in keras.'''
xception = Xception(input_tensor=inputs,
weights='imagenet',
include_top=False)
# input: 416 x 416 x 3
# block14_sepconv2_act: 13 x 13 x 2048
# block13_sepconv2_bn(middle in block13): 26 x 26 x 1024
# add_46(end of block12): 26 x 26 x 728
# block4_sepconv2_bn(middle in block4) : 52 x 52 x 728
# add_37(end of block3) : 52 x 52 x 256
# f1: 13 x 13 x 2048
f1 = xception.get_layer('block14_sepconv2_act').output
# f2: 26 x 26 x 1024
f2 = xception.get_layer('block13_sepconv2_bn').output
# f3: 52 x 52 x 728
f3 = xception.get_layer('block4_sepconv2_bn').output
#f1_channel_num = 2048
#f2_channel_num = 1024
#f3_channel_num = 728
f1_channel_num = 1024
f2_channel_num = 512
f3_channel_num = 256
#feature map 1 head & output (13x13 for 416 input)
x, y1 = make_depthwise_separable_last_layers(f1,
f1_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='14')
#upsample fpn merge for feature map 1 & 2
x = compose(DarknetConv2D_BN_Leaky(f2_channel_num // 2, (1, 1)),
UpSampling2D(2))(x)
x = Concatenate()([x, f2])
#feature map 2 head & output (26x26 for 416 input)
x, y2 = make_depthwise_separable_last_layers(x,
f2_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='15')
#upsample fpn merge for feature map 2 & 3
x = compose(DarknetConv2D_BN_Leaky(f3_channel_num // 2, (1, 1)),
UpSampling2D(2))(x)
x = Concatenate()([x, f3])
#feature map 3 head & output (52x52 for 416 input)
x, y3 = make_depthwise_separable_last_layers(x,
f3_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='16')
return Model(inputs=inputs, outputs=[y1, y2, y3])
def yolo3lite_spp_shufflenetv2_body(inputs, num_anchors, num_classes):
'''Create YOLO_v3 Lite SPP ShuffleNetV2 model CNN body in keras.'''
shufflenetv2 = ShuffleNetV2(input_tensor=inputs,
weights=None,
include_top=False)
# input: 416 x 416 x 3
# 1x1conv5_out: 13 x 13 x 1024
# stage4/block1/relu_1x1conv_1: 26 x 26 x 464
# stage3/block1/relu_1x1conv_1: 52 x 52 x 232
# f1: 13 x 13 x 1024
f1 = shufflenetv2.get_layer('1x1conv5_out').output
# f2: 26 x 26 x 464
f2 = shufflenetv2.get_layer('stage4/block1/relu_1x1conv_1').output
# f3: 52 x 52 x 232
f3 = shufflenetv2.get_layer('stage3/block1/relu_1x1conv_1').output
f1_channel_num = 1024
f2_channel_num = 464
f3_channel_num = 232
#f1_channel_num = 1024
#f2_channel_num = 512
#f3_channel_num = 256
#feature map 1 head & output (13x13 for 416 input)
#x, y1 = make_depthwise_separable_last_layers(f1, f1_channel_num//2, num_anchors * (num_classes + 5), block_id_str='14')
x, y1 = make_spp_depthwise_separable_last_layers(f1,
f1_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='17')
#upsample fpn merge for feature map 1 & 2
x = compose(DarknetConv2D_BN_Leaky(f2_channel_num // 2, (1, 1)),
UpSampling2D(2))(x)
x = Concatenate()([x, f2])
#feature map 2 head & output (26x26 for 416 input)
x, y2 = make_depthwise_separable_last_layers(x,
f2_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='18')
#upsample fpn merge for feature map 2 & 3
x = compose(DarknetConv2D_BN_Leaky(f3_channel_num // 2, (1, 1)),
UpSampling2D(2))(x)
x = Concatenate()([x, f3])
#feature map 3 head & output (52x52 for 416 input)
x, y3 = make_depthwise_separable_last_layers(x,
f3_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='19')
return Model(inputs=inputs, outputs=[y1, y2, y3])
def yolo3lite_spp_efficientnet_body(inputs, num_anchors, num_classes, level=3):
'''
Create YOLO_v3 Lite SPP EfficientNet model CNN body in keras.
# Arguments
level: EfficientNet level number.
by default we use basic EfficientNetB3 as backbone
'''
efficientnet, feature_map_info = get_efficientnet_backbone_info(
inputs, level=level)
f1 = efficientnet.get_layer('top_activation').output
f1_channel_num = feature_map_info['f1_channel_num']
f2 = efficientnet.get_layer('block6a_expand_activation').output
f2_channel_num = feature_map_info['f2_channel_num']
f3 = efficientnet.get_layer('block4a_expand_activation').output
f3_channel_num = feature_map_info['f3_channel_num']
#feature map 1 head & output (19x19 for 608 input)
#x, y1 = make_spp_depthwise_separable_last_layers(f1, 672, num_anchors * (num_classes + 5), block_id_str='8')
x, y1 = make_spp_depthwise_separable_last_layers(f1,
f1_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='8')
#upsample fpn merge for feature map 1 & 2
x = compose(
#DarknetConv2D_BN_Leaky(336, (1,1)),
DarknetConv2D_BN_Leaky(f2_channel_num // 2, (1, 1)),
UpSampling2D(2))(x)
x = Concatenate()([x, f2])
#feature map 2 head & output (38x38 for 608 input)
#x, y2 = make_depthwise_separable_last_layers(x, 240, num_anchors*(num_classes+5), block_id_str='9')
x, y2 = make_depthwise_separable_last_layers(x,
f2_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='9')
#upsample fpn merge for feature map 2 & 3
x = compose(
#DarknetConv2D_BN_Leaky(120, (1,1)),
DarknetConv2D_BN_Leaky(f3_channel_num // 2, (1, 1)),
UpSampling2D(2))(x)
x = Concatenate()([x, f3])
#feature map 3 head & output (76x76 for 608 input)
#x, y3 = make_depthwise_separable_last_layers(x, 120, num_anchors*(num_classes+5), block_id_str='10')
x, y3 = make_depthwise_separable_last_layers(x,
f3_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='10')
return Model(inputs=inputs, outputs=[y1, y2, y3])
def yolo3lite_efficientnet_body(inputs, num_anchors, num_classes, level=0):
'''
Create YOLO_v3 Lite EfficientNet model CNN body in keras.
# Arguments
level: EfficientNet level number.
by default we use basic EfficientNetB0 as backbone
'''
efficientnet, feature_map_info = get_efficientnet_backbone_info(
inputs, level=level)
f1 = efficientnet.get_layer('top_activation').output
#x, y1 = make_depthwise_separable_last_layers(f1, 672, num_anchors * (num_classes + 5), block_id_str='8')
f2_channel_num = feature_map_info['f2_channel_num']
x, y1 = make_depthwise_separable_last_layers(f1,
f2_channel_num,
num_anchors *
(num_classes + 5),
block_id_str='8')
x = compose(
#DarknetConv2D_BN_Leaky(336, (1,1)),
DarknetConv2D_BN_Leaky(f2_channel_num // 2, (1, 1)),
UpSampling2D(2))(x)
f2 = efficientnet.get_layer('block6a_expand_activation').output
x = Concatenate()([x, f2])
#x, y2 = make_depthwise_separable_last_layers(x, 240, num_anchors*(num_classes+5), block_id_str='9')
f3_channel_num = feature_map_info['f3_channel_num']
x, y2 = make_depthwise_separable_last_layers(x,
f3_channel_num,
num_anchors *
(num_classes + 5),
block_id_str='9')
x = compose(
#DarknetConv2D_BN_Leaky(120, (1,1)),
DarknetConv2D_BN_Leaky(f3_channel_num // 2, (1, 1)),
UpSampling2D(2))(x)
f3 = efficientnet.get_layer('block4a_expand_activation').output
x = Concatenate()([x, f3])
#x, y3 = make_depthwise_separable_last_layers(x, 120, num_anchors*(num_classes+5), block_id_str='10')
x, y3 = make_depthwise_separable_last_layers(x,
f3_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='10')
return Model(inputs=inputs, outputs=[y1, y2, y3])
def yolo3lite_spp_mobilenetv2_body(inputs,
num_anchors,
num_classes,
alpha=1.0):
'''Create YOLO_v3 Lite SPP MobileNetV2 model CNN body in keras.'''
mobilenetv2 = MobileNetV2(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
# input: 416 x 416 x 3
# out_relu: 13 x 13 x 1280
# block_13_expand_relu: 26 x 26 x (576*alpha)
# block_6_expand_relu: 52 x 52 x (192*alpha)
f1 = mobilenetv2.get_layer('out_relu').output
# f1 :13 x 13 x 1280
x, y1 = make_spp_depthwise_separable_last_layers(
f1,
int(576 * alpha),
num_anchors * (num_classes + 5),
block_id_str='17',
predict_filters=int(1024 * alpha))
x = compose(DarknetConv2D_BN_Leaky(int(288 * alpha), (1, 1)),
UpSampling2D(2))(x)
f2 = mobilenetv2.get_layer('block_13_expand_relu').output
# f2: 26 x 26 x (576*alpha)
x = Concatenate()([x, f2])
x, y2 = make_depthwise_separable_last_layers(
x,
int(192 * alpha),
num_anchors * (num_classes + 5),
block_id_str='18',
predict_filters=int(512 * alpha))
x = compose(DarknetConv2D_BN_Leaky(int(96 * alpha), (1, 1)),
UpSampling2D(2))(x)
f3 = mobilenetv2.get_layer('block_6_expand_relu').output
# f3 : 52 x 52 x (192*alpha)
x = Concatenate()([x, f3])
x, y3 = make_depthwise_separable_last_layers(
x,
int(96 * alpha),
num_anchors * (num_classes + 5),
block_id_str='19',
predict_filters=int(256 * alpha))
return Model(inputs=inputs, outputs=[y1, y2, y3])
def yolo3lite_mobilenetv3small_body(inputs,
num_anchors,
num_classes,
alpha=1.0):
'''Create YOLO_v3 Lite MobileNetV3Small model CNN body in keras.'''
mobilenetv3small = MobileNetV3Small(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
# input: 416 x 416 x 3
# activation_31(layer 165, final feature map): 13 x 13 x (576*alpha)
# expanded_conv_10/Add(layer 162, end of block10): 13 x 13 x (96*alpha)
# activation_22(layer 117, middle in block8) : 26 x 26 x (288*alpha)
# expanded_conv_7/Add(layer 114, end of block7) : 26 x 26 x (48*alpha)
# activation_7(layer 38, middle in block3) : 52 x 52 x (96*alpha)
# expanded_conv_2/Add(layer 35, end of block2): 52 x 52 x (24*alpha)
# NOTE: activation layer name may different for TF1.x/2.x, so we
# use index to fetch layer
f1 = mobilenetv3small.layers[165].output
# f1 :13 x 13 x (576*alpha)
x, y1 = make_depthwise_separable_last_layers(
f1, int(288 * alpha), num_anchors * (num_classes + 5))
#x, y1 = make_depthwise_separable_last_layers(f1, int(288*alpha), num_anchors * (num_classes + 5), predict_filters=int(1024*alpha))
x = compose(DarknetConv2D_BN_Leaky(int(144 * alpha), (1, 1)),
UpSampling2D(2))(x)
f2 = mobilenetv3small.layers[117].output
# f2: 26 x 26 x (288*alpha)
x = Concatenate()([x, f2])
x, y2 = make_depthwise_separable_last_layers(
x, int(96 * alpha), num_anchors * (num_classes + 5))
#x, y2 = make_depthwise_separable_last_layers(x, int(96*alpha), num_anchors*(num_classes+5), predict_filters=int(512*alpha))
x = compose(DarknetConv2D_BN_Leaky(int(48 * alpha), (1, 1)),
UpSampling2D(2))(x)
f3 = mobilenetv3small.layers[38].output
# f3 : 52 x 52 x (96*alpha)
x = Concatenate()([x, f3])
x, y3 = make_depthwise_separable_last_layers(
x, int(48 * alpha), num_anchors * (num_classes + 5))
#x, y3 = make_depthwise_separable_last_layers(x, int(48*alpha), num_anchors*(num_classes+5), predict_filters=int(256*alpha))
return Model(inputs=inputs, outputs=[y1, y2, y3])
def yolo3lite_spp_mobilenet_body(inputs, num_anchors, num_classes, alpha=1.0):
'''Create YOLO_v3 Lite SPP MobileNet model CNN body in keras.'''
mobilenet = MobileNet(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
# input: 416 x 416 x 3
# conv_pw_13_relu :13 x 13 x (1024*alpha)
# conv_pw_11_relu :26 x 26 x (512*alpha)
# conv_pw_5_relu : 52 x 52 x (256*alpha)
f1 = mobilenet.get_layer('conv_pw_13_relu').output
# f1 :13 x 13 x (1024*alpha)
#x, y1 = make_depthwise_separable_last_layers(f1, int(512*alpha), num_anchors * (num_classes + 5), block_id_str='14')
x, y1 = make_spp_depthwise_separable_last_layers(f1,
int(512 * alpha),
num_anchors *
(num_classes + 5),
block_id_str='14')
x = compose(DarknetConv2D_BN_Leaky(int(256 * alpha), (1, 1)),
UpSampling2D(2))(x)
f2 = mobilenet.get_layer('conv_pw_11_relu').output
# f2: 26 x 26 x (512*alpha)
x = Concatenate()([x, f2])
x, y2 = make_depthwise_separable_last_layers(x,
int(256 * alpha),
num_anchors *
(num_classes + 5),
block_id_str='15')
x = compose(DarknetConv2D_BN_Leaky(int(128 * alpha), (1, 1)),
UpSampling2D(2))(x)
f3 = mobilenet.get_layer('conv_pw_5_relu').output
# f3 : 52 x 52 x (256*alpha)
x = Concatenate()([x, f3])
x, y3 = make_depthwise_separable_last_layers(x,
int(128 * alpha),
num_anchors *
(num_classes + 5),
block_id_str='16')
return Model(inputs=inputs, outputs=[y1, y2, y3])
def yolo3lite_body(inputs, num_anchors, num_classes):
"""Create YOLO_V3 Lite model CNN body in Keras."""
darknetlite = Model(inputs, darknet53lite_body(inputs))
x, y1 = make_depthwise_separable_last_layers(
darknetlite.output, 512, num_anchors * (num_classes + 5))
x = compose(DarknetConv2D_BN_Leaky(256, (1, 1)), UpSampling2D(2))(x)
x = Concatenate()([x, darknetlite.layers[152].output])
x, y2 = make_depthwise_separable_last_layers(
x, 256, num_anchors * (num_classes + 5))
x = compose(DarknetConv2D_BN_Leaky(128, (1, 1)), UpSampling2D(2))(x)
x = Concatenate()([x, darknetlite.layers[92].output])
x, y3 = make_depthwise_separable_last_layers(
x, 128, num_anchors * (num_classes + 5))
return Model(inputs, [y1, y2, y3])
def yolo3lite_xception_body(inputs, num_anchors, num_classes):
'''Create YOLO_v3 Lite Xception model CNN body in keras.'''
xception = Xception(input_tensor=inputs,
weights='imagenet',
include_top=False)
# input: 416 x 416 x 3
# block14_sepconv2_act: 13 x 13 x 2048
# block13_sepconv2_bn(middle in block13): 26 x 26 x 1024
# add_46(end of block12): 26 x 26 x 728
# block4_sepconv2_bn(middle in block4) : 52 x 52 x 728
# add_37(end of block3) : 52 x 52 x 256
f1 = xception.get_layer('block14_sepconv2_act').output
# f1 :13 x 13 x 2048
x, y1 = make_depthwise_separable_last_layers(f1,
1024,
num_anchors *
(num_classes + 5),
block_id_str='14')
x = compose(DarknetConv2D_BN_Leaky(512, (1, 1)), UpSampling2D(2))(x)
f2 = xception.get_layer('block13_sepconv2_bn').output
# f2: 26 x 26 x 1024
x = Concatenate()([x, f2])
x, y2 = make_depthwise_separable_last_layers(x,
512,
num_anchors *
(num_classes + 5),
block_id_str='15')
x = compose(DarknetConv2D_BN_Leaky(256, (1, 1)), UpSampling2D(2))(x)
f3 = xception.get_layer('block4_sepconv2_bn').output
# f3 : 52 x 52 x 728
x = Concatenate()([x, f3])
x, y3 = make_depthwise_separable_last_layers(x,
256,
num_anchors *
(num_classes + 5),
block_id_str='16')
return Model(inputs=inputs, outputs=[y1, y2, y3])
def yolo3lite_mobilenetv3large_body(inputs, num_anchors, num_classes, alpha=1.0):
'''Create YOLO_v3 Lite MobileNetV3Large model CNN body in keras.'''
mobilenetv3large = MobileNetV3Large(input_tensor=inputs, weights='imagenet', include_top=False, alpha=alpha)
# input: 416 x 416 x 3
# activation_38(layer 194, final feature map): 13 x 13 x (960*alpha)
# expanded_conv_14/Add(layer 191, end of block14): 13 x 13 x (160*alpha)
# activation_29(layer 146, middle in block12) : 26 x 26 x (672*alpha)
# expanded_conv_11/Add(layer 143, end of block11) : 26 x 26 x (112*alpha)
# activation_15(layer 79, middle in block6) : 52 x 52 x (240*alpha)
# expanded_conv_5/Add(layer 76, end of block5): 52 x 52 x (40*alpha)
# NOTE: activation layer name may different for TF1.x/2.x, so we
# use index to fetch layer
f1 = mobilenetv3large.layers[194].output
# f1 :13 x 13 x (960*alpha)
x, y1 = make_depthwise_separable_last_layers(f1, int(672*alpha), num_anchors * (num_classes + 5), block_id_str='15')
#x, y1 = make_depthwise_separable_last_layers(f1, int(672*alpha), num_anchors * (num_classes + 5), block_id_str='15', predict_filters=int(1024*alpha))
x = compose(
DarknetConv2D_BN_Leaky(int(336*alpha), (1,1)),
UpSampling2D(2))(x)
f2 = mobilenetv3large.layers[146].output
# f2: 26 x 26 x (672*alpha)
x = Concatenate()([x,f2])
x, y2 = make_depthwise_separable_last_layers(x, int(240*alpha), num_anchors*(num_classes+5), block_id_str='16')
#x, y2 = make_depthwise_separable_last_layers(x, int(240*alpha), num_anchors*(num_classes+5), block_id_str='16', predict_filters=int(512*alpha))
x = compose(
DarknetConv2D_BN_Leaky(int(120*alpha), (1,1)),
UpSampling2D(2))(x)
f3 = mobilenetv3large.layers[79].output
# f3 : 52 x 52 x (240*alpha)
x = Concatenate()([x, f3])
x, y3 = make_depthwise_separable_last_layers(x, int(120*alpha), num_anchors*(num_classes+5), block_id_str='17')
#x, y3 = make_depthwise_separable_last_layers(x, int(120*alpha), num_anchors*(num_classes+5), block_id_str='17', predict_filters=int(256*alpha))
return Model(inputs = inputs, outputs=[y1,y2,y3])
def yolo3lite_body(inputs, num_anchors, num_classes):
"""Create YOLO_V3 Lite model CNN body in Keras."""
darknetlite = Model(inputs, darknet53lite_body(inputs))
# f1: 13 x 13 x 1024
f1 = darknetlite.output
# f2: 26 x 26 x 512
f2 = darknetlite.layers[152].output
# f3: 52 x 52 x 256
f3 = darknetlite.layers[92].output
f1_channel_num = 1024
f2_channel_num = 512
f3_channel_num = 256
#feature map 1 head & output (19x19 for 608 input)
x, y1 = make_depthwise_separable_last_layers(
f1, f1_channel_num // 2, num_anchors * (num_classes + 5))
#upsample fpn merge for feature map 1 & 2
x = compose(DarknetConv2D_BN_Leaky(f2_channel_num // 2, (1, 1)),
UpSampling2D(2))(x)
x = Concatenate()([x, f2])
#feature map 2 head & output (38x38 for 608 input)
x, y2 = make_depthwise_separable_last_layers(
x, f2_channel_num // 2, num_anchors * (num_classes + 5))
#upsample fpn merge for feature map 2 & 3
x = compose(DarknetConv2D_BN_Leaky(f3_channel_num // 2, (1, 1)),
UpSampling2D(2))(x)
x = Concatenate()([x, f3])
#feature map 3 head & output (76x76 for 608 input)
x, y3 = make_depthwise_separable_last_layers(
x, f3_channel_num // 2, num_anchors * (num_classes + 5))
return Model(inputs, [y1, y2, y3])
def yolo3lite_mobilenetv2_body(inputs, num_anchors, num_classes, alpha=1.0):
'''Create YOLO_v3 Lite MobileNetV2 model CNN body in keras.'''
mobilenetv2 = MobileNetV2(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
# input: 416 x 416 x 3
# out_relu: 13 x 13 x 1280
# block_13_expand_relu: 26 x 26 x (576*alpha)
# block_6_expand_relu: 52 x 52 x (192*alpha)
# f1 :13 x 13 x 1280
f1 = mobilenetv2.get_layer('out_relu').output
# f2: 26 x 26 x (576*alpha)
f2 = mobilenetv2.get_layer('block_13_expand_relu').output
# f3 : 52 x 52 x (192*alpha)
f3 = mobilenetv2.get_layer('block_6_expand_relu').output
f1_channel_num = int(1280 * alpha)
f2_channel_num = int(576 * alpha)
f3_channel_num = int(192 * alpha)
#f1_channel_num = 1024
#f2_channel_num = 512
#f3_channel_num = 256
#feature map 1 head & output (13x13 for 416 input)
x, y1 = make_depthwise_separable_last_layers(f1,
f1_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='17')
#x, y1 = make_depthwise_separable_last_layers(f1, f1_channel_num//2, num_anchors * (num_classes + 5), block_id_str='17', predict_filters=int(1024*alpha))
#upsample fpn merge for feature map 1 & 2
x = compose(DarknetConv2D_BN_Leaky(f2_channel_num // 2, (1, 1)),
UpSampling2D(2))(x)
x = Concatenate()([x, f2])
#feature map 2 head & output (26x26 for 416 input)
x, y2 = make_depthwise_separable_last_layers(x,
f2_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='18')
#x, y2 = make_depthwise_separable_last_layers(x, f2_channel_num//2, num_anchors * (num_classes + 5), block_id_str='18', predict_filters=int(512*alpha))
#upsample fpn merge for feature map 2 & 3
x = compose(DarknetConv2D_BN_Leaky(f3_channel_num // 2, (1, 1)),
UpSampling2D(2))(x)
x = Concatenate()([x, f3])
#feature map 3 head & output (52x52 for 416 input)
x, y3 = make_depthwise_separable_last_layers(x,
f3_channel_num // 2,
num_anchors *
(num_classes + 5),
block_id_str='19')
#x, y3 = make_depthwise_separable_last_layers(x, f3_channel_num//2, num_anchors * (num_classes + 5), block_id_str='19', predict_filters=int(256*alpha))
return Model(inputs=inputs, outputs=[y1, y2, y3])
