def yolo4_body(inputs, num_anchors, num_classes, weights_path=None):
"""Create YOLO_V4 model CNN body in Keras."""
darknet = Model(inputs, csp_darknet53_body(inputs))
if weights_path is not None:
darknet.load_weights(weights_path, by_name=True)
print('Load weights {}.'.format(weights_path))
#feature map 1 head (19x19 for 608 input)
x1 = make_yolo_spp_head(darknet.output, 512)
#upsample fpn merge for feature map 1 & 2
x1_upsample = compose(DarknetConv2D_BN_Leaky(256, (1, 1)),
UpSampling2D(2))(x1)
x2 = DarknetConv2D_BN_Leaky(256, (1, 1))(darknet.layers[204].output)
x2 = Concatenate()([x2, x1_upsample])
#feature map 2 head (38x38 for 608 input)
x2 = make_yolo_head(x2, 256)
#upsample fpn merge for feature map 2 & 3
x2_upsample = compose(DarknetConv2D_BN_Leaky(128, (1, 1)),
UpSampling2D(2))(x2)
x3 = DarknetConv2D_BN_Leaky(128, (1, 1))(darknet.layers[131].output)
x3 = Concatenate()([x3, x2_upsample])
#feature map 3 head & output (76x76 for 608 input)
#x3, y3 = make_last_layers(x3, 128, num_anchors*(num_classes+5))
x3 = make_yolo_head(x3, 128)
y3 = compose(DarknetConv2D_BN_Leaky(256, (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x3)
#downsample fpn merge for feature map 3 & 2
x3_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
DarknetConv2D_BN_Leaky(256, (3, 3), strides=(2, 2)))(x3)
x2 = Concatenate()([x3_downsample, x2])
#feature map 2 output (38x38 for 608 input)
#x2, y2 = make_last_layers(x2, 256, num_anchors*(num_classes+5))
x2 = make_yolo_head(x2, 256)
y2 = compose(DarknetConv2D_BN_Leaky(512, (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x2)
#downsample fpn merge for feature map 2 & 1
x2_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
DarknetConv2D_BN_Leaky(512, (3, 3), strides=(2, 2)))(x2)
x1 = Concatenate()([x2_downsample, x1])
#feature map 1 output (19x19 for 608 input)
#x1, y1 = make_last_layers(x1, 512, num_anchors*(num_classes+5))
x1 = make_yolo_head(x1, 512)
y1 = compose(DarknetConv2D_BN_Leaky(1024, (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x1)
return Model(inputs, [y1, y2, y3])
def yolo4_mobilenetv3large_body(inputs, num_anchors, num_classes, alpha=1.0):
"""Create YOLO_V4 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)
# f1 :13 x 13 x (960*alpha)
# NOTE: activation layer name may different for TF1.x/2.x, so we
# use index to fetch layer
f1 = mobilenetv3large.layers[194].output
#feature map 1 head (13 x 13 x (480*alpha) for 416 input)
x1 = make_yolo_spp_head(f1, int(480 * alpha))
#upsample fpn merge for feature map 1 & 2
x1_upsample = compose(DarknetConv2D_BN_Leaky(int(336 * alpha), (1, 1)),
UpSampling2D(2))(x1)
f2 = mobilenetv3large.layers[146].output
# f2: 26 x 26 x (672*alpha) for 416 input
x2 = DarknetConv2D_BN_Leaky(int(336 * alpha), (1, 1))(f2)
x2 = Concatenate()([x2, x1_upsample])
#feature map 2 head (26 x 26 x (336*alpha) for 416 input)
x2 = make_yolo_head(x2, int(336 * alpha))
#upsample fpn merge for feature map 2 & 3
x2_upsample = compose(DarknetConv2D_BN_Leaky(int(120 * alpha), (1, 1)),
UpSampling2D(2))(x2)
f3 = mobilenetv3large.layers[79].output
# f3 : 52 x 52 x (240*alpha) for 416 input
x3 = DarknetConv2D_BN_Leaky(int(120 * alpha), (1, 1))(f3)
x3 = Concatenate()([x3, x2_upsample])
#feature map 3 head & output (52 x 52 x (240*alpha) for 416 input)
#x3, y3 = make_last_layers(x3, int(120*alpha), num_anchors*(num_classes+5))
x3 = make_yolo_head(x3, int(120 * alpha))
y3 = compose(DarknetConv2D_BN_Leaky(int(240 * alpha), (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x3)
#downsample fpn merge for feature map 3 & 2
x3_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
DarknetConv2D_BN_Leaky(int(336 * alpha), (3, 3), strides=(2, 2)))(x3)
x2 = Concatenate()([x3_downsample, x2])
#feature map 2 output (26 x 26 x (672*alpha) for 416 input)
#x2, y2 = make_last_layers(x2, int(336*alpha), num_anchors*(num_classes+5))
x2 = make_yolo_head(x2, int(336 * alpha))
y2 = compose(DarknetConv2D_BN_Leaky(int(672 * alpha), (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x2)
#downsample fpn merge for feature map 2 & 1
x2_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
DarknetConv2D_BN_Leaky(int(480 * alpha), (3, 3), strides=(2, 2)))(x2)
x1 = Concatenate()([x2_downsample, x1])
#feature map 1 output (13 x 13 x (960*alpha) for 416 input)
#x1, y1 = make_last_layers(x1, int(480*alpha), num_anchors*(num_classes+5))
x1 = make_yolo_head(x1, int(480 * alpha))
y1 = compose(DarknetConv2D_BN_Leaky(int(960 * alpha), (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x1)
return Model(inputs, [y1, y2, y3])
def yolo4_mobilenet_body(inputs, num_anchors, num_classes, alpha=1.0):
"""Create YOLO_V4 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) for 416 input
#feature map 1 head (13 x 13 x (512*alpha) for 416 input)
x1 = make_yolo_spp_head(f1, int(512 * alpha))
#upsample fpn merge for feature map 1 & 2
x1_upsample = compose(DarknetConv2D_BN_Leaky(int(256 * alpha), (1, 1)),
UpSampling2D(2))(x1)
f2 = mobilenet.get_layer('conv_pw_11_relu').output
# f2: 26 x 26 x (512*alpha) for 416 input
x2 = DarknetConv2D_BN_Leaky(int(256 * alpha), (1, 1))(f2)
x2 = Concatenate()([x2, x1_upsample])
#feature map 2 head (26 x 26 x (256*alpha) for 416 input)
x2 = make_yolo_head(x2, int(256 * alpha))
#upsample fpn merge for feature map 2 & 3
x2_upsample = compose(DarknetConv2D_BN_Leaky(int(128 * alpha), (1, 1)),
UpSampling2D(2))(x2)
f3 = mobilenet.get_layer('conv_pw_5_relu').output
# f3 : 52 x 52 x (256*alpha) for 416 input
x3 = DarknetConv2D_BN_Leaky(int(128 * alpha), (1, 1))(f3)
x3 = Concatenate()([x3, x2_upsample])
#feature map 3 head & output (52 x 52 x (256*alpha) for 416 input)
#x3, y3 = make_last_layers(x3, int(128*alpha), num_anchors*(num_classes+5))
x3 = make_yolo_head(x3, int(128 * alpha))
y3 = compose(DarknetConv2D_BN_Leaky(int(256 * alpha), (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x3)
#downsample fpn merge for feature map 3 & 2
x3_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
DarknetConv2D_BN_Leaky(int(256 * alpha), (3, 3), strides=(2, 2)))(x3)
x2 = Concatenate()([x3_downsample, x2])
#feature map 2 output (26 x 26 x (512*alpha) for 416 input)
#x2, y2 = make_last_layers(x2, int(256*alpha), num_anchors*(num_classes+5))
x2 = make_yolo_head(x2, int(256 * alpha))
y2 = compose(DarknetConv2D_BN_Leaky(int(512 * alpha), (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x2)
#downsample fpn merge for feature map 2 & 1
x2_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
DarknetConv2D_BN_Leaky(int(512 * alpha), (3, 3), strides=(2, 2)))(x2)
x1 = Concatenate()([x2_downsample, x1])
#feature map 1 output (13 x 13 x (1024*alpha) for 416 input)
#x1, y1 = make_last_layers(x1, int(512*alpha), num_anchors*(num_classes+5))
x1 = make_yolo_head(x1, int(512 * alpha))
y1 = compose(DarknetConv2D_BN_Leaky(int(1024 * alpha), (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x1)
return Model(inputs, [y1, y2, y3])
def yolo4_efficientnet_body(inputs, num_anchors, num_classes, level=1):
'''
Create YOLO_v4 EfficientNet model CNN body in keras.
# Arguments
level: EfficientNet level number.
by default we use basic EfficientNetB1 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 (13x13x(f1_channel_num//2) for 416 input)
x1 = make_yolo_spp_head(f1, f1_channel_num // 2)
#upsample fpn merge for feature map 1 & 2
x1_upsample = compose(DarknetConv2D_BN_Leaky(f2_channel_num // 2, (1, 1)),
UpSampling2D(2))(x1)
x2 = DarknetConv2D_BN_Leaky(f2_channel_num // 2, (1, 1))(f2)
x2 = Concatenate()([x2, x1_upsample])
#feature map 2 head (26x26x(f2_channel_num//2) for 416 input)
x2 = make_yolo_head(x2, f2_channel_num // 2)
#upsample fpn merge for feature map 2 & 3
x2_upsample = compose(DarknetConv2D_BN_Leaky(f3_channel_num // 2, (1, 1)),
UpSampling2D(2))(x2)
x3 = DarknetConv2D_BN_Leaky(f3_channel_num // 2, (1, 1))(f3)
x3 = Concatenate()([x3, x2_upsample])
#feature map 3 head & output (52x52xf3_channel_num for 416 input)
#x3, y3 = make_last_layers(x3, f3_channel_num//2, num_anchors*(num_classes+5))
x3 = make_yolo_head(x3, f3_channel_num // 2)
y3 = compose(DarknetConv2D_BN_Leaky(f3_channel_num, (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x3)
#downsample fpn merge for feature map 3 & 2
x3_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
DarknetConv2D_BN_Leaky(f2_channel_num // 2, (3, 3),
strides=(2, 2)))(x3)
x2 = Concatenate()([x3_downsample, x2])
#feature map 2 output (26x26xf2_channel_num for 416 input)
#x2, y2 = make_last_layers(x2, f2_channel_num//2, num_anchors*(num_classes+5))
x2 = make_yolo_head(x2, f2_channel_num // 2)
y2 = compose(DarknetConv2D_BN_Leaky(f2_channel_num, (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x2)
#downsample fpn merge for feature map 2 & 1
x2_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
DarknetConv2D_BN_Leaky(f1_channel_num // 2, (3, 3),
strides=(2, 2)))(x2)
x1 = Concatenate()([x2_downsample, x1])
#feature map 1 output (13x13xf1_channel_num for 416 input)
#x1, y1 = make_last_layers(x1, f1_channel_num//2, num_anchors*(num_classes+5))
x1 = make_yolo_head(x1, f1_channel_num // 2)
y1 = compose(DarknetConv2D_BN_Leaky(f1_channel_num, (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x1)
return Model(inputs, [y1, y2, y3])
def yolo4_mobilenetv3small_body(inputs, num_anchors, num_classes, alpha=1.0):
"""Create YOLO_V4 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)
# f1 :13 x 13 x (576*alpha)
# NOTE: activation layer name may different for TF1.x/2.x, so we
# use index to fetch layer
f1 = mobilenetv3small.layers[165].output
#feature map 1 head (13 x 13 x (288*alpha) for 416 input)
x1 = make_yolo_spp_head(f1, int(288 * alpha))
#upsample fpn merge for feature map 1 & 2
x1_upsample = compose(DarknetConv2D_BN_Leaky(int(144 * alpha), (1, 1)),
UpSampling2D(2))(x1)
f2 = mobilenetv3small.layers[117].output
# f2: 26 x 26 x (288*alpha) for 416 input
x2 = DarknetConv2D_BN_Leaky(int(144 * alpha), (1, 1))(f2)
x2 = Concatenate()([x2, x1_upsample])
#feature map 2 head (26 x 26 x (144*alpha) for 416 input)
x2 = make_yolo_head(x2, int(144 * alpha))
#upsample fpn merge for feature map 2 & 3
x2_upsample = compose(DarknetConv2D_BN_Leaky(int(48 * alpha), (1, 1)),
UpSampling2D(2))(x2)
f3 = mobilenetv3small.layers[38].output
# f3 : 52 x 52 x (96*alpha)
x3 = DarknetConv2D_BN_Leaky(int(48 * alpha), (1, 1))(f3)
x3 = Concatenate()([x3, x2_upsample])
#feature map 3 head & output (52 x 52 x (96*alpha) for 416 input)
#x3, y3 = make_last_layers(x3, int(48*alpha), num_anchors*(num_classes+5))
x3 = make_yolo_head(x3, int(48 * alpha))
y3 = compose(DarknetConv2D_BN_Leaky(int(96 * alpha), (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x3)
#downsample fpn merge for feature map 3 & 2
x3_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
DarknetConv2D_BN_Leaky(int(144 * alpha), (3, 3), strides=(2, 2)))(x3)
x2 = Concatenate()([x3_downsample, x2])
#feature map 2 output (26 x 26 x (288*alpha) for 416 input)
#x2, y2 = make_last_layers(x2, int(144*alpha), num_anchors*(num_classes+5))
x2 = make_yolo_head(x2, int(144 * alpha))
y2 = compose(DarknetConv2D_BN_Leaky(int(288 * alpha), (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x2)
#downsample fpn merge for feature map 2 & 1
x2_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
DarknetConv2D_BN_Leaky(int(288 * alpha), (3, 3), strides=(2, 2)))(x2)
x1 = Concatenate()([x2_downsample, x1])
#feature map 1 output (13 x 13 x (576*alpha) for 416 input)
#x1, y1 = make_last_layers(x1, int(288*alpha), num_anchors*(num_classes+5))
x1 = make_yolo_head(x1, int(288 * alpha))
y1 = compose(DarknetConv2D_BN_Leaky(int(576 * alpha), (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x1)
return Model(inputs, [y1, y2, y3])
