def tiny_yolo3_mobilenetv3small_body(inputs, num_anchors, num_classes, alpha=1.0):
'''Create Tiny YOLO_v3 MobileNetV3Small model CNN body in keras.'''
mobilenetv3small = MobileNetV3Small(input_tensor=inputs, weights='imagenet', include_top=False, alpha=alpha)
print('backbone layers number: {}'.format(len(mobilenetv3small.layers)))
# 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 :13 x 13 x (576*alpha)
f1 = mobilenetv3small.layers[165].output
# f2: 26 x 26 x (288*alpha)
f2 = mobilenetv3small.layers[117].output
f1_channel_num = int(576*alpha)
f2_channel_num = int(288*alpha)
#f1_channel_num = 1024
#f2_channel_num = 512
y1, y2 = tiny_yolo3_predictions((f1, f2), (f1_channel_num, f2_channel_num), num_anchors, num_classes)
return Model(inputs, [y1,y2])
def yolo2_mobilenetv3small_body(inputs, num_anchors, num_classes, alpha=1.0):
"""Create YOLO_V2 MobileNetV3Small model CNN body in Keras."""
mobilenetv3small = MobileNetV3Small(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
# input: 416 x 416 x 3
# mobilenetv3small.output(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)
conv_head1 = compose(DarknetConv2D_BN_Leaky(int(576 * alpha), (3, 3)),
DarknetConv2D_BN_Leaky(int(576 * alpha), (3, 3)))(
mobilenetv3small.output)
# activation_22(layer 117) output shape: 26 x 26 x (288*alpha)
activation_22 = mobilenetv3small.layers[117].output
conv_head2 = DarknetConv2D_BN_Leaky(int(64 * alpha), (1, 1))(activation_22)
# TODO: Allow Keras Lambda to use func arguments for output_shape?
conv_head2_reshaped = Lambda(space_to_depth_x2,
output_shape=space_to_depth_x2_output_shape,
name='space_to_depth')(conv_head2)
x = Concatenate()([conv_head2_reshaped, conv_head1])
x = DarknetConv2D_BN_Leaky(int(576 * alpha), (3, 3))(x)
x = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),
name='predict_conv')(x)
return Model(inputs, x)
def yolo3_mobilenetv3small_body(inputs, num_anchors, num_classes, alpha=1.0):
"""Create YOLO_V3 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 :13 x 13 x (576*alpha)
f1 = mobilenetv3small.layers[165].output
# f2: 26 x 26 x (288*alpha)
f2 = mobilenetv3small.layers[117].output
# f3 : 52 x 52 x (96*alpha)
f3 = mobilenetv3small.layers[38].output
f1_channel_num = int(576 * alpha)
f2_channel_num = int(288 * alpha)
f3_channel_num = int(96 * 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_last_layers(f1, f1_channel_num // 2,
num_anchors * (num_classes + 5))
#x, y1 = make_last_layers(f1, f1_channel_num//2, num_anchors * (num_classes + 5), 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_last_layers(x, f2_channel_num // 2,
num_anchors * (num_classes + 5))
#x, y2 = make_last_layers(x, f2_channel_num//2, num_anchors*(num_classes+5), 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_last_layers(x, f3_channel_num // 2,
num_anchors * (num_classes + 5))
#x, y3 = make_last_layers(x, f3_channel_num//2, num_anchors*(num_classes+5), predict_filters=int(256*alpha))
return Model(inputs=inputs, outputs=[y1, y2, y3])
def tiny_yolo4_mobilenetv3small_body(inputs,
num_anchors,
num_classes,
alpha=1.0,
use_spp=True):
'''Create Tiny 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
# f2: 26 x 26 x (288*alpha) for 416 input
f2 = mobilenetv3small.layers[117].output
#feature map 1 head (13 x 13 x (288*alpha) for 416 input)
x1 = DarknetConv2D_BN_Leaky(int(288 * alpha), (1, 1))(f1)
if use_spp:
x1 = Spp_Conv2D_BN_Leaky(x1, 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)
x2 = compose(
Concatenate(),
#Depthwise_Separable_Conv2D_BN_Leaky(filters=int(288*alpha), kernel_size=(3, 3), block_id_str='11'),
DarknetConv2D_BN_Leaky(int(288 * alpha), (3, 3)))([x1_upsample, f2])
#feature map 2 output (26 x 26 x (288*alpha) for 416 input)
y2 = 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))),
#Darknet_Depthwise_Separable_Conv2D_BN_Leaky(int(288*alpha), (3,3), strides=(2,2), block_id_str='12'),
DarknetConv2D_BN_Leaky(int(288 * alpha), (3, 3), strides=(2, 2)))(x2)
x1 = compose(
Concatenate(),
#Depthwise_Separable_Conv2D_BN_Leaky(filters=int(576*alpha), kernel_size=(3, 3), block_id_str='13'),
DarknetConv2D_BN_Leaky(int(576 * alpha), (3, 3)))([x2_downsample, x1])
#feature map 1 output (13 x 13 x (576*alpha) for 416 input)
y1 = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1))(x1)
return Model(inputs, [y1, y2])
def tiny_yolo3lite_mobilenetv3small_body(inputs,
num_anchors,
num_classes,
alpha=1.0):
'''Create Tiny 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 :13 x 13 x (576*alpha)
f1 = mobilenetv3small.layers[165].output
# f2: 26 x 26 x (288*alpha)
f2 = mobilenetv3small.layers[117].output
f1_channel_num = int(576 * alpha)
f2_channel_num = int(288 * alpha)
#f1_channel_num = 1024
#f2_channel_num = 512
#feature map 1 transform
x1 = DarknetConv2D_BN_Leaky(f1_channel_num // 2, (1, 1))(f1)
#feature map 1 output (13x13 for 416 input)
y1 = compose(
#DarknetConv2D_BN_Leaky(f1_channel_num, (3,3)),
Depthwise_Separable_Conv2D_BN_Leaky(filters=f1_channel_num,
kernel_size=(3, 3),
block_id_str='15'),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x1)
#upsample fpn merge for feature map 1 & 2
x2 = compose(DarknetConv2D_BN_Leaky(f2_channel_num // 2, (1, 1)),
UpSampling2D(2))(x1)
#feature map 2 output (26x26 for 416 input)
y2 = compose(
Concatenate(),
#DarknetConv2D_BN_Leaky(f2_channel_num, (3,3)),
Depthwise_Separable_Conv2D_BN_Leaky(filters=f2_channel_num,
kernel_size=(3, 3),
block_id_str='16'),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))([x2, f2])
return Model(inputs, [y1, y2])
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 tiny_yolo2lite_mobilenetv3small_body(inputs, num_anchors, num_classes, alpha=1.0):
"""Create Tiny YOLO_V2 Lite MobileNetV3Small model CNN body in Keras."""
mobilenetv3small = MobileNetV3Small(input_tensor=inputs, weights='imagenet', include_top=False, alpha=alpha)
print('backbone layers number: {}'.format(len(mobilenetv3small.layers)))
# input: 416 x 416 x 3
# mobilenetv3small.output(layer 165, final feature map): 13 x 13 x (576*alpha)
# f1: 13 x 13 x (576*alpha)
f1 = mobilenetv3small.output
f1_channel_num = int(576*alpha)
y = compose(
Depthwise_Separable_Conv2D_BN_Leaky(f1_channel_num, (3,3), block_id_str='pred_1'),
DarknetConv2D(num_anchors*(num_classes+5), (1,1), name='predict_conv'))(f1)
return Model(inputs, y)
def tiny_yolo2_mobilenetv3small_body(inputs,
num_anchors,
num_classes,
alpha=1.0):
"""Create Tiny YOLO_V2 MobileNetV3Small model CNN body in Keras."""
mobilenetv3small = MobileNetV3Small(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
# input: 416 x 416 x 3
# mobilenetv3small.output(layer 165, final feature map): 13 x 13 x (576*alpha)
y = compose(
DarknetConv2D_BN_Leaky(int(576 * alpha), (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),
name='predict_conv'))(mobilenetv3small.output)
return Model(inputs, y)
def tiny_yolo3_mobilenetv3small_body(inputs,
num_anchors,
num_classes,
alpha=1.0):
'''Create Tiny YOLO_v3 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
x1 = mobilenetv3small.layers[117].output
x2 = mobilenetv3small.layers[165].output
x2 = DarknetConv2D_BN_Leaky(int(288 * alpha), (1, 1))(x2)
y1 = compose(
DarknetConv2D_BN_Leaky(int(576 * alpha), (3, 3)),
#Depthwise_Separable_Conv2D_BN_Leaky(filters=int(576*alpha), kernel_size=(3, 3), block_id_str='15'),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x2)
x2 = compose(DarknetConv2D_BN_Leaky(int(144 * alpha), (1, 1)),
UpSampling2D(2))(x2)
y2 = compose(
Concatenate(),
DarknetConv2D_BN_Leaky(int(288 * alpha), (3, 3)),
#Depthwise_Separable_Conv2D_BN_Leaky(filters=int(288*alpha), kernel_size=(3, 3), block_id_str='16'),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))([x2, x1])
return Model(inputs, [y1, y2])
def yolo4lite_mobilenetv3small_body(inputs,
num_anchors,
num_classes,
alpha=1.0):
'''Create YOLO_v4 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: 13 x 13 x (576*alpha)
f1 = mobilenetv3small.layers[165].output
# f2: 26 x 26 x (288*alpha) for 416 input
f2 = mobilenetv3small.layers[117].output
# f3: 52 x 52 x (96*alpha)
f3 = mobilenetv3small.layers[38].output
f1_channel_num = int(576 * alpha)
f2_channel_num = int(288 * alpha)
f3_channel_num = int(96 * alpha)
#f1_channel_num = 1024
#f2_channel_num = 512
#f3_channel_num = 256
y1, y2, y3 = yolo4lite_predictions(
(f1, f2, f3), (f1_channel_num, f2_channel_num, f3_channel_num),
num_anchors, num_classes)
return Model(inputs, [y1, y2, y3])
def yolo2lite_mobilenetv3small_body(inputs, num_anchors, num_classes, alpha=1.0):
"""Create YOLO_V2 Lite MobileNetV3Small model CNN body in Keras."""
mobilenetv3small = MobileNetV3Small(input_tensor=inputs, weights='imagenet', include_top=False, alpha=alpha)
print('backbone layers number: {}'.format(len(mobilenetv3small.layers)))
# input: 416 x 416 x 3
# mobilenetv3small.output(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)
# NOTE: activation layer name may different for TF1.x/2.x, so we
# use index to fetch layer
# f1: 13 x 13 x (960*alpha)
f1 = mobilenetv3small.output
# f2: 26 x 26 x (672*alpha)
f2 = mobilenetv3small.layers[117].output
f1_channel_num = int(576*alpha)
f2_channel_num = int(288*alpha)
y = yolo2lite_predictions((f1, f2), (f1_channel_num, f2_channel_num), num_anchors, num_classes)
return Model(inputs, y)
def yolo4lite_mobilenetv3small_body(inputs,
num_anchors,
num_classes,
alpha=1.0):
'''Create YOLO_v4 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)
# 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_depthwise_separable_head(f1,
int(288 * alpha),
block_id_str='11')
#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_depthwise_separable_head(x2,
int(144 * alpha),
block_id_str='12')
#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_depthwise_separable_last_layers(x3, int(48*alpha), num_anchors*(num_classes+5), block_id_str='13')
x3 = make_yolo_depthwise_separable_head(x3,
int(48 * alpha),
block_id_str='13')
y3 = compose(
Depthwise_Separable_Conv2D_BN_Leaky(int(96 * alpha), (3, 3),
block_id_str='13_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))),
Darknet_Depthwise_Separable_Conv2D_BN_Leaky(int(144 * alpha), (3, 3),
strides=(2, 2),
block_id_str='13_4'))(x3)
x2 = Concatenate()([x3_downsample, x2])
#feature map 2 output (26 x 26 x (288*alpha) for 416 input)
#x2, y2 = make_depthwise_separable_last_layers(x2, int(144*alpha), num_anchors*(num_classes+5), block_id_str='14')
x2 = make_yolo_depthwise_separable_head(x2,
int(144 * alpha),
block_id_str='14')
y2 = compose(
Depthwise_Separable_Conv2D_BN_Leaky(int(288 * alpha), (3, 3),
block_id_str='14_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))),
Darknet_Depthwise_Separable_Conv2D_BN_Leaky(int(288 * alpha), (3, 3),
strides=(2, 2),
block_id_str='14_4'))(x2)
x1 = Concatenate()([x2_downsample, x1])
#feature map 1 output (13 x 13 x (576*alpha) for 416 input)
#x1, y1 = make_depthwise_separable_last_layers(x1, int(288*alpha), num_anchors*(num_classes+5), block_id_str='15')
x1 = make_yolo_depthwise_separable_head(x1,
int(288 * alpha),
block_id_str='15')
y1 = compose(
Depthwise_Separable_Conv2D_BN_Leaky(int(576 * alpha), (3, 3),
block_id_str='15_3'),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x1)
return Model(inputs, [y1, y2, y3])
def get_base_model(model_type, model_input_shape, weights='imagenet'):
input_tensor = Input(shape=model_input_shape + (3, ), name='image_input')
if model_type == 'mobilenet':
model = MobileNet(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=weights,
pooling=None,
include_top=False,
alpha=0.5)
elif model_type == 'mobilenetv2':
model = MobileNetV2(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=weights,
pooling=None,
include_top=False,
alpha=0.5)
elif model_type == 'mobilenetv3large':
model = MobileNetV3Large(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=weights,
pooling=None,
include_top=False,
alpha=0.75)
elif model_type == 'mobilenetv3small':
model = MobileNetV3Small(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=weights,
pooling=None,
include_top=False,
alpha=0.75)
elif model_type == 'peleenet':
model = PeleeNet(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=weights,
pooling=None,
include_top=False)
elif model_type == 'ghostnet':
model = GhostNet(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=weights,
pooling=None,
include_top=False)
elif model_type == 'squeezenet':
model = SqueezeNet(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=weights,
pooling=None,
include_top=False)
elif model_type == 'mobilevit_s':
model = MobileViT_S(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=weights,
pooling=None,
include_top=False)
elif model_type == 'mobilevit_xs':
model = MobileViT_XS(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=weights,
pooling=None,
include_top=False)
elif model_type == 'mobilevit_xxs':
model = MobileViT_XXS(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=weights,
pooling=None,
include_top=False)
elif model_type == 'resnet50':
model = ResNet50(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=weights,
pooling=None,
include_top=False)
elif model_type == 'simple_cnn':
model = SimpleCNN(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=None,
pooling=None,
include_top=False)
elif model_type == 'simple_cnn_lite':
model = SimpleCNNLite(input_tensor=input_tensor,
input_shape=model_input_shape + (3, ),
weights=None,
pooling=None,
include_top=False)
else:
raise ValueError('Unsupported model type')
return model
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)
# NOTE: activation layer name may different for TF1.x/2.x, so we
# use index to fetch layer
# f1: 13 x 13 x (576*alpha)
f1 = mobilenetv3small.layers[165].output
# f2: 26 x 26 x (288*alpha) for 416 input
f2 = mobilenetv3small.layers[117].output
# f3: 52 x 52 x (96*alpha)
f3 = mobilenetv3small.layers[38].output
f1_channel_num = int(576 * alpha)
f2_channel_num = int(288 * alpha)
f3_channel_num = int(96 * alpha)
#f1_channel_num = 1024
#f2_channel_num = 512
#f3_channel_num = 256
#feature map 1 head (13 x 13 x 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 (26 x 26 x 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 (52 x 52 x f3_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 (26 x 26 x f2_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 (13 x 13 x f1_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])