def tiny_yolo2_body(inputs, num_anchors, num_classes):
'''Create Tiny YOLO_v2 model CNN body in keras.'''
x = compose(DarknetConv2D_BN_Leaky(16, (3, 3)),
MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),
DarknetConv2D_BN_Leaky(32, (3, 3)),
MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),
DarknetConv2D_BN_Leaky(64, (3, 3)),
MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),
DarknetConv2D_BN_Leaky(128, (3, 3)),
MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),
DarknetConv2D_BN_Leaky(256, (3, 3)),
MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='same'),
DarknetConv2D_BN_Leaky(512, (3, 3)),
MaxPooling2D(pool_size=(2, 2), strides=(1, 1), padding='same'),
DarknetConv2D_BN_Leaky(1024, (3, 3)))(inputs)
# TODO: darknet tiny YOLOv2 use different filter number for COCO and VOC
if num_classes == 80:
y = compose(
DarknetConv2D_BN_Leaky(512, (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),
name='predict_conv'))(x)
else:
y = compose(
DarknetConv2D_BN_Leaky(1024, (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),
name='predict_conv'))(x)
return Model(inputs, y)
def yolo2lite_efficientnet_body(inputs, num_anchors, num_classes, level=0):
'''
Create YOLO_v2 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_channel_num = feature_map_info['f1_channel_num']
conv_head1 = compose(
Depthwise_Separable_Conv2D_BN_Leaky(f1_channel_num, (3, 3)),
Depthwise_Separable_Conv2D_BN_Leaky(f1_channel_num, (3, 3)))(efficientnet.output)
f2 = efficientnet.get_layer('block6a_expand_activation').output
conv_head2 = DarknetConv2D_BN_Leaky(int(64*(f1_channel_num//1024)), (1, 1))(f2)
# 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 = Depthwise_Separable_Conv2D_BN_Leaky(f1_channel_num, (3, 3))(x)
x = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1), name='predict_conv')(x)
return Model(inputs, x)
def yolo2lite_mobilenetv2_body(inputs, num_anchors, num_classes, alpha=1.0):
"""Create YOLO_V2 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
# mobilenetv2.output : 13 x 13 x 1280
# block_13_expand_relu(layers[119]) : 26 x 26 x (576*alpha)
conv_head1 = compose(
Depthwise_Separable_Conv2D_BN_Leaky(1280, (3, 3)),
Depthwise_Separable_Conv2D_BN_Leaky(1280, (3, 3)))(mobilenetv2.output)
# block_13_expand_relu output shape: 26 x 26 x (576*alpha)
block_13_expand_relu = mobilenetv2.layers[119].output
conv_head2 = DarknetConv2D_BN_Leaky(int(64*alpha), (1, 1))(block_13_expand_relu)
# 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 = Depthwise_Separable_Conv2D_BN_Leaky(1280, (3, 3))(x)
x = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1), name='predict_conv')(x)
return Model(inputs, x)
def yolo2_body(inputs, num_anchors, num_classes, weights_path=None):
"""Create YOLO_V2 model CNN body in Keras."""
darknet19 = Model(inputs, darknet19_body()(inputs))
if weights_path is not None:
darknet19.load_weights(weights_path, by_name=True)
print('Load weights {}.'.format(weights_path))
# input: 416 x 416 x 3
# darknet19.output : 13 x 13 x 1024
# conv13(layers[43]) : 26 x 26 x 512
conv20 = compose(DarknetConv2D_BN_Leaky(1024, (3, 3)),
DarknetConv2D_BN_Leaky(1024, (3, 3)))(darknet19.output)
# conv13 output shape: 26 x 26 x 512
conv13 = darknet19.layers[43].output
conv21 = DarknetConv2D_BN_Leaky(64, (1, 1))(conv13)
# TODO: Allow Keras Lambda to use func arguments for output_shape?
conv21_reshaped = Lambda(space_to_depth_x2,
output_shape=space_to_depth_x2_output_shape,
name='space_to_depth')(conv21)
x = Concatenate()([conv21_reshaped, conv20])
x = DarknetConv2D_BN_Leaky(1024, (3, 3))(x)
x = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),
name='predict_conv')(x)
return Model(inputs, x)
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 tiny_yolo2lite_mobilenetv3large_body(inputs,
num_anchors,
num_classes,
alpha=1.0):
"""Create Tiny YOLO_V2 Lite MobileNetV3Large model CNN body in Keras."""
mobilenetv3large = MobileNetV3Large(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
print('backbone layers number: {}'.format(len(mobilenetv3large.layers)))
# input: 416 x 416 x 3
# mobilenetv3large.output(layer 194, final feature map): 13 x 13 x (960*alpha)
# f1: 13 x 13 x (960*alpha)
f1 = mobilenetv3large.output
f1_channel_num = int(960 * 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 yolo2_mobilenetv3large_body(inputs, num_anchors, num_classes, alpha=1.0):
"""Create YOLO_V2 MobileNetV3Large model CNN body in Keras."""
mobilenetv3large = MobileNetV3Large(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
# input: 416 x 416 x 3
# mobilenetv3large.output(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)
conv_head1 = compose(DarknetConv2D_BN_Leaky(int(960 * alpha), (3, 3)),
DarknetConv2D_BN_Leaky(int(960 * alpha), (3, 3)))(
mobilenetv3large.output)
# activation_29(layer 146) output shape: 26 x 26 x (672*alpha)
activation_29 = mobilenetv3large.layers[146].output
conv_head2 = DarknetConv2D_BN_Leaky(int(64 * alpha), (1, 1))(activation_29)
# 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(960 * alpha), (3, 3))(x)
x = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),
name='predict_conv')(x)
return Model(inputs, x)
def yolo2_mobilenet_body(inputs, num_anchors, num_classes, alpha=1.0):
"""Create YOLO_V2 MobileNet model CNN body in Keras."""
mobilenet = MobileNet(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
# input: 416 x 416 x 3
# mobilenet.output : 13 x 13 x (1024*alpha)
# conv_pw_11_relu(layers[73]) : 26 x 26 x (512*alpha)
conv_head1 = compose(DarknetConv2D_BN_Leaky(int(1024 * alpha), (3, 3)),
DarknetConv2D_BN_Leaky(int(1024 * alpha),
(3, 3)))(mobilenet.output)
# conv_pw_11_relu output shape: 26 x 26 x (512*alpha)
conv_pw_11_relu = mobilenet.layers[73].output
conv_head2 = DarknetConv2D_BN_Leaky(int(64 * alpha),
(1, 1))(conv_pw_11_relu)
# 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(1024 * alpha), (3, 3))(x)
x = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),
name='predict_conv')(x)
return Model(inputs, x)
def darknet19(inputs):
"""Generate Darknet-19 model for Imagenet classification."""
body = darknet19_body()(inputs)
x = DarknetConv2D(1000, (1, 1))(body)
x = GlobalAveragePooling2D()(x)
logits = Softmax()(x)
return Model(inputs, logits)
def yolo2lite_xception_body(inputs, num_anchors, num_classes):
"""Create YOLO_V2 Lite Xception model CNN body in Keras."""
xception = Xception(input_tensor=inputs,
weights='imagenet',
include_top=False)
# input: 416 x 416 x 3
# xception.output: 13 x 13 x 2048
# block13_sepconv2_bn(middle in block13, layers[121]): 26 x 26 x 1024
# add_46(end of block12, layers[115]): 26 x 26 x 728
conv_head1 = compose(Depthwise_Separable_Conv2D_BN_Leaky(2048, (3, 3)),
Depthwise_Separable_Conv2D_BN_Leaky(2048, (3, 3)))(
xception.output)
# block13_sepconv2_bn output shape: 26 x 26 x 1024
block13_sepconv2_bn = xception.layers[121].output
conv_head2 = DarknetConv2D_BN_Leaky(128, (1, 1))(block13_sepconv2_bn)
# 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 = Depthwise_Separable_Conv2D_BN_Leaky(2048, (3, 3))(x)
x = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),
name='predict_conv')(x)
return Model(inputs, x)
def tiny_yolo2lite_mobilenetv2_body(inputs, num_anchors, num_classes):
"""Create Tiny YOLO_V2 Lite MobileNetV2 model CNN body in Keras."""
mobilenetv2 = MobileNetV2(input_tensor=inputs, weights='imagenet', include_top=False, alpha=1.0)
# input: 416 x 416 x 3
# mobilenetv2.output : 13 x 13 x 1280
y = compose(
Depthwise_Separable_Conv2D_BN_Leaky(1280, (3,3)),
DarknetConv2D(num_anchors*(num_classes+5), (1,1), name='predict_conv'))(mobilenetv2.output)
return Model(inputs, y)
def tiny_yolo2lite_efficientnet_body(inputs, num_anchors, num_classes, level=0):
'''
Create Tiny YOLO_v2 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_channel_num = feature_map_info['f1_channel_num']
y = compose(
Depthwise_Separable_Conv2D_BN_Leaky(f1_channel_num, (3,3)),
DarknetConv2D(num_anchors*(num_classes+5), (1,1), name='predict_conv'))(efficientnet.output)
return Model(inputs, y)
def tiny_yolo2_mobilenet_body(inputs, num_anchors, num_classes, alpha=1.0):
"""Create Tiny YOLO_V2 MobileNet model CNN body in Keras."""
mobilenet = MobileNet(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
# input: 416 x 416 x 3
# mobilenet.output : 13 x 13 x (1024*alpha)
y = compose(
DarknetConv2D_BN_Leaky(int(1024 * alpha), (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),
name='predict_conv'))(mobilenet.output)
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)
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(
DarknetConv2D_BN_Leaky(f1_channel_num, (3,3)),
DarknetConv2D(num_anchors*(num_classes+5), (1,1), name='predict_conv'))(f1)
return Model(inputs, y)
def tiny_yolo2lite_efficientnet_body(inputs, num_anchors, num_classes, level=0):
'''
Create Tiny YOLO_v2 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)
print('backbone layers number: {}'.format(len(efficientnet.layers)))
f1 = efficientnet.get_layer('top_activation').output
f1_channel_num = feature_map_info['f1_channel_num']
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_mobilenetv3large_body(inputs,
num_anchors,
num_classes,
alpha=1.0):
"""Create Tiny YOLO_V2 MobileNetV3Large model CNN body in Keras."""
mobilenetv3large = MobileNetV3Large(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
# input: 416 x 416 x 3
# mobilenetv3large.output(layer 194, final feature map): 13 x 13 x (960*alpha)
y = compose(
DarknetConv2D_BN_Leaky(int(960 * alpha), (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),
name='predict_conv'))(mobilenetv3large.output)
return Model(inputs, y)
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)
# input: 416 x 416 x 3
# mobilenetv3small.output(layer 165, final feature map): 13 x 13 x (576*alpha)
y = compose(
Depthwise_Separable_Conv2D_BN_Leaky(int(576 * alpha), (3, 3),
block_id_str='11'),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),
name='predict_conv'))(mobilenetv3small.output)
return Model(inputs, y)
def tiny_yolo2_mobilenetv2_body(inputs, num_anchors, num_classes):
"""Create Tiny YOLO_V2 MobileNetV2 model CNN body in Keras."""
mobilenetv2 = MobileNetV2(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=1.0)
print('backbone layers number: {}'.format(len(mobilenetv2.layers)))
# input: 416 x 416 x 3
# mobilenetv2.output : 13 x 13 x 1280
# f1: 13 x 13 x 1280
f1 = mobilenetv2.output
f1_channel_num = 1280
y = compose(
DarknetConv2D_BN_Leaky(f1_channel_num, (3, 3)),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),
name='predict_conv'))(f1)
return Model(inputs, y)