"""
MobileNetV3 for ImageNet-1K, implemented in Keras.
Original paper: 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244.
"""
__all__ = ['mobilenetv3', 'mobilenetv3_small_w7d20', 'mobilenetv3_small_wd2', 'mobilenetv3_small_w3d4',
'mobilenetv3_small_w1', 'mobilenetv3_small_w5d4', 'mobilenetv3_large_w7d20', 'mobilenetv3_large_wd2',
'mobilenetv3_large_w3d4', 'mobilenetv3_large_w1', 'mobilenetv3_large_w5d4']
import os
from keras import layers as nn
from keras.models import Model
from .common import round_channels, conv1x1, conv1x1_block, conv3x3_block, dwconv3x3_block, dwconv5x5_block,\
se_block, HSwish, is_channels_first, flatten
def mobilenetv3_unit(x,
in_channels,
out_channels,
exp_channels,
strides,
use_kernel3,
activation,
use_se,
name="mobilenetv3_unit"):
"""
MobileNetV3 unit.
Parameters:
----------
x : keras.backend tensor/variable/symbol
Input tensor/variable/symbol.
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
exp_channels : int
Number of middle (expanded) channels.
strides : int or tuple/list of 2 int
Strides of the second convolution layer.
use_kernel3 : bool
Whether to use 3x3 (instead of 5x5) kernel.
activation : str
Activation function or name of activation function.
use_se : bool
Whether to use SE-module.
name : str, default 'mobilenetv3_unit'
Unit name.
Returns
-------
keras.backend tensor/variable/symbol
Resulted tensor/variable/symbol.
"""
assert (exp_channels >= out_channels)
residual = (in_channels == out_channels) and (strides == 1)
use_exp_conv = exp_channels != out_channels
mid_channels = exp_channels
if residual:
identity = x
if use_exp_conv:
x = conv1x1_block(
x=x,
in_channels=in_channels,
out_channels=mid_channels,
activation=activation,
name=name + "/exp_conv")
if use_kernel3:
x = dwconv3x3_block(
x=x,
in_channels=mid_channels,
out_channels=mid_channels,
strides=strides,
activation=activation,
name=name + "/conv1")
else:
x = dwconv5x5_block(
x=x,
in_channels=mid_channels,
out_channels=mid_channels,
strides=strides,
activation=activation,
name=name + "/conv1")
if use_se:
x = se_block(
x=x,
channels=mid_channels,
reduction=4,
approx_sigmoid=True,
round_mid=True,
name=name + "/se")
x = conv1x1_block(
x=x,
in_channels=mid_channels,
out_channels=out_channels,
activation=None,
name=name + "/conv2")
if residual:
x = nn.add([x, identity], name=name + "/add")
return x
def mobilenetv3_final_block(x,
in_channels,
out_channels,
use_se,
name="mobilenetv3_final_block"):
"""
MobileNetV3 final block.
Parameters:
----------
x : keras.backend tensor/variable/symbol
Input tensor/variable/symbol.
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
use_se : bool
Whether to use SE-module.
name : str, default 'mobilenetv3_final_block'
Unit name.
Returns
-------
keras.backend tensor/variable/symbol
Resulted tensor/variable/symbol.
"""
x = conv1x1_block(
x=x,
in_channels=in_channels,
out_channels=out_channels,
activation="hswish",
name=name + "/conv")
if use_se:
x = se_block(
x=x,
channels=out_channels,
reduction=4,
approx_sigmoid=True,
round_mid=True,
name=name + "/se")
return x
def mobilenetv3_classifier(x,
in_channels,
out_channels,
mid_channels,
dropout_rate,
name="mobilenetv3_final_block"):
"""
MobileNetV3 classifier.
Parameters:
----------
x : keras.backend tensor/variable/symbol
Input tensor/variable/symbol.
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
mid_channels : int
Number of middle channels.
dropout_rate : float
Parameter of Dropout layer. Faction of the input units to drop.
name : str, default 'mobilenetv3_classifier'
Unit name.
Returns
-------
keras.backend tensor/variable/symbol
Resulted tensor/variable/symbol.
"""
x = conv1x1(
x=x,
in_channels=in_channels,
out_channels=mid_channels,
name=name + "/conv1")
x = HSwish(name=name + "/hswish")(x)
use_dropout = (dropout_rate != 0.0)
if use_dropout:
x = nn.Dropout(
rate=dropout_rate,
name=name + "dropout")(x)
x = conv1x1(
x=x,
in_channels=mid_channels,
out_channels=out_channels,
use_bias=True,
name=name + "/conv2")
return x
def mobilenetv3(channels,
exp_channels,
init_block_channels,
final_block_channels,
classifier_mid_channels,
kernels3,
use_relu,
use_se,
first_stride,
final_use_se,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
MobileNetV3 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
exp_channels : list of list of int
Number of middle (expanded) channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
final_block_channels : int
Number of output channels for the final block of the feature extractor.
classifier_mid_channels : int
Number of middle channels for classifier.
kernels3 : list of list of int/bool
Using 3x3 (instead of 5x5) kernel for each unit.
use_relu : list of list of int/bool
Using ReLU activation flag for each unit.
use_se : list of list of int/bool
Using SE-block flag for each unit.
first_stride : bool
Whether to use stride for the first stage.
final_use_se : bool
Whether to use SE-module in the final block.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, in_size[0], in_size[1]) if is_channels_first() else\
(in_size[0], in_size[1], in_channels)
input = nn.Input(shape=input_shape)
x = conv3x3_block(
x=input,
in_channels=in_channels,
out_channels=init_block_channels,
strides=2,
activation="hswish",
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
exp_channels_ij = exp_channels[i][j]
strides = 2 if (j == 0) and ((i != 0) or first_stride) else 1
use_kernel3 = kernels3[i][j] == 1
activation = "relu" if use_relu[i][j] == 1 else "hswish"
use_se_flag = use_se[i][j] == 1
x = mobilenetv3_unit(
x=x,
in_channels=in_channels,
out_channels=out_channels,
exp_channels=exp_channels_ij,
use_kernel3=use_kernel3,
strides=strides,
activation=activation,
use_se=use_se_flag,
name="features/stage{}/unit{}".format(i + 1, j + 1))
in_channels = out_channels
x = mobilenetv3_final_block(
x=x,
in_channels=in_channels,
out_channels=final_block_channels,
use_se=final_use_se,
name="features/final_block")
in_channels = final_block_channels
x = nn.AvgPool2D(
pool_size=7,
strides=1,
name="features/final_pool")(x)
x = mobilenetv3_classifier(
x=x,
in_channels=in_channels,
out_channels=classes,
mid_channels=classifier_mid_channels,
dropout_rate=0.2,
name="output")
x = flatten(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def get_mobilenetv3(version,
width_scale,
model_name=None,
pretrained=False,
root=os.path.join("~", ".keras", "models"),
**kwargs):
"""
Create MobileNetV3 model with specific parameters.
Parameters:
----------
version : str
Version of MobileNetV3 ('small' or 'large').
width_scale : float
Scale factor for width of layers.
model_name : str or None, default None
Model name for loading pretrained model.
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.keras/models'
Location for keeping the model parameters.
"""
if version == "small":
init_block_channels = 16
channels = [[16], [24, 24], [40, 40, 40, 48, 48], [96, 96, 96]]
exp_channels = [[16], [72, 88], [96, 240, 240, 120, 144], [288, 576, 576]]
kernels3 = [[1], [1, 1], [0, 0, 0, 0, 0], [0, 0, 0]]
use_relu = [[1], [1, 1], [0, 0, 0, 0, 0], [0, 0, 0]]
use_se = [[1], [0, 0], [1, 1, 1, 1, 1], [1, 1, 1]]
first_stride = True
final_block_channels = 576
elif version == "large":
init_block_channels = 16
channels = [[16], [24, 24], [40, 40, 40], [80, 80, 80, 80, 112, 112], [160, 160, 160]]
exp_channels = [[16], [64, 72], [72, 120, 120], [240, 200, 184, 184, 480, 672], [672, 960, 960]]
kernels3 = [[1], [1, 1], [0, 0, 0], [1, 1, 1, 1, 1, 1], [0, 0, 0]]
use_relu = [[1], [1, 1], [1, 1, 1], [0, 0, 0, 0, 0, 0], [0, 0, 0]]
use_se = [[0], [0, 0], [1, 1, 1], [0, 0, 0, 0, 1, 1], [1, 1, 1]]
first_stride = False
final_block_channels = 960
else:
raise ValueError("Unsupported MobileNetV3 version {}".format(version))
final_use_se = False
classifier_mid_channels = 1280
if width_scale != 1.0:
channels = [[round_channels(cij * width_scale) for cij in ci] for ci in channels]
exp_channels = [[round_channels(cij * width_scale) for cij in ci] for ci in exp_channels]
init_block_channels = round_channels(init_block_channels * width_scale)
if width_scale > 1.0:
final_block_channels = round_channels(final_block_channels * width_scale)
net = mobilenetv3(
channels=channels,
exp_channels=exp_channels,
init_block_channels=init_block_channels,
final_block_channels=final_block_channels,
classifier_mid_channels=classifier_mid_channels,
kernels3=kernels3,
use_relu=use_relu,
use_se=use_se,
first_stride=first_stride,
final_use_se=final_use_se,
**kwargs)
if pretrained:
if (model_name is None) or (not model_name):
raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.")
from .model_store import download_model
download_model(
net=net,
model_name=model_name,
local_model_store_dir_path=root)
return net
def mobilenetv3_small_w7d20(**kwargs):
"""
MobileNetV3 Small 224/0.35 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.keras/models'
Location for keeping the model parameters.
"""
return get_mobilenetv3(version="small", width_scale=0.35, model_name="mobilenetv3_small_w7d20", **kwargs)
def mobilenetv3_small_wd2(**kwargs):
"""
MobileNetV3 Small 224/0.5 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.keras/models'
Location for keeping the model parameters.
"""
return get_mobilenetv3(version="small", width_scale=0.5, model_name="mobilenetv3_small_wd2", **kwargs)
def mobilenetv3_small_w3d4(**kwargs):
"""
MobileNetV3 Small 224/0.75 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.keras/models'
Location for keeping the model parameters.
"""
return get_mobilenetv3(version="small", width_scale=0.75, model_name="mobilenetv3_small_w3d4", **kwargs)
def mobilenetv3_small_w1(**kwargs):
"""
MobileNetV3 Small 224/1.0 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.keras/models'
Location for keeping the model parameters.
"""
return get_mobilenetv3(version="small", width_scale=1.0, model_name="mobilenetv3_small_w1", **kwargs)
def mobilenetv3_small_w5d4(**kwargs):
"""
MobileNetV3 Small 224/1.25 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.keras/models'
Location for keeping the model parameters.
"""
return get_mobilenetv3(version="small", width_scale=1.25, model_name="mobilenetv3_small_w5d4", **kwargs)
def mobilenetv3_large_w7d20(**kwargs):
"""
MobileNetV3 Small 224/0.35 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.keras/models'
Location for keeping the model parameters.
"""
return get_mobilenetv3(version="large", width_scale=0.35, model_name="mobilenetv3_small_w7d20", **kwargs)
def mobilenetv3_large_wd2(**kwargs):
"""
MobileNetV3 Large 224/0.5 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.keras/models'
Location for keeping the model parameters.
"""
return get_mobilenetv3(version="large", width_scale=0.5, model_name="mobilenetv3_large_wd2", **kwargs)
def mobilenetv3_large_w3d4(**kwargs):
"""
MobileNetV3 Large 224/0.75 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.keras/models'
Location for keeping the model parameters.
"""
return get_mobilenetv3(version="large", width_scale=0.75, model_name="mobilenetv3_large_w3d4", **kwargs)
def mobilenetv3_large_w1(**kwargs):
"""
MobileNetV3 Large 224/1.0 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.keras/models'
Location for keeping the model parameters.
"""
return get_mobilenetv3(version="large", width_scale=1.0, model_name="mobilenetv3_large_w1", **kwargs)
def mobilenetv3_large_w5d4(**kwargs):
"""
MobileNetV3 Large 224/1.25 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.keras/models'
Location for keeping the model parameters.
"""
return get_mobilenetv3(version="large", width_scale=1.25, model_name="mobilenetv3_large_w5d4", **kwargs)
def _test():
import numpy as np
import keras
pretrained = False
models = [
mobilenetv3_small_w7d20,
mobilenetv3_small_wd2,
mobilenetv3_small_w3d4,
mobilenetv3_small_w1,
mobilenetv3_small_w5d4,
mobilenetv3_large_w7d20,
mobilenetv3_large_wd2,
mobilenetv3_large_w3d4,
mobilenetv3_large_w1,
mobilenetv3_large_w5d4,
]
for model in models:
net = model(pretrained=pretrained)
# net.summary()
weight_count = keras.utils.layer_utils.count_params(net.trainable_weights)
print("m={}, {}".format(model.__name__, weight_count))
assert (model != mobilenetv3_small_w7d20 or weight_count == 2159600)
assert (model != mobilenetv3_small_wd2 or weight_count == 2288976)
assert (model != mobilenetv3_small_w3d4 or weight_count == 2581312)
assert (model != mobilenetv3_small_w1 or weight_count == 2945288)
assert (model != mobilenetv3_small_w5d4 or weight_count == 3643632)
assert (model != mobilenetv3_large_w7d20 or weight_count == 2943080)
assert (model != mobilenetv3_large_wd2 or weight_count == 3334896)
assert (model != mobilenetv3_large_w3d4 or weight_count == 4263496)
assert (model != mobilenetv3_large_w1 or weight_count == 5481752)
assert (model != mobilenetv3_large_w5d4 or weight_count == 7459144)
if is_channels_first():
x = np.zeros((1, 3, 224, 224), np.float32)
else:
x = np.zeros((1, 224, 224, 3), np.float32)
y = net.predict(x)
assert (y.shape == (1, 1000))
if __name__ == "__main__":
_test()
