godofpdog / mobilenetv3_keras

This is a keras implementation of MobileNetV3 architecture as described in the paper "Searching for MobileNetV3".

MobileNet V3

---

Requirments

numpy 1.15.1
keras 2.2.4
tensorflow-gpu 1.9.0
opencv-python 3.4.3.18
imgaug 0.2.8

Training

• You can define taining learning rate schedule by edit src/learning_rate_schedule.py.

python train.py -c config/train.ini

Arguments

• data
• model
• train
• gpu

data

Argument	Description	Type	Default
train	Training dataset directory.	str	None
valid	Validation dataset directory.	str	None

model

Argument	Description	Type	Default
input_size	Input size of MobileNet V3 model.	int	224
model_size	"large" or "small" version of MobileNet V3 model.	str	large
pooling_type	Pooling type of MobileNet V3 model. (avg or depthwith)	str	avg
num_classes	Number of classes.	int	1000

train

Argument	Description	Type	Default
epochs	Maximun number of training epochs.	int	200
batch_size	Batch size of data generator.	int	32
save_path	Saved weights path.	str	weights/*.h5
pretrained_path	Pre-trained model path of MobileNet V3 model.	str	None

gpu

Argument	Description	Type	Default
gpu	Specify a GPU.	str	-1

---

bottleneck structure configuation

• You can define custom bottleneck structure by edit large_config_list and small_config_list in MobileNet_V3.py

Argument	Description	Type	Code
out_dim	Output chennal dimension.	int	out
kernel	Kernel size of filter.	tuple	kernel
strides	Strides of the converlutional operation.	tuple	stride
expansion_dim	Expansion dimension of the bottleneck block.	int	exp
is_use_bias	Use bias or not.	bool	bias
res	Use shortcut operation or not.	bool	res
is_use_se	Use SE block or not.	bool	se
activation	Activative functions. ('RE' or 'HS')	str	active
num_layers	Layer index number.	int	id

example

# NOTE               out   kernel  stride  exp  bias   res    se     active id  
large_config_list = [[16,  (3, 3), (1, 1), 16,  False, False, False, 'RE',  0],
                     [24,  (3, 3), (2, 2), 64,  False, False, False, 'RE',  1],
                     [24,  (3, 3), (1, 1), 72,  False, True,  False, 'RE',  2],
                     [40,  (5, 5), (2, 2), 72,  False, False, True,  'RE',  3],
                     [40,  (5, 5), (1, 1), 120, False, True,  True,  'RE',  4],
                     [40,  (5, 5), (1, 1), 120, False, True,  True,  'RE',  5],
                     [80,  (3, 3), (2, 2), 240, False, False, False, 'HS',  6],
                     [80,  (3, 3), (1, 1), 200, False, True,  False, 'HS',  7],
                     [80,  (3, 3), (1, 1), 184, False, True,  False, 'HS',  8],
                     [80,  (3, 3), (1, 1), 184, False, True,  False, 'HS',  9],
                     [112, (3, 3), (1, 1), 480, False, False, True,  'HS', 10],
                     [112, (3, 3), (1, 1), 672, False, True,  True,  'HS', 11],
                     [160, (5, 5), (1, 1), 672, False, False, True,  'HS', 12],
                     [160, (5, 5), (2, 2), 672, False, True,  True,  'HS', 13],
                     [160, (5, 5), (1, 1), 960, False, True,  True,  'HS', 14]]

mobilenetv3_keras's Issues

tensorflow-gpu 1.9.0 not available

Hi godofpdog,

Will this work for any other versions of tensorflow-gpu?
ERROR: Could not find a version that satisfies the requirement tensorflow-gpu==1.9.0 (from versions: 2.2.0rc1, 2.2.0rc2, 2.2.0rc3, 2.2.0rc4, 2.2.0, 2.2.1, 2.3.0rc0, 2.3.0rc1, 2.3.0rc2, 2.3.0, 2.3.1)
ERROR: No matching distribution found for tensorflow-gpu==1.9.0

`Sanket

line 72 in .py file

x = __depthwise_block(_inputs, kernel=kernel, strides=strides, is_use_se=is_use_se, activation=activation)

should be

x = __depthwise_block(x, kernel=kernel, strides=strides, is_use_se=is_use_se, activation=activation)