tf ResNeSt and RegNet

Introduction

Currently support tensorflow in

• ResNeSt
• RegNet

model only, no pertrain model for download. easy to read and modified.
welcome for using it, ask question, test it, find some bugs maybe.

ResNeSt based on offical github .

Usage

usage is simple:

from models.model_factory import get_model

input_shape = [224,244,3]
n_classes = 81
fc_activation = 'softmax'

model = get_model(model_name="ResNest50",input_shape=input_shape,n_classes=n_classes,
                verbose=False,fc_activation=fc_activation)
model.compile(optimizer='adam', loss=tf.keras.losses.BinaryCrossentropy())

if you want use Mish as activation (default is relu):

#it imporve the results, but come with high memory usage
model = get_model(model_name="ResNest50",input_shape=input_shape,n_classes=n_classes,
                verbose=False,fc_activation=fc_activation,active='mish')

Models

models now support:

ResNest50
ResNest101
ResNest200
ResNest269

RegNetX400
RegNetX1.6
RegNetY400
RegNetY1.6
AnyOther RegNetX/Y

RegNet

for RegNet, cause there are various version, you can easily set it by stage_depth,stage_width,stage_G.

#RegNetY600
model = get_model(model_name="RegNet",input_shape=input_shape,n_classes=n_classes,
                verbose=True,fc_activation=fc_activation,stage_depth=[1,3,7,4],
                stage_width=[48,112,256,608],stage_G=16,SEstyle_atten="SE")

#RegNetX600
model = get_model(model_name="RegNet",input_shape=input_shape,n_classes=n_classes,
                verbose=True,fc_activation=fc_activation,stage_depth=[1,3,5,7],
                stage_width=[48,96,240,528],stage_G=24,SEstyle_atten="noSE")

details seting (from orginal paper ):

• facebookresearch/pycls

Discussion

I compared ResNeSt50 and some RegNet(below 4.0GF) in my own project, also compared to EfficientNet b0/b1/b2. it seems EfficientNet is still good at balance in size/speed and accuracy, and ResNeSt50 performe well at accuarcy also lower in size/speed, And RegNet not that fast and acuracy not that good, seems normal.