def buildin_models(name,
dropout=1,
emb_shape=512,
input_shape=(112, 112, 3),
output_layer="GDC",
bn_momentum=0.99,
bn_epsilon=0.001,
add_pointwise_conv=False,
**kwargs):
name_lower = name.lower()
""" Basic model """
if name_lower == "mobilenet":
xx = keras.applications.MobileNet(input_shape=input_shape,
include_top=False,
weights="imagenet",
**kwargs)
elif name_lower == "mobilenetv2":
xx = keras.applications.MobileNetV2(input_shape=input_shape,
include_top=False,
weights="imagenet",
**kwargs)
elif name_lower == "r34" or name_lower == "r50" or name_lower == "r100" or name_lower == "r101":
from backbones import resnet # MXNet insightface version resnet
model_name = "ResNet" + name_lower[1:]
model_class = getattr(resnet, model_name)
xx = model_class(input_shape=input_shape,
include_top=False,
weights=None,
**kwargs)
elif name_lower.startswith("resnet"): # keras.applications.ResNetxxx
if name_lower.endswith("v2"):
model_name = "ResNet" + name_lower[len("resnet"):-2] + "V2"
else:
model_name = "ResNet" + name_lower[len("resnet"):]
model_class = getattr(keras.applications, model_name)
xx = model_class(weights="imagenet",
include_top=False,
input_shape=input_shape,
**kwargs)
elif name_lower.startswith("efficientnet"):
# import tensorflow.keras.applications.efficientnet as efficientnet
from backbones import efficientnet
model_name = "EfficientNet" + name_lower[-2:].upper()
model_class = getattr(efficientnet, model_name)
xx = model_class(weights="imagenet",
include_top=False,
input_shape=input_shape,
**kwargs) # or weights='imagenet'
elif name_lower.startswith("se_resnext"):
from keras_squeeze_excite_network import se_resnext
if name_lower.endswith("101"): # se_resnext101
depth = [3, 4, 23, 3]
else: # se_resnext50
depth = [3, 4, 6, 3]
xx = se_resnext.SEResNextImageNet(weights="imagenet",
input_shape=input_shape,
include_top=False,
depth=depth)
elif name_lower.startswith("resnest"):
from backbones import resnest
if name_lower == "resnest50":
xx = resnest.ResNest50(input_shape=input_shape)
else:
xx = resnest.ResNest101(input_shape=input_shape)
elif name_lower.startswith("mobilenetv3"):
from backbones import mobilenet_v3
model_class = mobilenet_v3.MobileNetV3Small if "small" in name_lower else mobilenet_v3.MobileNetV3Large
# from tensorflow.keras.layers.experimental.preprocessing import Rescaling
# model_class = keras.applications.MobileNetV3Small if "small" in name_lower else keras.applications.MobileNetV3Large
xx = model_class(input_shape=input_shape,
include_top=False,
weights="imagenet")
# xx = keras.models.clone_model(xx, clone_function=lambda layer: Rescaling(1.) if isinstance(layer, Rescaling) else layer)
elif "mobilefacenet" in name_lower or "mobile_facenet" in name_lower:
from backbones import mobile_facenet
use_se = True if "se" in name_lower else False
xx = mobile_facenet.mobile_facenet(input_shape=input_shape,
include_top=False,
name=name,
use_se=use_se)
elif name_lower == "ghostnet":
from backbones import ghost_model
xx = ghost_model.GhostNet(input_shape=input_shape,
include_top=False,
width=1.3)
elif hasattr(keras.applications, name):
model_class = getattr(keras.applications, name)
xx = model_class(weights="imagenet",
include_top=False,
input_shape=input_shape)
else:
return None
xx.trainable = True
inputs = xx.inputs[0]
nn = xx.outputs[0]
if add_pointwise_conv: # Model using `pointwise_conv + GDC` / `pointwise_conv + E` is smaller than `E`
nn = keras.layers.Conv2D(512, 1, use_bias=False, padding="same")(nn)
nn = keras.layers.BatchNormalization(momentum=bn_momentum,
epsilon=bn_epsilon)(nn)
nn = keras.layers.PReLU(shared_axes=[1, 2])(nn)
if output_layer == "E":
""" Fully Connected """
nn = keras.layers.BatchNormalization(momentum=bn_momentum,
epsilon=bn_epsilon)(nn)
if dropout > 0 and dropout < 1:
nn = keras.layers.Dropout(dropout)(nn)
nn = keras.layers.Flatten()(nn)
nn = keras.layers.Dense(emb_shape,
activation=None,
use_bias=True,
kernel_initializer="glorot_normal")(nn)
else:
""" GDC """
nn = keras.layers.DepthwiseConv2D(int(nn.shape[1]),
depth_multiplier=1,
use_bias=False)(nn)
nn = keras.layers.BatchNormalization(momentum=bn_momentum,
epsilon=bn_epsilon)(nn)
if dropout > 0 and dropout < 1:
nn = keras.layers.Dropout(dropout)(nn)
nn = keras.layers.Conv2D(emb_shape,
1,
use_bias=True,
activation=None,
kernel_initializer="glorot_normal")(nn)
nn = keras.layers.Flatten()(nn)
# nn = keras.layers.Dense(emb_shape, activation=None, use_bias=True, kernel_initializer="glorot_normal")(nn)
# `fix_gamma=True` in MXNet means `scale=False` in Keras
embedding = keras.layers.BatchNormalization(momentum=bn_momentum,
epsilon=bn_epsilon,
name="embedding",
scale=False)(nn)
basic_model = keras.models.Model(inputs, embedding, name=xx.name)
return basic_model
def buildin_models(name,
dropout=1,
emb_shape=512,
input_shape=(112, 112, 3),
**kwargs):
name = name.lower()
""" Basic model """
if name == "mobilenet":
xx = keras.applications.MobileNet(input_shape=input_shape,
include_top=False,
weights=None,
**kwargs)
elif name == "mobilenetv2":
xx = keras.applications.MobileNetV2(input_shape=input_shape,
include_top=False,
weights=None,
**kwargs)
elif name == "resnet50v2":
xx = keras.applications.ResNet50V2(input_shape=input_shape,
include_top=False,
weights="imagenet",
**kwargs)
elif name == "resnet101v2":
xx = keras.applications.ResNet101V2(input_shape=input_shape,
include_top=False,
weights="imagenet",
**kwargs)
elif name == "nasnetmobile":
xx = keras.applications.NASNetMobile(input_shape=input_shape,
include_top=False,
weights=None,
**kwargs)
elif name.startswith("efficientnet"):
if "-dev" in tf.__version__:
# import tensorflow.keras.applications.efficientnet as efntf
import efficientnet.tfkeras as efntf
else:
import efficientnet.tfkeras as efntf
if name[-2] == "b":
compound_scale = int(name[-1])
models = [
efntf.EfficientNetB0,
efntf.EfficientNetB1,
efntf.EfficientNetB2,
efntf.EfficientNetB3,
efntf.EfficientNetB4,
efntf.EfficientNetB5,
efntf.EfficientNetB6,
efntf.EfficientNetB7,
]
model = models[compound_scale]
else:
model = efntf.EfficientNetL2
xx = model(weights="noisy-student",
include_top=False,
input_shape=input_shape) # or weights='imagenet'
elif name.startswith("se_resnext"):
from keras_squeeze_excite_network import se_resnext
if name.endswith("101"): # se_resnext101
depth = [3, 4, 23, 3]
else: # se_resnext50
depth = [3, 4, 6, 3]
xx = se_resnext.SEResNextImageNet(weights="imagenet",
input_shape=input_shape,
include_top=False,
depth=depth)
elif name.startswith("resnest"):
from backbones import resnest
if name == "resnest50":
xx = resnest.ResNest50(input_shape=input_shape)
else:
xx = resnest.ResNest101(input_shape=input_shape)
elif name.startswith("mobilenetv3"):
from backbones import mobilenetv3
size = 'small' if 'small' in name else 'large'
xx = mobilenetv3.MobilenetV3(input_shape=input_shape,
include_top=False,
size=size)
elif "mobilefacenet" in name or "mobile_facenet" in name:
from backbones import mobile_facenet
use_se = True if 'se' in name else False
xx = mobile_facenet.mobile_facenet(input_shape=input_shape,
include_top=False,
name=name,
use_se=use_se)
else:
return None
# xx = keras.models.load_model('checkpoints/mobilnet_v1_basic_922667.h5', compile=False)
xx.trainable = True
inputs = xx.inputs[0]
nn = xx.outputs[0]
# nn = keras.layers.Conv2D(emb_shape, xx.output_shape[1], use_bias=False)(nn)
""" GDC """
nn = keras.layers.Conv2D(512, 1, use_bias=False)(nn)
nn = keras.layers.BatchNormalization()(nn)
nn = keras.layers.PReLU(shared_axes=[1, 2])(nn)
nn = keras.layers.DepthwiseConv2D(int(nn.shape[1]),
depth_multiplier=1,
use_bias=False)(nn)
nn = keras.layers.BatchNormalization()(nn)
nn = keras.layers.Conv2D(emb_shape, 1, use_bias=False, activation=None)(nn)
if dropout > 0 and dropout < 1:
nn = keras.layers.Dropout(dropout)(nn)
nn = keras.layers.Flatten()(nn)
# nn = keras.layers.Dense(emb_shape, activation=None, use_bias=False, kernel_initializer="glorot_normal")(nn)
embedding = keras.layers.BatchNormalization(name="embedding")(nn)
# norm_emb = layers.Lambda(tf.nn.l2_normalize, name='norm_embedding', arguments={'axis': 1})(embedding)
basic_model = keras.models.Model(inputs, embedding, name=xx.name)
return basic_model
def buildin_models(name,
dropout=1,
emb_shape=512,
input_shape=(112, 112, 3),
output_layer="GDC",
bn_momentum=0.99,
bn_epsilon=0.001,
**kwargs):
name = name.lower()
""" Basic model """
if name == "mobilenet":
xx = keras.applications.MobileNet(input_shape=input_shape,
include_top=False,
weights="imagenet",
**kwargs)
elif name == "mobilenetv2":
xx = keras.applications.MobileNetV2(input_shape=input_shape,
include_top=False,
weights="imagenet",
**kwargs)
elif name == "resnet34":
from backbones import resnet
xx = resnet.ResNet34(input_shape=input_shape,
include_top=False,
weights=None,
**kwargs)
elif name == "r50":
from backbones import resnet
xx = resnet.ResNet50(input_shape=input_shape,
include_top=False,
weights=None,
**kwargs)
elif name == "resnet50":
xx = keras.applications.ResNet50(input_shape=input_shape,
include_top=False,
weights="imagenet",
**kwargs)
elif name == "resnet50v2":
xx = keras.applications.ResNet50V2(input_shape=input_shape,
include_top=False,
weights="imagenet",
**kwargs)
elif name == "resnet101":
# xx = ResNet101(input_shape=input_shape, include_top=False, weights=None, **kwargs)
xx = keras.applications.ResNet101(input_shape=input_shape,
include_top=False,
weights="imagenet",
**kwargs)
elif name == "resnet101v2":
xx = keras.applications.ResNet101V2(input_shape=input_shape,
include_top=False,
weights="imagenet",
**kwargs)
elif name == "nasnetmobile":
xx = keras.applications.NASNetMobile(input_shape=input_shape,
include_top=False,
weights=None,
**kwargs)
elif name.startswith("efficientnet"):
if "-dev" in tf.__version__:
# import tensorflow.keras.applications.efficientnet as efficientnet
from backbones import efficientnet
else:
# import efficientnet.tfkeras as efficientnet
from backbones import efficientnet
if name[-2] == "b":
compound_scale = int(name[-1])
models = [
efficientnet.EfficientNetB0,
efficientnet.EfficientNetB1,
efficientnet.EfficientNetB2,
efficientnet.EfficientNetB3,
efficientnet.EfficientNetB4,
efficientnet.EfficientNetB5,
efficientnet.EfficientNetB6,
efficientnet.EfficientNetB7,
]
model = models[compound_scale]
else:
model = efficientnet.EfficientNetL2
xx = model(weights="imagenet",
include_top=False,
input_shape=input_shape) # or weights='imagenet'
elif name.startswith("se_resnext"):
from keras_squeeze_excite_network import se_resnext
if name.endswith("101"): # se_resnext101
depth = [3, 4, 23, 3]
else: # se_resnext50
depth = [3, 4, 6, 3]
xx = se_resnext.SEResNextImageNet(weights="imagenet",
input_shape=input_shape,
include_top=False,
depth=depth)
elif name.startswith("resnest"):
from backbones import resnest
if name == "resnest50":
xx = resnest.ResNest50(input_shape=input_shape)
else:
xx = resnest.ResNest101(input_shape=input_shape)
elif name.startswith("mobilenetv3"):
from backbones import mobilenet_v3
# size = "small" if "small" in name else "large"
# xx = mobilenetv3.MobilenetV3(input_shape=input_shape, include_top=False, size=size)
if "small" in name:
xx = mobilenet_v3.MobileNetV3Small(input_shape=input_shape,
include_top=False,
weights="imagenet")
else:
xx = mobilenet_v3.MobileNetV3Large(input_shape=input_shape,
include_top=False,
weights="imagenet")
elif "mobilefacenet" in name or "mobile_facenet" in name:
from backbones import mobile_facenet
use_se = True if "se" in name else False
xx = mobile_facenet.mobile_facenet(input_shape=input_shape,
include_top=False,
name=name,
use_se=use_se)
else:
return None
# xx = keras.models.load_model('checkpoints/mobilnet_v1_basic_922667.h5', compile=False)
xx.trainable = True
inputs = xx.inputs[0]
nn = xx.outputs[0]
# nn = keras.layers.Conv2D(emb_shape, xx.output_shape[1], use_bias=False)(nn)
if output_layer == "E":
""" Fully Connected """
nn = keras.layers.BatchNormalization(momentum=bn_momentum,
epsilon=bn_epsilon)(nn)
if dropout > 0 and dropout < 1:
nn = keras.layers.Dropout(dropout)(nn)
nn = keras.layers.Flatten()(nn)
nn = keras.layers.Dense(emb_shape,
activation=None,
use_bias=True,
kernel_initializer="glorot_normal")(nn)
else:
""" GDC """
# nn = keras.layers.Conv2D(512, 1, use_bias=False)(nn)
# nn = keras.layers.BatchNormalization(momentum=bn_momentum, epsilon=bn_epsilon)(nn)
# nn = keras.layers.PReLU(shared_axes=[1, 2])(nn)
nn = keras.layers.DepthwiseConv2D(int(nn.shape[1]),
depth_multiplier=1,
use_bias=False)(nn)
nn = keras.layers.BatchNormalization(momentum=bn_momentum,
epsilon=bn_epsilon)(nn)
if dropout > 0 and dropout < 1:
nn = keras.layers.Dropout(dropout)(nn)
nn = keras.layers.Conv2D(emb_shape,
1,
use_bias=False,
activation=None,
kernel_initializer="glorot_normal")(nn)
nn = keras.layers.Flatten()(nn)
# nn = keras.layers.Dense(emb_shape, activation=None, use_bias=True, kernel_initializer="glorot_normal")(nn)
# `fix_gamma=True` in MXNet means `scale=False` in Keras
embedding = keras.layers.BatchNormalization(momentum=bn_momentum,
epsilon=bn_epsilon,
name="embedding",
scale=False)(nn)
basic_model = keras.models.Model(inputs, embedding, name=xx.name)
return basic_model
def __init_model_from_name__(name, input_shape=(112, 112, 3), weights="imagenet", **kwargs):
name_lower = name.lower()
""" Basic model """
if name_lower == "mobilenet":
xx = keras.applications.MobileNet(input_shape=input_shape, include_top=False, weights=weights, **kwargs)
elif name_lower == "mobilenet_m1":
from backbones import mobilenet
xx = mobilenet.MobileNet(input_shape=input_shape, include_top=False, weights=None, **kwargs)
elif name_lower == "mobilenetv2":
xx = keras.applications.MobileNetV2(input_shape=input_shape, include_top=False, weights=weights, **kwargs)
elif name_lower == "r34" or name_lower == "r50" or name_lower == "r100" or name_lower == "r101":
from backbones import resnet # MXNet insightface version resnet
model_name = "ResNet" + name_lower[1:]
model_class = getattr(resnet, model_name)
xx = model_class(input_shape=input_shape, include_top=False, weights=None, **kwargs)
elif name_lower.startswith("resnet"): # keras.applications.ResNetxxx
if name_lower.endswith("v2"):
model_name = "ResNet" + name_lower[len("resnet") : -2] + "V2"
else:
model_name = "ResNet" + name_lower[len("resnet") :]
model_class = getattr(keras.applications, model_name)
xx = model_class(weights=weights, include_top=False, input_shape=input_shape, **kwargs)
elif name_lower.startswith("efficientnet"):
# import tensorflow.keras.applications.efficientnet as efficientnet
from backbones import efficientnet
model_name = "EfficientNet" + name_lower[-2:].upper()
model_class = getattr(efficientnet, model_name)
xx = model_class(weights=weights, include_top=False, input_shape=input_shape, **kwargs) # or weights='imagenet'
elif name_lower.startswith("se_resnext"):
from keras_squeeze_excite_network import se_resnext
if name_lower.endswith("101"): # se_resnext101
depth = [3, 4, 23, 3]
else: # se_resnext50
depth = [3, 4, 6, 3]
xx = se_resnext.SEResNextImageNet(weights=weights, input_shape=input_shape, include_top=False, depth=depth)
elif name_lower.startswith("resnest"):
from backbones import resnest
if name_lower == "resnest50":
xx = resnest.ResNest50(input_shape=input_shape)
else:
xx = resnest.ResNest101(input_shape=input_shape)
elif name_lower.startswith("mobilenetv3"):
from backbones import mobilenet_v3
# from backbones import mobilenetv3 as mobilenet_v3
model_class = mobilenet_v3.MobileNetV3Small if "small" in name_lower else mobilenet_v3.MobileNetV3Large
# from tensorflow.keras.layers.experimental.preprocessing import Rescaling
# model_class = keras.applications.MobileNetV3Small if "small" in name_lower else keras.applications.MobileNetV3Large
xx = model_class(input_shape=input_shape, include_top=False, weights=weights)
# xx = keras.models.clone_model(xx, clone_function=lambda layer: Rescaling(1.) if isinstance(layer, Rescaling) else layer)
elif "mobilefacenet" in name_lower or "mobile_facenet" in name_lower:
from backbones import mobile_facenet
use_se = True if "se" in name_lower else False
xx = mobile_facenet.mobile_facenet(input_shape=input_shape, include_top=False, name=name, use_se=use_se)
elif name_lower == "ghostnet":
from backbones import ghost_model
xx = ghost_model.GhostNet(input_shape=input_shape, include_top=False, width=1.3, **kwargs)
elif name_lower.startswith("botnet"):
from backbones import botnet
model_name = "BotNet" + name_lower[len("botnet") :]
model_class = getattr(botnet, model_name)
xx = model_class(include_top=False, input_shape=input_shape, strides=1, **kwargs)
elif hasattr(keras.applications, name):
model_class = getattr(keras.applications, name)
xx = model_class(weights=weights, include_top=False, input_shape=input_shape, **kwargs)
else:
return None
xx.trainable = True
return xx