def get_model(input_shape, class_num):
args = parse_args()
if args.model == 'resnet50':
model = ResNet50(include_top=True,
weights=None,
input_tensor=None,
input_shape=input_shape,
pooling=None,
mode=args.mode,
classes=class_num)
elif args.model == 'resnet50_imagenet':
model = ResNet50(include_top=True,
weights='imagenet',
input_tensor=None,
input_shape=input_shape,
pooling=None,
mode=args.mode,
classes=class_num)
elif args.model == 'pretrained':
modelName = os.path.join(os.getcwd(), 'weights', args.pretrained)
model = load_model(modelName)
rmsprop = RMSprop(lr=1e-05, rho=0.99, epsilon=1e-08, decay=0.001)
model.compile(optimizer=rmsprop,
loss='categorical_crossentropy',
metrics=['categorical_accuracy'])
print model.summary()
return model
def calc_features(input_path, output_path, n_iterations=100, overwrite=True):
model = ResNet50(weights='imagenet')
layer_name = 'avg_pool'
intermediate_layer_model = Model(input=model.input,
output=model.get_layer(layer_name).output)
i = 0
for folder in glob(input_path + '*'):
print(i)
if i >= n_iterations:
break
output_basename = os.path.basename(os.path.normpath(folder))
output_name = output_path + output_basename + "_features"
if not overwrite and os.path.exists(output_name + ".npy"):
i += 1
continue
batch = get_data_id(folder)
img = np.ndarray([len(batch), 3, 224, 224], dtype=np.float32)
img = batch
intermediate_output = intermediate_layer_model.predict(img,
batch_size=20)
np.save(output_name, intermediate_output)
i += 1
def finetuned_resnet(include_top, weights_dir):
"""
:param include_top: True for training, False for generating intermediate results for
LSTM cell
:param weights_dir: path to load finetune_resnet.h5
:return:
"""
base_model = ResNet50(include_top=False,
weights='imagenet',
input_shape=IMSIZE)
for layer in base_model.layers:
layer.trainable = False
x = base_model.output
x = Flatten()(x)
x = Dense(2048, activation='relu')(x)
x = Dropout(0.5)(x)
x = Dense(1024, activation='relu')(x)
x = Dropout(0.5)(x)
if include_top:
x = Dense(N_CLASSES, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=x)
if os.path.exists(weights_dir):
model.load_weights(weights_dir, by_name=True)
return model
def model_fn(FLAGS, objective, optimizer, metrics):
"""
pre-trained resnet50 model
"""
input_layer = keras.layers.Input(shape=(FLAGS.input_size, FLAGS.input_size,
3))
base_model = ResNet50(weights="imagenet",
include_top=False,
pooling=None,
input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
classes=FLAGS.num_classes,
input_layer=input_layer)
base_model2 = Inceptionv3(weights="imagenet",
include_top=False,
pooling=None,
input_shape=(FLAGS.input_size, FLAGS.input_size,
3),
classes=FLAGS.num_classes,
input_layer=input_layer)
x = base_model.output
x2 = base_model2.output
x = keras.layers.GlobalAveragePooling2D()(x)
x2 = keras.layers.GlobalAveragePooling2D()(x2)
t = keras.layers.Concatenate(axis=1)([x, x2])
t = keras.layers.Dense(500, use_bias=True, activation='relu')(t)
t = keras.layers.Dropout(0.3)(t)
predictions = Dense(FLAGS.num_classes, activation='softmax')(t)
model = Model(inputs=base_model.input, outputs=predictions)
model.compile(loss=objective, optimizer=optimizer, metrics=metrics)
return model
def model_fn(FLAGS, objective, optimizer, metrics):
inputs_dim = Input(shape=(FLAGS.input_size, FLAGS.input_size, 3))
x = ResNet50(weights="imagenet",
include_top=False,
pooling=max,
input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
classes=FLAGS.num_classes)(inputs_dim)
squeeze = GlobalAveragePooling2D()(x)
excitation = Dense(units=2048 // 16)(squeeze)
excitation = Activation('relu')(excitation)
excitation = Dense(units=2048)(excitation)
excitation = Activation('sigmoid')(excitation)
excitation = Reshape((1, 1, 2048))(excitation)
scale = multiply([x, excitation])
x = GlobalAveragePooling2D()(scale)
# x = Dropout(0.3)(x)
fc2 = Dense(FLAGS.num_classes)(x)
fc2 = Activation('sigmoid')(fc2) #此处注意，为sigmoid函数
model = Model(inputs=inputs_dim, outputs=fc2)
# model.load_weights('/home/work/user-job-dir/src/SE-Xception.h5',by_name=True)
# model = load_model('/home/work/user-job-dir/src/SE-Xception.h5')
model.compile(loss=objective, optimizer=optimizer, metrics=metrics)
return model
def get_by_name(model_name, *args, **kwargs):
model = None
if model_name == 'resnet50':
from models.resnet50 import ResNet50
model = ResNet50(*args, **kwargs)
else:
raise ValueError("Model %s not recognized." % model_name)
print("Model %s was created" % model.name)
return model
def model_fn(FLAGS, objective, optimizer, metrics):
base_model = ResNet50(weights="imagenet",
include_top=False,
pooling=None,
input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
classes=FLAGS.num_classes)
base_model = multi_gpu_model(base_model, 4)
for layer in base_model.layers:
layer.trainable = False
x = base_model.output
x = Flatten()(x)
predictions = Dense(FLAGS.num_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
model.compile(loss=objective, optimizer=optimizer, metrics=metrics)
return model
def model_fn(FLAGS, objective, optimizer, metrics):
"""
pre-trained resnet50 model
"""
base_model = ResNet50(weights="imagenet",
include_top=False,
pooling='avg',
input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
classes=FLAGS.num_classes)
# for layer in base_model.layers:
# layer.trainable = False
x = base_model.output
x = Dropout(0.5)(x)
predictions = Dense(FLAGS.num_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
model.compile(loss=objective, optimizer=optimizer, metrics=metrics)
return model
def model_fn(FLAGS, objective, optimizer, metrics):
"""
pre-trained resnet50 model
"""
base_model = ResNet50(weights="imagenet",
include_top=False,
pooling=None,
input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
classes=FLAGS.num_classes)
for layer in base_model.layers:
layer.trainable = False
x = base_model.output
x = Flatten()(x) # 把多维的输入一维化，常用在从卷积层到全连接层的过渡
predictions = Dense(FLAGS.num_classes, activation='softmax')(x) # 全连接神经网络层
model = Model(inputs=base_model.input, outputs=predictions) # 定义模型的输入输出
model.compile(loss=objective, optimizer=optimizer,
metrics=metrics) # 编译创建好的模型
return model
def load_model(model_name):
print("Loading {} model...".format(model_name))
if model_name == "resnet50":
model = ResNet50(weights='imagenet')
elif model_name == "inceptionv3":
model = InceptionV3(weights='imagenet')
elif model_name == "vgg19":
model = VGG19(weights='imagenet')
elif model_name == "vgg16":
model = VGG16(weights='imagenet')
elif model_name == "alexnet":
sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)
model = convnet('alexnet',
weights_path="../alexnet/alexnet_weights.h5",
heatmap=False)
model.compile(optimizer=sgd, loss='mse')
else:
raise ValueError("Wrong model name")
print("Loaded!")
return model, model_name
def _load_model(model_name, include_top=True, weights='imagenet'):
if model_name == 'resnet50':
from models.resnet50 import ResNet50
model = ResNet50(weights=weights, include_top=include_top)
elif model_name == 'vgg16':
from models.vgg16 import VGG16
model = VGG16(weights=weights, include_top=include_top)
if weights == 'hybrid1365':
model.load_weights(join(weights_dir,
'vgg16_hybrid1365_weights.h5'))
elif weights == 'places365':
model.load_weights(join(weights_dir, 'vgg16_places365_weights.h5'))
elif model_name == 'vgg19':
from models.vgg19 import VGG19
model = VGG19(weights=weights, include_top=include_top)
elif model_name == 'inception':
from models.inception_v3 import InceptionV3
model = InceptionV3(weights=weights, include_top=include_top)
elif model_name == 'vgg16_hybrid1365':
from models.vgg16 import VGG16
model = VGG16(weights=weights, include_top=include_top)
model.load_weights(join(weights_dir, 'vgg16_hybrid1365_weights.h5'),
by_name=True)
elif model_name == 'vgg16_places365':
from models.vgg16 import VGG16
model = VGG16(weights=weights, include_top=include_top)
model.load_weights(join(weights_dir, 'vgg16_places365_weights.h5'),
by_name=True)
elif model_name == 'resnet152_hybrid1365':
from models.resnet import ResnetBuilder
model = ResnetBuilder.build_resnet_152((3, 224, 224), 100)
print(join(weights_dir, 'resnet152_hybrid1365.h5'))
model.load_weights(join(weights_dir, 'resnet152_hybrid1365.h5'),
by_name=True)
else:
print(
'Not a valid model. Valid models are vgg16, vgg19, resnet50, inception, vgg16_hybrid1365, and vgg16_places365'
)
sys.exit(1)
return model
def model_fn(FLAGS, objective, optimizer, metrics):
"""
pre-trained resnet50 model
"""
base_model = ResNet50(weights="imagenet",
include_top=False,
pooling=None,
input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
classes=FLAGS.num_classes)
for layer in base_model.layers:
layer.trainable = False
x = base_model.output
x = Flatten()(x)
x = Dense(256,
activation='sigmoid',
kernel_regularizer=regularizers.l1(0.0001))(x)
x = Dropout(rate=0.3)(x)
predictions = Dense(FLAGS.num_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
model.compile(loss=objective, optimizer=optimizer, metrics=metrics)
return model
def model_fn(FLAGS):
# K.set_learning_phase(0)
# setup model
base_model = ResNet50(weights="imagenet",
include_top=False,
pooling=None,
input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
classes=FLAGS.num_classes)
for layer in base_model.layers:
layer.trainable = False
# if FLAGS.mode == 'train':
# K.set_learning_phase(1)
model = add_new_last_layer(base_model, FLAGS.num_classes)
# print(model.summary())
# print(model.layers[84].name)
# exit()
# Adam = adam(lr=FLAGS.learning_rate,clipnorm=0.001)
model.compile(optimizer="adam",
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def model_fn(FLAGS, objective, optimizer, metrics, input_shape=None):
"""
pre-trained resnet50 model
"""
# base_model = keras.applications.resnet50.ResNet50(include_top=False, weights='imagenet', input_tensor=None, input_shape=(FLAGS.input_size, FLAGS.input_size, 3), pooling=None, classes=FLAGS.num_classes)
# base_model = ResNext50(weights=None,
# include_top=False,
# pooling=None,
# input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
# classes=FLAGS.num_classes)
# base_model = ResNet50(weights='imagenet',
# include_top=False,
# pooling=None,
# input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
# classes=FLAGS.num_classes)
input_shape = _obtain_input_shape(input_shape,
default_size=224,
min_size=32,
data_format=backend.image_data_format(),
require_flatten=True,
weights='imagenet')
# # inp=Input(shape=(FLAGS.batch_size,FLAGS.input_size,FLAGS.input_size,3))
inp = Input(input_shape)
base_model = EfficientNetB5(
# input_tensor=inp,
weights='/home/work/user-job-dir/src/weights/efficientnet-b5_notop.h5',
include_top=False,
pooling=None,
input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
classes=FLAGS.num_classes)
# base_model.load_weights('/home/work/user-job-dir/src/weights/efficientnet-b5_notop.h5')
base_model1 = InceptionV3(input_tensor=inp,
weights='imagenet',
include_top=False,
pooling=None,
input_shape=(FLAGS.input_size, FLAGS.input_size,
3),
classes=FLAGS.num_classes)
base_model2 = Xception(input_tensor=inp,
weights='imagenet',
include_top=False,
pooling=None,
input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
classes=FLAGS.num_classes)
base_model3 = ResNet50(input_tensor=inp,
weights='imagenet',
include_top=False,
pooling=None,
input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
classes=FLAGS.num_classes)
for layer in base_model1.layers:
layer.trainable = True
for layer in base_model2.layers:
layer.trainable = True
for layer in base_model3.layers:
layer.trainable = True
# x1 = base_model1.output
# x1 = Flatten()(x1)
inception = base_model1.output
xception = base_model2.output
densenet = base_model3.output
top1_model = GlobalMaxPooling2D(data_format='channels_last')(inception)
top2_model = GlobalMaxPooling2D(data_format='channels_last')(xception)
top3_model = GlobalMaxPooling2D(data_format='channels_last')(densenet)
top1_model = Flatten()(top1_model)
top2_model = Flatten()(top2_model)
top3_model = Flatten()(top3_model)
t = concatenate([top1_model, top2_model, top3_model], axis=1)
# # # t=Flatten()(t)
# top_model=Dense(512,activation='relu')(t)
# top_model=Dropout(rate=0.5)(top_model)
for layer in base_model.layers:
layer.trainable = True
out = base_model.output
out = Flatten()(out)
predictions = Dense(FLAGS.num_classes, activation='softmax')(out)
model = Model(inputs=base_model.input, outputs=predictions)
model.compile(loss=objective, optimizer=optimizer, metrics=metrics)
return model
def __init__(self):
"""Creates the VGGFace2 neural network, MTCNN and dlib face detectors."""
self.vggface = ResNet50(n_classes=8631)
self.vggface.load_weights(WEIGHTS)
self.mtcnn = MTCNN()
self.dlib = Dlib()
def init_resnet(self):
from models.resnet50 import ResNet50
self.base_model = ResNet50(weights='imagenet', include_top=False)
self.input_shape = 224
self.feature_dim = 2048
self.output_shape = self.base_model.output_shape