def get_named_model_func(model_name):
if model_name == 'Xception':
return applications.xception.Xception(weights='imagenet', include_top=False, pooling='avg')
if model_name == 'VGG16':
return applications.vgg16.VGG16(weights='imagenet', include_top=False, pooling='avg')
if model_name == 'VGG19':
return applications.vgg19.VGG19(weights='imagenet', include_top=False, pooling='avg')
if model_name == 'InceptionV3':
return applications.inception_v3.InceptionV3(weights='imagenet', include_top=False, pooling='avg')
if model_name == 'InceptionResNetV2':
return applications.inception_resnet_v2.InceptionResNetV2(weights='imagenet', include_top=False, pooling='avg')
if model_name == 'MobileNet':
return applications.mobilenet.MobileNet(weights='imagenet', include_top=False, pooling='avg')
if model_name == 'ResNet50':
return applications.resnet50.ResNet50(weights='imagenet', include_top=False, pooling='avg')
if model_name == 'DenseNet201':
return applications.DenseNet201(weights='imagenet', include_top=False, pooling='avg')
raise ValueError('Unknown model')
def save_bottlebeck_features():
train_datagen = ImageDataGenerator(
rescale=1. / 255,
rotation_range=0.2,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True)
# train_datagen = ImageDataGenerator(rescale=1. / 255)
test_datagen = ImageDataGenerator(rescale=1. / 255)
# build the DenseNet201 network
model = applications.DenseNet201(include_top=False, weights='imagenet')
train_generator = train_datagen.flow_from_directory(
train_data_dir,
target_size=(img_width, img_height),
batch_size=batch_size,
class_mode=None,
shuffle=False)
bottleneck_features_train = model.predict_generator(
train_generator, aug_factor * nb_classes * nb_train_samples // batch_size)
train_labels = np.tile(np.repeat(np.arange(nb_classes), nb_train_samples), aug_factor)
np.savez(bn_train_path, data=bottleneck_features_train, label=train_labels)
test_generator = test_datagen.flow_from_directory(
validation_data_dir,
target_size=(img_width, img_height),
batch_size=batch_size,
class_mode=None,
shuffle=False)
bottleneck_features_validation = model.predict_generator(
test_generator, nb_classes * nb_validation_samples // batch_size)
validation_labels = np.repeat(np.arange(nb_classes), nb_validation_samples)
np.savez(bn_validation_path, data=bottleneck_features_validation, label=validation_labels)
def get_imagenet_cnn(architecture,
variant,
size,
alpha,
output_layer,
include_top=False,
weights='imagenet'):
from keras import applications, Model
if include_top:
assert output_layer == '*last*'
if size == '*auto*':
size = get_image_size(architecture, variant, size)
shape = (size, size, 3)
if architecture == 'densenet':
if variant == '*auto*':
variant = 'densenet-121'
if variant == 'densenet-121':
model = applications.DenseNet121(weights=weights,
include_top=include_top,
input_shape=shape)
elif variant == 'densenet-169':
model = applications.DenseNet169(weights=weights,
include_top=include_top,
input_shape=shape)
elif variant == 'densenet-201':
model = applications.DenseNet201(weights=weights,
include_top=include_top,
input_shape=shape)
elif architecture == 'inception-resnet-v2':
model = applications.InceptionResNetV2(weights=weights,
include_top=include_top,
input_shape=shape)
elif architecture == 'mobilenet':
model = applications.MobileNet(weights=weights,
include_top=include_top,
input_shape=shape,
alpha=alpha)
elif architecture == 'mobilenet-v2':
model = applications.MobileNetV2(weights=weights,
include_top=include_top,
input_shape=shape,
alpha=alpha)
elif architecture == 'nasnet':
if variant == '*auto*':
variant = 'large'
if variant == 'large':
model = applications.NASNetLarge(weights=weights,
include_top=include_top,
input_shape=shape)
else:
model = applications.NASNetMobile(weights=weights,
include_top=include_top,
input_shape=shape)
elif architecture == 'resnet-50':
model = applications.ResNet50(weights=weights,
include_top=include_top,
input_shape=shape)
elif architecture == 'vgg-16':
model = applications.VGG16(weights=weights,
include_top=include_top,
input_shape=shape)
elif architecture == 'vgg-19':
model = applications.VGG19(weights=weights,
include_top=include_top,
input_shape=shape)
elif architecture == 'xception':
model = applications.Xception(weights=weights,
include_top=include_top,
input_shape=shape)
elif architecture == 'inception-v3':
model = applications.InceptionV3(weights=weights,
include_top=include_top,
input_shape=shape)
if output_layer != '*last*':
try:
if isinstance(output_layer, int):
layer = model.layers[output_layer]
else:
layer = model.get_layer(output_layer)
except Exception:
raise VergeMLError('layer not found: {}'.format(output_layer))
model = Model(inputs=model.input, outputs=layer.output)
return model
def create_model(existing='', is_twohundred=False, is_halffeatures=True):
if len(existing) == 0:
print('Loading base model (DenseNet)..')
# Encoder Layers
if is_twohundred:
base_model = applications.DenseNet201(input_shape=(None, None, 3),
include_top=False)
else:
base_model = applications.DenseNet169(input_shape=(None, None, 3),
include_top=False)
print('Base model loaded.')
# Starting point for decoder
base_model_output_shape = base_model.layers[-1].output.shape
# Layer freezing?
for layer in base_model.layers:
layer.trainable = True
# Starting number of decoder filters
if is_halffeatures:
decode_filters = int(int(base_model_output_shape[-1]) / 2)
else:
decode_filters = int(base_model_output_shape[-1])
# Define upsampling layer
def upproject(tensor, filters, name, concat_with):
up_i = BilinearUpSampling2D((2, 2),
name=name + '_upsampling2d')(tensor)
up_i = Concatenate(name=name + '_concat')(
[up_i,
base_model.get_layer(concat_with).output]) # Skip connection
up_i = Conv2D(filters=filters,
kernel_size=3,
strides=1,
padding='same',
name=name + '_convA')(up_i)
up_i = LeakyReLU(alpha=0.2)(up_i)
up_i = Conv2D(filters=filters,
kernel_size=3,
strides=1,
padding='same',
name=name + '_convB')(up_i)
up_i = LeakyReLU(alpha=0.2)(up_i)
return up_i
# Decoder Layers
decoder = Conv2D(filters=decode_filters,
kernel_size=1,
padding='same',
input_shape=base_model_output_shape,
name='conv2')(base_model.output)
decoder = upproject(decoder,
int(decode_filters / 2),
'up1',
concat_with='pool3_pool')
decoder = upproject(decoder,
int(decode_filters / 4),
'up2',
concat_with='pool2_pool')
decoder = upproject(decoder,
int(decode_filters / 8),
'up3',
concat_with='pool1')
decoder = upproject(decoder,
int(decode_filters / 16),
'up4',
concat_with='conv1/relu')
if False:
decoder = upproject(decoder,
int(decode_filters / 32),
'up5',
concat_with='input_1')
# Extract depths (final layer)
conv3 = Conv2D(filters=1,
kernel_size=3,
strides=1,
padding='same',
name='conv3')(decoder)
# Create the model
model = Model(inputs=base_model.input, outputs=conv3)
else:
# Load model from file
if not existing.endswith('.h5'):
sys.exit(
'Please provide a correct model file when using [existing] argument.'
)
custom_objects = {
'BilinearUpSampling2D': BilinearUpSampling2D,
'depth_loss_function': depth_loss_function
}
model = load_model(existing, custom_objects=custom_objects)
print('\nExisting model loaded.\n')
print('Model created.')
return model
def train(train_path,
val_path,
test_path,
batch_size=32,
epochs=50,
network='InceptionResNetV2',
data_augmentation=True,
mode='finetune',
optimizer='Adadelta',
fc=1,
classes=5,
gpu=1):
'''
Inputs:
train_path: data path for train set (data should be stored like train/DR, train/Normal)
val_path: data path for validation set
test_path: data path for test set
batch_size: data sizes per step
epochs: loop counts over whole train set
network: {
'InceptionResNetV2': fine-tune mode will train last 2 inception blocks
'DenseNet201': fine-tune mode will train last Dense block
'InceptionV3': fine-tune mode will train last 2 inception blocks
'Xception'
'NASNet'
'MobileNetV2'
'ResNet50': According to https://arxiv.org/pdf/1805.08974.pdf, it is most suitable for transfer learning?
}
data_augmentation: whether to do data augmentation or not
mode: {
'retrain': randomly initialize all layers and retrain the whole model
'finetune': train specified layers
'transfer' train fc layer(s)
}
optimizer: {
'Adadelta'
'RMSprop'
}
fc: {
1: only one fc layer at last
2: include two fc layers at last
}
classes: category counts
'''
if mode == 'retrain':
include_top = False
weights = None
pooling = 'avg'
else:
include_top = False
weights = 'imagenet'
pooling = 'avg'
if network == 'DenseNet201':
from keras.applications.densenet import preprocess_input
img_width, img_height = 224, 224
base_model = applications.DenseNet201(include_top=include_top,
weights=weights,
pooling=pooling)
# train last Dense Block
if mode == 'finetune':
trainable = False
for layer in base_model.layers:
if layer.name == 'conv5_block1_0_bn':
trainable = True
layer.trainable = trainable
if network == 'Xception':
from keras.applications.xception import preprocess_input
img_width, img_height = 299, 299
base_model = applications.Xception(include_top=include_top,
weights=weights,
pooling=pooling)
if network == 'InceptionV3':
from keras.applications.inception_v3 import preprocess_input
img_width, img_height = 299, 299
base_model = applications.InceptionV3(include_top=include_top,
weights=weights,
pooling=pooling)
# train top 2 inception blocks
if mode == 'finetune':
for layer in base_model.layers[:249]:
layer.trainable = False
for layer in base_model.layers[249:]:
#print(layer.name)
layer.trainable = True
if network == 'InceptionResNetV2':
from keras.applications.inception_resnet_v2 import preprocess_input
img_width, img_height = 299, 299
base_model = applications.InceptionResNetV2(include_top=include_top,
weights=weights,
pooling=pooling)
# train top 1 inception blocks
if mode == 'finetune':
trainable = True
for layer in base_model.layers:
#print(layer.name)
if layer.name == 'conv2d_9':
trainable = False
if layer.name == 'conv2d_201':
trainable = True
layer.trainable = trainable
if network == 'NASNet':
from keras.applications.nasnet import preprocess_input
img_width, img_height = 331, 331
base_model = applications.NASNetLarge(include_top=include_top,
weights=weights,
pooling=pooling)
if network == 'MoblieNetV2':
from keras.applications.mobilenetv2 import preprocess_input
img_width, img_height = 224, 224
base_model = applications.MobileNetV2(include_top=include_top,
weights=weights,
pooling=pooling)
if network == 'ResNet50':
from keras.applications.resnet50 import preprocess_input
img_width, img_height = 224, 224
base_model = applications.ResNet50(include_top=include_top,
weights=weights,
pooling=pooling)
bottleneck = base_model.output
if fc == 2:
bottleneck = Dense(
512,
activation='relu',
kernel_regularizer=keras.regularizers.l2(l=0.001))(bottleneck)
predictions = Dense(
classes,
kernel_regularizer=keras.regularizers.l2(l=0.001),
activation='softmax',
bias_regularizer=keras.regularizers.l2(l=0.001))(bottleneck)
model = Model(inputs=base_model.input, outputs=predictions)
if mode == 'transfer':
# train only the top layers (which were randomly initialized)
# freeze all convolutional layers
for layer in base_model.layers:
layer.trainable = False
if mode == 'retrain':
# train a complete model
for layer in base_model.layers:
layer.trainable = True
if optimizer == 'Adadelta':
opt = optimizers.Adadelta()
if optimizer == 'Adam':
opt = optimizers.Adam()
if optimizer == 'RMSprop':
opt = optimizers.RMSprop(lr=0.005, rho=0.9, epsilon=1.0, decay=0.94)
if gpu > 1:
batch_size *= gpu
model = multi_gpu_model(model, gpus=gpu)
model.compile(optimizer=opt,
loss='categorical_crossentropy',
metrics=['accuracy'])
if data_augmentation:
# Initialize the train and test generators with data Augumentation
train_datagen = ImageDataGenerator(
preprocessing_function=preprocess_input,
horizontal_flip=True,
fill_mode="nearest",
zoom_range=0.3,
width_shift_range=0.3,
height_shift_range=0.3,
rotation_range=30)
val_datagen = ImageDataGenerator(
preprocessing_function=preprocess_input)
else:
train_datagen = ImageDataGenerator(
preprocessing_function=preprocess_input)
val_datagen = ImageDataGenerator(
preprocessing_function=preprocess_input)
train_generator = train_datagen.flow_from_directory(
train_path,
target_size=(img_height, img_width),
batch_size=batch_size,
class_mode="categorical")
validation_generator = val_datagen.flow_from_directory(
val_path,
target_size=(img_height, img_width),
class_mode="categorical")
test_generator = val_datagen.flow_from_directory(test_path,
target_size=(img_height,
img_width),
class_mode="categorical")
checkpoint = ModelCheckpoint("{}_{}_{}.h5".format(network, mode,
optimizer),
monitor='val_acc',
verbose=1,
save_best_only=True,
save_weights_only=False,
mode='auto',
period=1)
early = EarlyStopping(monitor='val_acc',
min_delta=0,
patience=10,
verbose=1,
mode='auto')
model.fit_generator(train_generator,
epochs=epochs,
validation_data=validation_generator,
callbacks=[checkpoint, early])
score = model.evaluate_generator(test_generator)
print(score)
input_shape=(256, 256, 3))
elif args.base_model == 'xception':
base_model = applications.Xception(weights='imagenet',
include_top=False,
input_shape=(256, 256, 3))
elif args.base_model == 'densenet121':
base_model = applications.DenseNet121(weights='imagenet',
include_top=False,
input_shape=(256, 256, 3))
elif args.base_model == 'densenet169':
base_model = applications.DenseNet169(weights='imagenet',
include_top=False,
input_shape=(256, 256, 3))
elif args.base_model == 'densenet201':
base_model = applications.DenseNet201(weights='imagenet',
include_top=False,
input_shape=(256, 256, 3))
elif args.base_model == 'nasnetmobile':
base_model = applications.NASNetMobile(weights='imagenet',
include_top=False,
input_shape=(256, 256, 3))
elif args.base_model == 'nasnetlarge':
base_model = applications.NASNetLarge(weights='imagenet',
include_top=False,
input_shape=(256, 256, 3))
else:
raise ValueError(
"Model you entered is not present in the model zoo thats offered")
if args.bottleneck_tensorname is None:
# Taking the last tensor (Just before the softmax layer)
def create_two_branch_model(existing='', is_twohundred=False, is_halffeatures=True, channels=4):
if len(existing) == 0:
print('Loading base model (DenseNet)..')
def crop(dimension, start, end):
# Crops (or slices) a Tensor on a given dimension from start to end
# example : to crop tensor x[:, :, 5:10]
# call slice(2, 5, 10) as you want to crop on the second dimension
def func(x):
if dimension == 0:
return x[start: end]
if dimension == 1:
return x[:, start: end]
if dimension == 2:
return x[:, :, start: end]
if dimension == 3:
return x[:, :, :, start: end]
if dimension == 4:
return x[:, :, :, :, start: end]
return Lambda(func)
input_rgbd = Input(shape=(None, None, channels), name='input_rgbd')
input_rgb = crop(3, 0, 3)(input_rgbd)
input_sparse = crop(3, 3, channels)(input_rgbd)
#base_model_sz_input = Conv2D(3, (3,3), padding='same')(input_sparse)
# Encoder Layers
if is_twohundred:
base_model = applications.DenseNet201(input_shape=(None, None, 3), include_top=False, weights='imagenet', input_tensor=input_rgb)
#base_model_sz = applications.DenseNet201(input_shape=(None, None, 3), include_top=False, weights='imagenet', input_tensor=concatenate([input_sparse, input_sparse, input_sparse], axis=-1))
base_model_sz = applications.DenseNet201(input_shape=(None, None, channels-3), include_top=False, weights=None, input_tensor=input_sparse)
else:
base_model = applications.DenseNet169(input_shape=(None, None, 3), include_top=False, weights='imagenet', input_tensor=input_rgb)
#base_model_sz = applications.DenseNet169(input_shape=(None, None, 3), include_top=False, weights='imagenet', input_tensor=concatenate([input_sparse, input_sparse, input_sparse], axis=-1))
base_model_sz = applications.DenseNet121(input_shape=(None, None, channels-3), include_top=False, weights=None, input_tensor=input_sparse)
pretrained = applications.DenseNet121(input_shape=(None, None, 3), include_top=False, weights='imagenet')
for layer in pretrained.layers:
print(layer.name)
if layer.get_weights() != []: # Skip input, pooling and no weights layers
target_layer = base_model_sz.get_layer(name=layer.name)
if layer.name != 'conv1/conv': # Initialize imagenet weights in all layers except the first conv1 layer where the channels do not match
target_layer.set_weights(layer.get_weights())
print('Base model loaded.')
# Layer freezing?
for layer in base_model.layers:
layer.trainable = True
for layer in base_model_sz.layers:
layer.trainable = True
layer.name = layer.name + str("_sz")
# Starting point for decoder
encoder_output = concatenate([base_model.output, base_model_sz.output], axis=-1)
base_model_output_shape = encoder_output.shape
#base_model_output_shape = base_model.layers[-1].output.shape
#base_model_output_shape_sz = base_model_sz.layers[-1].output.shape
# Starting number of decoder filters
if is_halffeatures:
decode_filters = int(int(base_model_output_shape[-1])/2)
else:
decode_filters = int(base_model_output_shape[-1])
# Define upsampling layer
def upproject(tensor, filters, name, concat_with):
up_i = BilinearUpSampling2D((2, 2), name=name+'_upsampling2d')(tensor)
up_i = Concatenate(name=name+'_concat')([up_i, base_model.get_layer(concat_with).output, base_model_sz.get_layer(concat_with+str("_sz")).output]) # Skip connection
up_i = Conv2D(filters=filters, kernel_size=3, strides=1, padding='same', name=name+'_convA')(up_i)
up_i = LeakyReLU(alpha=0.2)(up_i)
up_i = Conv2D(filters=filters, kernel_size=3, strides=1, padding='same', name=name+'_convB')(up_i)
up_i = LeakyReLU(alpha=0.2)(up_i)
return up_i
# Decoder Layers
decoder = Conv2D(filters=decode_filters, kernel_size=1, padding='same', input_shape=base_model_output_shape, name='conv2')(encoder_output)
decoder = upproject(decoder, int(decode_filters/2), 'up1', concat_with='pool3_pool')
decoder = upproject(decoder, int(decode_filters/4), 'up2', concat_with='pool2_pool')
decoder = upproject(decoder, int(decode_filters/8), 'up3', concat_with='pool1')
decoder = upproject(decoder, int(decode_filters/16), 'up4', concat_with='conv1/relu')
if False: decoder = upproject(decoder, int(decode_filters/32), 'up5', concat_with='input_1')
# Extract depths (final layer)
conv3 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv3')(decoder)
# Create the model
model = Model(inputs=input_rgbd, outputs=conv3)
else:
# Load model from file
if not existing.endswith('.h5'):
sys.exit('Please provide a correct model file when using [existing] argument.')
custom_objects = {'BilinearUpSampling2D': BilinearUpSampling2D, 'depth_loss_function': depth_loss_function}
model = load_model(existing, custom_objects=custom_objects, compile=False)
print('\nExisting model loaded.\n')
print('Model created.')
return model
def mySpatialModel(model_name,
spatial_size,
nb_classes,
channels,
weights_path=None):
input_tensor = Input(shape=(channels, spatial_size, spatial_size))
input_shape = (channels, spatial_size, spatial_size)
base_model = None
predictions = None
data_dim = 1024
if model_name == 'ResNet50':
input_tensor = Input(shape=(spatial_size, spatial_size, channels))
input_shape = (spatial_size, spatial_size, channels)
base_model = kerasApp.ResNet50(include_top=False,
input_tensor=input_tensor,
input_shape=input_shape,
weights=weights_path,
classes=nb_classes,
pooling=None)
x = base_model.output
# 添加自己的全链接分类层 method 1
#x = Flatten()(x)
#predictions = Dense(nb_classes, activation='softmax')(x)
#method 2
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(nb_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
elif model_name == 'VGG16':
input_tensor = Input(shape=(spatial_size, spatial_size, channels))
input_shape = (spatial_size, spatial_size, channels)
base_model = kerasApp.VGG16(include_top=False,
input_tensor=input_tensor,
input_shape=input_shape,
weights=weights_path,
classes=nb_classes,
pooling=None)
x = base_model.output
x = GlobalAveragePooling2D()(
x) # add a global spatial average pooling layer
x = Dense(1024,
activation='relu')(x) # let's add a fully-connected layer
predictions = Dense(nb_classes,
activation='softmax')(x) # and a logistic layer
model = Model(inputs=base_model.input, outputs=predictions)
elif model_name == 'VGG19':
base_model = kerasApp.VGG19(include_top=False,
input_tensor=input_tensor,
input_shape=input_shape,
weights=weights_path,
classes=2,
pooling=None)
x = base_model.output
# 添加自己的全链接分类层
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(nb_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
elif model_name == 'InceptionV3':
input_tensor = Input(shape=(spatial_size, spatial_size, channels))
input_shape = (spatial_size, spatial_size, channels)
base_model = kerasApp.InceptionV3(weights=weights_path,
include_top=False,
pooling=None,
input_shape=input_shape,
classes=nb_classes)
x = base_model.output
# 添加自己的全链接分类层
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(nb_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
elif model_name == 'InceptionResNetV2':
input_tensor = Input(shape=(spatial_size, spatial_size, channels))
input_shape = (
spatial_size,
spatial_size,
channels,
)
base_model = kerasApp.InceptionResNetV2(weights=weights_path,
include_top=False,
pooling=None,
input_shape=input_shape,
classes=nb_classes)
x = base_model.output
# 添加自己的全链接分类层
x = GlobalAveragePooling2D()(x)
data_dim = 1536
predictions = Dense(nb_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
elif model_name == 'Xception':
input_shape_xception = (spatial_size, spatial_size, channels)
base_model = kerasApp.Xception(weights=weights_path,
include_top=False,
pooling="avg",
input_shape=input_shape_xception,
classes=nb_classes)
x = base_model.output
predictions = Dense(nb_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
elif model_name == 'DenseNet121':
base_model = kerasApp.DenseNet121(weights=weights_path,
include_top=False,
pooling=None,
input_shape=input_shape,
classes=nb_classes)
x = base_model.output
# 添加自己的全链接分类层
x = GlobalAveragePooling2D()(x)
predictions = Dense(nb_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
elif model_name == 'DenseNet169':
base_model = kerasApp.DenseNet169(weights=weights_path,
include_top=False,
pooling=None,
input_shape=input_shape,
classes=nb_classes)
x = base_model.output
# 添加自己的全链接分类层
x = GlobalAveragePooling2D()(x)
predictions = Dense(nb_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
elif model_name == 'DenseNet201':
base_model = kerasApp.DenseNet201(weights=weights_path,
include_top=False,
pooling=None,
input_shape=input_shape,
classes=nb_classes)
x = base_model.output
# 添加自己的全链接分类层
x = GlobalAveragePooling2D()(x)
predictions = Dense(nb_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
elif model_name == 'MobileNet':
base_model = kerasApp.MobileNet(weights=weights_path,
include_top=False,
pooling=None,
input_shape=input_shape,
classes=nb_classes)
x = base_model.output
# 添加自己的全链接分类层
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
x = Dense(1024, activation='relu')(x)
x = Dense(512, activation='relu')(x)
data_dim = 512
predictions = Dense(nb_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
else:
print("this model--[" + model_name + "]-- doesnt exist!")
# 冻结base_model所有层，这样就可以正确获得bottleneck特征
for layer in base_model.layers:
layer.trainable = True
# 训练模型
model = Model(inputs=base_model.input, outputs=predictions)
print('-------------当前base_model模型[' + model_name +
"]-------------------\n")
print('base_model层数目:' + str(len(base_model.layers)))
print('model模型层数目:' + str(len(model.layers)))
featureLayer = model.layers[len(model.layers) - 2]
print(featureLayer.output_shape)
print("data_dim:" + str(featureLayer.output_shape[1]))
print("---------------------------------------------\n")
#sgd = SGD(lr=lr, decay=decay, momentum=momentum, nesterov=True)
# 绘制模型
#if plot_model:
# plot_model(model, to_file=model_name+'.png', show_shapes=True)
return model
def create_model(is_twohundred=False, is_halffeatures=True):
print('Loading base model (DenseNet)..')
# Encoder Layers
if is_twohundred:
base_model = applications.DenseNet201(input_shape=(480, 640, 3), include_top=False)
else:
base_model = applications.DenseNet169(input_shape=(480, 640, 3), include_top=False)
print('Base model loaded.')
# Starting point for decoder
base_model_output_shape = base_model.layers[-1].output.shape
# Layer freezing?
for layer in base_model.layers: layer.trainable = True
# Starting number of decoder filters
if is_halffeatures:
decode_filters = int(int(base_model_output_shape[-1]) / 2)
else:
decode_filters = int(base_model_output_shape[-1])
def expend_as(tensor, rep):
return layers.Lambda(lambda x, repnum: K.repeat_elements(x, repnum, axis=3),
arguments={'repnum': rep})(tensor)
# Define upsampling layer
def upproject(tensor, filters, name, concat_with):
up_i = BilinearUpSampling2D((2, 2), name=name + '_upsampling2d')(tensor)
up_i = Concatenate(name=name + '_concat')(
[up_i, concat_with]) # Skip connection
up_i = Conv2D(filters=filters, kernel_size=3, strides=1, padding='same', name=name + '_convA')(up_i)
up_i = LeakyReLU(alpha=0.2)(up_i)
up_i = Conv2D(filters=filters, kernel_size=3, strides=1, padding='same', name=name + '_convB')(up_i)
up_i = LeakyReLU(alpha=0.2)(up_i)
return up_i
def global_non_local(X,cc):
h, w , c = list(X.shape)[1], list(X.shape)[2], list(X.shape)[3]
c=cc
theta = Conv2D(c, kernel_size=(1,1), padding='same')(X)
theta_rsh = Reshape((h*w, c))(theta)
phi = Conv2D(c, kernel_size=(1,1), padding='same')(X)
phi_rsh = Reshape((c, h*w))(phi)
g = Conv2D(c, kernel_size=(1,1), padding='same')(X)
g_rsh = Reshape((h*w, c))(g)
theta_phi = tf.matmul(theta_rsh, phi_rsh)
theta_phi = tf.keras.layers.Softmax()(theta_phi)
theta_phi_g = tf.matmul(theta_phi, g_rsh)
theta_phi_g = Reshape((h, w, c))(theta_phi_g)
theta_phi_g = Conv2D(c*2, kernel_size=(1,1), padding='same')(theta_phi_g)
out = Add()([theta_phi_g, X])
return out
def gating_signal(input, out_size, batch_norm=False):
"""
resize the down layer feature map into the same dimension as the up layer feature map
using 1x1 conv
:param input: down-dim feature map
:param out_size:output channel number
:return: the gating feature map with the same dimension of the up layer feature map
"""
x = keras.layers.Conv2D(out_size, (1, 1), padding='same')(input)
if batch_norm:
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
return x
def attention_block(x, gating, inter_shape):
shape_x = K.int_shape(x)
shape_g = K.int_shape(gating)
theta_x = layers.Conv2D(inter_shape, (2, 2), strides=(2, 2), padding='same')(x) # 16
shape_theta_x = K.int_shape(theta_x)
phi_g = layers.Conv2D(inter_shape, (1, 1), padding='same')(gating)
upsample_g = layers.Conv2DTranspose(inter_shape, (3, 3),
strides=(shape_theta_x[1] // shape_g[1], shape_theta_x[2] // shape_g[2]),
padding='same')(phi_g) # 16
concat_xg = layers.add([upsample_g, theta_x])
act_xg = layers.Activation('relu')(concat_xg)
psi = layers.Conv2D(1, (1, 1), padding='same')(act_xg)
sigmoid_xg = layers.Activation('sigmoid')(psi)
shape_sigmoid = K.int_shape(sigmoid_xg)
upsample_psi = layers.UpSampling2D(size=(shape_x[1] // shape_sigmoid[1], shape_x[2] // shape_sigmoid[2]))(
sigmoid_xg) # 32
upsample_psi = expend_as(upsample_psi, shape_x[3])
y = layers.multiply([upsample_psi, x])
result = layers.Conv2D(shape_x[3], (1, 1), padding='same')(y)
result_bn = layers.BatchNormalization()(result)
return result_bn
print('decode_filters=',decode_filters)
non_local=global_non_local(base_model.output,decode_filters)
# Decoder Layers
decoder_1 = Conv2D(filters=decode_filters, kernel_size=1, padding='same', input_shape=base_model_output_shape,
name='conv2')(non_local)
gating_2=gating_signal(decoder_1,int(decode_filters/2),False)
att2=attention_block(base_model.get_layer('pool3_pool').output,gating_2,int(decode_filters/2))
decoder_2 = upproject(decoder_1, int(decode_filters / 2), 'up1', concat_with=att2)
gating_3 = gating_signal(decoder_2, int(decode_filters / 4), False)
att3 = attention_block(base_model.get_layer('pool2_pool').output, gating_3, int(decode_filters / 4))
decoder_3 = upproject(decoder_2, int(decode_filters / 4), 'up2', concat_with=att3)
# conv3 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv3_')(decoder)
gating_4 = gating_signal(decoder_3,int(decode_filters/8),False)
att4 = attention_block(base_model.get_layer('pool1').output,gating_4,int(decode_filters/8))
decoder_4 = upproject(decoder_3, int(decode_filters / 8), 'up3', concat_with=att4)
# conv2 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv2_')(decoder)
gating_5 = gating_signal(decoder_4,int(decode_filters/16),False)
att5 = attention_block(base_model.get_layer('conv1/relu').output,gating_5,int(decode_filters/16))
decoder_5 = upproject(decoder_4, int(decode_filters / 16), 'up4', concat_with=att5)
# conv1 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv1_')(decoder)
gating_6 = gating_signal(decoder_5,int(decode_filters/32),False)
att6=attention_block(base_model.get_layer('input_1').output,gating_6,int(decode_filters/32))
decoder_6 = upproject(decoder_5, int(decode_filters / 32), 'up5', concat_with=att6)
conv0 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv0_')(decoder_6)
predictions_raw_0 = conv0 * 1000
# predictions_raw_1 = conv1 * 1000
# predictions_raw_2 = conv2 * 1000
# predictions_raw_3 = conv3 * 1000
predictions_0 = tf.clip_by_value(predictions_raw_0, 1.0, 65535.0)
# predictions_1 = tf.clip_by_value(predictions_raw_1, 1.0, 65535.0)
# predictions_2 = tf.clip_by_value(predictions_raw_2, 1.0, 65535.0)
# predictions_3 = tf.clip_by_value(predictions_raw_3, 1.0, 65535.0)
final_outputs_0 = tf.cast(tf.clip_by_value(predictions_raw_0, 0, 65535.0), tf.int16)
eval_outputs__0 = tf.clip_by_value(predictions_raw_0, 0.1, 65535.0)
# final_outputs_1 = tf.cast(tf.clip_by_value(predictions_raw_1, 0, 65535.0), tf.int16)
# final_outputs_2 = tf.cast(tf.clip_by_value(predictions_raw_2, 0, 65535.0), tf.int16)
# final_outputs_3 = tf.cast(tf.clip_by_value(predictions_raw_3, 0, 65535.0), tf.int16)
# Create the model
model = Model(inputs=base_model.input, outputs=
[predictions_0, final_outputs_0, eval_outputs__0])
model.summary()
print('Model created.')
return model
def mySpatialModelChannelTest(model_name,spatial_size, nb_classes, channels, channel_first=True, weights_path=None, lr=0.005, decay=1e-6, momentum=0.9,plot_model=True): input_tensor = Input(shape=(channels, spatial_size, spatial_size)) input_shape = (channels, spatial_size, spatial_size) base_model=None predictions=None data_dim=1024 base_model = kerasApp.ResNet50(include_top=False, input_tensor=input_tensor, input_shape=input_shape, weights=None, classes=nb_classes, pooling=None) x = base_model.output x = GlobalAveragePooling2D()(x) x = Dense(1024, activation='relu')(x) predictions = Dense(nb_classes, activation='softmax')(x) # 训练模型 model = Model(inputs=base_model.input, outputs=predictions) print_shape(model,model_name) base_model = kerasApp.VGG16(include_top=False, input_tensor=input_tensor, input_shape=input_shape, weights=None, classes=nb_classes, pooling=None) x = base_model.output # 添加自己的全链接分类层 x = GlobalAveragePooling2D()(x) # add a global spatial average pooling layer x = Dense(1024, activation='relu')(x) # let's add a fully-connected layer predictions = Dense(nb_classes, activation='softmax')(x) # 训练模型 model = Model(inputs=base_model.input, outputs=predictions) print_shape(model, model_name) base_model = kerasApp.VGG19(include_top=False, input_tensor=input_tensor, input_shape=input_shape, weights=None, classes=2, pooling='avg') print_shape(base_model, model_name) base_model = kerasApp.InceptionV3(weights=None, include_top=False, pooling=None, input_shape=input_shape, classes=nb_classes) print_shape(base_model, model_name) base_model = kerasApp.InceptionResNetV2(weights=None, include_top=False, pooling=None, input_shape=input_shape, classes=nb_classes) x = base_model.output # 添加自己的全链接分类层 x = GlobalAveragePooling2D()(x) predictions = Dense(nb_classes, activation='softmax')(x) # 训练模型 model = Model(inputs=base_model.input, outputs=predictions) print_shape(model, model_name) #channel last input_tensor_Xception = Input(shape=( spatial_size, spatial_size,channels)) input_shape__Xception = (spatial_size, spatial_size,channels) base_model = kerasApp.Xception(weights=None, include_top=False, pooling=None, input_shape=input_shape__Xception, classes=nb_classes) print_shape(base_model, model_name) base_model = kerasApp.DenseNet121(weights=None, include_top=False, pooling=None, input_shape=input_shape, classes=nb_classes) print_shape(base_model, model_name) base_model = kerasApp.DenseNet169(weights=None, include_top=False, pooling=None, input_shape=input_shape, classes=nb_classes) print_shape(base_model, model_name) base_model = kerasApp.DenseNet201(weights=None, include_top=False, pooling=None, input_shape=input_shape, classes=nb_classes) print_shape(base_model, model_name) input_shape = (channels, spatial_size, spatial_size) base_model = kerasApp.MobileNet(weights=None, include_top=False, pooling=None, input_shape=input_shape, classes=nb_classes)
def create_model(is_twohundred=False, is_halffeatures=True):
print('Loading base model (DenseNet)..')
# Encoder Layers
if is_twohundred:
base_model = applications.DenseNet201(input_shape=(480, 640, 3), include_top=False)
else:
#base_model = applications.DenseNet169(input_shape=(480, 640, 3), include_top=False)
base_model = efn.EfficientNetB1(weights='imagenet',include_top=False,input_shape=((480, 640, 3)))
print('Base model loaded.')
# Starting point for decoder
base_model_output_shape = base_model.layers[-1].output.shape
# Layer freezing?
for layer in base_model.layers: layer.trainable = True
# Starting number of decoder filters
if is_halffeatures:
decode_filters = int(int(base_model_output_shape[-1]) / 2)
else:
decode_filters = int(base_model_output_shape[-1])
base_model.summary()
# Define upsampling layer
def upproject(tensor, filters, name, concat_with):
up_i = BilinearUpSampling2D((2, 2), name=name + '_upsampling2d')(tensor)
up_i = Concatenate(name=name + '_concat')(
[up_i, base_model.get_layer(concat_with).output]) # Skip connection
up_i = Conv2D(filters=filters, kernel_size=3, strides=1, padding='same', name=name + '_convA')(up_i)
up_i = LeakyReLU(alpha=0.2)(up_i)
up_i = Conv2D(filters=filters, kernel_size=3, strides=1, padding='same', name=name + '_convB')(up_i)
up_i = LeakyReLU(alpha=0.2)(up_i)
return up_i
def global_non_local(X,cc):
h, w , c = list(X.shape)[1], list(X.shape)[2], list(X.shape)[3]
c=cc
theta = Conv2D(c, kernel_size=(1,1), padding='same')(X)
theta_rsh = Reshape((h*w, c))(theta)
phi = Conv2D(c, kernel_size=(1,1), padding='same')(X)
phi_rsh = Reshape((c, h*w))(phi)
g = Conv2D(c, kernel_size=(1,1), padding='same')(X)
g_rsh = Reshape((h*w, c))(g)
theta_phi = tf.matmul(theta_rsh, phi_rsh)
theta_phi = tf.keras.layers.Softmax()(theta_phi)
theta_phi_g = tf.matmul(theta_phi, g_rsh)
theta_phi_g = Reshape((h, w, c))(theta_phi_g)
theta_phi_g = Conv2D(c*2, kernel_size=(1,1), padding='same')(theta_phi_g)
out = Add()([theta_phi_g, X])
return out
non_local=global_non_local(base_model.output,decode_filters)
# Decoder Layers
decoder = Conv2D(filters=decode_filters, kernel_size=1, padding='same', input_shape=base_model_output_shape,
name='conv2')(non_local)
decoder = upproject(decoder, int(decode_filters / 2), 'up1', concat_with='block5d_add')
decoder = upproject(decoder, int(decode_filters / 4), 'up2', concat_with='block3c_add')
# conv3 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv3_')(decoder)
decoder = upproject(decoder, int(decode_filters / 8), 'up3', concat_with='block2c_add')
# conv2 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv2_')(decoder)
decoder = upproject(decoder, int(decode_filters / 16), 'up4', concat_with='block1b_add')
# conv1 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv1_')(decoder)
decoder = upproject(decoder, int(decode_filters / 32), 'up5', concat_with='input_1')
conv0 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv0_')(decoder)
'''
decoder = Conv2D(filters=decode_filters, kernel_size=1, padding='same', input_shape=base_model_output_shape,
name='conv2')(base_model.output)
decoder = upproject(decoder, int(decode_filters / 2), 'up1', concat_with='pool3_pool')
decoder = upproject(decoder, int(decode_filters / 4), 'up2', concat_with='pool2_pool')
decoder = upproject(decoder, int(decode_filters / 8), 'up3', concat_with='pool1')
decoder = upproject(decoder, int(decode_filters / 16), 'up4', concat_with='conv1/relu')
decoder = upproject(decoder, int(decode_filters / 32), 'up5', concat_with='input_1')
conv0 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv0_')(decoder)
'''
predictions_raw_0 = conv0 * 1000
# predictions_raw_1 = conv1 * 1000
# predictions_raw_2 = conv2 * 1000
# predictions_raw_3 = conv3 * 1000
predictions_0 = tf.clip_by_value(predictions_raw_0, 1.0, 65535.0)
# predictions_1 = tf.clip_by_value(predictions_raw_1, 1.0, 65535.0)
# predictions_2 = tf.clip_by_value(predictions_raw_2, 1.0, 65535.0)
# predictions_3 = tf.clip_by_value(predictions_raw_3, 1.0, 65535.0)
final_outputs_0 = tf.cast(tf.clip_by_value(predictions_raw_0, 0, 65535.0), tf.int16)
eval_outputs__0 = tf.clip_by_value(predictions_raw_0, 0.1, 65535.0)
# final_outputs_1 = tf.cast(tf.clip_by_value(predictions_raw_1, 0, 65535.0), tf.int16)
# final_outputs_2 = tf.cast(tf.clip_by_value(predictions_raw_2, 0, 65535.0), tf.int16)
# final_outputs_3 = tf.cast(tf.clip_by_value(predictions_raw_3, 0, 65535.0), tf.int16)
# Create the model
model = Model(inputs=base_model.input, outputs=
[predictions_0, final_outputs_0, eval_outputs__0])
model.summary()
#print(base_model.output)
#print(base_model.layers[-1].output)
print('Model created.')
return model
#a=create_model()
def create_model(args, is_twohundred=False, is_halffeatures=True):
print('Loading base model (DenseNet)..')
# Encoder Layers
if is_twohundred:
base_model = applications.DenseNet201(input_shape=(480, 640, 3), include_top=False)
else:
base_model = applications.DenseNet169(input_shape=(480, 640, 3), include_top=False)
print('Base model loaded.')
# Starting point for decoder
base_model_output_shape = base_model.layers[-1].output.shape
# Layer freezing?
for layer in base_model.layers: layer.trainable = True
# Starting number of decoder filters
if is_halffeatures:
decode_filters = int(int(base_model_output_shape[-1]) / 2)
else:
decode_filters = int(base_model_output_shape[-1])
# Define upsampling layer
def upproject(tensor, filters, name, concat_with):
up_i = BilinearUpSampling2D((2, 2), name=name + '_upsampling2d')(tensor)
up_i = Concatenate(name=name + '_concat')(
[up_i, base_model.get_layer(concat_with).output]) # Skip connection
up_i = Conv2D(filters=filters, kernel_size=3, strides=1, padding='same', name=name + '_convA')(up_i)
up_i = LeakyReLU(alpha=0.2)(up_i)
up_i = Conv2D(filters=filters, kernel_size=3, strides=1, padding='same', name=name + '_convB')(up_i)
up_i = LeakyReLU(alpha=0.2)(up_i)
return up_i
# Decoder Layers
decoder = Conv2D(filters=decode_filters, kernel_size=1, padding='same', input_shape=base_model_output_shape,
name='conv2')(base_model.output)
decoder = upproject(decoder, int(decode_filters / 2), 'up1', concat_with='pool3_pool')
decoder = upproject(decoder, int(decode_filters / 4), 'up2', concat_with='pool2_pool')
# conv3 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv3_')(decoder)
decoder = upproject(decoder, int(decode_filters / 8), 'up3', concat_with='pool1')
# conv2 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv2_')(decoder)
decoder = upproject(decoder, int(decode_filters / 16), 'up4', concat_with='conv1/relu')
# conv1 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv1_')(decoder)
decoder = upproject(decoder, int(decode_filters / 32), 'up5', concat_with='input_1')
conv0 = Conv2D(filters=1, kernel_size=3, strides=1, padding='same', name='conv0_')(decoder)
predictions_raw_0 = conv0 * 1000
# predictions_raw_1 = conv1 * 1000
# predictions_raw_2 = conv2 * 1000
# predictions_raw_3 = conv3 * 1000
predictions_0 = tf.clip_by_value(predictions_raw_0, 1.0, 65535.0)
# predictions_1 = tf.clip_by_value(predictions_raw_1, 1.0, 65535.0)
# predictions_2 = tf.clip_by_value(predictions_raw_2, 1.0, 65535.0)
# predictions_3 = tf.clip_by_value(predictions_raw_3, 1.0, 65535.0)
final_outputs_0 = tf.cast(tf.clip_by_value(predictions_raw_0, 0, 65535.0), tf.int16)
eval_outputs__0 = tf.clip_by_value(predictions_raw_0, 0.1, 65535.0)
# final_outputs_1 = tf.cast(tf.clip_by_value(predictions_raw_1, 0, 65535.0), tf.int16)
# final_outputs_2 = tf.cast(tf.clip_by_value(predictions_raw_2, 0, 65535.0), tf.int16)
# final_outputs_3 = tf.cast(tf.clip_by_value(predictions_raw_3, 0, 65535.0), tf.int16)
# Create the model
model = Model(inputs=base_model.input, outputs=
[predictions_0, final_outputs_0, eval_outputs__0])
model.summary()
print('Model created.')
return model
