def nasnet_retinanet(num_classes,
backbone='nasnet',
inputs=None,
modifier=None,
**kwargs):
k.clear_session()
# choose default input
if inputs is None:
if keras.backend.image_data_format() == 'channels_first':
inputs = keras.layers.Input(shape=(3, None, None))
else:
inputs = keras.layers.Input(shape=(None, None, 3))
# create the resnet backbone
if backbone == 'nasnet':
nasnet_model = NASNetLarge(weights=None,
include_top=False,
input_tensor=inputs)
else:
raise ValueError('Backbone (\'{}\') is invalid.'.format(backbone))
# invoke modifier if given
if modifier:
nasnet_model = modifier(nasnet_model)
concatenated_features = [
nasnet_model.get_layer('add_4').output,
nasnet_model.get_layer('activation_204').output, nasnet_model.output
]
# create the full model
return retinanet.retinanet(inputs=inputs,
num_classes=num_classes,
backbone_layers=concatenated_features,
**kwargs)
def save_model12(new_model_path, conv_model_path):
model = NASNetLarge(
input_shape=(img_width, img_height, 3),
include_top=False,
weights=None
)
if pretrained:
model = NASNetLarge(
input_shape=(img_width, img_height, 3),
include_top=False,
weights='imagenet'
)
model.summary()
transfer_layer = model.get_layer('?')
conv_model = Model(inputs=model.input,
outputs=transfer_layer.output)
new_model = Sequential()
new_model.add(conv_model)
new_model.add(GlobalAveragePooling2D())
if num_fc_layers>=1:
new_model.add(Dense(num_fc_neurons, activation='relu'))
if num_fc_layers>=2:
new_model.add(Dropout(dropout))
new_model.add(Dense(num_fc_neurons, activation='relu'))
if num_fc_layers>=3:
new_model.add(Dropout(dropout))
new_model.add(Dense(num_fc_neurons, activation='relu'))
new_model.add(Dense(num_classes, activation='softmax'))
print(new_model.summary())
new_model.save(new_model_path)
conv_model.save(conv_model_path)
return
print(x_val.shape) #print(np.asarray(range(len(uc)))) #print(y_val[0,:]) # ## Adapting the NasNetLarge Architecture to Regression Problems # In[8]: from keras.applications.nasnet import NASNetLarge from keras.models import Model model = NASNetLarge(weights='imagenet', include_top=True, input_shape=(331, 331, 3)) x = model.get_layer(index=len(model.layers)-2).output print(x) x = Dense(1)(x) model = Model(inputs=model.input, outputs=x) model.summary() # **Using RMSprop optimizer, mean absolute error with metrics, and mean square erro with loss** # In[ ]: opt = RMSprop(lr=0.0001) model.compile(loss='mean_squared_error', optimizer=opt, metrics=['mae'])
def NASNet_large_FCN(input_image, weights=None):
"""
return Model instance
"""
input_tensor = Input(shape=(input_image))
model = NASNetLarge(input_shape=input_image,
input_tensor=input_tensor,
include_top=False,
weights=weights)
#print model.summary()
#return
#reduce_stem_1 = model.get_layer()
normal_18 = model.get_layer(name='normal_concat_18').output
activation_normal_concat_7 = model.get_layer(name='activation_118').output
activation_normal_concat_8 = model.get_layer(name='activation_130').output
activation_normal_concat_9 = model.get_layer(name='activation_142').output
activation_normal_concat_10 = model.get_layer(name='activation_154').output
activation_normal_concat_11 = model.get_layer(name='activation_166').output
activation_normal_concat_12 = model.get_layer(name='activation_178').output
"""
activation_normal_concat_7 = ScaledLayer()(activation_normal_concat_7)
activation_normal_concat_8 = ScaledLayer()(activation_normal_concat_8)
activation_normal_concat_9 = ScaledLayer()(activation_normal_concat_9)
activation_normal_concat_10 = ScaledLayer()(activation_normal_concat_10)
activation_normal_concat_11 = ScaledLayer()(activation_normal_concat_11)
activation_normal_concat_12 = ScaledLayer()(activation_normal_concat_12)
"""
fuse_activation_7_10 = Add()([activation_normal_concat_7, activation_normal_concat_8, activation_normal_concat_9, \
activation_normal_concat_10, activation_normal_concat_11, activation_normal_concat_12])
activation_normal_concat_0 = model.get_layer(name='activation_35').output
activation_normal_concat_1 = model.get_layer(name='activation_47').output
activation_normal_concat_2 = model.get_layer(name='activation_59').output
activation_normal_concat_3 = model.get_layer(name='activation_71').output
activation_normal_concat_4 = model.get_layer(name='activation_83').output
activation_normal_concat_5 = model.get_layer(name='activation_95').output
"""
activation_normal_concat_0 = ScaledLayer()(activation_normal_concat_0)
activation_normal_concat_1 = ScaledLayer()(activation_normal_concat_1)
activation_normal_concat_2 = ScaledLayer()(activation_normal_concat_2)
activation_normal_concat_3 = ScaledLayer()(activation_normal_concat_3)
activation_normal_concat_4 = ScaledLayer()(activation_normal_concat_4)
activation_normal_concat_5 = ScaledLayer()(activation_normal_concat_5)
"""
fuse_activation_0_5 = Add()([activation_normal_concat_0, activation_normal_concat_1, activation_normal_concat_2, \
activation_normal_concat_3, activation_normal_concat_4, activation_normal_concat_5])
conv_normal_18 = Conv2D(filters=6, kernel_size=(1, 1))(normal_18)
upscore_normal_18 = Conv2DTranspose(filters=6,
kernel_size=(4, 4),
strides=(2, 2),
padding='same')(conv_normal_18)
conv_fuse_7_10 = Conv2D(filters=6,
kernel_size=(1, 1))(fuse_activation_7_10)
conv_fuse_7_10 = Add()([conv_fuse_7_10, upscore_normal_18])
upscore_fuse_7_10 = Conv2DTranspose(filters=6,
kernel_size=(4, 4),
strides=(2, 2),
padding='same')(conv_fuse_7_10)
conv_fuse_0_5 = Conv2D(filters=6, kernel_size=(1, 1))(fuse_activation_0_5)
conv_fuse_0_5 = Add()([conv_fuse_0_5, upscore_fuse_7_10])
upscore = Conv2DTranspose(filters=6,
kernel_size=(16, 16),
strides=(8, 8),
padding='same')(conv_fuse_0_5)
model = Model(inputs=input_tensor, outputs=upscore)
print model.summary()
return model
def NASNet_large_ensemble_FCN(input_image, weights=None, fine_tune=False):
input_tensor = Input(shape=(input_image))
model = NASNetLarge(input_shape=input_image,
input_tensor=input_tensor,
include_top=False,
weights=weights)
#print model.summary()
#reduce_stem_1 = model.get_layer()
normal_18 = model.get_layer(name='normal_concat_18').output
activation_normal_concat_7 = model.get_layer(name='activation_118').output
activation_normal_concat_8 = model.get_layer(name='activation_130').output
activation_normal_concat_9 = model.get_layer(name='activation_142').output
activation_normal_concat_10 = model.get_layer(name='activation_154').output
activation_normal_concat_11 = model.get_layer(name='activation_166').output
activation_normal_concat_12 = model.get_layer(name='activation_178').output
fuse_activation_7_10 = Add()([activation_normal_concat_7, activation_normal_concat_8, activation_normal_concat_9, \
activation_normal_concat_10, activation_normal_concat_11, activation_normal_concat_12])
activation_normal_concat_0 = model.get_layer(name='activation_35').output
activation_normal_concat_1 = model.get_layer(name='activation_47').output
activation_normal_concat_2 = model.get_layer(name='activation_59').output
activation_normal_concat_3 = model.get_layer(name='activation_71').output
activation_normal_concat_4 = model.get_layer(name='activation_83').output
activation_normal_concat_5 = model.get_layer(name='activation_95').output
fuse_activation_0_5 = Add()([activation_normal_concat_0, activation_normal_concat_1, activation_normal_concat_2, \
activation_normal_concat_3, activation_normal_concat_4, activation_normal_concat_5])
conv_normal_18 = Conv2D(filters=6, kernel_size=(1, 1))(normal_18)
upscore_normal_18 = Conv2DTranspose(filters=6,
kernel_size=(4, 4),
strides=(2, 2),
padding='same')(conv_normal_18)
conv_fuse_7_10 = Conv2D(filters=6,
kernel_size=(1, 1))(fuse_activation_7_10)
conv_fuse_7_10 = Add()([conv_fuse_7_10, upscore_normal_18])
upscore_fuse_7_10 = Conv2DTranspose(filters=6,
kernel_size=(4, 4),
strides=(2, 2),
padding='same')(conv_fuse_7_10)
conv_fuse_0_5 = Conv2D(filters=6, kernel_size=(1, 1))(fuse_activation_0_5)
conv_fuse_0_5 = Add()([conv_fuse_0_5, upscore_fuse_7_10])
output_8 = Conv2DTranspose(filters=6,
kernel_size=(16, 16),
strides=(8, 8),
padding='same')(conv_fuse_0_5)
output_16 = Conv2DTranspose(filters=6,
kernel_size=(32, 32),
strides=(16, 16),
padding='same')(conv_fuse_7_10)
output_32 = Conv2DTranspose(filters=6,
kernel_size=(64, 64),
strides=(32, 32),
padding='same')(conv_normal_18)
model = Model(inputs=input_tensor,
outputs=[output_8, output_16, output_32])
#print model.summary()
return model
def TransferNet(input_shape: Tuple[int, int],
num_classes: int,
feature_extractor: FeatureExtractor = None,
dense_layers: int = 3,
dropout_rate: float = 0.3) -> Model:
"""
Exploits a pre-trained model as feature extractor, feeding its output into a fully-connected NN.
The feature extractor model is NOT fine-tuned for the specific task.
Dropout and batch normalization are used throughout the trainable portion of the network.
:param input_shape: Shape of the input tensor as a 2-dimensional int tuple
:param num_classes: Number of classes for the final FC layer
:param feature_extractor: FeatureExtractor instance representing which pre-trained model to use as feature extractor
:param dense_layers: Number of layers for the FC NN
:param dropout_rate: Dropout rate
:return: a Keras model
"""
adam_opt = Adam(lr=0.1)
model_input = Input(shape=input_shape)
# load pre-trained model on ImageNet
if feature_extractor == FeatureExtractor.Dense121:
fe_model = DenseNet121(weights="imagenet",
include_top=False,
input_tensor=model_input)
elif feature_extractor == FeatureExtractor.Dense169:
fe_model = DenseNet169(weights="imagenet",
include_top=False,
input_tensor=model_input)
elif feature_extractor == FeatureExtractor.Dense201:
fe_model = DenseNet201(weights="imagenet",
include_top=False,
input_tensor=model_input)
elif feature_extractor == FeatureExtractor.NASNetLarge:
fe_model = NASNetLarge(weights="imagenet",
include_top=False,
input_tensor=model_input)
else:
# default: NASNetMobile
fe_model = NASNetMobile(weights="imagenet",
include_top=False,
input_tensor=model_input)
fe_model = Model(input=model_input,
output=fe_model.output,
name="FeatureExtractor")
fe_model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=adam_opt,
metrics=["accuracy"])
# get handles to the model (input, output tensors)
fe_input = fe_model.get_layer(index=0).input
fe_output = fe_model.get_layer(index=-1).output
# freeze layers
for _, layer in enumerate(fe_model.layers):
layer.trainable = False
# final fully-connected layers
dense = Flatten()(fe_output)
dense = BatchNormalization()(dense)
dense = Dropout(rate=dropout_rate)(dense)
num_units = 128
for i in range(1, dense_layers + 1):
dense = Dense(units=int(num_units / i), activation="relu")(dense)
dense = BatchNormalization()(dense)
dense = Dropout(rate=dropout_rate)(dense)
output_dense = Dense(units=num_classes, activation="softmax")(dense)
model = Model(input=fe_input, output=output_dense, name="TransferNet")
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=adam_opt,
metrics=["accuracy"])
return model
