def NN(train_X):
# Read Heng-Tze Cheng 2016 paper - Wide & Deep
input = keras.layers.Input(shape=train_X.shape[1:])
get_custom_objects().update({'Swish': Swish(swish)})
# # Architecture 1
# hidden1 = keras.layers.Dense(33, activation='Swish')(input)
# hidden2 = keras.layers.Dense(33, activation='Swish')(hidden1)
# concat = keras.layers.Concatenate()([input, hidden2])
# # Architecture 2
# hidden1 = keras.layers.Dense(33, activation='Swish')(input)
# hidden2 = keras.layers.Dense(50, activation='Swish')(hidden1)
# hidden3 = keras.layers.Dense(20, activation='Swish')(hidden2)
# concat = keras.layers.Concatenate()([input, hidden2, hidden3])
#
# There are a few different weight constraints to choose from. A good simple constraint for this model is to simply normalize the weights so that the norm is equal to 1.0.
# This constraint has the effect of forcing all incoming weights to be small.
# unit_norm in Keras
#
# Architecture 3
hidden1 = keras.layers.Dense(33,
activation='relu',
kernel_initializer="he_normal",
kernel_constraint=unit_norm())(input)
hidden2 = keras.layers.Dense(33,
activation='relu',
kernel_initializer="he_normal",
kernel_constraint=unit_norm())(hidden1)
hidden3 = keras.layers.Dense(33,
activation='relu',
kernel_initializer="he_normal",
kernel_constraint=unit_norm())(hidden2)
concat = keras.layers.Concatenate()([input, hidden2, hidden3])
output = keras.layers.Dense(1)(concat)
model = keras.models.Model(inputs=[input], outputs=[output])
model.compile(loss='mean_squared_error', optimizer='adam', metrics=['mse'])
# model.compile(loss = euclidean_distance_loss,
# optimizer='adam',
# metrics=['mse'])
return model
def design_CNN_3D(self):
inputs = Input(shape=self.X_train.shape[1:])
x = Conv3D(filters=32,
kernel_size=(3, 3, 11),
bias_constraint=unit_norm())(inputs)
x = LeakyReLU()(x)
x = Conv3D(filters=32,
kernel_size=(3, 3, 5),
bias_constraint=unit_norm())(x)
x = LeakyReLU()(x)
x = Flatten()(x)
x = Dense(units=256, activation='relu')(x)
x = Dropout(rate=0.4)(x)
x = Dense(units=128, activation='relu')(x)
x = Dropout(rate=0.4)(x)
outputs = Dense(units=len(self.labels), activation='softmax')(x)
self.model = Model(inputs=inputs, outputs=outputs)
def compile_3D_CNN(self):
inputs = Input(shape=self.X_train.shape[1:])
x = Conv3D(filters=64,
kernel_size=(3, 3, 11),
bias_constraint=unit_norm())(inputs)
x = LeakyReLU(alpha=0.2)(x)
x = Conv3D(filters=32,
kernel_size=(2, 2, 5),
bias_constraint=unit_norm())(x)
x = LeakyReLU(alpha=0.2)(x)
x = Flatten()(x)
x = Dense(units=256, activation='relu')(x)
x = Dropout(rate=0.4)(x)
x = Dense(units=128, activation='relu')(x)
x = Dropout(rate=0.4)(x)
outputs = Dense(units=len(self.labels), activation='softmax')(x)
self.model = Model(inputs=inputs, outputs=outputs)
self.model.compile(optimizer='Adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
def build_dense_model(num_features):
""" Simple two layer MLP """
regression_target = layers.Input(shape=(1,), name='ground_truth')
feature_input = layers.Input(shape=(num_features,), name='feature_input')
encoder_output = layers.GaussianDropout(0.1)(feature_input)
encoder_output = layers.Dense(64, activation='tanh', name='encoder_hidden')(encoder_output)
encoder_output = layers.Dense(64, activation='tanh', name='encoder_hidden_2')(encoder_output)
z_mean, z_log_var = layers.Dense(LATENT_DIM, name='z_mean')(encoder_output), \
layers.Dense(LATENT_DIM, name='z_log_var')(encoder_output)
r_mean, r_log_var = layers.Dense(1, name='r_mean')(encoder_output), \
layers.Dense(1, name='r_log_var')(encoder_output)
# Sample latent and regression target
z = layers.Lambda(sampling, output_shape=(LATENT_DIM,), name='z')([z_mean, z_log_var])
r = layers.Lambda(sampling, output_shape=(1,), name='r')([r_mean, r_log_var])
# Latent generator
pz_mean = layers.Dense(LATENT_DIM,
kernel_constraint=constraints.unit_norm(),
name='pz_mean')(r)
encoder = Model([feature_input, regression_target],
[z_mean, z_log_var, z,
r_mean, r_log_var, r,
pz_mean],
name='encoder')
latent_input = layers.Input(shape=(LATENT_DIM,), name='decoder_input')
decoder_output = layers.Dense(64, activation='tanh', name='decoder_hidden')(latent_input)
decoder_output = layers.Dense(64, activation='tanh', name='decoder_hidden_2')(decoder_output)
decoder_output = layers.Dense(num_features, name='decoder_output')(decoder_output)
decoder = Model(latent_input, decoder_output, name='decoder')
encoder_decoder_output = decoder(encoder([feature_input, regression_target])[2])
vae = Model([feature_input, regression_target], encoder_decoder_output, name='vae')
# Manually write up losses
reconstruction_loss = mse(feature_input, encoder_decoder_output)
kl_loss = 1 + z_log_var - K.square(z_mean - pz_mean) - K.exp(z_log_var)
kl_loss = -0.5 * K.sum(kl_loss, axis=-1)
label_loss = tf.divide(0.5 * K.square(r_mean - regression_target), K.exp(r_log_var)) + 0.5 * r_log_var
vae_loss = K.mean(reconstruction_loss + kl_loss + label_loss)
vae.add_loss(vae_loss)
vae.compile(optimizer=Adam())
regressor = Model(feature_input, r_mean, name='regressor')
return vae, regressor
def build_net3(CATES=12, height=64, width=64, channel=3, using_white_norm=True, using_SE=True):
base_model = build_shared_plain_network(using_white_norm=using_white_norm, using_SE=using_SE)
print(base_model.summary())
x1 = Input(shape=(height, width, channel))
x2 = Input(shape=(height, width, channel))
x3 = Input(shape=(height, width, channel))
y1 = base_model(x1)
y2 = base_model(x2)
y3 = base_model(x3)
cfeat = Concatenate(axis=-1)([l2_normalize(y1, axis=-1), l2_normalize(y2, axis=-1), l2_normalize(y3, axis=-1)])
cfeat = BatchNormalization()(cfeat)
cfeat = Dropout(0.5)(cfeat)
cfeat = Dense(512, use_bias=False)(cfeat)
cfeat = BatchNormalization()(cfeat)
cfeat = l2_normalize(cfeat, axis=-1)
print("cates->", CATES, cfeat.shape)
bulk_feat = Dense(CATES, use_bias=False, kernel_constraint=unit_norm(axis=0), activation=softmax, name="W1")(16 * cfeat)
age = Dense(1, use_bias=False, name="age")(bulk_feat)
gender = Dense(2, use_bias=False, kernel_constraint=unit_norm(axis=0), activation=softmax, name="gender")(cfeat)
return Model(inputs=[x1, x2, x3], outputs=[age, bulk_feat, gender])
def build(width, height, depth):
model = Sequential()
inputShape = (height, width, depth)
# if we are using "channels first", update the input shape
if K.image_data_format() == "channels_first":
inputShape = (depth, height, width)
# 1st layer convolutional
model.add(
Conv2D(32, (3, 3), input_shape=(150, 150, 1), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
# 2nd layer convolutional
model.add(
Conv2D(32, (3, 3),
activation='relu',
kernel_constraint=max_norm(3., axis=[0, 1, 2])))
model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
# 3rd layer convolutional
model.add(
Conv2D(64, (3, 3),
activation='relu',
kernel_constraint=max_norm(3., axis=[0, 1, 2])))
model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
# Flatten
model.add(Flatten())
# 4th layer fully connected
model.add(Dense(64, activation='relu', kernel_constraint=unit_norm()))
model.add(Dropout(0.5))
# 5th layer fully connected
model.add(Dense(1, activation='sigmoid'))
return model
def build_keras_model(model_type: str = None) -> keras.Sequential:
"""
Creates the desired CNN model. Passing in a known string will
use preconfigured models taken from papers.
"""
initializer = keras.initializers.HeNormal()
model = keras.Sequential()
if model_type == "davies":
opts = {
"strides": 2,
"activation": "relu",
"kernel_initializer": initializer,
"padding": "same",
}
#lots of small conv w/ max pooling
model.add(Conv2D(8, 15, **opts))
model.add(MaxPooling2D(pool_size=2, strides=2, padding="same"))
model.add(Conv2D(8, 15, **opts))
model.add(MaxPooling2D(pool_size=2, strides=2, padding="same"))
model.add(Conv2D(16, 5, **opts))
model.add(Conv2D(16, 5, **opts))
model.add(Flatten())
model.add(Dense(512, kernel_initializer=initializer))
model.add(
Dense(1, kernel_initializer=initializer, activation="sigmoid"))
else:
#my current model
restraint = max_norm(3)
convOpts = {
"strides": 2,
"activation": "relu",
"kernel_initializer": initializer,
"padding": "same",
"kernel_constraint": restraint
}
maxOpts = {
"strides": 2,
"padding": "same",
}
model.add(Conv2D(64, 4, **convOpts))
model.add(MaxPooling2D(4, **maxOpts))
model.add(Conv2D(64, 4, **convOpts))
model.add(MaxPooling2D(4, **maxOpts))
model.add(Dropout(.5))
model.add(Conv2D(64, 4, **convOpts))
model.add(MaxPooling2D(4, **maxOpts))
model.add(Dropout(.25))
model.add(Conv2D(128, 2, **convOpts))
model.add(Flatten())
model.add(
Dense(512,
kernel_initializer=initializer,
kernel_constraint=unit_norm()))
model.add(BatchNormalization())
model.add(
Dense(1, kernel_initializer=initializer, activation="sigmoid"))
model.compile(
"adam", # gradient optimizer
loss="binary_crossentropy", # loss function
metrics=["accuracy"], # what we are optimizing against
)
return model
def meteor_sort() -> None:
"""
Meteor sort training main script. In this case we carry out the following tasks:
- Generate the training and validation generator.
- Create the tensorflow (keras) model.
- If enabled, run the `meteor_sort_learning_rate()` function.
- Train the model.
- If enabled, convert the model to tensorflow lite and run `get_performance_measures()` and
`plot_acc_and_loss()` methods to get the performance measures (precision, recall and F1-Score) and plot
the accuracy and loss over the training iterations in both the training and the validation sets.
:return: None
"""
tf.keras.backend.clear_session()
# Data
data_dir = join(getcwd(), "meteor_data")
train_dir = join(data_dir, 'train')
validation_dir = join(data_dir, 'validation')
# Model handling
model_to_convert = ""
model_name = 'model_v2_1'
results_dir = join(getcwd(), 'results')
results_dir_weights = join(results_dir, 'weights')
# Hyperparameters for the training
image_resolution: tuple = (256, 256)
image_resolution_gray_scale: tuple = (256, 256, 1)
epochs: int = 10
learning_rate: float = 5e-4
get_ideal_learning_rate: bool = False
train_set_threshold: float = 0.92
validation_set_threshold: float = 0.93
num_training_images = len(listdir(join(train_dir, 'meteors'))) + len(
listdir(join(train_dir, 'non_meteors')))
num_validation_images = len(listdir(join(validation_dir, 'meteors'))) \
+ len(listdir(join(validation_dir, 'non_meteors')))
batch_size: int = 64
steps_per_epoch: int = int(num_training_images / batch_size)
validation_steps: int = int(num_validation_images / batch_size)
# Rescale all images by 1./255
train_datagen = ImageDataGenerator(
rescale=1.0 / 255,
rotation_range=
10, # Range from 0 to 180 degrees to randomly rotate images
width_shift_range=0.05,
height_shift_range=0.05,
shear_range=5, # Shear the image by 5 degrees
zoom_range=0.1,
horizontal_flip=True,
vertical_flip=True,
fill_mode='nearest')
validation_datagen = ImageDataGenerator(rescale=1.0 / 255.)
train_generator = train_datagen.flow_from_directory(
train_dir,
batch_size=batch_size,
class_mode='binary',
color_mode='grayscale',
target_size=image_resolution)
validation_generator = validation_datagen.flow_from_directory(
validation_dir,
batch_size=batch_size,
class_mode='binary',
color_mode='grayscale',
target_size=image_resolution)
model = Sequential([
Conv2D(8, (7, 7),
activation='elu',
input_shape=image_resolution_gray_scale,
strides=1,
kernel_initializer='he_uniform',
kernel_constraint=unit_norm()),
MaxPooling2D(pool_size=(3, 3)),
BatchNormalization(),
Conv2D(12, (5, 5),
activation='elu',
kernel_initializer='he_uniform',
kernel_constraint=unit_norm()),
MaxPooling2D(pool_size=(3, 3)),
BatchNormalization(),
Conv2D(12, (3, 3),
activation='elu',
kernel_initializer='he_uniform',
kernel_constraint=unit_norm()),
MaxPooling2D(pool_size=(2, 2)),
BatchNormalization(),
Conv2D(8, (3, 3),
activation='elu',
kernel_initializer='he_uniform',
kernel_constraint=unit_norm()),
MaxPooling2D(pool_size=(2, 2)),
BatchNormalization(),
Flatten(),
Dense(200,
activation='elu',
kernel_initializer='he_uniform',
kernel_constraint=unit_norm()),
BatchNormalization(),
Dense(16,
activation='elu',
kernel_initializer='he_uniform',
kernel_constraint=unit_norm()),
BatchNormalization(),
Dense(1, activation='sigmoid', kernel_initializer='he_uniform')
])
if get_ideal_learning_rate:
meteor_sort_learning_rate(model, train_dir, image_resolution,
batch_size, epochs, steps_per_epoch)
print(model.summary())
optimizer = Adam(learning_rate=learning_rate)
model.compile(optimizer=optimizer,
loss='binary_crossentropy',
metrics=['accuracy'])
my_callback = MeteorSortCallback(train_set_threshold,
validation_set_threshold, model,
model_name, results_dir_weights)
history = model.fit(train_generator,
validation_data=validation_generator,
steps_per_epoch=steps_per_epoch,
epochs=epochs,
validation_steps=validation_steps,
shuffle=True,
verbose=1,
callbacks=[my_callback])
# Print model performance and get performance measures
if model_to_convert != "":
# Load model to convert weights:
model.load_weights(join(results_dir, model_to_convert))
# Convert model to tflite:
converter = lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()
open("meteorLiteModel.tflite", "wb").write(tflite_model)
# Get performance measures:
get_performance_measures(model,
train_dir,
image_resolution,
join(results_dir,
'performance_' + model_name + '.txt'),
threshold=0.50)
# Plot Accuracy and Loss in both train and validation sets
plot_acc_and_loss(history, results_dir, model_name[-5:])
print(opt_layers)
model = tf.keras.Sequential()
for i in range(num_layers):
if entropy:
regularizer = entropy_reg(C)
else:
regularizer = regularizers.l1(C)
if initializer == "orthogonal":
keras_initializer_arg = "orthogonal"
else:
print("Using hypothesized optimal solution for layer %d initializer" %
i)
keras_initializer_arg = Constant(value=opt_layers[i])
model.add(
layers.Dense(n,
use_bias=False,
kernel_initializer=keras_initializer_arg,
kernel_constraint=constraints.unit_norm(axis=1),
kernel_regularizer=regularizer))
model.compile(optimizer='adam', loss='mse', metrics=['mae'])
Xtrain = np.random.normal(size=(ntrain, n))
Ytrain = np.matmul(Xtrain, had)
model.fit(Xtrain, Ytrain, epochs=epochs, batch_size=batch)
print("Solution:")
for i in range(num_layers):
print(model.layers[i].get_weights())
