for activation in ['exponential', 'relu']:
for other_activation in ['relu']:
for trial in range(num_trials):
if reg:
name = base_name + '_reg'
else:
name = base_name + '_noreg'
name += '_' + str(activation)
name += '_' + str(trial)
model = genome_model.model(input_shape=(L, A),
num_labels=1,
activation=activation,
other_activation=other_activation,
dropout=dropout,
bn=bn,
l2=None)
loss = keras.losses.BinaryCrossentropy(from_logits=False,
label_smoothing=0.0)
optimizer = keras.optimizers.Adam(learning_rate=0.0003)
if reg:
history, trainer = tfomics.fit.fit_lr_decay(
model,
loss,
optimizer,
x_train,
y_train,
validation_data=(x_valid, y_valid),
#-----------------------------------------------------------------
# load data
data_path = '../data'
filepath = os.path.join(data_path, 'synthetic_code_dataset.h5')
x_train, y_train, x_valid, y_valid, x_test, y_test, model_test = helper.load_data(
filepath)
N, L, A = x_train.shape
num_labels = y_train.shape[1]
#-----------------------------------------------------------------
# create model
model = genome_model.model(input_shape=(L, A),
num_labels=1,
activation=activation,
pool_size=4,
dropout=dropout,
bn=bn,
l2=None)
loss = keras.losses.BinaryCrossentropy(from_logits=False, label_smoothing=0.0)
optimizer = keras.optimizers.Adam(learning_rate=0.001)
#-----------------------------------------------------------------
# Fit model
attacker = tfomics.attack.PGDAttack((batch_size, L, A),
model,
loss,
learning_rate=0.01,
epsilon=epsilon,
num_steps=num_steps)
history, trainer = tfomics.fit.fit_robust(model,