EPOCHS = 10
FGSM_PARAMS = {'eps': 0.05,
'clip_min': 0.,
'clip_max': 1.,
}
# initialize keras/tf session
sess = tf.Session(graph=tf.get_default_graph())
K.set_session(sess)
# get dataset
(train_images, train_labels), (test_images, test_labels) = get_scaled_fashion_mnist()
# load models
model_1 = get_vanilla_NN()
model_2 = get_QNN(32)
# train models
model_1.fit(train_images, train_labels, epochs=EPOCHS, verbose=0)
model_2.fit(train_images, train_labels, epochs=EPOCHS, verbose=0)
# evaluate models on the test set
_, test_acc_1 = model_1.evaluate(test_images, test_labels, verbose=0)
_, test_acc_2 = model_2.evaluate(test_images, test_labels, verbose=0)
print("Test accuracy NN_1: " + str(test_acc_1))
print("Test accuracy NN_2: " + str(test_acc_2))
# filter samples correctly classified by both models
test_images, test_labels = filter_correctly_classified_samples(test_images, test_labels, [model_1, model_2])
print("From now on using " + str(test_images.shape[0]) + " samples that are correctly classified by both networks.")
'abort_early': True, # if we stop improving, abort gradient descent early
'learning_rate':
1e-2, # larger values converge faster to less accurate results
'initial_const': 1e-3, # the initial constant c to pick as a first guess
}
# initialize keras/tf session
sess = tf.Session(graph=tf.get_default_graph())
K.set_session(sess)
# get dataset
(train_images, train_labels), (test_images, test_labels) = get_fashion_mnist()
#(train_images, train_labels), (test_images, test_labels) = get_mnist()
# load models
model_2bits_1 = get_QNN(2)
model_4bits_1 = get_QNN(4)
model_8bits_1 = get_QNN(8)
model_16bits_1 = get_QNN(16)
model_32bits_1 = get_QNN(32)
model_vanilla_nn_1 = get_vanilla_NN()
model_2bits_2 = get_QNN(2)
model_4bits_2 = get_QNN(4)
model_8bits_2 = get_QNN(8)
model_16bits_2 = get_QNN(16)
model_32bits_2 = get_QNN(32)
model_vanilla_nn_2 = get_vanilla_NN()
# train models
print("Training models...")
FGSM_PARAMS = {
'eps': 0.05,
'clip_min': 0.,
'clip_max': 1.,
}
# initialize keras/tf session
sess = tf.Session(graph=tf.get_default_graph())
K.set_session(sess)
# get dataset
(train_images, train_labels), (test_images,
test_labels) = get_scaled_fashion_mnist()
# load models
model_2bits = get_QNN(2)
# train models
model_2bits.fit(train_images, train_labels, epochs=EPOCHS, verbose=0)
_, test_acc = model_2bits.evaluate(test_images, test_labels, verbose=0)
print("Test accuracy of NN with 2 bits: " + str(test_acc))
# save weights
print("Saving weights")
model_2bits.save_weights("2bits-weights.h5")
# load not quantized model and quantized weights
print(
"Loading the weights in same architecture NN but no quantization involved")
model = get_vanilla_NN()