print("y_A shape", y_A.shape)
print("X_B shape", X_B.shape)
print("y_B shape", y_B.shape)
print("overlap_indexes len", len(overlap_indexes))
print("non_overlap_indexes len", len(non_overlap_indexes))
print("validate_indexes len", len(validate_indexes))
print("test_indexes len", len(test_indexes))
print(
"################################ Build Federated Models ############################"
)
tf.reset_default_graph()
autoencoder_A = Autoencoder(1)
autoencoder_B = Autoencoder(2)
autoencoder_A.build(X_A.shape[-1], 200, learning_rate=0.01)
autoencoder_B.build(X_B.shape[-1], 200, learning_rate=0.01)
# alpha = 100
fake_model_param = FakeFTLModelParam()
partyA = PlainFTLGuestModel(autoencoder_A, fake_model_param)
partyB = PlainFTLHostModel(autoencoder_B, fake_model_param)
federatedLearning = LocalPlainFederatedTransferLearning(partyA, partyB)
print(
"################################ Train Federated Models ############################"
)
def _create_local_model(self, local_model_param):
autoencoder = Autoencoder("local_host_model_01")
autoencoder.build(input_dim=local_model_param.input_dim, hidden_dim=local_model_param.encode_dim,
learning_rate=local_model_param.learning_rate)
return autoencoder
def _create_local_model(self, ftl_local_model_param, ftl_data_param):
autoencoder = Autoencoder("local_ftl_host_model_01")
autoencoder.build(input_dim=ftl_data_param.n_feature_host,
hidden_dim=ftl_local_model_param.encode_dim,
learning_rate=ftl_local_model_param.learning_rate)
return autoencoder
def test_autoencoder_restore_model(self):
X = np.array([[4, 2, 3], [6, 5, 1], [3, 4, 1], [1, 2, 3]])
_, D = X.shape
tf.reset_default_graph()
autoencoder = Autoencoder(0)
autoencoder.build(D, 5)
init_op = tf.global_variables_initializer()
with tf.Session() as session:
autoencoder.set_session(session)
session.run(init_op)
autoencoder.fit(X, epoch=10)
model_parameters = autoencoder.get_model_parameters()
tf.reset_default_graph()
autoencoder.restore_model(model_parameters)
init_op = tf.global_variables_initializer()
with tf.Session() as session:
autoencoder.set_session(session)
session.run(init_op)
Wh = autoencoder.Wh.eval()
Wo = autoencoder.Wo.eval()
bh = autoencoder.bh.eval()
bo = autoencoder.bo.eval()
assert_matrix(model_parameters["Wh"], Wh)
assert_matrix(model_parameters["Wo"], Wo)
assert_matrix(model_parameters["bh"], bh)
assert_matrix(model_parameters["bo"], bo)
def test_single_autoencoder():
# To run this test, you may first download MINST dataset from kaggle:
# https://www.kaggle.com/ngbolin/mnist-dataset-digit-recognizer
file_path = '../../../../data/MINST/train.csv'
Xtrain, Ytrain, Xtest, Ytest = getKaggleMNIST(file_path)
Xtrain = Xtrain.astype(np.float32)
Xtest = Xtest.astype(np.float32)
_, D = Xtrain.shape
autoencoder = Autoencoder(0)
autoencoder.build(D, 200)
init_op = tf.global_variables_initializer()
with tf.Session() as session:
autoencoder.set_session(session)
session.run(init_op)
autoencoder.fit(Xtrain, epoch=1, show_fig=True)
i = np.random.choice(len(Xtest))
x = Xtest[i]
y = autoencoder.predict([x])
plt.subplot(1, 2, 1)
plt.imshow(x.reshape(28, 28), cmap='gray')
plt.title('Original')
plt.subplot(1, 2, 2)
plt.imshow(y.reshape(28, 28), cmap='gray')
plt.title('Reconstructed')
plt.show()
model_parameters = autoencoder.get_model_parameters()
# test whether autoencoder can be restored from stored model parameters
tf.reset_default_graph()
autoencoder_2 = Autoencoder(0)
autoencoder_2.restore_model(model_parameters)
init_op = tf.global_variables_initializer()
with tf.Session() as session:
autoencoder_2.set_session(session)
session.run(init_op)
y_hat = autoencoder_2.predict([x])
plt.subplot(1, 2, 1)
plt.imshow(y.reshape(28, 28), cmap='gray')
plt.title('Original')
plt.subplot(1, 2, 2)
plt.imshow(y_hat.reshape(28, 28), cmap='gray')
plt.title('Reconstructed')
plt.show()
assert_matrix(y, y_hat)