epochs) + '_v' + str(VERSION_NUMBER)
keras_model_name = description + '.h5'
if SAVE_ALT_MODEL_CPU:
keras_model_name = 'alt_' + keras_model_name
file_name = description
seq_length = 2000
if num_channels < 2:
x_shape = [seq_length, 1]
input_shape = seq_length
else:
x_shape = [seq_length, 2]
input_shape = (seq_length, num_channels)
y_shape = [seq_length, num_classes]
# Start Timer:
start_time_ms = tfs.current_time_ms()
# Import Data:
x_tt, y_tt = tfs.load_data_v2('data/extended_5_class/mit_bih_tlabeled_w8s_fixed_all', [seq_length, 2], y_shape,
'relevant_data', 'Y')
if num_channels < 2:
x_tt = np.reshape(x_tt[:, :, 0], [-1, seq_length, 1])
xx_flex, y_flex = tfs.load_data_v2('data/flexEcg_8s_normal', [seq_length, 1], [1], 'relevant_data', 'Y')
x_train, x_test, y_train, y_test = train_test_split(x_tt, y_tt, train_size=0.75, random_state=1) # 0.66
ptb_data_lead_ii = tfs.load_mat('_adversarial/data/lead_ii_all/all_y.mat', key='Y', shape=[seq_length, 1])
def get_model_alt():
k_model = Sequential()
k_model.add(Reshape((seq_length, num_channels), input_shape=(input_shape, 1)))
outputs, prediction, accuracy = tfs.get_outputs(y, y_conv, output_node_name)
# 9. Initialize tensorflow for training & evaluation:
saver, init_op, config = tfs.tf_initialize()
# Enter Training Routine:
with tf.Session(config=config) as sess:
sess.run(init_op)
# Print Model Information Before Starting.
model_dims = tfs.get_model_dimensions(h, h_flat, h_fc, y_conv, NUM_LAYERS)
filter_dims = tfs.get_filter_dimensions(W_x, W_y, Str_X, Str_Y, Alphas,
NUM_LAYERS)
print(model_dims, '\n', filter_dims)
# Save model as pbtxt:
tf.train.write_graph(sess.graph_def, EXPORT_DIRECTORY,
Model_description + '.pbtxt', True)
start_time_ms = tfs.current_time_ms()
# Train Model:
val_accuracy_rate = tfs.train(x, y, keep_prob, accuracy, train_step,
x_train, y_train, x_test, y_test,
keep_prob_feed, train_steps)
elapsed_time_ms = (tfs.current_time_ms() - start_time_ms)
# Test Accuracy: (Test/Train Split)
tt_acc = tfs.test(sess,
x,
y,
accuracy,
x_test,
y_test,
keep_prob,
test_type='Train-Split')
# Validation Accuracy:
d = InstanceNormalization()(d)
return d
input_samples = Input(shape=(input_length, 1))
d1 = discriminator_layer(input_samples, 64, 5, 2, normalization=False)
d2 = discriminator_layer(d1, 128, 5, 2)
d3 = discriminator_layer(d2, 256, 5, 2)
d4 = discriminator_layer(d3, 512, 5, 2)
validity = Conv1D(1, kernel_size=5, strides=1, padding='same')(d4)
return Model(input_samples, validity)
# Train:
batch_size = 128
start_time_ms = tfs.current_time_ms()
# Initialize:
lambda_cycle = 10.0 # Cycle-consistency loss
lambda_id = 0.1 * lambda_cycle # Identity loss
optimizer = Adam(learn_rate, beta_1=0.50)
# Build and compile the discriminators
d_A = build_discriminator()
d_B = build_discriminator()
print('Discriminator: ')
print(d_A.summary())
d_A.compile(loss='mse', optimizer=optimizer, metrics=['accuracy'])
d_B.compile(loss='mse', optimizer=optimizer, metrics=['accuracy'])