def train( self, x_labelled, y, x_unlabelled,
epochs,
x_valid, y_valid,
print_every = 1,
learning_rate = 3e-4,
beta1 = 0.9,
beta2 = 0.999,
seed = 31415,
stop_iter = 100,
save_path = None,
load_path = None ):
''' Session and Summary '''
if save_path is None:
self.save_path = 'checkpoints/model_GC_{}-{}-{}_{}.cpkt'.format(
self.num_lab,learning_rate,self.batch_size,time.time())
else:
self.save_path = save_path
np.random.seed(seed)
tf.set_random_seed(seed)
with self.G.as_default():
self.optimiser = tf.train.AdamOptimizer( learning_rate = learning_rate, beta1 = beta1, beta2 = beta2 )
self.train_op = self.optimiser.minimize( self.cost )
init = tf.initialize_all_variables()
self._test_vars = None
_data_labelled = np.hstack( [x_labelled, y] )
_data_unlabelled = x_unlabelled
x_valid_mu, x_valid_lsgms = x_valid[ :, :int(self.dim_x) ], x_valid[ :, int(self.dim_x): int(2*self.dim_x) ]
with self.session as sess:
sess.run(init)
if load_path == 'default': self.saver.restore( sess, self.save_path )
elif load_path is not None: self.saver.restore( sess, load_path )
best_eval_accuracy = 0.
best_train_accuracy = 0.
stop_counter = 0
print("****lab_clf", self.weights['lab_clf'])
print("****lab_recon_clf", self.weights['lab_recon_clf'])
print("****ulab_clf", self.weights['ulab_clf'])
print("****ulab_recon_clf", self.weights['ulab_recon_clf'])
print("****L_loss", self.weights['L_loss'])
print("****U_loss", self.weights['U_loss'])
for epoch in range(epochs):
''' Shuffle Data '''
np.random.shuffle( _data_labelled )
np.random.shuffle( _data_unlabelled )
''' Training '''
for x_l_mu, x_l_lsgms, y, x_u_mu, x_u_lsgms in utils.feed_numpy_semisupervised(
self.num_lab_batch, self.num_ulab_batch,
_data_labelled[:,:2*self.dim_x], _data_labelled[:,2*self.dim_x:],_data_unlabelled ):
training_result = sess.run( [self.train_op, self.cost],
feed_dict = { self.x_labelled_mu: x_l_mu,
self.x_labelled_lsgms: x_l_lsgms,
self.y_lab: y,
self.x_unlabelled_mu: x_u_mu,
self.x_unlabelled_lsgms: x_u_lsgms} )
training_cost = training_result[1]
''' Evaluation '''
stop_counter += 1
if epoch % print_every == 0:
test_vars = tf.get_collection(bookkeeper.GraphKeys.TEST_VARIABLES)
if test_vars:
if test_vars != self._test_vars:
self._test_vars = list(test_vars)
self._test_var_init_op = tf.initialize_variables(test_vars)
self._test_var_init_op.run()
eval_accuracy, eval_cross_entropy = \
sess.run( [self.eval_accuracy, self.eval_cross_entropy],
feed_dict = { self.x_labelled_mu: x_valid_mu,
self.x_labelled_lsgms: x_valid_lsgms,
self.y_lab: y_valid } )
if eval_accuracy > best_eval_accuracy:
best_eval_accuracy = eval_accuracy
self.saver.save( sess, self.save_path )
stop_counter = 0
print("accuracy update: ", best_eval_accuracy)
utils.print_metrics( epoch+1,
['Training', 'cost', training_cost],
['Validation', 'accuracy', eval_accuracy],
['Validation', 'cross-entropy', eval_cross_entropy] )
if stop_counter >= stop_iter:
print('Stopping GC training')
print('No change in validation accuracy for {} iterations'.format(stop_iter))
print('Best validation accuracy: {}'.format(best_eval_accuracy))
print('Model saved in {}'.format(self.save_path))
break
def train( self, x_labelled, y, x_unlabelled,
epochs,
x_valid, y_valid,
print_every = 1,
learning_rate = 3e-4,
beta1 = 0.9,
beta2 = 0.999,
seed = 31415,
stop_iter = 100,
save_path = None,
load_path = None ):
''' Session and Summary '''
if save_path is None:
self.save_path = 'checkpoints/model_GC_{}-{}-{}_{}.cpkt'.format(
self.num_lab,learning_rate,self.batch_size,time.time())
else:
self.save_path = save_path
np.random.seed(seed)
tf.set_random_seed(seed)
with self.G.as_default():
self.optimiser = tf.train.AdamOptimizer( learning_rate = learning_rate, beta1 = beta1, beta2 = beta2 )
self.train_op = self.optimiser.minimize( self.cost )
init = tf.initialize_all_variables()
self._test_vars = None
_data_labelled = np.hstack( [x_labelled, y] )
_data_unlabelled = x_unlabelled
x_valid_mu, x_valid_lsgms = x_valid[ :, :self.dim_x ], x_valid[ :, self.dim_x:2*self.dim_x ]
with self.session as sess:
sess.run(init)
if load_path == 'default': self.saver.restore( sess, self.save_path )
elif load_path is not None: self.saver.restore( sess, load_path )
best_eval_accuracy = 0.
stop_counter = 0
for epoch in range(epochs):
''' Shuffle Data '''
np.random.shuffle( _data_labelled )
np.random.shuffle( _data_unlabelled )
''' Training '''
for x_l_mu, x_l_lsgms, y, x_u_mu, x_u_lsgms in utils.feed_numpy_semisupervised(
self.num_lab_batch, self.num_ulab_batch,
_data_labelled[:,:2*self.dim_x], _data_labelled[:,2*self.dim_x:],_data_unlabelled ):
training_result = sess.run( [self.train_op, self.cost],
feed_dict = { self.x_labelled_mu: x_l_mu,
self.x_labelled_lsgms: x_l_lsgms,
self.y_lab: y,
self.x_unlabelled_mu: x_u_mu,
self.x_unlabelled_lsgms: x_u_lsgms } )
training_cost = training_result[1]
''' Evaluation '''
stop_counter += 1
if epoch % print_every == 0:
test_vars = tf.get_collection(bookkeeper.GraphKeys.TEST_VARIABLES)
if test_vars:
if test_vars != self._test_vars:
self._test_vars = list(test_vars)
self._test_var_init_op = tf.initialize_variables(test_vars)
self._test_var_init_op.run()
eval_accuracy, eval_cross_entropy = \
sess.run( [self.eval_accuracy, self.eval_cross_entropy],
feed_dict = { self.x_labelled_mu: x_valid_mu,
self.x_labelled_lsgms: x_valid_lsgms,
self.y_lab: y_valid } )
if eval_accuracy > best_eval_accuracy:
best_eval_accuracy = eval_accuracy
self.saver.save( sess, self.save_path )
stop_counter = 0
utils.print_metrics( epoch+1,
['Training', 'cost', training_cost],
['Validation', 'accuracy', eval_accuracy],
['Validation', 'cross-entropy', eval_cross_entropy] )
if stop_counter >= stop_iter:
print('Stopping GC training')
print('No change in validation accuracy for {} iterations'.format(stop_iter))
print('Best validation accuracy: {}'.format(best_eval_accuracy))
print('Model saved in {}'.format(self.save_path))
break