def testClassifierGraph(self):
FLAGS.rnn_num_layers = 2
model = graphs.VatxtModel()
train_op, _, _ = model.classifier_training()
# Pretrained vars: embedding + LSTM layers
self.assertEqual(len(model.pretrained_variables),
1 + 2 * FLAGS.rnn_num_layers)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
tf.train.start_queue_runners(sess)
sess.run(train_op)
def testATMethods(self):
at_methods = [None, 'rp', 'at', 'vat', 'atvat']
for method in at_methods:
FLAGS.adv_training_method = method
with tf.Graph().as_default():
graphs.VatxtModel().classifier_graph()
# Ensure variables have been reused
# Embedding + LSTM layers + hidden layers + logits layer
expected_num_vars = 1 + 2 * FLAGS.rnn_num_layers + 2 * (
FLAGS.cl_num_layers) + 2
self.assertEqual(len(tf.trainable_variables()),
expected_num_vars)
def testEvalGraph(self):
_, _ = graphs.VatxtModel().eval_graph()
def testSeqAE(self):
FLAGS.use_seq2seq_autoencoder = True
graphs.VatxtModel().language_model_training()
def testCandidateSampling(self):
FLAGS.num_candidate_samples = 10
graphs.VatxtModel().language_model_training()
def testSyncReplicas(self):
FLAGS.sync_replicas = True
graphs.VatxtModel().language_model_training()
def testMulticlass(self):
FLAGS.num_classes = 10
graphs.VatxtModel().classifier_graph()
def testLanguageModelGraph(self):
train_op, _, _ = graphs.VatxtModel().language_model_training()
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
tf.train.start_queue_runners(sess)
sess.run(train_op)
