def main(_):
if not FLAGS.config_file:
raise ValueError("Must set --config_file to set model's hyperparams")
config = Config(FLAGS.config_file)
data = dataloader.Dataloader(config.data_file)
dataset = data.load_data()
config.num_steps = data.sentence_max_len
config.vocab_size = data.vocab_size
config.label_size = data.label_size
initializer = tf.random_uniform_initializer(-config.init_scale,
config.init_scale)
kfold = config.kfold
total_data = len(dataset)
for i in range(kfold):
test_data = dataset[int(total_data / kfold * i):int(total_data /
kfold * (i + 1))]
train_data = dataset[0:int(total_data / kfold * i)]
train_data.extend(dataset[int(total_data / kfold *
(i + 1)):total_data])
tf.reset_default_graph()
with tf.name_scope("Train"):
train_input = ModelInput(raw_data=train_data,
batch_size=config.batch_size)
with tf.variable_scope("Model",
reuse=None,
initializer=initializer):
train_model = Model(is_training=True,
config=config,
input_=train_input)
with tf.name_scope("Test"):
test_input = ModelInput(raw_data=test_data,
batch_size=config.batch_size)
with tf.variable_scope("Model",
reuse=True,
initializer=initializer):
test_model = Model(is_training=False,
config=config,
input_=test_input)
with tf.name_scope("Test_train"):
test_train_input = ModelInput(raw_data=train_data,
batch_size=config.batch_size)
with tf.variable_scope("Model",
reuse=True,
initializer=initializer):
test_train_model = Model(is_training=False,
config=config,
input_=test_train_input)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
train(sess, train_model, test_model, config)
test(sess, test_model, config)
test(sess, test_train_model, config)
print("""]}""")
