def main(_):
if FLAGS.build_data:
build_data()
exit()
_, vocab_freq = semeval.load_vocab_and_freq()
word_embed = util.load_embedding(word_dim=FLAGS.word_dim)
with tf.Graph().as_default():
semeval_train_iter, semeval_test_iter = semeval.read_tfrecord(
FLAGS.num_epochs, FLAGS.batch_size)
model_name = 'cnn-%d-%d' % (FLAGS.word_dim, FLAGS.num_epochs)
semeval_train = semeval_train_iter.get_next()
semeval_test = semeval_test_iter.get_next()
m_train, m_valid = cnn_model.build_train_valid_model(
model_name, word_embed, vocab_freq, semeval_train, semeval_test,
FLAGS.is_adv, FLAGS.is_test)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()) # for file queue
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
sess.run(init_op)
print('=' * 80)
# for t in sess.run(m_train.layers):
# print(t.shape)
# exit()
if FLAGS.is_test:
test(sess, m_valid, semeval_test_iter)
else:
train_semeval(sess, m_train, m_valid, semeval_test_iter)
def main(_):
if FLAGS.build_data:
build_data()
exit()
word_embed = util.load_embedding(word_dim=FLAGS.word_dim)
with tf.Graph().as_default():
train_iter, test_iter = fudan.read_tfrecord(FLAGS.num_epochs,
FLAGS.batch_size)
train_data = train_iter.get_next()
test_data = test_iter.get_next()
m_train, m_valid = mtl_model.build_train_valid_model(
'fudan-mtl', word_embed, train_data, test_data, FLAGS.adv)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()) # for file queue
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
sess.run(init_op)
print('=' * 80)
if FLAGS.test:
test(sess, models)
else:
train(sess, m_train, m_valid, test_iter)
def main(_):
if FLAGS.build_data:
build_data()
exit()
word_embed = util.load_embedding(word_dim=FLAGS.word_dim)
with tf.Graph().as_default():
all_train = []
all_test = []
data_iter = fudan.read_tfrecord(FLAGS.num_epochs, FLAGS.batch_size)
for task_id, (train_data, test_data) in enumerate(data_iter):
task_name = fudan.get_task_name(task_id)
all_train.append((task_name, train_data))
all_test.append((task_name, test_data))
model_name = 'fudan-mtl'
if FLAGS.adv:
model_name += '-adv'
m_train, m_valid = mtl_model.build_train_valid_model(
model_name, word_embed, all_train, all_test, FLAGS.adv, FLAGS.test)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()) # for file queue
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
sess.run(init_op)
print('=' * 80)
if FLAGS.test:
test(sess, m_valid)
else:
train(sess, m_train, m_valid)
def main(_):
if FLAGS.build_data:
build_data()
exit()
word_embed = util.load_embedding(word_dim=FLAGS.word_dim)
with tf.Graph().as_default():
semeval_train_iter, semeval_test_iter = semeval.read_tfrecord(
FLAGS.num_epochs, FLAGS.batch_size)
dbpedia_train_iter, dbpedia_test_iter = dbpedia.read_tfrecord(
FLAGS.num_epochs, FLAGS.batch_size)
model_name = 'mtl-dbpedia-%d' % FLAGS.word_dim
semeval_train = semeval_train_iter.get_next()
semeval_test = semeval_test_iter.get_next()
dbpedia_train = dbpedia_train_iter.get_next()
dbpedia_test = dbpedia_test_iter.get_next()
m_train, m_valid = mtl_model.build_train_valid_model(
model_name, word_embed, semeval_train, semeval_test, dbpedia_train,
dbpedia_test, FLAGS.is_mtl, FLAGS.is_adv, FLAGS.test)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()) # for file queue
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
sess.run(init_op)
print('=' * 80)
if FLAGS.test:
test(sess, m_valid, semeval_test_iter)
else:
# train_dbpedia(sess, m_train, m_valid, dbpedia_test_iter)
train_semeval(sess, m_train, m_valid, semeval_test_iter)
def run_model(num_examples, is_train, inspect_data=False, plot=False):
if FLAGS.subword:
word_embed = None
else:
word_embed = util.load_embedding(word_dim=FLAGS.hidden_size)
name = "train" if is_train else "eval"
with tf.Graph().as_default():
global_step = tf.train.get_or_create_global_step()
model = mtl_model.MTLModel(word_embed, input_fn(is_train=is_train), FLAGS.adv, is_train=is_train)
model.build_train_op()
model.set_saver(model_name())
with tf.Session() as sess:
try:
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()) # for file queue
sess.run(init_op)
model.restore(sess)
except Exception as e:
tf.logging.warning("restore failed: {}".format(str(e)))
summary_prefix = os.path.join(get_logdir(), model_name())
writer = tf.summary.FileWriter(summary_prefix + '/{}'.format(name), sess.graph)
n_task = len(model.tensors)
batches = num_examples / FLAGS.batch_size
merged = model.merged_summary(name)
all_loss, all_acc = 0., 0.
if inspect_data:
eval_errors = collections.defaultdict(list)
for batch in range(int(batches)):
if inspect_data:
eval_fetch = [model.tensors, model.data, model.alignments, model.pred, model.separate_acc]
res, data, align, pred, separate_acc = sess.run(eval_fetch) # res = [[acc], [loss]]
if plot:
inspect(data, align, pred)
if FLAGS.vader:
for i, ((acc, _), (private_acc, shared_acc, vader_acc)) in enumerate(zip(res, separate_acc.values())):
eval_errors[fudan.get_task_name(i)].append([float(acc), float(private_acc), float(shared_acc), float(vader_acc)])
else:
for i, ((acc, _), (private_acc, shared_acc)) in enumerate(zip(res, separate_acc.values())):
eval_errors[fudan.get_task_name(i)].append(
[float(acc), float(private_acc), float(shared_acc)])
else:
if is_train:
train_fetch = [model.tensors, model.train_ops, merged, global_step]
res, _, summary, gs = sess.run(train_fetch) # res = [[acc], [loss]]
writer.add_summary(summary, gs)
else:
eval_fetch = [model.tensors, merged, global_step]
res, summary, gs = sess.run(eval_fetch) # res = [[acc], [loss]]
global eval_step
writer.add_summary(summary, eval_step)
eval_step = eval_step + 1
res = np.array(res)
all_loss += sum(res[:, 1].astype(np.float))
all_acc += sum(res[:, 0].astype(np.float))
all_loss /= (batches * n_task)
all_acc /= (batches * n_task)
if is_train:
model.save(sess, global_step)
if inspect_data:
columns = ['err', 'private_err', 'shared_err', 'vader_err'] if FLAGS.vader else ['err', 'private_err', 'shared_err', 'vader_err']
df = 1- pd.DataFrame(
data=np.array(list(eval_errors.values())).mean(axis=1),
index=list(eval_errors.keys()),
columns=columns
)
print(df)
print(df.mean())
return all_loss, all_acc
