def run_func():
config = Config()
# ========= Load Dataset =========
# You can change this code to load dataset in your own way
vocab, rev_vocab = initialize_vocab(config.vocab_path)
dev_path = "data/squad/fuse.json"
dev_dirname = os.path.dirname(os.path.abspath(dev_path))
dev_filename = os.path.basename(dev_path)
context_data, question_data, question_uuid_data = prepare_dev(
dev_dirname, dev_filename, vocab)
ques_len = len(question_data)
answers = [[0, 0] for _ in xrange(ques_len)]
dataset = [question_data, context_data, answers]
embed_path = config.embed_path
embeddings = get_trimmed_glove_vectors(embed_path)
encoder = Encoder(config.hidden_state_size)
decoder = Decoder(config.hidden_state_size)
qa = QASystem(encoder, decoder, embeddings, config)
with tf.Session() as sess:
qa.initialize_model(sess, config.train_dir)
answers, _ = generate_answers(sess, qa, dataset, question_uuid_data,
rev_vocab)
# write to json file to root dir
with io.open('temp/fuse-answer.json', 'w', encoding='utf-8') as f:
f.write(unicode(json.dumps(answers, ensure_ascii=False)))
def run_func2(dataset, config):
vocab, rev_vocab = initialize_vocab(config.vocab_path)
q, c, a = zip(*[[_q, _c, _a] for (_q, _c, _a) in dataset])
dataset = [q, c, a]
embed_path = config.embed_path
embeddings = get_trimmed_glove_vectors(embed_path)
encoder = Encoder(config.hidden_state_size)
decoder = Decoder(config.hidden_state_size)
qa = QASystem(encoder, decoder, embeddings, config)
question_uuid_data = [i for i in xrange(len(a))]
with tf.Session() as sess:
qa.initialize_model(sess, config.train_dir)
answers, answers_canonical = generate_answers(sess, qa, dataset,
question_uuid_data,
rev_vocab)
# write to json file to root dir
with io.open('dev-prediction.txt', 'w', encoding='utf-8') as f:
for i in xrange(len(a)):
curr_ans = unicode(answers[i], "utf-8")
f.write("%s\n" % (curr_ans))
def run_func():
config = Config()
vocab, rev_vocab = initialize_vocab(config.vocab_path)
dev_path = "download/squad/test.json"
dev_dirname = os.path.dirname(os.path.abspath(dev_path))
dev_filename = os.path.basename(dev_path)
context_data, question_data, question_uuid_data = prepare_dev(
dev_dirname, dev_filename, vocab)
ques_len = len(question_data)
answers = [[0, 0] for _ in xrange(ques_len)]
dataset = [question_data, context_data, answers]
embed_path = config.embed_path
embeddings = get_trimmed_glove_vectors(embed_path)
encoder = Encoder(config.hidden_state_size)
decoder = Decoder(config.hidden_state_size)
qa = QASystem(encoder, decoder, embeddings, config)
data = "Id,Answer\n"
with tf.Session() as sess:
qa.initialize_model(sess, config.train_dir)
answers, _ = generate_answers(sess, qa, dataset, question_uuid_data,
rev_vocab)
for a in answers:
ans = answers[a]
data += a + "," + normalize_answer(ans).replace(" s ", "s ") + "\n"
with open('submission.csv', 'wb') as file:
file.write(data)
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
datasetTrain = initialize_datasets(FLAGS.data_dir, 'train')
datasetVal = initialize_datasets(FLAGS.data_dir, 'val')
#datasetTrain = datasetTrain[0:100]
embed_path = FLAGS.embed_path or pjoin("data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = Encoder(size=FLAGS.state_size, vocab_dim=FLAGS.embedding_size)
decoder = Decoder(output_size=FLAGS.output_size)
qa = QASystem(encoder, decoder, embed_path, rev_vocab, FLAGS)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
qa.train(sess, datasetTrain, save_train_dir)
qa.evaluate_answer(sess, datasetVal, sample=1000, log=True)
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
# load all in once, maybe better to try batch by batch
question_path = "./data/squad/train.ids.question"
context_path = "./data/squad/train.ids.context"
answer_path = "./data/squad/train.span"
val_q = "./data/squad/val.ids.question"
val_c = "./data/squad/val.ids.context"
val_a = "./data/squad/val.span"
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
# embeddings is a matrix of shape [vocab_size, embedding_size]
embeddings = np.load(embed_path)['glove'].astype(np.float32)
val_data = load_and_pad_val_data(val_q, val_c, val_a)
# vocab is the mapping from word -> token id
# rev_vocab is the reverse mapping, from id -> word
vocab, rev_vocab = initialize_vocab(vocab_path)
# someone posted that the max length of question is 766
info = (question_path, context_path, answer_path, FLAGS.batch_size,
FLAGS.max_length, FLAGS.output_size)
'''
batch_gen = batch_generator(question_path, context_path, answer_path, FLAGS.batch_size, FLAGS.max_length, FLAGS.output_size)
i = 0;
while True:
batch_gen.next()
i += 1
logging.info(i)
'''
encoder = Encoder(size=FLAGS.state_size, vocab_dim=FLAGS.embedding_size)
decoder = Decoder(output_size=FLAGS.output_size)
qa = QASystem(encoder, decoder, FLAGS, embeddings)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
qa.train(sess, batch_generator, info, save_train_dir, val_data,
rev_vocab)
def main(FLAGS):
print(80 * "=")
print("INITIALIZING")
print(80 * "=")
# Do what you need to load datasets from FLAGS.data_dir
#parser, embeddings, train_examples, dev_set, test_set = load_and_preprocess_data(debug)
if not os.path.exists('./data/weights/'):
os.makedirs('./data/weights/')
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
print("Loading Embedding Matrix")
embeddings = np.load(embed_path)['glove']
encoder = Encoder(size=FLAGS.output_size, vocab_dim=FLAGS.embedding_size)
decoder = Decoder(output_size=FLAGS.output_size)
qa = QASystem(encoder, decoder, FLAGS, embeddings)
#qa = QASystem(encoder, FLAGS, embeddings)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
print("Building Network ... ")
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
print("Load Training Data")
dataset = initialize_datasets(FLAGS.data_dir,
dataset='train',
debugMode=True)
# encoder.encode_question(
# dataset['Questions'], question['Questions_masks'])
print(80 * "=")
print("Training")
print(80 * "=")
qa.train(sess, dataset, save_train_dir)
print("Finished Training")
print("Load Validation Data")
dataset = initialize_datasets(FLAGS.data_dir,
dataset='val',
debugMode=True)
print(80 * "=")
print("Evaluation")
print(80 * "=")
qa.evaluate_answer(sess, dataset, vocab, FLAGS.evaluate, log=True)
def main(_):
FLAGS.config = int(sys.argv[1])
load_config(current_config=FLAGS.config)
vocab, rev_vocab = initialize_vocab(FLAGS.vocab_path)
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.vocab_dim))
global_train_dir = '/tmp/cs224n-squad-train'
# Adds symlink to {train_dir} from /tmp/cs224n-squad-train to canonicalize the
# file paths saved in the checkpoint. This allows the model to be reloaded even
# if the location of the checkpoint files has moved, allowing usage with CodaLab.
# This must be done on both train.py and qa_answer.py in order to work.
if not os.path.exists(FLAGS.train_dir):
os.makedirs(FLAGS.train_dir)
if os.path.exists(global_train_dir):
os.unlink(global_train_dir)
os.symlink(os.path.abspath(FLAGS.train_dir), global_train_dir)
train_dir = global_train_dir
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
# ========= Load Dataset =========
# You can change this code to load dataset in your own way
dev_dirname = os.path.dirname(os.path.abspath(FLAGS.dev_path))
dev_filename = os.path.basename(FLAGS.dev_path)
context_data, question_data, question_uuid_data = prepare_dev(
dev_dirname, dev_filename, vocab)
dataset = (context_data, question_data, question_uuid_data)
# ========= Model-specific =========
# You must change the following code to adjust to your model
encoder = Encoder(size=FLAGS.state_size, vocab_dim=FLAGS.embedding_size)
decoder = Decoder(output_size=FLAGS.output_size)
qa = QASystem(encoder, decoder, train_dir)
with tf.Session() as sess:
initialize_model(sess, qa, train_dir)
answers = generate_answers(sess, qa, dataset, rev_vocab)
# write to json file to root dir
with io.open('dev-prediction.json', 'w', encoding='utf-8') as f:
f.write(unicode(json.dumps(answers, ensure_ascii=False)))
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
#dataset = load_data(FLAGS.data_dir) # None
dataset = {}
num_train = load_data_dq(dataset, 'train', FLAGS.data_dir)
num_val = load_data_dq(dataset, 'val', FLAGS.data_dir)
load_data_sa(dataset, 'train', FLAGS.data_dir, num_train)
load_data_sa(dataset, 'val', FLAGS.data_dir, num_val)
trim(dataset['train'])
trim(dataset['val'])
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = Encoder(size=FLAGS.state_size, vocab_dim=FLAGS.embedding_size)
decoder = Decoder(output_size=FLAGS.output_size)
# Session moved upfront to set the ifgpu flag before QASystem
with tf.Session() as sess:
pass
local_device_protos = device_lib.list_local_devices() # 38559755
for x in local_device_protos:
if x.device_type == 'GPU':
FLAGS.ifgpu = True
break
qa = QASystem(encoder, decoder, embed_path, rev_vocab)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
#print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
# tfdbg
#with tf.Session() as sess:
# pass
#sess = tfdbg.LocalCLIDebugWrapperSession(sess)
#sess.add_tensor_filter("has_inf_or_nan", tfdbg.has_inf_or_nan)
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
qa.train(sess, dataset, save_train_dir)
#qa.evaluate_answer(sess, dataset, vocab, FLAGS.evaluate, log=True)
sess.close() #tfdbg
def main(_):
dataset = {}
question_max_len = 40
context_max_len = 600
# Preprocess and collect small test data
test_q_path = pjoin(FLAGS.data_dir, "test.ids.question")
test_q_data, test_q_seq_len = pad_sentences(test_q_path, question_max_len)
assert not any(
test_q_seq_len > test_q_data.shape[1]
), 'Some questions have length greater than max question length'
test_c_path = pjoin(FLAGS.data_dir, "test.ids.context")
test_c_data, test_c_seq_len = pad_sentences(test_c_path, context_max_len)
assert not any(
test_c_seq_len > test_c_data.shape[1]
), 'Some contexts have length greater than max context length'
test_s_path = pjoin(FLAGS.data_dir, "test.span")
test_s_e_id = get_answer_span(test_s_path, context_max_len)
dataset['test'] = [
test_q_data, test_q_seq_len, test_c_data, test_c_seq_len, test_s_e_id
]
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = Encoder(state_size=FLAGS.state_size,
vocab_dim=FLAGS.embedding_size)
decoder = Decoder(state_size=FLAGS.state_size)
qa = QASystem(encoder, decoder, question_max_len, context_max_len,
embed_path, FLAGS.learning_rate, FLAGS.batch_size,
FLAGS.dropout, FLAGS.optimizer,
FLAGS.max_gradient_norm) # try without dropout
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
qa.train(sess, dataset['test'], save_train_dir,
small_data_test=True) #
def main(_):
vocab, rev_vocab = initialize_vocab(FLAGS.vocab_path)
FLAGS.embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
# ========= Load Dataset =========
# You can change this code to load dataset in your own way
dev_dirname = os.path.dirname(os.path.abspath(FLAGS.dev_path))
dev_filename = os.path.basename(FLAGS.dev_path)
context_data, question_data, question_uuid_data = prepare_dev(
dev_dirname, dev_filename, vocab)
dataset = {
"val_context": context_data,
"val_questions": question_data,
"val_question_uuids": question_uuid_data
}
# ========= Model-specific =========
# You must change the following code to adjust to your model
encoder = Encoder(size=FLAGS.state_size,
vocab_dim=FLAGS.embedding_size,
FLAGS=FLAGS)
decoder = Decoder(FLAGS=FLAGS)
qa = QASystem(encoder, decoder, FLAGS)
with tf.Session() as sess:
#train_dir = get_normalized_train_dir(FLAGS.train_dir)
train_dir = FLAGS.train_dir
print("train_dir: ", train_dir)
initialize_model(sess, qa, train_dir)
print("Generating Answers")
answers = generate_answers(sess, qa, dataset, rev_vocab)
# write to json file to root dir
print("Writing to json file")
with io.open('dev-prediction.json', 'w', encoding='utf-8') as f:
f.write(unicode(json.dumps(answers, ensure_ascii=False)))
def main(_):
# TODO
# Do what you need to load datasets from FLAGS.data_dir
dataset = dict()
for dataset_type in ['train', 'val']:
with open(os.path.join(FLAGS.data_dir,
"%s.ids.context" % dataset_type)) as f:
data_context = [
map(int, line.split()) for line in f.read().splitlines()
]
with open(
os.path.join(FLAGS.data_dir,
"%s.ids.question" % dataset_type)) as f:
data_question = [
map(int, line.split()) for line in f.read().splitlines()
]
with open(os.path.join(FLAGS.data_dir, "%s.span" % dataset_type)) as f:
data_span = [
map(int, line.split()) for line in f.read().splitlines()
]
dataset[dataset_type] = (data_context, data_question, data_span)
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = Encoder(size=FLAGS.state_size,
vocab_dim=FLAGS.embedding_size,
config=config)
decoder = Decoder(output_size=FLAGS.output_size, config=config)
qa = QASystem(encoder, decoder, config=config)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
qa.train(sess, dataset, save_train_dir)
qa.evaluate_answer(sess, dataset, FLAGS.evaluate, log=True)
def main(_):
data_dir = cfg.DATA_DIR
vocab, rev_vocab = initialize_vocab(FLAGS.vocab)
# gpu setting
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
tf.reset_default_graph()
encoder = Encoder(size=2 * cfg.lstm_num_hidden)
decoder = Decoder(output_size=2 * cfg.lstm_num_hidden)
qa = QASystem(encoder, decoder, FLAGS.embed)
with tf.Session(config=config) as sess:
init = tf.global_variables_initializer()
sess.run(init)
load_train_dir = get_normalized_train_dir(FLAGS.ckpt)
initialize_model(sess, qa, load_train_dir)
print(
'*********************************************************************'
)
print(
"Welcome! You can use this to explore the behavior of the model.")
print(
'*********************************************************************'
)
while True:
print('-------------------')
print('Input the context: ')
print('-------------------')
sentence = raw_input()
print('-------------------')
print('Input the question: ')
print('-------------------')
query = raw_input()
raw_context = nltk.word_tokenize(sentence)
context = sentence_to_token_ids(sentence,
vocab,
tokenizer=nltk.word_tokenize)
question = sentence_to_token_ids(query,
vocab,
tokenizer=nltk.word_tokenize)
context_in = mask_input(context, cfg.context_max_len)
question_in = mask_input(question, cfg.question_max_len)
start, end = qa.answer(sess, [context_in], [question_in])
answer = ' '.join(raw_context[start[0]:end[0] + 1])
print('==========================================')
print('ANSWER: {}'.format(answer))
print('==========================================')
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
# use .readlines() to load file ourselves
# use python generator
question_path = pjoin(FLAGS.data_dir, "data_train/train_tokenH.txt")
paragraph_path = pjoin(FLAGS.data_dir, "data_train/train_tokenP.txt")
answer_path = pjoin(FLAGS.data_dir, "data_train/train_index.txt")
val_question_path = pjoin(FLAGS.data_dir, "data_dev/dev_tokenH.txt")
val_paragraph_path = pjoin(FLAGS.data_dir, "data_dev/dev_tokenP.txt")
val_answer_path = pjoin(FLAGS.data_dir, "data_dev/dev_index.txt")
# for testing
# dataset = [(1,1,1), (1,1,1)]
dataset = load_dataset(question_path, paragraph_path, answer_path,
FLAGS.batch_size)
val_dataset = load_dataset(val_question_path, val_paragraph_path,
val_answer_path, FLAGS.batch_size)
#generate_histograms(dataset)
#generate_histograms(val_dataset)
# loads embedding
FLAGS.embed_path = FLAGS.embed_path or pjoin("data", "sgns.merge.word.npz")
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.txt")
vocab, rev_vocab = initialize_vocab(
vocab_path) # one is list and one is dict
encoder = Encoder(size=FLAGS.state_size, vocab_dim=FLAGS.embedding_size)
decoder = Decoder(size=FLAGS.state_size, output_size=FLAGS.output_size)
qa = QASystem(encoder, decoder, FLAGS)
# log file
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
# start training
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
qa.train(sess, dataset, val_dataset, save_train_dir, rev_vocab)
def main(_):
# TODO maybe pass as loaded dataset abstraction instead of
# file_paths?
default_hparams = create_hparams(FLAGS)
context_file_path = FLAGS.data_dir + '/train.ids.context'
question_file_path = FLAGS.data_dir + '/train.ids.question'
span_file_path = FLAGS.data_dir + '/train.span'
dataset = (context_file_path, question_file_path, span_file_path)
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
default_hparams.add_hparam('vocab_size', len(vocab))
encoder = Encoder(size=FLAGS.state_size, vocab_dim=FLAGS.embedding_size)
decoder = Decoder(state_size=FLAGS.state_size)
qa = QASystem(encoder, decoder, default_hparams)
# Setup embeddings
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
np_embeddings = np.float32(np.load(embed_path)['glove'])
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
session_config = tf.ConfigProto(log_device_placement=False,
allow_soft_placement=True)
session_config.gpu_options.allow_growth = True
with tf.Session(config=session_config) as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
qa.train(sess, dataset, save_train_dir, np_embeddings)
qa.evaluate_answer(sess, dataset, vocab, FLAGS.evaluate, log=True)
def run_func():
config = Config()
train = squad_dataset(config.question_train, config.context_train,
config.answer_train)
dev = squad_dataset(config.question_dev, config.context_dev,
config.answer_dev)
# print(config.question_train)
embed_path = config.embed_path
vocab_path = config.vocab_path
# print(config.embed_path, config.vocab_path)
vocab, rev_vocab = initialize_vocab(vocab_path)
embeddings = get_trimmed_glove_vectors(embed_path)
encoder = Encoder(config.hidden_size)
decoder = Decoder(config.hidden_size)
qa = QASystem(encoder, decoder, embeddings, config)
with tf.Session() as sess:
# ====== Load a pretrained model if it exists or create a new one if no pretrained available ======
qa.initialize_model(sess, config.train_dir)
# train process
# qa.train(sess, [train, dev], config.train_dir)
# em = qa.evaluate_model(sess, dev)
# run process
while True:
question = input('please input question: ')
if question == 'exit':
break
raw_context = input('please input context: ')
if raw_context == 'exit':
break
question = [
vocab[x] if x in vocab.keys() else 2 for x in question.split()
]
context = [
vocab[x] if x in vocab.keys() else 2
for x in raw_context.split()
]
test = [[question], [context], [[1, 2]]]
a_s, a_e = qa.answer(sess, test)
if a_e == a_s:
print("answer: ", raw_context.split()[a_s[0]])
else:
print("answer: ",
' '.join(raw_context.split()[a_s[0]:a_e[0] + 1]))
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
datasetTrain = initialize_datasets(FLAGS.data_dir,
'train.',
debugMode=False)
datasetVal = initialize_datasets(FLAGS.data_dir, 'val.', debugMode=False)
datasetTrain.extend(datasetVal)
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = Encoder(size=FLAGS.state_size, vocab_dim=FLAGS.embedding_size)
decoder = Decoder(output_size=FLAGS.output_size)
#This is taking a long time
tic = datetime.now()
qa = QASystem(encoder, decoder, embed_path, FLAGS, rev_vocab)
print('Time to setup the model: ', datetime.now() - tic)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
#saver = tf.train.Saver()
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
# Get directory to save model
#save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
results_path = "results/{:%Y%m%d_%H%M%S}/".format(datetime.now())
save_train_dir = results_path + "model.weights/"
if not os.path.exists(save_train_dir):
os.makedirs(save_train_dir)
qa.train(sess, datasetTrain, save_train_dir) #, saver)
qa.evaluate_answer(sess, datasetVal, rev_vocab, sample=1000, log=True)
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
dataset = load_dataset()
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = Encoder(state_size=FLAGS.state_size,
embedding_size=FLAGS.embedding_size,
output_size=FLAGS.output_size)
decoder = Decoder(state_size=FLAGS.state_size,
output_size=FLAGS.output_size)
qa_args = {
"embed_path": embed_path,
"embedding_size": FLAGS.embedding_size,
"output_size": FLAGS.output_size,
"optimizer": FLAGS.optimizer,
"learning_rate": FLAGS.learning_rate,
"epochs": FLAGS.epochs,
"batch_size": FLAGS.batch_size,
"max_gradient_norm": FLAGS.max_gradient_norm,
"dropout_keep_prob": 1.0 - FLAGS.dropout,
"train_dir": FLAGS.train_dir,
"state_size": FLAGS.state_size
}
qa = QASystem(encoder, decoder, **qa_args)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
qa.train(sess, dataset, save_train_dir)
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
dataset = {"train": load_data(FLAGS.data_dir, mode="train"), \
"val": load_data(FLAGS.data_dir, mode="val")}
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = Encoder(state_size=FLAGS.state_size,
vocab_dim=FLAGS.embedding_size)
matcher = Matcher(perspective_dim=25,
input_size=FLAGS.state_size) # add flag
decoder = Decoder(output_size=FLAGS.output_size,
state_size=FLAGS.state_size,
n_perspective_dim=50 * 2) # add flag
qa = QASystem(encoder, matcher, decoder, \
vocab=vocab, vocab_dim=FLAGS.embedding_size, rev_vocab=rev_vocab, embed_path=embed_path)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
tf.global_variables_initializer().run()
graph_writer = tf.summary.FileWriter("qa-graph")
graph_writer.add_graph(sess.graph)
qa.train(sess, dataset, save_train_dir)
qa.evaluate_answer(sess, dataset, 500, log=True)
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
training_question_data_path = pjoin(FLAGS.data_dir, 'train.question')
dataset = load_dataset(FLAGS.data_dir)
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
pretrained_embeddings = np.load(embed_path)['glove']
encoder = Encoder(size=FLAGS.state_size,
vocab_dim=FLAGS.embedding_size,
pretrained_embeddings=pretrained_embeddings,
max_question_length=FLAGS.max_question_length,
max_context_length=FLAGS.max_context_length)
decoder = Decoder(output_size=FLAGS.output_size,
size=FLAGS.state_size,
max_context_length=FLAGS.max_context_length)
qa = QASystem(encoder, decoder)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
qa.train(sess, dataset, save_train_dir)
qa.evaluate_answer(sess, dataset, vocab, FLAGS.evaluate, log=True)
def main(_):
dataset = [
load_dataset([
"embedding/train.ids.question", "embedding/train.ids.context",
"data/train.span"
]),
load_dataset([
"embedding/val.ids.question", "embedding/val.ids.context",
"data/val.span"
])
]
embed_path = pjoin("embedding",
"glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
tf.reset_default_graph()
encoder = Encoder(FLAGS.num_hidden_unit, tf.contrib.rnn.GRUCell)
decoder = Decoder(FLAGS.num_hidden_unit, tf.contrib.rnn.BasicLSTMCell)
embedding = np.load(embed_path)["glove"]
qa = QASystem(encoder, decoder, embedding, FLAGS.keep_prob)
if not os.path.exists(FLAGS.train_dir):
os.makedirs(FLAGS.train_dir)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
initialize_model(sess, qa, FLAGS.train_dir)
qa.train(sess, dataset, FLAGS.epochs, FLAGS.batch_size,
FLAGS.train_dir)
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
dataset = None
embed_path = FLAGS.embed_path or pjoin("data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
with np.load(embed_path) as data:
glove_embeddings = np.asfarray(data["glove"], dtype=np.float32)
dataset = load_and_preprocess_data()
# print(train_data)
encoder = Encoder(size=FLAGS.state_size, vocab_dim=FLAGS.embedding_size, config=FLAGS)
decoder = Decoder(output_size=FLAGS.output_size, config=FLAGS)
qa = QASystem(encoder, decoder, embeddings=glove_embeddings, config=FLAGS, vocab=(vocab, rev_vocab))
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
qa.train(sess, dataset, save_train_dir)
qa.evaluate_answer(sess, FLAGS.evaluate, log=True)
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
dataset = None
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = Encoder(size=FLAGS.size)
deocder = Decoder(output_size=FLAGS.output_size)
qa = QASystem(encoder, deocder)
global_train_dir = '/tmp/cs224n-squad-train'
# Adds symlink to {train_dir} from /tmp/cs224n-squad-train to canonicalize the
# file paths saved in the checkpoint. This allows the model to be reloaded even
# if the location of the checkpoint files has moved, allowing usage with CodaLab.
# This must be done on both train.py and qa_answer.py in order to work.
if os.path.exists(global_train_dir):
os.unlink(global_train_dir)
if not os.path.exists(FLAGS.train_dir):
os.makedirs(FLAGS.train_dir)
os.symlink(os.path.abspath(FLAGS.train_dir), global_train_dir)
train_dir = global_train_dir
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
initialize_model(sess, qa, train_dir)
qa.train(sess, dataset)
qa.evaluate_answer(sess, dataset, vocab, FLAGS.evaluate, log=True)
def run_func():
config = Config()
train = squad_dataset(config.question_train, config.context_train,
config.answer_train)
dev = squad_dataset(config.question_dev, config.context_dev,
config.answer_dev)
embed_path = config.embed_path
vocab_path = config.vocab_path
vocab, rev_vocab = initialize_vocab(vocab_path)
embeddings = get_trimmed_glove_vectors(embed_path)
encoder = Encoder(config.hidden_state_size)
decoder = Decoder(config.hidden_state_size)
qa = QASystem(encoder, decoder, embeddings, config)
with tf.Session() as sess:
# ====== Load a pretrained model if it exists or create a new one if no pretrained available ======
qa.initialize_model(sess, config.train_dir)
qa.train(sess, [train, dev], config.train_dir)
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
dataset = get_dataset(FLAGS.data_dir, FLAGS.max_question_size,
FLAGS.max_paragraph_size)
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
FLAGS.embed_path = embed_path
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = Encoder(size=FLAGS.state_size,
vocab_dim=FLAGS.embedding_size,
FLAGS=FLAGS)
decoder = Decoder(FLAGS=FLAGS)
qa = QASystem(encoder, decoder, FLAGS)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
#load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir or FLAGS.train_dir) #Change these back for final submission
load_train_dir = FLAGS.load_train_dir or FLAGS.train_dir
print("load_train_dir: ", load_train_dir)
initialize_model(sess, qa, load_train_dir)
#save_train_dir = get_normalized_train_dir(FLAGS.train_dir) #Change back for final submission
save_train_dir = FLAGS.train_dir
print("save_train_dir: ", save_train_dir)
qa.train(sess, dataset, save_train_dir, rev_vocab)
def main(_):
FLAGS.config = int(sys.argv[1])
load_config(current_config=FLAGS.config)
# Do what you need to load datasets from FLAGS.data_dir
dataset = load_data(FLAGS.data_dir) # ((question, context), answer)
train_data = preprocess_dataset(dataset['train'],
FLAGS.output_size, FLAGS.question_size)
val_data = preprocess_dataset(dataset['val'],
FLAGS.output_size, FLAGS.question_size)
# print(dataset)
embed_path = FLAGS.embed_path or pjoin("data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = Encoder(size=FLAGS.state_size, vocab_dim=FLAGS.embedding_size)
decoder = Decoder(output_size=FLAGS.output_size)
qa = QASystem(encoder, decoder)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir,
"log" + '_config_' + str(FLAGS.config) + ".txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags" + '_config_' +\
str(FLAGS.config) + ".json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
qa.train(sess, train_data, val_data, save_train_dir)
qa.evaluate_answer(sess, train_data, val_data, FLAGS.evaluate, log=True)
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
dataset = []
dataset.append(pjoin(FLAGS.data_dir, "train.ids.question"))
dataset.append(pjoin(FLAGS.data_dir, "train.ids.context"))
dataset.append(pjoin(FLAGS.data_dir, "train.span"))
dataset.append(pjoin(FLAGS.data_dir, "val.ids.question"))
dataset.append(pjoin(FLAGS.data_dir, "val.ids.context"))
dataset.append(pjoin(FLAGS.data_dir, "val.span"))
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = Encoder(size=FLAGS.state_size, vocab_dim=FLAGS.embedding_size)
decoder = Decoder(output_size=FLAGS.paragraph_output_size)
qa = QASystem(encoder, decoder)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
qa.train(sess, dataset, save_train_dir)
def main(_):
config = Config()
dataset = None # TODO ;load dateset ??? - look at dataset and seenhow it loooks - change model.py accordingly
embed_path = config.embed_path or pjoin("data", "squad", "glove.trimmed.{}.npz".format(config.embed_size))
embeddingz = np.load(embed_path)
embeddings = embeddingz['glove']
embeddingz.close()
vocab_len = embeddings.shape[0]
train = load_squad(config.data_dir, "train", vocab_len, config.data_dir, max_samples=config.max_train_samples)
val = load_squad(config.data_dir, "val", vocab_len, config.data_dir, max_samples=config.max_val_samples)
print('train size: ', len(train), ' val size: ', len(val))
vocab_path = config.vocab_path or pjoin(config.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
# print_samples(train,1, rev_vocab) #train is triplets of (context,question,answer)
encoder = Encoder(state_size=config.hidden_size, embedding_size=config.embed_size)
decoder = Decoder(state_size=config.hidden_size, embedding_size=config.embed_size)
qa = QASystem(encoder, decoder)
with tf.Session() as sess:
load_train_dir = (config.load_train_dir or config.train_dir) # put "" here if you want to build a new model
initialize_model(sess, qa, load_train_dir)
save_train_dir = config.train_dir
ds_train = qa.pad_sequences(train)
ret_q, ret_p, ret_labels = ds_train
qa.train(sess, ds_train, save_train_dir)
ds_val = qa.pad_sequences(val)
print('train error')
qa.evaluate_answer(sess, ds_train, log=True)
print('val error')
qa.evaluate_answer(sess, ds_val, log=True)
def main(_):
# Do what you need to load datasets from FLAGS.data_dir
dataset = {}
question_max_len = 40
context_max_len = 600
# Preprocess and collect train data
train_q_path = pjoin(FLAGS.data_dir, "train.ids.question")
train_q_data, train_q_seq_len = pad_sentences(train_q_path,
question_max_len)
assert not any(
train_q_seq_len > train_q_data.shape[1]
), 'Some questions have length greater than max question length'
train_c_path = pjoin(FLAGS.data_dir, "train.ids.context")
train_c_data, train_c_seq_len = pad_sentences(train_c_path,
context_max_len)
assert not any(
train_c_seq_len > train_c_data.shape[1]
), 'Some contexts have length greater than max context length'
train_s_path = pjoin(FLAGS.data_dir, "train.span")
train_s_e_id = get_answer_span(train_s_path, context_max_len)
dataset['train'] = [
train_q_data, train_q_seq_len, train_c_data, train_c_seq_len,
train_s_e_id
]
# Preprocess and collect validation data
val_q_path = pjoin(FLAGS.data_dir, "val.ids.question")
val_q_data, val_q_seq_len = pad_sentences(val_q_path, question_max_len)
val_c_path = pjoin(FLAGS.data_dir, "val.ids.context")
val_c_data, val_c_seq_len = pad_sentences(val_c_path, context_max_len)
val_s_path = pjoin(FLAGS.data_dir, "val.span")
val_s_e_id = get_answer_span(val_s_path, context_max_len)
dataset['val'] = [
val_q_data, val_q_seq_len, val_c_data, val_c_seq_len, val_s_e_id
]
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = Encoder(state_size=FLAGS.state_size,
vocab_dim=FLAGS.embedding_size)
decoder = Decoder(state_size=FLAGS.state_size)
qa = QASystem(encoder, decoder, question_max_len, context_max_len,
embed_path, FLAGS.learning_rate, FLAGS.batch_size,
FLAGS.dropout, FLAGS.optimizer,
FLAGS.max_gradient_norm) # try without dropout
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
for i in range(FLAGS.epochs):
qa.train(sess, dataset['train'], save_train_dir) #
print('Finish training epoch {}'.format(i))
qa.evaluate_answer(sess, dataset['val']) # vocab, FLAGS.evaluate
def main(_):
# Reads file name for paragraph, parses paragraphs, changes form to list of lists of word embeddings
# and stores length of each paragraph in paraLens. Next 5 blocks of code do the same for questions and get the indices
# of answers within the paragraph. Also, does the same for all three elements of the validation dataset.
paraFileName = pjoin(FLAGS.data_dir, "train") + ".ids.context"
paraFile = open(paraFileName, "r")
paras = [[int(a) for a in b.strip().split(" ") if a != '']
for b in paraFile.read().strip().split("\n")]
paras, paraLens = zip(*[padList(para, FLAGS.paraLen) for para in paras])
questionFileName = pjoin(FLAGS.data_dir, "train") + ".ids.question"
questionFile = open(questionFileName, "r")
questions = [[int(a) for a in b.strip().split(" ") if a != '']
for b in questionFile.read().strip().split("\n")]
questions, questionLens = zip(
*[padList(question, FLAGS.qLen) for question in questions])
ansFileName = pjoin(FLAGS.data_dir, "train") + ".span"
ansFile = open(ansFileName, "r")
ans = [[int(a) for a in b.strip().split(" ") if a != '']
for b in ansFile.read().strip().split("\n")]
paraFileNameVal = pjoin(FLAGS.data_dir, "val") + ".ids.context"
paraFileVal = open(paraFileNameVal, "r")
parasVal = [[int(a) for a in b.strip().split(" ") if a != '']
for b in paraFileVal.read().strip().split("\n")]
parasVal, paraLensVal = zip(
*[padList(para, FLAGS.paraLen) for para in parasVal])
questionFileNameVal = pjoin(FLAGS.data_dir, "val") + ".ids.question"
questionFileVal = open(questionFileNameVal, "r")
questionsVal = [[int(a) for a in b.strip().split(" ") if a != '']
for b in questionFileVal.read().strip().split("\n")]
questionsVal, questionLensVal = zip(
*[padList(question, FLAGS.qLen) for question in questionsVal])
ansFileNameVal = pjoin(FLAGS.data_dir, "val") + ".span"
ansFileVal = open(ansFileNameVal, "r")
ansVal = [[int(a) for a in b.strip().split(" ") if a != '']
for b in ansFileVal.read().strip().split("\n")]
train_remove = []
val_remove = []
for i in range(len(ans)):
if ans[i][0] >= FLAGS.paraLen or ans[i][1] >= FLAGS.paraLen:
train_remove.append(i)
for i in range(len(ansVal)):
if ansVal[i][0] >= FLAGS.paraLen or ans[i][1] >= FLAGS.paraLen:
val_remove.append(i)
train_remove = list(reversed(train_remove))
val_remove = list(reversed(val_remove))
paras = list(paras)
questions = list(questions)
paraLens = list(paraLens)
questionLens = list(questionLens)
parasVal = list(parasVal)
questionsVal = list(questionsVal)
paraLensVal = list(paraLensVal)
questionLensVal = list(questionLensVal)
for i in range(len(train_remove)):
ind = train_remove[i]
paras.pop(ind)
questions.pop(ind)
paraLens.pop(ind)
questionLens.pop(ind)
ans.pop(ind)
for i in range(len(val_remove)):
ind = val_remove[i]
parasVal.pop(ind)
questionsVal.pop(ind)
paraLensVal.pop(ind)
questionLensVal.pop(ind)
ansVal.pop(ind)
paras = tuple(paras)
questions = tuple(questions)
paraLens = tuple(paraLens)
questionLens = tuple(questionLens)
parasVal = tuple(parasVal)
questionsVal = tuple(questionsVal)
paraLensVal = tuple(paraLensVal)
questionLensVal = tuple(questionLensVal)
#valDataset has paragraphs and questions as word embeddings, answers as the start and end indices in the
#paragraph and length of each paragraph and question for the validation dataset
valDataset = (parasVal, questionsVal, ansVal, paraLensVal, questionLensVal)
#dataset is the same format as valDataset but for the training dataset. It also has valDataset as the last element
dataset = (paras, questions, ans, paraLens, questionLens, valDataset)
embed_path = FLAGS.embed_path or pjoin(
"data", "squad", "glove.trimmed.{}.npz".format(FLAGS.embedding_size))
vocab_path = FLAGS.vocab_path or pjoin(FLAGS.data_dir, "vocab.dat")
vocab, rev_vocab = initialize_vocab(vocab_path)
encoder = CoattentionEncoder(size=FLAGS.state_size,
vocab_dim=FLAGS.embedding_size)
#encoder = Encoder(size=FLAGS.state_size, vocab_dim=FLAGS.embedding_size)
decoder = Decoder(output_size=FLAGS.output_size)
#decoder = ChunkDecoder(output_size=FLAGS.output_size)
embeddings = np.load(embed_path)['glove']
#print(embeddings.keys())
qa = QASystem(encoder, decoder, embeddings, vocab, rev_vocab)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
file_handler = logging.FileHandler(pjoin(FLAGS.log_dir, "log.txt"))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.log_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
saver = tf.train.Saver()
with tf.Session() as sess:
load_train_dir = get_normalized_train_dir(FLAGS.load_train_dir
or FLAGS.train_dir)
initialize_model(sess, qa, load_train_dir, saver)
save_train_dir = get_normalized_train_dir(FLAGS.train_dir)
sample_rates = [1, 2, 5, 10]
for rate in sample_rates:
print("rate: ", rate)
f1_l = []
em_l = []
for i in range(10):
f1, em = qa.evaluate_answer(sess,
valDataset,
sample=len(valDataset[0]),
sample_rate=rate)
f1_l.append(f1)
em_l.append(em)
#print("f1: ", f1, "; em: ", em)
print(rate)
print(f1_l)
print(em_l)
def main(_):
'''Check the Config.py to set up models pathes to be ensembled.'''
data_dir = cfg.DATA_DIR
set_names = cfg.set_names
suffixes = cfg.suffixes
dataset = mask_dataset(data_dir, set_names, suffixes)
raw_answers = read_raw_answers(data_dir)
vocab_path = pjoin(data_dir, cfg.vocab_file)
vocab, rev_vocab = initialize_vocab(vocab_path)
if not os.path.exists(cfg.log_dir):
os.makedirs(cfg.log_dir)
if not os.path.exists(cfg.cache_dir):
os.makedirs(cfg.cache_dir)
if not os.path.exists(cfg.fig_dir):
os.makedirs(cfg.fig_dir)
c_time = time.strftime('%Y%m%d_%H%M', time.localtime())
file_handler = logging.FileHandler(
pjoin(cfg.log_dir, 'ensemble_log' + c_time + '.txt'))
logging.getLogger().addHandler(file_handler)
model_pathes = cfg.model_pathes
num_m = len(model_pathes)
train_s = np.zeros((cfg.num_eval, num_m), dtype=np.int32)
train_e = np.zeros((cfg.num_eval, num_m), dtype=np.int32)
val_s = np.zeros((cfg.num_eval, num_m), dtype=np.int32)
val_e = np.zeros((cfg.num_eval, num_m), dtype=np.int32)
# gpu setting
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
for i in xrange(num_m):
tf.reset_default_graph()
with tf.Session(config=config) as sess:
encoder = Encoder(size=2 * cfg.lstm_num_hidden)
decoder = Decoder(output_size=2 * cfg.lstm_num_hidden)
qa = QASystem(encoder, decoder)
init = tf.global_variables_initializer()
sess.run(init)
load_train_dir = get_normalized_train_dir(model_pathes[i])
initialize_model(sess, qa, load_train_dir)
ts, te, vs, ve = qa.evaluate_answer(sess,
dataset,
raw_answers,
rev_vocab,
log=True,
ensemble=True,
training=True,
sample=cfg.num_eval)
train_s[:, i] = ts
train_e[:, i] = te
val_s[:, i] = vs
val_e[:, i] = ve
if i == num_m - 1:
# np.save('cache/ensemble.npy', [train_s, train_e, val_s, val_e])
train_s = bin_count(train_s)
train_e = bin_count(train_e)
val_s = bin_count(val_s)
val_e = bin_count(val_e)
qa.evaluate_answer(sess,
dataset,
raw_answers,
rev_vocab,
log=True,
training=True,
sendin=(train_s, train_e, val_s, val_e),
sample=cfg.num_eval)
