def main():
qanet = QANet(50)
init1 = filter(lambda p: p.requires_grad and p.dim() >= 2,
qanet.parameters())
init2 = filter(lambda p: p.requires_grad and p.dim() <= 2,
qanet.parameters())
# Parameter initialization
for param in init1:
nn.init.xavier_uniform_(param)
for param in init2:
nn.init.normal_(param)
train = SQuAD(TRAIN_JSON)
val = SQuAD(DEV_JSON)
# trainSet = DataLoader(dataset=train, batch_size=4, shuffle=True, collate_fn=collate)
valSet = DataLoader(dataset=val,
batch_size=4,
shuffle=True,
collate_fn=collate)
trainSet = DataLoader(dataset=train,
batch_size=4,
shuffle=True,
collate_fn=collate)
print('length of dataloader', len(trainSet))
optimizer = torch.optim.Adam(qanet.parameters(), lr=LEARNING_RATE)
loss_list = []
for epoch in range(10):
print('epoch ', epoch)
for i, (c, q, a) in enumerate(trainSet):
y_pred = qanet(c, q)
loss = utils.loss(y_pred, a)
loss_list.append(loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
if i % 200 == 0:
print('loss ', loss.item())
with open('your_file.txt', 'w') as f:
for item in loss_list:
f.write("%s\n" % item)
print('loss file written.')
torch.save(qanet, 'qanet')
print('model saved.')
def demo(config):
with open(config.word_emb_file, "r") as fh:
word_mat = np.array(json.load(fh), dtype=np.float32)
with open(config.char_emb_file, "r") as fh:
char_mat = np.array(json.load(fh), dtype=np.float32)
with open(config.test_meta, "r") as fh:
meta = json.load(fh)
model = QANet(config, None, word_mat, char_mat, trainable=False, demo=True)
demo = Demo(model, config)
def main():
r"""
Main function.
"""
model = QANet()
# initial parameters
logging.info('Initial paramters...')
if opt.load_trained_model:
model.load_parameters(opt.trained_model_name, ctx=CTX)
else:
logging.info('Initial model parameters...')
initial_model_parameters(model)
print(model)
if opt.is_train:
loss_function = MySoftmaxCrossEntropy()
ema = ExponentialMovingAverage(decay=opt.ema_decay)
# initial trainer
trainer = gluon.Trainer(
model.collect_params(), 'adam', {
'learning_rate': opt.init_learning_rate,
'beta1': opt.beta1,
'beta2': opt.beta2,
'epsilon': opt.epsilon
})
if opt.load_trained_model:
trainer.load_states(opt.trained_trainer_name)
# initial dataloader
train_data_loader = DataLoader(batch_size=opt.train_batch_size,
dev_set=False)
# train
logging.info('Train')
train(model, train_data_loader, trainer, loss_function, ema)
else:
logging.info('Evaluating dev set...')
f1_score, em_score = evaluate(model, dataset_type='dev', ema=None)
logging.debug('The dev dataset F1 is:%.5f, and EM is: %.5f', f1_score,
em_score)
def test(config):
os.environ["CUDA_VISIBLE_DEVICES"] = config.choose_gpu
with open(config.word_emb_file, "r") as fh:
word_mat = np.array(json.load(fh), dtype=np.float32)
with open(config.char_emb_file, "r") as fh:
char_mat = np.array(json.load(fh), dtype=np.float32)
with open(config.test_eval_file, "r") as fh:
eval_file = json.load(fh)
with open(config.test_meta, "r") as fh:
meta = json.load(fh)
total = meta["total"]
graph = tf.Graph()
print("Loading model...")
with graph.as_default() as g:
test_batch = get_dataset(config.test_record_file,
get_record_parser(config, is_test=True),
config).make_one_shot_iterator()
model = QANet(config,
test_batch,
word_mat,
char_mat,
trainable=False,
graph=g)
sess_config = tf.ConfigProto(allow_soft_placement=True)
sess_config.gpu_options.allow_growth = True
sess_config.gpu_options.per_process_gpu_memory_fraction = config.gpu_memory_fraction
with tf.Session(config=sess_config) as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.save_dir))
if config.decay < 1.0:
sess.run(model.assign_vars)
losses = []
answer_dict = {}
remapped_dict = {}
for step in tqdm(range(total // config.batch_size + 1)):
qa_id, loss, yp1, yp2 = sess.run(
[model.qa_id, model.loss, model.yp1, model.yp2])
answer_dict_, remapped_dict_ = convert_tokens(
eval_file, qa_id.tolist(), yp1.tolist(), yp2.tolist())
answer_dict.update(answer_dict_)
remapped_dict.update(remapped_dict_)
losses.append(loss)
loss = np.mean(losses)
metrics = evaluate(eval_file, answer_dict)
with open(config.answer_file, "w") as fh:
json.dump(remapped_dict, fh)
print("Exact Match: {}, F1: {}".format(metrics['exact_match'],
metrics['f1']))
def main():
r"""
Main function.
"""
model = QANet()
# initial parameters
print('Initial paramters...')
if NEED_LOAD_TRAINED_MODEL:
model.load_parameters(TARGET_MODEL_FILE_NAME, ctx=CTX)
else:
print('Initial model parameters...')
initial_model_parameters(model)
print(model)
if TRAIN_FLAG is True:
loss_function = MySoftmaxCrossEntropy()
ema = ExponentialMovingAverage(decay=EXPONENTIAL_MOVING_AVERAGE_DECAY)
# initial trainer
trainer = gluon.Trainer(
model.collect_params(), 'adam', {
'learning_rate': INIT_LEARNING_RATE,
'beta1': BETA1,
'beta2': BETA2,
'epsilon': EPSILON
})
if NEED_LOAD_TRAINED_MODEL:
trainer.load_states(TARGET_TRAINER_FILE_NAME)
# initial dataloader
train_data_loader = DataLoader(batch_size=TRAIN_BATCH_SIZE,
dev_set=False)
# train
print('Train...')
train(model, train_data_loader, trainer, loss_function, ema)
else:
print('Evaluating dev set...')
f1_score, em_score = evaluate(model, dataset_type='dev', ema=None)
print('The dev dataset F1 is:%s, and EM is: %s' % (f1_score, em_score))
def test(cfg):
logging.info('Model is loading...')
with open(cfg['dev_eval_file'], "r") as fh:
dev_eval_file = json.load(fh)
dev_dataset = SQuADDataset(cfg['dev_record_file'], -1, cfg['batch_size'], cfg['word2ind_file'])
model_args = pickle.load(open(cfg['args_filename'], 'rb'))
model = QANet(**model_args)
model.load_state_dict(torch.load(cfg['dump_filename']))
model.to(device)
metrics, answer_dict = evaluation(model, dev_dataset, dev_eval_file, len(dev_dataset))
with open('logs/answers.json', 'w') as f:
json.dump(answer_dict, f)
logging.info("TEST loss %f F1 %f EM %f\n", metrics["loss"], metrics["f1"], metrics["exact_match"])
def train(model_params, launch_params):
with open(launch_params['word_emb_file'], "r") as fh:
word_mat = np.array(json.load(fh), dtype=np.float32)
with open(launch_params['char_emb_file'], "r") as fh:
char_mat = np.array(json.load(fh), dtype=np.float32)
with open(launch_params['train_eval_file'], "r") as fh:
train_eval_file = json.load(fh)
with open(launch_params['dev_eval_file'], "r") as fh:
dev_eval_file = json.load(fh)
writer = SummaryWriter(os.path.join(launch_params['log'], launch_params['prefix']))
lr = launch_params['learning_rate']
base_lr = 1.0
warm_up = launch_params['lr_warm_up_num']
model_params['word_mat'] = word_mat
model_params['char_mat'] = char_mat
logging.info('Load dataset and create model.')
dev_dataset = SQuADDataset(launch_params['dev_record_file'], launch_params['test_num_batches'],
launch_params['batch_size'], launch_params['word2ind_file'])
if launch_params['fine_tuning']:
train_dataset = SQuADDataset(launch_params['train_record_file'], launch_params['fine_tuning_steps'],
launch_params['batch_size'], launch_params['word2ind_file'])
model_args = pickle.load(open(launch_params['args_filename'], 'rb'))
model = QANet(**model_args)
model.load_state_dict(torch.load(launch_params['dump_filename']))
model.to(device)
else:
train_dataset = SQuADDataset(launch_params['train_record_file'], launch_params['num_steps'],
launch_params['batch_size'], launch_params['word2ind_file'])
model = QANet(**model_params).to(device)
launch_params['fine_tuning_steps'] = 0
params = filter(lambda param: param.requires_grad, model.parameters())
optimizer = optim.Adam(params, lr=base_lr, betas=(launch_params['beta1'], launch_params['beta2']), eps=1e-7, weight_decay=3e-7)
cr = lr / log2(warm_up)
scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda ee: cr * log2(ee + 1) if ee < warm_up else lr)
qt = False
logging.info('Start training.')
for iter in range(launch_params['num_steps']):
try:
passage_w, passage_c, question_w, question_c, y1, y2, ids = train_dataset[iter]
passage_w, passage_c = passage_w.to(device), passage_c.to(device)
question_w, question_c = question_w.to(device), question_c.to(device)
y1, y2 = y1.to(device), y2.to(device)
loss, p1, p2 = model.train_step([passage_w, passage_c, question_w, question_c], y1, y2, optimizer, scheduler)
if iter % launch_params['train_interval'] == 0:
logging.info('Iteration %d; Loss: %f', iter+launch_params['fine_tuning_steps'], loss)
writer.add_scalar('Loss', loss, iter+launch_params['fine_tuning_steps'])
if iter % launch_params['train_sample_interval'] == 0:
start = torch.argmax(p1[0, :]).item()
end = torch.argmax(p2[0, start:]).item()+start
passage = train_dataset.decode(passage_w)
question = train_dataset.decode(question_w)
generated_answer = train_dataset.decode(passage_w[:, start:end+1])
real_answer = train_dataset.decode(passage_w[:, y1[0]:y2[0]+1])
logging.info('Train Sample:\n Passage: %s\nQuestion: %s\nOriginal answer: %s\nGenerated answer: %s',
passage, question, real_answer, generated_answer)
if iter % launch_params['test_interval'] == 0:
metrics, _ = evaluation(model, train_dataset, train_eval_file, launch_params['val_num_batches'])
logging.info("VALID loss %f F1 %f EM %f", metrics['loss'], metrics['f1'], metrics['exact_match'])
writer.add_scalar('Valid_loss', metrics['loss'], iter)
writer.add_scalar('Valid_f1', metrics['f1'], iter)
writer.add_scalar('Valid_em', metrics['exact_match'], iter)
if iter % launch_params['test_interval'] == 0:
metrics, _ = evaluation(model, dev_dataset, dev_eval_file, launch_params['test_num_batches'])
logging.info("TEST loss %f F1 %f EM %f", metrics['loss'], metrics['f1'], metrics['exact_match'])
writer.add_scalar('Test_loss', metrics['loss'], iter)
writer.add_scalar('Test_f1', metrics['f1'], iter)
writer.add_scalar('Test_em', metrics['exact_match'], iter)
except RuntimeError as e:
logging.error(str(e))
except KeyboardInterrupt:
break
torch.save(model.cpu().state_dict(), launch_params['dump_filename'])
pickle.dump(model_params, open(launch_params['args_filename'], 'wb'))
logging.info('Model has been saved.')
def train(config):
print(dict(config.__dict__['__flags']))
print()
print(sys.stdout.flush())
os.environ["CUDA_VISIBLE_DEVICES"] = config.choose_gpu
with open(config.word_emb_file, "r") as fh:
word_mat = np.array(json.load(fh), dtype=np.float32)
with open(config.char_emb_file, "r") as fh:
char_mat = np.array(json.load(fh), dtype=np.float32)
with open(config.train_eval_file, "r") as fh:
train_eval_file = json.load(fh)
with open(config.dev_eval_file, "r") as fh:
dev_eval_file = json.load(fh)
with open(config.dev_meta, "r") as fh:
meta = json.load(fh)
dev_total = meta["total"]
print("Building model...")
parser = get_record_parser(config)
graph = tf.Graph()
with graph.as_default() as g:
train_dataset = get_batch_dataset(config.train_record_file, parser,
config)
dev_dataset = get_dataset(config.dev_record_file, parser, config)
handle = tf.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(
handle, train_dataset.output_types, train_dataset.output_shapes)
train_iterator = train_dataset.make_one_shot_iterator()
dev_iterator = dev_dataset.make_one_shot_iterator()
model = QANet(config, iterator, word_mat, char_mat, graph=g)
sess_config = tf.ConfigProto(allow_soft_placement=True)
sess_config.gpu_options.allow_growth = True
sess_config.gpu_options.per_process_gpu_memory_fraction = config.gpu_memory_fraction
loss_save = 100.0
patience = 0
best_f1 = 0.
best_em = 0.
with tf.Session(config=sess_config) as sess:
writer = tf.summary.FileWriter(config.log_dir)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
train_handle = sess.run(train_iterator.string_handle())
dev_handle = sess.run(dev_iterator.string_handle())
if os.path.exists(os.path.join(config.save_dir, "checkpoint")):
saver.restore(sess,
tf.train.latest_checkpoint(config.save_dir))
global_step = max(sess.run(model.global_step), 1)
for _ in tqdm(range(global_step, config.num_steps + 1)):
global_step = sess.run(model.global_step) + 1
loss, train_op = sess.run([model.loss, model.train_op],
feed_dict={
handle: train_handle,
model.dropout: config.dropout
})
if global_step % config.period == 0:
loss_sum = tf.Summary(value=[
tf.Summary.Value(tag="model/loss", simple_value=loss),
])
writer.add_summary(loss_sum, global_step)
if global_step % config.checkpoint == 0:
_, summ = evaluate_batch(model, config.val_num_batches,
train_eval_file, sess, "train",
handle, train_handle)
for s in summ:
writer.add_summary(s, global_step)
metrics, summ = evaluate_batch(
model, dev_total // config.batch_size + 1,
dev_eval_file, sess, "dev", handle, dev_handle)
dev_f1 = metrics["f1"]
dev_em = metrics["exact_match"]
if dev_f1 < best_f1 and dev_em < best_em:
patience += 1
if patience > config.early_stop:
break
else:
patience = 0
best_em = max(best_em, dev_em)
best_f1 = max(best_f1, dev_f1)
for s in summ:
writer.add_summary(s, global_step)
writer.flush()
filename = os.path.join(
config.save_dir, "model_{}.ckpt".format(global_step))
saver.save(sess, filename)
eval_dataset = SQuADDataset('data/dev.npz')
train_dataset = SQuADDataset('data/train.npz')
print('Loading Embeddigs..')
import numpy as np
import json
char_emb_matrix = np.array(json.load(open('data/char_emb.json')),
dtype=np.float32)
word_emb_matrix = np.array(json.load(open('data/word_emb.json')),
dtype=np.float32)
print('Create Model..')
from model import QANet
model = QANet(128, 400, 50, word_emb_matrix, char_emb_matrix,
droprate=0.1).to(device)
optimizer = optim.Adam(model.parameters(),
lr=0.001,
betas=(0.8, 0.999),
eps=1e-08,
weight_decay=3e-07,
amsgrad=False)
del char_emb_matrix, word_emb_matrix
import math
warm_up = 1000
warm_up_f = lambda x: math.log(x + 1) / math.log(warm_up
) if x < warm_up else 1
scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=[warm_up_f])
def train(num_units, batch_size, sentence_size, embedding_size, ctx):
net = QANet(num_units, batch_size, sentence_size, embedding_size)
net.collect_params().initialize(ctx)
loss = LogLoss()
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': .1})
def get_model(opt, word_mat, char_mat):
model = QANet(word_mat, char_mat,
opt.dropout, opt.dropout_char, opt.max_passage_len,
opt.encode_size)
