def exp1():
# create dataset
dataset = Dataset(pipe1)
# create word2vec model
w2v = Word2Vocab(dim=200)
trainfeed = DataFeed(op1, batch_size=64,
datapoints=dataset.trainset, w2v=w2v)
testfeed = DataFeed(op1, batch_size=32,
datapoints=dataset.testset, w2v=w2v)
# instantiate model
asreader = AttentionSumReader(hdim=200, emb_dim=200,
embedding=w2v.load_embeddings(),
vocab_size=w2v.vocab_size(),
num_layers=3)
# training
with tf.Session() as sess:
# init variables
sess.run(tf.global_variables_initializer())
# create trainer instance
trainer = Trainer(model=asreader,
trainfeed=trainfeed,
testfeed=testfeed,
verbose=True,
dump_cache=False)
# let the training begin
trainer.train()
def experiment(epochs=10, checkpoint=1, train_datapoints=train_datapoints):
model = Model(Config(), len(INPUT_VOCAB), len(OUTPUT_VOCAB))
if Config().cuda: model = model.cuda()
split_index = int(len(train_datapoints) * 0.85)
train_feed = DataFeed(train_datapoints[:split_index],
batchop=batchop,
batch_size=128)
test_feed = DataFeed(train_datapoints[split_index:],
batchop=batchop,
batch_size=120)
trainer = Trainer(model=model,
loss_function=loss,
accuracy_function=accuracy,
checkpoint=checkpoint,
epochs=epochs,
feeder=Feeder(train_feed, test_feed))
predictor = Predictor(model=model,
repr_function=repr_function,
feed=test_feed)
for e in range(1):
output, results = predictor.predict(
random.choice(range(test_feed.num_batch)))
display(HTML(results._repr_html_()))
del output, results
trainer.train()
def eval_joint_model(model, model_params):
# get test data for evaluation
data, metadata = gather('10k', 0)
# build data format
dformat = ['contexts', 'questions', 'answers']
accuracies = []
with tf.Session() as sess:
# reload model params
sess.run(set_op(model_params))
# build trainer
trainer = Trainer(sess, model, batch_size=128)
for i in range(1, 21):
# test feed for task 'i'
testfeed = DataFeed(dformat, data=data['test'][i])
# evaluate
loss, acc = trainer.evaluate(feed=testfeed)
# note down task accuracy
accuracies.append(acc)
print('\n:: Evaluation on individual tasks\n:: Accuracy')
for i, acc in enumerate(accuracies):
print(':: \t[{}] {}'.format(i, acc))
def __init__(self, instrument, strategies, start_date, end_date, opening_bal,time_stamp=None):
'''
constructs message queues
initialises brokers and traders
'''
self.instrument = instrument
self.start_date = start_date
self.end_date = end_date
self.opening_bal = opening_bal
self.datafeed = DataFeed(instrument)
self.orderQ = MsgQ()
self.receiptQ = MsgQ()
self.term_req_Q = MsgQ()
self.term_notice_Q = MsgQ()
self.broker = Broker(self.datafeed, self.orderQ, self.receiptQ, self.term_req_Q, self.term_notice_Q)
self.traders = []
for strategy in strategies:
trader = Trader(self.datafeed, self.broker, self.opening_bal, self.instrument, strategy, self.start_date, self.end_date)
self.traders.append(trader)
self.time_stamp = time_stamp
def experiment(eons=1000, epochs=10, checkpoint=5):
try:
try:
model = BiLSTMDecoderModel(Config(), len(INPUT_VOCAB), len(CHAR_VOCAB), len(OUTPUT_VOCAB))
if Config().cuda: model = model.cuda()
model.load_state_dict(torch.load('attn_model.pth'))
log.info('loaded the old image for the model')
except:
log.exception('failed to load the model')
model = BiLSTMDecoderModel(Config(), len(INPUT_VOCAB), len(CHAR_VOCAB), len(OUTPUT_VOCAB))
if Config().cuda: model = model.cuda()
print('**** the model', model)
train_feed, test_feed, predictor_feed = {}, {}, {}
trainer, predictor = {}, {}
max_size = max( sorted( [len(i[0]) for i in classified_datapoints.values()] )[:-1] )
#max_size = max( sorted( [len(i[0]) for i in classified_datapoints.values()] ) )
for label in classified_datapoints.keys():
if len(classified_datapoints[label][0]) < 1: continue
label_desc = '-'.join([OUTPUT_VOCAB[l] for l in [i for i, x in enumerate(label) if x == 1]] )
print('label: {} and size: {}'.format(label, len(classified_datapoints[label][0])))
train_feed[label] = DataFeed(label_desc, classified_datapoints[label][0], batchop=batchop, batch_size=max(128, int(len(classified_datapoints[label][0])/600)) )
test_feed[label] = DataFeed(label_desc, classified_datapoints[label][1], batchop=batchop, batch_size=32)
predictor_feed[label] = DataFeed(label_desc, classified_datapoints[label][1], batchop=batchop, batch_size=12)
turns = int(max_size/train_feed[label].size) + 1
trainer[label] = Trainer(name=label_desc,
model=model,
loss_function=partial(loss, scale=1), accuracy_function=accuracy, f1score_function=f1score_function,
checkpoint=checkpoint, epochs=epochs,
feeder = Feeder(train_feed[label], test_feed[label]))
predictor[label] = Predictor(model=model, feed=predictor_feed[label], repr_function=repr_function)
test_predictor_feed = DataFeed('test', test_datapoints, batchop=test_batchop, batch_size=128)
test_predictor = Predictor(model=model, feed=test_predictor_feed, repr_function=test_repr_function)
all_class_train_feed = MultiplexedDataFeed('atrain', train_feed.values(), batchop=batchop, batch_size=256)
all_class_test_feed = MultiplexedDataFeed('atest', test_feed.values(), batchop=batchop, batch_size=256)
all_class_predictor_feed = MultiplexedDataFeed('apredict',predictor_feed.values(), batchop=batchop, batch_size=256)
all_class_trainer = Trainer(name='all_class_trainer',
model=model,
loss_function=partial(loss, scale=1), accuracy_function=accuracy, f1score_function=f1score_function,
checkpoint=checkpoint, epochs=epochs,
feeder = Feeder(all_class_train_feed, all_class_test_feed))
all_class_predictor = Predictor(model=model, feed=all_class_predictor_feed, repr_function=repr_function)
label_trainer_triples = sorted( [(l, t, train_feed[l].size) for l, t in trainer.items()], key=lambda x: x[2] )
log.info('trainers built {}'.format(pformat(label_trainer_triples)))
dump = open('results/experiment_attn.csv', 'w').close()
for e in range(eons):
dump = open('results/experiment_attn.csv', 'a')
dump.write('#========================after eon: {}\n'.format(e))
dump.close()
log.info('on {}th eon'.format(e))
if e and not e % 1:
test_results = ListTable()
test_dump = open('results/experiment_attn_over_test_{}.csv'.format(e), 'w')
test_dump.write('|'.join(['id', 'toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate']) + '\n')
log.info('running over test')
for i in tqdm(range(test_predictor_feed.num_batch)):
log.debug('i: {}'.format(i))
output, results = test_predictor.predict(i)
test_results += results
test_dump.write(repr(test_results))
test_dump.close()
with open('results/experiment_attn.csv', 'a') as dump:
output, results = all_class_predictor.predict(random.choice(range(all_class_predictor_feed.num_batch)))
dump.write(repr(results))
del output, results
all_class_trainer.train()
"""
for label, _, _ in reversed(label_trainer_triples):
if not sum(label) and e and not e % 10: #Avoid neutral classes in every epoch
continue
label_desc = '-'.join([OUTPUT_VOCAB[l] for l in [i for i, x in enumerate(label) if x == 1]] )
log.info('=================================== training for {} datapoints ========================================'.format(label_desc))
with open('results/experiment_attn.csv', 'a') as dump:
output, results = predictor[label].predict(random.choice(range(predictor_feed[label].num_batch)))
dump.write(repr(results))
del output, results
turns = int(max_size/train_feed[label].size/6) + 1
log.info('======================== size: {} and turns: {}==========================================='.format(train_feed[label].size, turns))
for turn in range(turns):
log.info('================================== label: {} and turn: {}/{}====================================='.format(label_desc, turn, turns))
trainer[label].train()
"""
except:
log.exception('####################')
torch.save(model.state_dict(), open('attn_model.pth', 'wb'))
return locals()
def train_separate(task, dataset='1k', iterations=1, batch_size=32):
# get data for task
data, metadata = gather(dataset, task)
# build data format
dformat = ['contexts', 'questions', 'answers']
# create feeds
trainfeed = DataFeed(dformat, data=data['train'])
testfeed = DataFeed(dformat, data=data['test'])
hdim = 20 if task else 50
eval_interval = 100 if task else 10
batch_size = 32 if task else 128
# instantiate model
model = MemoryNet(hdim=20,
num_hops=3,
memsize=metadata['clen'],
sentence_size=metadata['slen'],
qlen=metadata['qlen'],
vocab_size=metadata['vocab_size'],
num_candidates=metadata['candidates']['vocab_size'])
# info
print(':: <task {}> [0/2] Info')
print(':: \t memory size : {}, #candidates : {}'.format(
metadata['clen'], metadata['candidates']['vocab_size']))
with tf.Session() as sess:
# run for multiple initializations
i, accuracy, model_params = 0, [0.], [None]
while accuracy[-1] < 0.95 and i < iterations:
# init session
sess.run(tf.global_variables_initializer())
# create trainer
trainer = Trainer(sess,
model,
trainfeed,
testfeed,
batch_size=batch_size)
print('\n:: <task {}> ({}) [1/2] Pretraining'.format(task, i))
# pretrain
acc = trainer.fit(epochs=100000,
eval_interval=1,
mode=Trainer.PRETRAIN,
verbose=False,
lr=0.0005)
print(':: \tAccuracy after pretraining: ', acc)
print('\n:: <task {}> ({}) [2/2] Training'.format(task, i))
# train
acc = trainer.fit(epochs=1000000,
eval_interval=eval_interval,
mode=Trainer.TRAIN,
verbose=False,
lr=0.0005)
print(':: \tAccuracy after training: ', acc)
# next iteration
i += 1
# add accuracy to list
accuracy.append(acc)
model_params.append(sess.run(tf.trainable_variables()))
print(acc)
print(':: [x/x] End of training')
print(':: Max accuracy :', max(accuracy))
# return model and best model params
return model, model_params[accuracy.index(max(accuracy))]
def train_separate_all(dataset='1k'):
batch_size = 64
task_max_acc = []
for task in range(1, 21):
# get task 1
#task = 18
data, metadata = gather('1k', task)
# gather info from metadata
num_candidates = metadata['candidates']['vocab_size']
vocab_size = metadata['vocab_size']
memsize = metadata['clen']
sentence_size = metadata['slen']
qlen = metadata['qlen']
print(':: <task {}> memory size : {}'.format(task, memsize))
# build data format
dformat = ['contexts', 'questions', 'answers']
# create feeds
trainfeed = DataFeed(dformat, data=data['train'])
testfeed = DataFeed(dformat, data=data['test'])
# instantiate model
model = MemoryNet(hdim=20,
num_hops=3,
memsize=memsize,
sentence_size=sentence_size,
qlen=qlen,
vocab_size=vocab_size,
num_candidates=num_candidates)
with tf.Session() as sess:
# run for multiple initializations
i, accuracy = 0, [0.]
while accuracy[-1] < 0.95 and i < 5:
# init session
sess.run(tf.global_variables_initializer())
# create trainer
trainer = Trainer(sess,
model,
trainfeed,
testfeed,
batch_size=batch_size)
print('\n:: <task {}> ({}) [1/2] Pretraining'.format(task, i))
# pretrain
acc = trainer.fit(epochs=100000,
eval_interval=1,
mode=Trainer.PRETRAIN,
verbose=False,
batch_size=64,
lr=0.0005)
print(':: \tAccuracy after pretraining: ', acc)
print('\n:: <task {}> ({}) [2/2] Training'.format(task, i))
# train
acc = trainer.fit(epochs=1000000,
eval_interval=10,
mode=Trainer.TRAIN,
verbose=False,
batch_size=64,
lr=0.0005)
print(':: \tAccuracy after training: ', acc)
# next iteration
i += 1
# add accuracy to list
accuracy.append(acc)
print(acc)
print('Experiment Results : ')
for i, a in enumerate(accuracy[1:]):
print(i, a)
task_max_acc.append(max(accuracy))
print('____________________________________________')
for i, acc in enumerate(task_max_acc):
print('Task ({}) : {}'.format(i + 1, acc))
print('____________________________________________')
def experiment(VOCAB,
LABELS,
WORD2INDEX,
LABEL2INDEX,
raw_samples,
loss_weight,
datapoints=[[], []],
eons=1000,
epochs=10,
checkpoint=5):
try:
try:
model = BiLSTMDecoderModel(Config(), len(VOCAB), len(LABELS))
if Config().cuda: model = model.cuda()
log.info('loaded the old image for the model, from {}'.format(
'{}.{}'.format(SELF_NAME, 'pth')))
model.load_state_dict(torch.load('{}.{}'.format(SELF_NAME, 'pth')))
log.info('loaded the old image for the model')
except:
log.exception('failed to load the model')
model = BiLSTMDecoderModel(Config(), len(VOCAB), len(LABELS))
if Config().cuda: model = model.cuda()
print('**** the model', model)
name = SELF_NAME
_batchop = partial(batchop, LABEL2INDEX=LABEL2INDEX)
train_feed = DataFeed(name,
datapoints[0],
batchop=_batchop,
vocab=WORD2INDEX,
batch_size=1024)
test_feed = DataFeed(name,
datapoints[1],
batchop=_batchop,
vocab=WORD2INDEX,
batch_size=256)
predictor_feed = DataFeed(name,
datapoints[1],
batchop=_batchop,
vocab=WORD2INDEX,
batch_size=128)
loss_weight = Variable(torch.Tensor(loss_weight))
if Config.cuda: loss_weight = loss_weight.cuda()
_loss = partial(loss, loss_function=nn.NLLLoss(loss_weight))
trainer = Trainer(name=name,
model=model,
loss_function=_loss,
accuracy_function=accuracy,
checkpoint=checkpoint,
epochs=epochs,
feeder=Feeder(train_feed, test_feed))
predictor = Predictor(model=model,
feed=predictor_feed,
repr_function=partial(repr_function,
VOCAB=VOCAB,
LABELS=LABELS,
raw_samples=raw_samples))
for e in range(eons):
dump = open('results/experiment_attn.csv', 'a')
dump.write('#========================after eon: {}\n'.format(e))
dump.close()
log.info('on {}th eon'.format(e))
with open('results/experiment_attn.csv', 'a') as dump:
results = ListTable()
for ri in range(predictor_feed.num_batch):
output, _results = predictor.predict(ri)
results.extend(_results)
dump.write(repr(results))
if not trainer.train():
raise Exception
except:
log.exception('####################')
trainer.save_best_model()
return locals()
def train_separate(task, dataset='1k', iterations=1, batch_size=128):
# get data for task
data, metadata = gather(dataset, task)
# build data format
dformat = ['contexts', 'questions', 'answers']
# create feeds
trainfeed = DataFeed(dformat, data=data['train'])
testfeed = DataFeed(dformat, data=data['test'])
# instantiate model
model = RelationNet(clen=metadata['clen'],
qlen=metadata['qlen'],
slen=metadata['slen'],
vocab_size=metadata['vocab_size'],
num_candidates=metadata['candidates']['vocab_size'])
# info
print(':: <task {}> [0/2] Info')
print(':: \t memory size : {}, #candidates : {}'.format(
metadata['clen'], metadata['candidates']['vocab_size']))
# create visualizer
vis = Visualizer()
vis.attach_scalars(model)
with tf.Session() as sess:
# run for multiple initializations
i, accuracy, model_params = 0, [0.], [None]
while accuracy[-1] < 0.95 and i < iterations:
# init session
sess.run(tf.global_variables_initializer())
# add graph to visualizer
vis.attach_graph(sess.graph)
# create trainer
trainer = Trainer(sess,
model,
trainfeed,
testfeed,
batch_size=batch_size)
print('\n:: <task {}> ({}) [1/1] Training'.format(task, i))
# train
acc = trainer.fit(epochs=1000000,
eval_interval=1,
mode=Trainer.TRAIN,
verbose=True,
lr=0.0002)
print(':: \tAccuracy after training: ', acc)
# next iteration
i += 1
# add accuracy to list
accuracy.append(acc)
model_params.append(sess.run(tf.trainable_variables()))
print(acc)
print(':: [x/x] End of training')
print(':: Max accuracy :', max(accuracy))
# return model and model params
return model, sess.run(tf.trainable_variables())
def create_datafeed(self, args):
self.datafeed = DataFeed()
self.init_observers_and_markets(args)
def experiment(eons=1000, epochs=1, checkpoint=1):
try:
model = BiLSTMDecoderModel(Config(), len(INPUT_VOCAB), len(CHAR_VOCAB),
len(OUTPUT_VOCAB))
if Config().cuda: model = model.cuda()
classified_train_feed = DataFeed(classified_datapoints[0],
batchop=batchop,
batch_size=128)
classified_test_feed = DataFeed(classified_datapoints[1],
batchop=batchop,
batch_size=128)
classified_trainer = Trainer(model=model,
loss_function=loss,
accuracy_function=accuracy,
checkpoint=checkpoint,
epochs=epochs,
feeder=Feeder(classified_train_feed,
classified_test_feed))
classified_predictor_feed = DataFeed(classified_datapoints[1],
batchop=batchop,
batch_size=12)
classified_predictor = Predictor(model=model,
feed=classified_predictor_feed,
repr_function=repr_function)
non_classified_train_feed = DataFeed(non_classified_datapoints[0],
batchop=batchop,
batch_size=128)
non_classified_test_feed = DataFeed(non_classified_datapoints[1],
batchop=batchop,
batch_size=128)
non_classified_trainer = Trainer(model=model,
loss_function=loss,
accuracy_function=accuracy,
checkpoint=checkpoint,
epochs=epochs,
feeder=Feeder(
non_classified_train_feed,
non_classified_test_feed))
non_classified_predictor_feed = DataFeed(non_classified_datapoints[1],
batchop=batchop,
batch_size=12)
non_classified_predictor = Predictor(
model=model,
feed=non_classified_predictor_feed,
repr_function=repr_function)
test_predictor_feed = DataFeed(test_datapoints,
batchop=test_batchop,
batch_size=128)
test_predictor = Predictor(model=model,
feed=test_predictor_feed,
repr_function=test_repr_function)
dump = open('results/experiment_attn.csv', 'w')
for e in range(eons):
dump.write('#========================after eon: {}\n'.format(e))
log.info('on {}th eon'.format(e))
output, results = classified_predictor.predict(
random.choice(range(classified_predictor_feed.num_batch)))
dump.write(repr(results))
del output, results
output, results = non_classified_predictor.predict(
random.choice(range(non_classified_predictor_feed.num_batch)))
dump.write(repr(results))
del output, results
#non_classified_trainer.train()
#for i in range(int(non_classified_train_feed.size/classified_train_feed.size)):
classified_trainer.train()
if e and not e % 10:
test_results = ListTable()
test_dump = open(
'results/experiment_attn_over_test_{}.csv'.format(e), 'w')
test_dump.write('|'.join([
'id', 'toxic', 'severe_toxic', 'obscene', 'threat',
'insult', 'identity_hate'
]) + '\n')
log.info('running over test')
for i in tqdm(range(test_predictor_feed.num_batch)):
log.debug('i: {}'.format(i))
output, results = test_predictor.predict(i)
test_results += results
test_dump.write(repr(test_results))
test_dump.close()
except KeyboardInterrupt:
torch.save(model.state_dict(), open('attn_model.pth', 'wb'))
dump.close()
return locals()
def __call__(self, dset='training'):
return DataFeed(self.dformat, self.vdata[dset])
class Simulator(object):
'''
simulation manager
'''
def __init__(self, instrument, strategies, start_date, end_date, opening_bal,time_stamp=None):
'''
constructs message queues
initialises brokers and traders
'''
self.instrument = instrument
self.start_date = start_date
self.end_date = end_date
self.opening_bal = opening_bal
self.datafeed = DataFeed(instrument)
self.orderQ = MsgQ()
self.receiptQ = MsgQ()
self.term_req_Q = MsgQ()
self.term_notice_Q = MsgQ()
self.broker = Broker(self.datafeed, self.orderQ, self.receiptQ, self.term_req_Q, self.term_notice_Q)
self.traders = []
for strategy in strategies:
trader = Trader(self.datafeed, self.broker, self.opening_bal, self.instrument, strategy, self.start_date, self.end_date)
self.traders.append(trader)
self.time_stamp = time_stamp
def run(self):
'''
simulate event series
'''
current_date = date(self.start_date.year, self.start_date.month, self.start_date.day)
length = self.end_date - self.start_date
d_total = length.days
display_int = d_total / 10
while (current_date <= self.end_date):
# PROCESS TRADING DAYS
if (self.datafeed.date_is_trading_day(current_date) == True):
self.broker.open_manage_and_close(current_date)
# book keeping
for trader in self.traders:
trader.ac.tally_individual_open_positions(current_date)
trader.ac.record_net_end_of_day_pos(current_date)
trader.ac.record_end_of_day_balances(current_date)
for trader in self.traders:
trader.execute_strategy(current_date)
#self.broker.log_closed_positions()
self.broker.log_all_positions(current_date)
# IGNORE NON-TRADING DAYS
else:
pass
current_date = current_date + timedelta(days=1)
elapsed = (self.end_date - current_date)
d_elapsed = elapsed.days
progress = (float(d_total) - float(d_elapsed)) / float(d_total) * 100.0
if (d_elapsed % display_int == 0):
print('%i/100' % int(progress))
self.traders[0].strategy.log_self()
def plot(self):
'''
analyse & report on simulation path and outcome
'''
d = date(self.start_date.year, self.start_date.month, self.start_date.day)
dates = []
prices = []
cash_bal = []
margin_bal = []
net_booked_position = []
net_open_position = []
daily_high = []
daily_low = []
mavg_band_ceiling = []
mavg_band_floor = []
trader = self.broker.traders[0]
ac = trader.ac
df = self.datafeed
pMin = None
pMax = None
while (d <= self.end_date):
# TRADING DAYS
if (self.datafeed.date_is_trading_day(d) == True):
dates.append(d)
mavg_top = df.n_day_moving_avg(None, d, 'high', Strategy.n)
mavg_bottom = df.n_day_moving_avg(None, d, 'low', Strategy.n)
mavg_band_ceiling.append(mavg_top)
mavg_band_floor.append(mavg_bottom)
pinfo = df.get_price_info(None, d)
prices.append(pinfo['close'])
daily_high.append(pinfo['high'])
daily_low.append(pinfo['low'])
s = str(d) + ',' + str(mavg_band_ceiling[len(mavg_band_ceiling) - 1]) + ',' + str(mavg_band_floor[len(mavg_band_floor) - 1]) + ',' + str(pinfo['close'])
logging.info(s)
cash_bal.append(ac.d_cash_bal[d])
margin_bal.append(ac.d_margin_bal[d])
net_booked_position.append(ac.d_net_booked_position[d])
net_open_position.append(ac.net_open_position[d])
if (pMin == None):
pMin = pinfo['low']
pMax = pinfo['high']
else:
if pinfo['low'] < pMin:
pMin = pinfo['low']
if pinfo['high'] > pMax:
pMax = pinfo['high']
# NON-TRADING DAYS
else:
pass
d = d + timedelta(days=1)
aDate = np.array(dates)
aPrice = np.array(prices)
fig = plt.figure(figsize=(20, 20))
ax = fig.add_subplot(111)
#ax.plot(aDate, aPrice, color='blue')
for series in [mavg_band_ceiling, mavg_band_floor]:
y = np.array(series)
t = np.array(dates)
ax.plot(t, y, color='red')
for series in [daily_high, daily_low]:
y = np.array(series)
t = np.array(dates)
ax.plot(t, y, color='blue')
plt.ylim([float(pMin), float(pMax)])
for series in [net_booked_position]:
y = np.array(series)
t = np.array(dates)
ax2 = ax.twinx()
ax2.plot(t, y, color='green')
ax.grid(False)
fig.autofmt_xdate(rotation=90)
fname = 'plot/plot_' + self.time_stamp
fig.savefig(fname)
from train.trainer import Trainer
from models.model import Model
from datafeed import DataFeed
from graph import *
if __name__ == '__main__':
# get data for task
data, lookup, metadata = gather()
# build data format
dformat = FIELDS + ['cmask']
# create feeds
trainfeed = DataFeed(dformat, data=data['valid'])
testfeed = DataFeed(dformat, data=data['test'])
# training params
batch_size = 32
# instantiate model
model = Model(ASReaderGraph,
dformat=dformat,
n=1,
optimizer=tf.train.AdamOptimizer,
lr=0.001,
vocab_size=metadata['vocab_size'],
max_candidates=metadata['max_candidates'],
demb=384,
dhdim=384,
