def main():
Logger.GLOBAL_LOG_LEVEL = LogLevel.INFO
#Config.backend = Backends.TENSORFLOW
Config.backend = Backends.TORCH
Config.cuda = True
Config.dropout = 0.1
Config.hidden_size = 128
Config.embedding_size = 256
Config.L2 = 0.00003
do_process = False
if do_process:
preprocess_SNLI(delete_data=True)
p = Pipeline('snli_example')
vocab = p.state['vocab']
vocab.load_from_disk()
batch_size = 128
if Config.backend == Backends.TENSORFLOW:
from spodernet.backends.tfbackend import TensorFlowConfig
TensorFlowConfig.init_batch_size(batch_size)
train_batcher = StreamBatcher('snli_example', 'snli_train', batch_size, randomize=True, loader_threads=8)
#train_batcher.subscribe_to_batch_prepared_event(SomeExpensivePreprocessing())
dev_batcher = StreamBatcher('snli_example', 'snli_dev', batch_size)
test_batcher = StreamBatcher('snli_example', 'snli_test', batch_size)
train_batcher.subscribe_to_events(AccuracyHook('Train', print_every_x_batches=1000))
dev_batcher.subscribe_to_events(AccuracyHook('Dev', print_every_x_batches=1000))
eta = ETAHook(print_every_x_batches=1000)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
model = Model()
model.add(Embedding(128, vocab.num_embeddings))
model.add(PairedBiDirectionalLSTM(128, hidden_size=256, variable_length=True, conditional_encoding=False))
model.add(SoftmaxCrossEntropy(input_size=256*4, num_labels=3))
t = Trainer(model)
for i in range(10):
t.train(train_batcher, epochs=1)
t.evaluate(dev_batcher)
def main(args, model_path):
if args.preprocess: preprocess(args.data, delete_data=True)
input_keys = ['e1', 'rel', 'rel_eval', 'e2', 'e2_multi1', 'e2_multi2']
p = Pipeline(args.data, keys=input_keys)
p.load_vocabs()
vocab = p.state['vocab']
num_entities = vocab['e1'].num_token
train_batcher = StreamBatcher(args.data, 'train', args.batch_size, randomize=True, keys=input_keys, loader_threads=args.loader_threads)
dev_rank_batcher = StreamBatcher(args.data, 'dev_ranking', args.test_batch_size, randomize=False, loader_threads=args.loader_threads, keys=input_keys)
test_rank_batcher = StreamBatcher(args.data, 'test_ranking', args.test_batch_size, randomize=False, loader_threads=args.loader_threads, keys=input_keys)
if args.model is None:
model = ConvE(args, vocab['e1'].num_token, vocab['rel'].num_token)
elif args.model == 'conve':
model = ConvE(args, vocab['e1'].num_token, vocab['rel'].num_token)
elif args.model == 'distmult':
model = DistMult(args, vocab['e1'].num_token, vocab['rel'].num_token)
elif args.model == 'complex':
model = Complex(args, vocab['e1'].num_token, vocab['rel'].num_token)
elif args.model == 'interacte':
model = InteractE(args, vocab['e1'].num_token, vocab['rel'].num_token)
else:
log.info('Unknown model: {0}', args.model)
raise Exception("Unknown model!")
train_batcher.at_batch_prepared_observers.insert(1,TargetIdx2MultiTarget(num_entities, 'e2_multi1', 'e2_multi1_binary'))
eta = ETAHook('train', print_every_x_batches=args.log_interval)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
train_batcher.subscribe_to_events(LossHook('train', print_every_x_batches=args.log_interval))
model.cuda()
if args.resume:
model_params = torch.load(model_path)
print(model)
total_param_size = []
params = [(key, value.size(), value.numel()) for key, value in model_params.items()]
for key, size, count in params:
total_param_size.append(count)
print(key, size, count)
print(np.array(total_param_size).sum())
model.load_state_dict(model_params)
model.eval()
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation')
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
else:
model.init()
total_param_size = []
params = [value.numel() for value in model.parameters()]
print(params)
print(np.sum(params))
opt = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.l2)
for epoch in range(args.epochs):
model.train()
for i, str2var in enumerate(train_batcher):
opt.zero_grad()
e1 = str2var['e1']
rel = str2var['rel']
e2_multi = str2var['e2_multi1_binary'].float()
# label smoothing
e2_multi = ((1.0-args.label_smoothing)*e2_multi) + (1.0/e2_multi.size(1))
pred = model.forward(e1, rel)
loss = model.loss(pred, e2_multi)
loss.backward()
opt.step()
train_batcher.state.loss = loss.cpu()
print('saving to {0}'.format(model_path))
torch.save(model.state_dict(), model_path)
model.eval()
with torch.no_grad():
if epoch % 5 == 0 and epoch > 0:
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
if epoch % 5 == 0:
if epoch > 0:
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation')
def preprocess(dataset_name, delete_data=False):
full_path = 'data/{0}/e1rel_to_e2_full.json'.format(dataset_name)
train_path = 'data/{0}/e1rel_to_e2_train.json'.format(dataset_name)
dev_ranking_path = 'data/{0}/e1rel_to_e2_ranking_dev.json'.format(dataset_name)
test_ranking_path = 'data/{0}/e1rel_to_e2_ranking_test.json'.format(dataset_name)
keys2keys = {}
keys2keys['e1'] = 'e1' # entities
keys2keys['rel'] = 'rel' # relations
keys2keys['rel_eval'] = 'rel' # relations
keys2keys['e2'] = 'e1' # entities
keys2keys['e2_multi1'] = 'e1' # entity
keys2keys['e2_multi2'] = 'e1' # entity
input_keys = ['e1', 'rel', 'rel_eval', 'e2', 'e2_multi1', 'e2_multi2']
d = DatasetStreamer(input_keys)
d.add_stream_processor(JsonLoaderProcessors())
d.add_stream_processor(DictKey2ListMapper(input_keys))
# process full vocabulary and save it to disk
d.set_path(full_path)
p = Pipeline(args.data, delete_data, keys=input_keys, skip_transformation=True)
p.add_sent_processor(ToLower())
p.add_sent_processor(CustomTokenizer(lambda x: x.split(' ')),keys=['e2_multi1', 'e2_multi2'])
p.add_token_processor(AddToVocab())
p.execute(d)
p.save_vocabs()
# process train, dev and test sets and save them to hdf5
p.skip_transformation = False
for path, name in zip([train_path, dev_ranking_path, test_ranking_path], ['train', 'dev_ranking', 'test_ranking']):
d.set_path(path)
p.clear_processors()
p.add_sent_processor(ToLower())
p.add_sent_processor(CustomTokenizer(lambda x: x.split(' ')),keys=['e2_multi1', 'e2_multi2'])
p.add_post_processor(ConvertTokenToIdx(keys2keys=keys2keys), keys=['e1', 'rel', 'rel_eval', 'e2', 'e2_multi1', 'e2_multi2'])
p.add_post_processor(StreamToHDF5(name, samples_per_file=1000, keys=input_keys))
p.execute(d)
def main():
if Config.process: preprocess(Config.dataset, delete_data=True)
input_keys = ['e1', 'rel', 'e2', 'e2_multi1', 'e2_multi2']
p = Pipeline(Config.dataset, keys=input_keys)
p.load_vocabs()
vocab = p.state['vocab']
num_entities = vocab['e1'].num_token
dict_tokentoid, dict_idtotoken = vocab['e1'].tokendicts()
dict_reltoid, dict_idtorel = vocab['rel'].tokendicts()
num_rel = vocab['rel'].num_token
train_batcher = StreamBatcher(Config.dataset, 'train', Config.batch_size, randomize=True, keys=input_keys)
dev_rank_batcher = StreamBatcher(Config.dataset, 'dev_ranking', Config.batch_size, randomize=False, loader_threads=4, keys=input_keys, is_volatile=True)
test_rank_batcher = StreamBatcher(Config.dataset, 'test_ranking', Config.batch_size, randomize=False, loader_threads=4, keys=input_keys, is_volatile=True)
if Config.model_name is None:
model = ConvE(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'ConvE':
model = ConvE(vocab['e1'].num_token, num_rel)
elif Config.model_name == 'DistMult':
model = DistMult(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'ComplEx':
model = Complex(vocab['e1'].num_token, vocab['rel'].num_token)
else:
log.info('Unknown model: {0}', Config.model_name)
raise Exception("Unknown model!")
train_batcher.at_batch_prepared_observers.insert(1,TargetIdx2MultiTarget(num_entities, 'e2_multi1', 'e2_multi1_binary'))
eta = ETAHook('train', print_every_x_batches=100)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
train_batcher.subscribe_to_events(LossHook('train', print_every_x_batches=100))
if Config.cuda:
model.cuda()
if load:
model_params = torch.load(model_path)
print(model)
total_param_size = []
params = [(key, value.size(), value.numel()) for key, value in model_params.items()]
for key, size, count in params:
total_param_size.append(count)
print(key, size, count)
print(np.array(total_param_size).sum())
model.load_state_dict(model_params)
model.eval()
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation')
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
else:
model.init()
total_param_size = []
params = [value.numel() for value in model.parameters()]
print(params)
print(np.sum(params))
################################################ loading
model.load_state_dict(torch.load('embeddings/auto-embeddings.pt'))
opt = torch.optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=Config.learning_rate, weight_decay=Config.L2)
# One hot encoding buffer that you create out of the loop and just keep reusing
y_onehot_e1 = torch.FloatTensor(Config.batch_size, num_entities)
# One hot encoding buffer that you create out of the loop and just keep reusing
y_onehot_r = torch.FloatTensor(Config.batch_size, num_rel)
model.eval()
train_data =[]
with open('data/'+Config.dataset+'/train.txt', 'r') as f:
for i, line in enumerate(f):
e1, rel, e2 = line.decode('utf-8').split('\t')
e1 = e1.strip()#.lower()
e2 = e2.strip()#.lower()
rel = rel.strip()#.lower()
train_data += [[e1, rel, e2]]
print len(train_data)
attack_list = []
E2_list = []
with open('data/'+Config.dataset+'/test.txt', 'r') as f:
for i, line in enumerate(f):
e1, rel, e2 = line.decode('utf-8').split('\t')
e1 = e1.strip().lower()
e2 = e2.strip().lower()
rel = rel.strip().lower()
attack_list += [[dict_tokentoid[e1], dict_reltoid[rel], dict_tokentoid[e2]]]
E2_list += [e2]
print len(attack_list)
E2_list = set(E2_list)
E2_dict = {}
for i in train_data:
if i[2].lower() in E2_list:
if dict_tokentoid[i[2].lower()] in E2_dict:
E2_dict[dict_tokentoid[i[2].lower()]] += [[dict_tokentoid[i[0].lower()], dict_reltoid[i[1].lower()]]]
else:
E2_dict[dict_tokentoid[i[2].lower()]] = [[dict_tokentoid[i[0].lower()], dict_reltoid[i[1].lower()]]]
str_at = []
embd_e = model.emb_e.weight.data.cpu().numpy()
embd_rel = model.emb_rel.weight.data.cpu().numpy()
n_t = 0
for trip in attack_list:
if n_t % 500 == 0:
print 'Number of processed triple: ', n_t
n_t += 1
e1 = trip[0]
rel = trip[1]
e2_or = trip[2]
e1 = torch.cuda.LongTensor([e1])
rel = torch.cuda.LongTensor([rel])
e2 = torch.cuda.LongTensor([e2_or])
pred = model.encoder(e1, rel)
E2 = model.encoder_2(e2)
A, B = find_best_at(-pred, E2)
attack_ext = -A*pred+B*E2
if e2_or in E2_dict:
nei = E2_dict[e2_or]
#attack = find_best_attack(E2.data.cpu().numpy(), pred.data.cpu().numpy(), nei, embd_e, embd_rel, attack_ext)
#attack = torch.autograd.Variable(torch.from_numpy(attack)).cuda().float()
attack = attack_ext
else:
attack = attack_ext
E1, R = model.decoder(attack)
_, predicted_e1 = torch.max(E1, 1)
_, predicted_R = torch.max(R, 1)
str_at += [[str(dict_idtotoken[predicted_e1.data.cpu().numpy()[0]]), str(dict_idtorel[predicted_R.data.cpu().numpy()[0]]), str(dict_idtotoken[e2_or])]]
new_train = str_at + train_data
print len(new_train)
with open('data/new_'+Config.dataset+'/train.txt', 'w') as f:
for item in new_train:
f.write("%s\n" % "\t".join(map(str, item)))
def main():
if Config.process: preprocess(Config.dataset, delete_data=True)
input_keys = ['e1', 'rel', 'rel_eval', 'e2', 'e2_multi1', 'e2_multi2']
p = Pipeline(Config.dataset, keys=input_keys)
p.load_vocabs()
vocab = p.state['vocab']
num_entities = vocab['e1'].num_token
train_batcher = StreamBatcher(Config.dataset,
'train',
Config.batch_size,
randomize=True,
keys=input_keys)
dev_rank_batcher = StreamBatcher(Config.dataset,
'dev_ranking',
Config.batch_size,
randomize=False,
loader_threads=4,
keys=input_keys)
test_rank_batcher = StreamBatcher(Config.dataset,
'test_ranking',
Config.batch_size,
randomize=False,
loader_threads=4,
keys=input_keys)
if Config.model_name is None:
model = ConvE(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'ConvE':
model = ConvE(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'DistMult':
model = DistMult(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'ComplEx':
model = Complex(vocab['e1'].num_token, vocab['rel'].num_token)
else:
log.info('Unknown model: {0}', Config.model_name)
raise Exception("Unknown model!")
train_batcher.at_batch_prepared_observers.insert(
1, TargetIdx2MultiTarget(num_entities, 'e2_multi1',
'e2_multi1_binary'))
eta = ETAHook('train', print_every_x_batches=100)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
train_batcher.subscribe_to_events(
LossHook('train', print_every_x_batches=100))
if Config.cuda:
model.cuda()
if load:
model_params = torch.load(model_path)
print(model)
total_param_size = []
params = [(key, value.size(), value.numel())
for key, value in model_params.items()]
for key, size, count in params:
total_param_size.append(count)
print(key, size, count)
print(np.array(total_param_size).sum())
model.load_state_dict(model_params)
model.eval()
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation')
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
else:
model.init()
total_param_size = []
params = [value.numel() for value in model.parameters()]
print(params)
print(np.sum(params))
max_mrr = 0
count = 0
max_count = 3
opt = torch.optim.Adam(model.parameters(),
lr=Config.learning_rate,
weight_decay=Config.L2)
for epoch in range(1, epochs + 1):
model.train()
for i, str2var in enumerate(train_batcher):
opt.zero_grad()
e1 = str2var['e1']
rel = str2var['rel']
e2_multi = str2var['e2_multi1_binary'].float()
# label smoothing
e2_multi = ((1.0 - Config.label_smoothing_epsilon) *
e2_multi) + (1.0 / e2_multi.size(1))
pred = model.forward(e1, rel)
loss = model.loss(pred, e2_multi)
loss.backward()
opt.step()
train_batcher.state.loss = loss.cpu()
print('saving to {0}'.format(model_path))
torch.save(model.state_dict(), model_path)
model.eval()
with torch.no_grad():
# ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
if epoch % 15 == 0:
mrr = ranking_and_hits(model, dev_rank_batcher, vocab,
'dev_evaluation')
if mrr <= max_mrr:
count += 1
if count > max_count:
break
else:
count = 0
max_mrr = mrr
mrr_test = ranking_and_hits(model, test_rank_batcher, vocab,
'test_evaluation')
def main(args):
if args.preprocess:
print('start preprocessing', flush=True)
preprocess(args, delete_data=True)
print('finish preprocessing', flush=True)
else:
input_keys = ['e1', 'rel', 'rel_eval', 'e2', 'e2_multi1', 'e2_multi2']
p = Pipeline(args.data, keys=input_keys)
print(time.strftime("%a, %d %b %Y %H:%M:%S +0000", time.localtime()) +
': start loading vocabs',
flush=True)
p.load_vocabs()
print(time.strftime("%a, %d %b %Y %H:%M:%S +0000", time.localtime()) +
': finish loading vocabs',
flush=True)
vocab = p.state['vocab']
num_entities = vocab['e1'].num_token
train_batcher = StreamBatcher(args.data,
'train',
args.batch_size,
randomize=True,
keys=input_keys,
loader_threads=args.loader_threads)
model = DistMult(args, vocab['e1'].num_token, vocab['rel'].num_token)
train_batcher.at_batch_prepared_observers.insert(
1,
TargetIdx2MultiTarget(num_entities, 'e2_multi1',
'e2_multi1_binary'))
# eta = ETAHook('train', print_every_x_batches=args.log_interval)
# train_batcher.subscribe_to_events(eta)
# train_batcher.subscribe_to_start_of_epoch_event(eta)
# train_batcher.subscribe_to_events(LossHook('train', print_every_x_batches=args.log_interval))
model.cuda()
model.init()
total_param_size = []
params = [value.numel() for value in model.parameters()]
print(params, flush=True)
print(np.sum(params), flush=True)
opt = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.l2)
print(time.strftime("%a, %d %b %Y %H:%M:%S +0000", time.localtime()) +
f': start training with epochs = {args.epochs}',
flush=True)
for epoch in range(args.epochs):
model.train()
# sampled_batches = set(np.random.choice(train_batcher.num_batches, args.num_batches, replace=False))
# print(time.strftime("%a, %d %b %Y %H:%M:%S +0000", time.localtime()) + f': start epoch {epoch} with batches = {len(sampled_batches)} out of {train_batcher.num_batches}', flush=True)
# processed_count = 0
for i, str2var in enumerate(train_batcher):
# if i not in sampled_batches: continue
# if processed_count%int(args.num_batches/1000)==0: print(time.strftime("%a, %d %b %Y %H:%M:%S +0000", time.localtime()) + f': start epoch {epoch} batch {i} = {processed_count}', flush=True)
# processed_count += 1
opt.zero_grad()
e1 = str2var['e1']
rel = str2var['rel']
e2_multi = str2var['e2_multi1_binary'].float()
e2_multi = ((1.0 - args.label_smoothing) *
e2_multi) + (1.0 / e2_multi.size(1))
pred = model.forward(e1, rel)
loss = model.loss(pred, e2_multi)
loss.backward()
opt.step()
# train_batcher.state.loss = loss.cpu()
print(time.strftime("%a, %d %b %Y %H:%M:%S +0000",
time.localtime()) +
f': finish training epoch {epoch}',
flush=True)
model.eval()
output(args, vocab['e1'], model.emb_e.weight.detach().cpu().numpy())
def preprocess(args, delete_data=False):
keys2keys = {}
keys2keys['e1'] = 'e1' # entities
keys2keys['rel'] = 'rel' # relations
keys2keys['rel_eval'] = 'rel' # relations
keys2keys['e2'] = 'e1' # entities
keys2keys['e2_multi1'] = 'e1' # entity
keys2keys['e2_multi2'] = 'e1' # entity
input_keys = ['e1', 'rel', 'rel_eval', 'e2', 'e2_multi1', 'e2_multi2']
print('create dataset streamer', flush=True)
d = DatasetStreamer(input_keys)
d.add_stream_processor(JsonLoaderProcessors())
d.add_stream_processor(DictKey2ListMapper(input_keys))
# process full vocabulary and save it to disk
d.set_path(args.train_path)
print('create pipeline', flush=True)
p = Pipeline(args.data,
delete_data,
keys=input_keys,
skip_transformation=True)
p.add_sent_processor(ToLower())
p.add_sent_processor(CustomTokenizer(lambda x: x.split(' ')),
keys=['e2_multi1', 'e2_multi2'])
p.add_token_processor(AddToVocab())
print('execute full vocabs', flush=True)
p.execute(d)
print('save full vocabs', flush=True)
p.save_vocabs()
# process train sets and save them to hdf5
p.skip_transformation = False
d.set_path(args.train_path)
p.clear_processors()
p.add_sent_processor(ToLower())
p.add_sent_processor(CustomTokenizer(lambda x: x.split(' ')),
keys=['e2_multi1', 'e2_multi2'])
p.add_post_processor(
ConvertTokenToIdx(keys2keys=keys2keys),
keys=['e1', 'rel', 'rel_eval', 'e2', 'e2_multi1', 'e2_multi2'])
p.add_post_processor(
StreamToHDF5('train', samples_per_file=1000, keys=input_keys))
print('execute and save train vocabs', flush=True)
p.execute(d)
def main():
if do_process: preprocess(dataset_name, delete_data=True)
input_keys = ['e1', 'rel', 'e2', 'e2_multi1', 'e2_multi2']
p = Pipeline(dataset_name, keys=input_keys)
p.load_vocabs()
vocab = p.state['vocab']
num_entities = vocab['e1'].num_token
train_batcher = StreamBatcher(dataset_name,
'train',
Config.batch_size,
randomize=True,
keys=input_keys)
dev_rank_batcher = StreamBatcher(dataset_name,
'dev_ranking',
Config.batch_size,
randomize=False,
loader_threads=4,
keys=input_keys,
is_volatile=True)
test_rank_batcher = StreamBatcher(dataset_name,
'test_ranking',
Config.batch_size,
randomize=False,
loader_threads=4,
keys=input_keys,
is_volatile=True)
#model = Complex(vocab['e1'].num_token, vocab['rel'].num_token)
#model = DistMult(vocab['e1'].num_token, vocab['rel'].num_token)
model = ConvE(vocab['e1'].num_token, vocab['rel'].num_token)
train_batcher.at_batch_prepared_observers.insert(
1, TargetIdx2MultiTarget(num_entities, 'e2_multi1',
'e2_multi1_binary'))
eta = ETAHook('train', print_every_x_batches=100)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
train_batcher.subscribe_to_events(
LossHook('train', print_every_x_batches=100))
if Config.cuda:
model.cuda()
if load:
model_params = torch.load(model_path)
print(model)
print([(key, value.size()) for key, value in model_params.items()])
model.load_state_dict(model_params)
model.eval()
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation')
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
else:
model.init()
opt = torch.optim.Adam(model.parameters(),
lr=Config.learning_rate,
weight_decay=Config.L2)
for epoch in range(epochs):
model.train()
for i, str2var in enumerate(train_batcher):
opt.zero_grad()
e1 = str2var['e1']
rel = str2var['rel']
e2_multi = str2var['e2_multi1_binary'].float()
# label smoothing
e2_multi = ((1.0 - Config.label_smoothing_epsilon) *
e2_multi) + (1.0 / e2_multi.size(1))
pred = model.forward(e1, rel)
loss = model.loss(pred, e2_multi)
loss.backward()
opt.step()
train_batcher.state.loss = loss
print('saving to {0}'.format(model_path))
torch.save(model.state_dict(), model_path)
model.eval()
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
if epoch % 3 == 0:
if epoch > 0:
ranking_and_hits(model, test_rank_batcher, vocab,
'test_evaluation')
def main():
if Config.process: preprocess(Config.dataset, delete_data=True)
train_triples_path = path_root + 'data/{0}/train.txt'.format(
Config.dataset)
# dev_triples_path = 'data/{0}/valid.txt'.format(Config.dataset) # used for development
test_triples_path = path_root + 'data/{0}/test.txt'.format(Config.dataset)
input_keys = ['e1', 'rel', 'rel_eval', 'e2', 'e2_multi1', 'e2_multi2']
p = Pipeline(Config.dataset, keys=input_keys)
p.load_vocabs()
vocab = p.state['vocab']
num_entities = vocab['e1'].num_token
train_batcher = StreamBatcher(Config.dataset,
'train',
Config.batch_size,
randomize=True,
keys=input_keys)
dev_rank_batcher = StreamBatcher(Config.dataset,
'dev_ranking',
Config.batch_size,
randomize=False,
loader_threads=4,
keys=input_keys)
test_rank_batcher = StreamBatcher(Config.dataset,
'test_ranking',
Config.batch_size,
randomize=False,
loader_threads=4,
keys=input_keys)
allRels = get_AllRels(vocab)
allEntTokens = get_AllEntities(vocab)
if Config.model_name is None:
model = ConvE(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'ConvE':
if not test:
model = ConvE(vocab['e1'].num_token, vocab['rel'].num_token,
allEntTokens, allRels)
else:
if testEntGraph:
types2E, types2rel2idx, types2rels = read_graphs(
gpath, f_post_fix, featIdx, isCCG, lower)
model = ConvE(vocab['e1'].num_token, vocab['rel'].num_token,
allEntTokens, allRels, types2E, types2rel2idx,
types2rels)
else:
model = ConvE(vocab['e1'].num_token, vocab['rel'].num_token,
allEntTokens, allRels)
elif Config.model_name == 'DistMult':
model = DistMult(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'ComplEx':
model = Complex(vocab['e1'].num_token, vocab['rel'].num_token)
else:
print('Unknown model: {0}', Config.model_name)
raise Exception("Unknown model!")
train_batcher.at_batch_prepared_observers.insert(
1, TargetIdx2MultiTarget(num_entities, 'e2_multi1',
'e2_multi1_binary'))
eta = ETAHook('train', print_every_x_batches=100)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
train_batcher.subscribe_to_events(
LossHook('train', print_every_x_batches=100))
if Config.cuda:
model.cuda()
if load:
model_params = torch.load(model_path)
print(model)
total_param_size = []
params = [(key, value.size(), value.numel())
for key, value in model_params.items()]
for key, size, count in params:
total_param_size.append(count)
print(key, size, count)
print(np.array(total_param_size).sum())
model.load_state_dict(model_params)
if test:
if computeAllProbs:
fout_probs = open(Config.probs_file_path, 'w')
model.eval()
with torch.no_grad():
compute_probs(model, test_rank_batcher, vocab,
'test_probs', fout_probs, test_triples_path)
else:
model.eval()
with torch.no_grad():
if testEntGraph:
# ranking_and_hits_entGraph(model, dev_rank_batcher, vocab, relW2idx, Config.model_name, 'dev_evaluation',train_triples_path, 20)
ranking_and_hits_entGraph(model, test_rank_batcher,
vocab, 'test_evaluation',
train_triples_path, 20)
else:
# ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation', train_triples_path, 20)
ranking_and_hits(model, test_rank_batcher, vocab,
'test_evaluation', train_triples_path,
20)
return
else:
model.init()
params = [value.numel() for value in model.parameters()]
print(params)
print(np.sum(params))
opt = torch.optim.Adam(model.parameters(),
lr=Config.learning_rate,
weight_decay=Config.L2)
for epoch in range(epochs):
model.train()
for i, str2var in enumerate(train_batcher):
opt.zero_grad()
e1 = str2var['e1']
rel = str2var['rel']
e2_multi = str2var['e2_multi1_binary'].float()
# label smoothing
e2_multi = ((1.0 - Config.label_smoothing_epsilon) *
e2_multi) + (1.0 / e2_multi.size(1))
pred = model.forward(e1, rel)
loss = model.loss(pred, e2_multi)
loss.backward()
opt.step()
train_batcher.state.loss = loss.cpu()
if save:
print('saving to {0}'.format(model_path))
if not os.path.isdir(path_root + 'saved_models'):
os.mkdir(path_root + 'saved_models')
torch.save(model.state_dict(), model_path)
model.eval()
if epoch % 5 == 0:
with torch.no_grad():
ranking_and_hits(model, dev_rank_batcher, vocab,
'dev_evaluation', train_triples_path, 20)
if epoch % 10 == 0: #This was 10
if epoch > 0:
ranking_and_hits(model, test_rank_batcher, vocab,
'test_evaluation', train_triples_path,
20)
# Let's write the rel embeddings!
if model_name == "ConvE":
fout = open(
'ents2emb_tmp_' + Config.model_name + '_' + Config.dataset +
'.txt', 'w')
lookup_tensor = torch.tensor(
[i for i in range(vocab['e1'].num_token)],
dtype=torch.long).to('cuda')
emb_e = model.emb_e(lookup_tensor).cpu().detach().numpy()
for i in range(vocab['e1'].num_token):
fout.write(vocab['e1'].idx2token[i] + '\t' + str(emb_e[i]) +
'\n')
fout.close()
fout = open(
'rels2emb_' + Config.model_name + '_' + Config.dataset +
'_tmp.txt', 'w')
for i in range(vocab['rel'].num_token):
if i in model.relIdx2Embed:
fout.write(vocab['rel'].idx2token[i] + '\t' +
str(model.relIdx2Embed[i]) + '\n')
fout.close()
if model_name == "ConvE":
#Let's write the final rel embeddings!
fout = open(
'rels2emb_' + Config.model_name + '_' + Config.dataset + '.txt',
'w')
for i in range(vocab['rel'].num_token):
if i in model.relIdx2Embed:
fout.write(vocab['rel'].idx2token[i] + '\t' +
str(model.relIdx2Embed[i]) + '\n')
else:
print("doesn't have: ", vocab['rel'].idx2token[i])
def main(args, model_path):
if args.preprocess: preprocess(args.data, delete_data=True)
input_keys = ['e1', 'rel', 'rel_eval', 'e2', 'e2_multi1', 'e2_multi2']
p = Pipeline(args.data, keys=input_keys)
p.load_vocabs()
vocab = p.state['vocab'] # 都要把数据转换成对象存储起来。这里用的是spodernet 中的Vocab对象
num_entities = vocab['e1'].num_token # 得到总共有多少个实体（sense）
# 生成三批数据
train_batcher = StreamBatcher(args.data,
'train',
args.batch_size,
randomize=True,
keys=input_keys,
loader_threads=args.loader_threads)
dev_rank_batcher = StreamBatcher(args.data,
'dev_ranking',
args.test_batch_size,
randomize=False,
loader_threads=args.loader_threads,
keys=input_keys)
test_rank_batcher = StreamBatcher(args.data,
'test_ranking',
args.test_batch_size,
randomize=False,
loader_threads=args.loader_threads,
keys=input_keys)
model = ConvE(args, vocab['e1'].num_token, vocab['rel'].num_token)
train_batcher.at_batch_prepared_observers.insert(
1, TargetIdx2MultiTarget(num_entities, 'e2_multi1',
'e2_multi1_binary'))
# 这部分功能应该是：在训练完之后使用一个回调
eta = ETAHook('train', print_every_x_batches=args.log_interval)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
train_batcher.subscribe_to_events(
LossHook('train', print_every_x_batches=args.log_interval))
P = Preprocessor("../external/wordnet-mlj12")
tokenidx_to_synset = vocab['e1'].idx2token
encoder = DefinitionEncoder()
encoder.cuda()
model.cuda()
if args.initialize:
model_params = torch.load(args.initialize)
print(model)
total_param_size = []
params = [(key, value.size(), value.numel())
for key, value in model_params.items()]
for key, size, count in params:
total_param_size.append(count)
print(key, size, count)
print(np.array(total_param_size).sum())
model.load_state_dict(model_params)
model.eval()
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation')
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
# 赋值definition encoder，但是在model的属性中，没有找到 encoder
model.encoder = encoder
model.encoder.init()
elif args.resume:
model.encoder = encoder
model_params = torch.load(model_path)
print(model)
total_param_size = []
params = [(key, value.size(), value.numel())
for key, value in model_params.items()]
for key, size, count in params:
total_param_size.append(count)
print(key, size, count)
print(np.array(total_param_size).sum())
model.load_state_dict(model_params)
model.eval()
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation',
tokenidx_to_synset, P.get_batch)
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation',
tokenidx_to_synset, P.get_batch)
else:
model.encoder = encoder
model.encoder.init()
model.init()
total_param_size = []
params = [value.numel() for value in model.parameters()]
print(params)
print(np.sum(params))
opt = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.l2)
best_dev_mrr = 0
model.eval()
dev_mrr = ranking_and_hits(model, dev_rank_batcher, vocab,
'dev_evaluation', tokenidx_to_synset,
P.get_batch)
# 准备训练
for epoch in range(args.epochs):
model.train()
for i, str2var in enumerate(train_batcher):
opt.zero_grad()
e1 = str2var['e1']
rel = str2var['rel']
e1_tokens = [
tokenidx_to_synset[idx]
for idx in e1.detach().cpu().numpy().ravel()
]
batch, lengths = P.get_batch(e1_tokens)
# e1_emb 就是使用 bilstm 得到的embedding
e1_emb = model.encoder((batch.cuda(), lengths))[0]
e2_multi = str2var['e2_multi1_binary'].float()
# label smoothing
e2_multi = ((1.0 - args.label_smoothing) *
e2_multi) + (1.0 / e2_multi.size(1))
# 放到
pred = model.forward(e1_emb, rel, e1_encoded=True)
loss = model.loss(pred, e2_multi)
loss.backward()
opt.step()
train_batcher.state.loss = loss.cpu()
#saving on improvement in dev score
#print('saving to {0}'.format(model_path))
#torch.save(model.state_dict(), model_path)
model.eval()
with torch.no_grad():
if epoch % 5 == 0 and epoch > 0:
dev_mrr = ranking_and_hits(model, dev_rank_batcher, vocab,
'dev_evaluation',
tokenidx_to_synset, P.get_batch)
if dev_mrr > best_dev_mrr:
print('saving to {} MRR {}->{}'.format(
model_path, best_dev_mrr, dev_mrr))
best_dev_mrr = dev_mrr
torch.save(model.state_dict(), model_path)
if epoch % 5 == 0:
if epoch > 0:
ranking_and_hits(model, test_rank_batcher, vocab,
'test_evaluation', tokenidx_to_synset,
P.get_batch)
if args.represent:
P = Preprocessor()
synsets = [P.idx_to_synset[idx] for idx in range(len(P.idx_to_synset))]
embeddings = []
embeddings_proj = []
for i in tqdm(range(0, len(synsets), args.test_batch_size)):
synsets_batch = synsets[i:i + args.test_batch_size]
with torch.no_grad():
batch, lengths = P.get_batch(synsets_batch)
emb_proj, emb = model.encoder((batch.cuda(), lengths))
embeddings_proj.append(emb_proj.detach().cpu())
embeddings.append(emb.detach().cpu())
embeddings = torch.cat(embeddings, 0).numpy()
embeddings_proj = torch.cat(embeddings_proj, 0).numpy()
print('embeddings', embeddings.shape, embeddings_proj.shape)
basename, ext = os.path.splitext(args.represent)
fname = args.represent
np.savez_compressed(fname, embeddings=embeddings, synsets=synsets)
fname = basename + '_projected' + ext
np.savez_compressed(fname, embeddings=embeddings_proj, synsets=synsets)
def main():
if Config.process: preprocess(Config.dataset, delete_data=True)
input_keys = ['e1', 'rel', 'e2', 'e2_multi1', 'e2_multi2']
p = Pipeline(Config.dataset, keys=input_keys)
p.load_vocabs()
vocab = p.state['vocab']
num_entities = vocab['e1'].num_token
train_batcher = StreamBatcher(Config.dataset, 'train', Config.batch_size, randomize=True, keys=input_keys)
dev_rank_batcher = StreamBatcher(Config.dataset, 'dev_ranking', Config.batch_size, randomize=False, loader_threads=4, keys=input_keys, is_volatile=True)
test_rank_batcher = StreamBatcher(Config.dataset, 'test_ranking', Config.batch_size, randomize=False, loader_threads=4, keys=input_keys, is_volatile=True)
# Load literals
numerical_literals = np.load(f'data/{Config.dataset}/literals/numerical_literals.npy')
# Initialize KBLN RBF parameters
X_train = np.load(f'data/{Config.dataset}/bin/train.npy')
h = X_train[:, 0]
t = X_train[:, 2]
n = numerical_literals[h, :] - numerical_literals[t, :]
c = np.mean(n, axis=0).astype('float32') # size: (n_literals)
var = np.var(n, axis=0) + 1e-6 # size: (n_literals), added eps to avoid degenerate case
# Get normalized literals (for LiteralE)
max_lit, min_lit = np.max(numerical_literals, axis=0), np.min(numerical_literals, axis=0)
numerical_literals_normalized = (numerical_literals - min_lit) / (max_lit - min_lit + 1e-8)
# Load literal models
if Config.model_name is None or Config.model_name == 'KBLN':
model = KBLN(vocab['e1'].num_token, vocab['rel'].num_token, numerical_literals, c, var)
print('Chosen model: KBLN')
elif Config.model_name == 'LiteralE_KBLN':
model = LiteralE_KBLN(vocab['e1'].num_token, vocab['rel'].num_token, numerical_literals, numerical_literals_normalized, c, var)
print('Chosen model: LiteralE_KBLN')
train_batcher.at_batch_prepared_observers.insert(1, TargetIdx2MultiTarget(num_entities, 'e2_multi1', 'e2_multi1_binary'))
eta = ETAHook('train', print_every_x_batches=100)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
train_batcher.subscribe_to_events(LossHook('train', print_every_x_batches=100))
if Config.cuda:
model.cuda()
if load:
model_params = torch.load(model_path)
print(model)
total_param_size = []
params = [(key, value.size(), value.numel()) for key, value in model_params.items()]
for key, size, count in params:
total_param_size.append(count)
print(key, size, count)
print(np.array(total_param_size).sum())
model.load_state_dict(model_params)
model.eval()
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation')
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
else:
model.init()
total_param_size = []
params = [value.numel() for value in model.parameters()]
print(params)
print(np.sum(params))
opt = torch.optim.Adam(model.parameters(), lr=Config.learning_rate, weight_decay=Config.L2)
for epoch in range(epochs):
model.train()
for i, str2var in enumerate(train_batcher):
opt.zero_grad()
e1 = str2var['e1']
rel = str2var['rel']
e2_multi = str2var['e2_multi1_binary'].float()
# label smoothing
e2_multi = ((1.0-Config.label_smoothing_epsilon)*e2_multi) + (1.0/e2_multi.size(1))
pred = model.forward(e1, rel)
loss = model.loss(pred, e2_multi)
loss.backward()
opt.step()
train_batcher.state.loss = loss
print('saving to {0}'.format(model_path))
torch.save(model.state_dict(), model_path)
model.eval()
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
if epoch % 3 == 0:
if epoch > 0:
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation')
def main():
if Config.process: preprocess(Config.dataset, delete_data=True)
input_keys = ['e1', 'rel', 'rel_eval', 'e2', 'e2_multi1', 'e2_multi2']
p = Pipeline(Config.dataset, keys=input_keys)
p.load_vocabs()
vocab = p.state['vocab']
num_entities = vocab['e1'].num_token
train_batcher = StreamBatcher(Config.dataset,
'train',
Config.batch_size,
randomize=True,
keys=input_keys)
dev_rank_batcher = StreamBatcher(Config.dataset,
'dev_ranking',
Config.batch_size,
randomize=False,
loader_threads=4,
keys=input_keys)
test_rank_batcher = StreamBatcher(Config.dataset,
'test_ranking',
Config.batch_size,
randomize=False,
loader_threads=4,
keys=input_keys)
if Config.model_name is None:
model = ConvE(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'ConvE':
model = ConvE(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'DistMult':
model = DistMult(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'ComplEx':
model = Complex(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'RNNDist':
model = RNNDist(vocab['e1'].num_token, vocab['rel'].num_token)
else:
log.info('Unknown model: {0}', Config.model_name)
raise Exception("Unknown model!")
train_batcher.at_batch_prepared_observers.insert(
1, TargetIdx2MultiTarget(num_entities, 'e2_multi1',
'e2_multi1_binary'))
eta = ETAHook('train', print_every_x_batches=100)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
train_batcher.subscribe_to_events(
LossHook('train', print_every_x_batches=100))
if Config.dataset == 'ICEWS18':
lengths = [
1618, 956, 815, 1461, 1634, 1596, 1754, 1494, 800, 979, 1588, 1779,
1831, 1762, 1566, 812, 820, 1707, 1988, 1845, 1670, 1695, 956, 930,
1641, 1813, 1759, 1664, 1616, 1021, 998, 1668, 1589, 1720
]
else:
lengths = [
1090, 730, 646, 939, 681, 783, 546, 526, 524, 586, 656, 741, 562,
474, 493, 487, 474, 477, 460, 532, 348, 530, 402, 493, 503, 452,
668, 512, 406, 467, 524, 563, 524, 418, 441, 487, 515, 475, 478,
532, 387, 479, 485, 417, 542, 496, 487, 445, 504, 350, 432, 445,
401, 570, 554, 504, 505, 483, 587, 441, 489, 501, 487, 513, 513,
524, 655, 545, 599, 702, 734, 519, 603, 579, 537, 635, 437, 422,
695, 575, 553, 485, 429, 663, 475, 673, 527, 559, 540, 591, 558,
698, 422, 1145, 969, 1074, 888, 683, 677, 910, 902, 644, 777, 695,
571, 656, 797, 576, 468, 676, 687, 549, 482, 1007, 778, 567, 813,
788, 879, 557, 724, 850, 809, 685, 714, 554, 799, 727, 208, 946,
979, 892, 859, 1092, 1038, 999, 1477, 1126, 1096, 1145, 955, 100,
1264, 1287, 962, 1031, 1603, 1662, 1179, 1064, 1179, 1105, 1465,
1176, 1219, 1137, 1112, 791, 829, 2347, 917, 913, 1107, 960, 850,
1005, 1045, 871, 972, 921, 1019, 984, 1033, 848, 918, 699, 1627,
1580, 1354, 1119, 1065, 1208, 1037, 1134, 980, 1249, 1031, 908,
787, 819, 804, 764, 959, 1057, 770, 691, 816, 620, 788, 829, 895,
1128, 1023, 1038, 1030, 1016, 991, 866, 878, 1013, 977, 914, 976,
717, 740, 904, 912, 1043, 1117, 930, 1116, 1028, 946, 922, 1151,
1092, 967, 1189, 1081, 1158, 943, 981, 1212, 1104, 941, 912, 1347,
1241, 1479, 1188, 1152, 1164, 1167, 1173, 1280, 979, 142, 1458,
910, 1126, 1053, 1083, 897, 1021, 1075, 881, 1054, 941, 927, 860,
1081, 876, 1952, 1576, 1560, 1599, 1226, 1083, 964, 1059, 1179,
982, 1032, 933, 877, 1032, 957, 884, 909, 846, 850, 798, 843, 1183,
1108, 1185, 797, 915, 952, 1181, 744, 86, 889, 1151, 925, 1119,
1115, 1036, 772, 1052, 837, 897, 1095, 926, 1034, 1031, 995, 907,
969, 981, 1135, 915, 1161, 100, 1269, 1244, 1331, 1124, 1074, 1162,
1159, 1078, 1311, 1210, 1308, 945, 1183, 1580, 1406, 1417, 1173,
1348, 1274, 1179, 893, 1107, 950, 1028, 1055, 1059, 1244, 1082,
1179, 1011, 955, 886, 865, 857
]
if Config.cuda:
model.cuda()
if load:
# if True:
model_params = torch.load(model_path)
print(model)
total_param_size = []
params = [(key, value.size(), value.numel())
for key, value in model_params.items()]
for key, size, count in params:
total_param_size.append(count)
print(key, size, count)
print(np.array(total_param_size).sum())
model.load_state_dict(model_params)
model.eval()
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation')
# ranks = ranking_and_hits2(model, test_rank_batcher, vocab, 'test_evaluation')
print(len(ranks))
mrr = []
curr_step = 0
for i in range(len(lengths)):
rr = np.array(ranks[curr_step:curr_step + 2 * lengths[i]])
mrr.append(np.mean(1 / rr))
curr_step += 2 * lengths[i]
with open(Config.dataset + 'mrr.txt', 'w') as f:
for i, mr in enumerate(mrr):
print("MRR (filtered) @ {}th day: {:.6f}".format(i, mr))
f.write(str(mr) + '\n')
h10 = []
curr_step = 0
for i in range(len(lengths)):
rr = np.array(ranks[curr_step:curr_step + 2 * lengths[i]])
h10.append(np.mean(rr <= 10))
with open(Config.dataset + 'h10.txt', 'w') as f:
for i, mr in enumerate(h10):
print("h10 (filtered) @ {}th day: {:.6f}".format(i, mr))
f.write(str(mr) + '\n')
h10 = []
for i in range(len(lengths)):
rr = np.array(ranks[curr_step:curr_step + 2 * lengths[i]])
h10.append(np.mean(rr <= 3))
with open(Config.dataset + 'h3.txt', 'w') as f:
for i, mr in enumerate(h10):
print("h10 (filtered) @ {}th day: {:.6f}".format(i, mr))
f.write(str(mr) + '\n')
h10 = []
for i in range(len(lengths)):
rr = np.array(ranks[curr_step:curr_step + 2 * lengths[i]])
h10.append(np.mean(rr <= 1))
with open(Config.dataset + 'h1.txt', 'w') as f:
for i, mr in enumerate(h10):
print("h10 (filtered) @ {}th day: {:.6f}".format(i, mr))
f.write(str(mr) + '\n')
print("length", len(ranks))
print("length_2", 2 * sum(lengths))
# ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
else:
model.init()
total_param_size = []
params = [value.numel() for value in model.parameters()]
print(params)
print(np.sum(params))
opt = torch.optim.Adam(model.parameters(),
lr=Config.learning_rate,
weight_decay=Config.L2)
for epoch in range(epochs):
# break
model.train()
for i, str2var in enumerate(train_batcher):
opt.zero_grad()
e1 = str2var['e1']
rel = str2var['rel']
e2_multi = str2var['e2_multi1_binary'].float()
# label smoothing
# e2_multi = ((1.0-Config.label_smoothing_epsilon)*e2_multi) + (1.0/e2_multi.size(1))
# print("this",Config.label_smoothing_epsilon, e2_multi.size(1))
pred = model.forward(e1, rel)
# loss = model.loss(pred, e2_multi)
# #
loss = torch.zeros(1).cuda()
for j in range(128):
position = torch.nonzero(e2_multi[j])[0].cuda()
label = torch.cat(
[torch.ones(len(position)),
torch.zeros(len(position))]).cuda()
neg_position = torch.randint(e2_multi.shape[1],
(len(position), )).long().cuda()
position = torch.cat([position, neg_position])
loss += model.loss(pred[j, position], label)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(),
1.0) # clip gradients
opt.step()
train_batcher.state.loss = loss.cpu()
print('saving to {0}'.format(model_path))
torch.save(model.state_dict(), model_path)
model.eval()
with torch.no_grad():
# ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
if epoch == 50:
ranks = ranking_and_hits(model, test_rank_batcher, vocab,
'test_evaluation')
def main():
if Config.process: preprocess(Config.dataset, delete_data=True)
input_keys = ['e1', 'rel', 'rel_eval', 'e2', 'e2_multi1', 'e2_multi2']
p = Pipeline(Config.dataset, keys=input_keys)
p.load_vocabs()
vocab = p.state['vocab']
num_entities = vocab['e1'].num_token
train_batcher = StreamBatcher(Config.dataset, 'train', Config.batch_size, randomize=True, keys=input_keys)
dev_rank_batcher = StreamBatcher(Config.dataset, 'dev_ranking', Config.batch_size, randomize=False, loader_threads=4, keys=input_keys)
test_rank_batcher = StreamBatcher(Config.dataset, 'test_ranking', Config.batch_size, randomize=False, loader_threads=4, keys=input_keys)
if Config.model_name is None:
model = ConvE(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'ConvE':
model = ConvE(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'DistMult':
model = DistMult(vocab['e1'].num_token, vocab['rel'].num_token)
elif Config.model_name == 'ComplEx':
model = Complex(vocab['e1'].num_token, vocab['rel'].num_token)
else:
log.info('Unknown model: {0}', Config.model_name)
raise Exception("Unknown model!")
train_batcher.at_batch_prepared_observers.insert(1,TargetIdx2MultiTarget(num_entities, 'e2_multi1', 'e2_multi1_binary'))
eta = ETAHook('train', print_every_x_batches=100)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
train_batcher.subscribe_to_events(LossHook('train', print_every_x_batches=100))
if Config.cuda:
model.cuda()
if load:
model_params = torch.load(model_path)
print(model)
total_param_size = []
params = [(key, value.size(), value.numel()) for key, value in model_params.items()]
for key, size, count in params:
total_param_size.append(count)
print(key, size, count)
print(np.array(total_param_size).sum())
model.load_state_dict(model_params)
model.eval()
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation',epochs,True)
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation',epochs,False)
else:
model.init()
total_param_size = []
params = [value.numel() for value in model.parameters()]
print(params)
print(np.sum(params))
opt = torch.optim.Adam(model.parameters(), lr=Config.learning_rate, weight_decay=Config.L2)
for epoch in range(epochs):
model.train()
for i, str2var in tqdm(enumerate(train_batcher)):
opt.zero_grad()
e1 = str2var['e1']
rel = str2var['rel']
e2_multi = str2var['e2_multi1_binary'].float()
# label smoothing
pred = model.forward(e1, rel)
loss = torch.zeros(1).cuda()
for j in range(128):
position = torch.nonzero(e2_multi[j])[0].cuda()
label = torch.cat([torch.ones(len(position)), torch.zeros(len(position))]).cuda()
neg_position = torch.randint(e2_multi.shape[1], (len(position),)).long().cuda()
position = torch.cat([position, neg_position])
loss += model.loss(pred[j, position], label)
loss.backward()
opt.step()
train_batcher.state.loss = loss.cpu()
print('saving to {0}'.format(model_path))
torch.save(model.state_dict(), model_path)
model.eval()
with torch.no_grad():
if epoch % 100 == 0:
if epoch > 0:
ranking_and_hits(model, test_rank_batcher, vocab, Config.dataset + "-" + Config.model_name,epoch,False)
if epoch + 1 == epochs:
ranking_and_hits(model, test_rank_batcher, vocab, Config.dataset,epoch,True)
def main():
if Config.process: preprocess(Config.dataset, delete_data=True)
input_keys = ['e1', 'rel', 'e2', 'e2_multi1', 'e2_multi2']
p = Pipeline(Config.dataset, keys=input_keys)
p.load_vocabs()
vocab = p.state['vocab']
num_entities = vocab['e1'].num_token
train_batcher = StreamBatcher(Config.dataset,
'train',
Config.batch_size,
randomize=True,
keys=input_keys)
dev_rank_batcher = StreamBatcher(Config.dataset,
'dev_ranking',
Config.batch_size,
randomize=False,
loader_threads=4,
keys=input_keys,
is_volatile=True)
test_rank_batcher = StreamBatcher(Config.dataset,
'test_ranking',
Config.batch_size,
randomize=False,
loader_threads=4,
keys=input_keys,
is_volatile=True)
# Load literals
numerical_literals = np.load(
f'data/{Config.dataset}/literals/numerical_literals.npy')
# Normalize literals
max_lit, min_lit = np.max(numerical_literals,
axis=0), np.min(numerical_literals, axis=0)
numerical_literals = (numerical_literals - min_lit) / (max_lit - min_lit +
1e-8)
# Load Multitask models
model = MTKGNN_DistMult(vocab['e1'].num_token, vocab['rel'].num_token,
numerical_literals)
train_batcher.at_batch_prepared_observers.insert(
1, TargetIdx2MultiTarget(num_entities, 'e2_multi1',
'e2_multi1_binary'))
eta = ETAHook('train', print_every_x_batches=100)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
train_batcher.subscribe_to_events(
LossHook('train', print_every_x_batches=100))
if Config.cuda:
model.cuda()
if load:
model_params = torch.load(model_path)
print(model)
total_param_size = []
params = [(key, value.size(), value.numel())
for key, value in model_params.items()]
for key, size, count in params:
total_param_size.append(count)
print(key, size, count)
print(np.array(total_param_size).sum())
model.load_state_dict(model_params)
model.eval()
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation')
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
else:
model.init()
total_param_size = []
params = [value.numel() for value in model.parameters()]
print(params)
print(np.sum(params))
opt_rel = torch.optim.Adam(model.rel_params,
lr=Config.learning_rate,
weight_decay=Config.L2)
opt_attr = torch.optim.Adam(model.attr_params,
lr=Config.learning_rate,
weight_decay=Config.L2)
for epoch in range(epochs):
model.train()
for i, str2var in enumerate(train_batcher):
opt_rel.zero_grad()
opt_attr.zero_grad()
e1 = str2var['e1']
e2 = str2var['e2']
rel = str2var['rel']
e2_multi = str2var['e2_multi1_binary'].float()
# label smoothing
e2_multi = ((1.0 - Config.label_smoothing_epsilon) *
e2_multi) + (1.0 / e2_multi.size(1))
pred = model.forward(e1, rel)
loss_rel = model.loss_rel(pred, e2_multi)
loss_rel.backward()
opt_rel.step()
pred_left, target_left = model.forward_attr(e1, 'left')
pred_right, target_right = model.forward_attr(e1, 'right')
loss_attr_left = model.loss_attr(pred_left, target_left)
loss_attr_right = model.loss_attr(pred_right, target_right)
loss_attr = loss_attr_left + loss_attr_right
loss_attr.backward()
opt_attr.step()
train_batcher.state.loss = loss_rel + loss_attr
# Attribute Specific Training
for k in range(4):
pred_left, pred_right, target = model.forward_AST()
loss_AST = model.loss_attr(pred_left, target) + model.loss_attr(
pred_right, target)
loss_AST.backward()
opt_attr.step()
print('saving to {0}'.format(model_path))
torch.save(model.state_dict(), model_path)
model.eval()
if epoch % 3 == 0:
if epoch > 0:
ranking_and_hits(model, dev_rank_batcher, vocab,
'dev_evaluation')
ranking_and_hits(model, test_rank_batcher, vocab,
'test_evaluation')
def preprocess_SNLI(delete_data=False):
# load data
#names, file_paths = snli2json()
#train_path, dev_path, test_path = file_paths
tokenizer = nltk.tokenize.WordPunctTokenizer()
zip_path = join(get_data_path(), 'snli_1.0.zip', 'snli_1.0')
file_paths = [
'snli_1.0_train.jsonl', 'snli_1.0_dev.jsonl', 'snli_1.0_test.jsonl'
]
not_t = []
t = ['input', 'support', 'target']
# tokenize and convert to hdf5
# 1. Setup pipeline to save lengths and generate vocabulary
p = Pipeline('snli_example', delete_data)
p.add_path(join(zip_path, file_paths[0]))
p.add_line_processor(JsonLoaderProcessors())
p.add_line_processor(
RemoveLineOnJsonValueCondition('gold_label',
lambda label: label == '-'))
p.add_line_processor(
DictKey2ListMapper(['sentence1', 'sentence2', 'gold_label']))
p.add_sent_processor(ToLower())
p.add_sent_processor(Tokenizer(tokenizer.tokenize), t)
#p.add_sent_processor(NaiveNCharTokenizer(3), not_t)
p.add_token_processor(AddToVocab())
p.add_post_processor(SaveLengthsToState())
p.execute()
p.clear_processors()
p.state['vocab'].save_to_disk()
# 2. Process the data further to stream it to hdf5
p.add_sent_processor(ToLower())
p.add_sent_processor(Tokenizer(tokenizer.tokenize), t)
#p.add_sent_processor(NaiveNCharTokenizer(3), not_t)
p.add_post_processor(ConvertTokenToIdx())
p.add_post_processor(
CreateBinsByNestedLength('snli_train', min_batch_size=128))
state = p.execute()
# dev and test data
p2 = Pipeline('snli_example')
p2.copy_vocab_from_pipeline(p)
p2.add_path(join(zip_path, file_paths[1]))
p2.add_line_processor(JsonLoaderProcessors())
p2.add_line_processor(
RemoveLineOnJsonValueCondition('gold_label',
lambda label: label == '-'))
p2.add_line_processor(
DictKey2ListMapper(['sentence1', 'sentence2', 'gold_label']))
p2.add_sent_processor(ToLower())
p2.add_sent_processor(Tokenizer(tokenizer.tokenize), t)
#p2.add_sent_processor(NaiveNCharTokenizer(3), not_t)
p2.add_post_processor(SaveLengthsToState())
p2.execute()
p2.clear_processors()
p2.add_sent_processor(ToLower())
p2.add_sent_processor(Tokenizer(tokenizer.tokenize), t)
#p2.add_sent_processor(NaiveNCharTokenizer(3), not_t)
p2.add_post_processor(ConvertTokenToIdx())
p2.add_post_processor(StreamToHDF5('snli_dev'))
p2.execute()
p3 = Pipeline('snli_example')
p3.copy_vocab_from_pipeline(p)
p3.add_path(join(zip_path, file_paths[2]))
p3.add_line_processor(JsonLoaderProcessors())
p3.add_line_processor(
RemoveLineOnJsonValueCondition('gold_label',
lambda label: label == '-'))
p3.add_line_processor(
DictKey2ListMapper(['sentence1', 'sentence2', 'gold_label']))
p3.add_sent_processor(ToLower())
p3.add_sent_processor(Tokenizer(tokenizer.tokenize), t)
#p3.add_sent_processor(NaiveNCharTokenizer(3), not_t)
p3.add_post_processor(SaveLengthsToState())
p3.execute()
p3.clear_processors()
p3.add_sent_processor(ToLower())
p3.add_sent_processor(Tokenizer(tokenizer.tokenize), t)
#p3.add_sent_processor(NaiveNCharTokenizer(3), not_t)
p3.add_post_processor(ConvertTokenToIdx())
p3.add_post_processor(StreamToHDF5('snli_test'))
p3.execute()
def main():
Logger.GLOBAL_LOG_LEVEL = LogLevel.INFO
#Config.backend = Backends.TENSORFLOW
Config.backend = Backends.TORCH
Config.cuda = True
Config.dropout = 0.1
Config.hidden_size = 128
Config.embedding_size = 256
Config.L2 = 0.00003
do_process = False
if do_process:
preprocess_SNLI(delete_data=True)
p = Pipeline('snli_example')
vocab = p.state['vocab']
vocab.load_from_disk()
batch_size = 128
if Config.backend == Backends.TENSORFLOW:
from spodernet.backends.tfbackend import TensorFlowConfig
TensorFlowConfig.init_batch_size(batch_size)
train_batcher = StreamBatcher('snli_example',
'snli_train',
batch_size,
randomize=True,
loader_threads=8)
#train_batcher.subscribe_to_batch_prepared_event(SomeExpensivePreprocessing())
dev_batcher = StreamBatcher('snli_example', 'snli_dev', batch_size)
test_batcher = StreamBatcher('snli_example', 'snli_test', batch_size)
#train_batcher.subscribe_to_events(AccuracyHook('Train', print_every_x_batches=1000))
train_batcher.subscribe_to_events(
LossHook('Train', print_every_x_batches=100))
train_batcher.subscribe_to_events(
AccuracyHook('Train', print_every_x_batches=100))
dev_batcher.subscribe_to_events(
AccuracyHook('Dev', print_every_x_batches=100))
dev_batcher.subscribe_to_events(LossHook('Dev', print_every_x_batches=100))
eta = ETAHook(print_every_x_batches=100)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
net = Net(vocab.num_embeddings, vocab.num_labels)
if Config.cuda:
net.cuda()
epochs = 10
opt = torch.optim.Adam(net.parameters(), lr=0.001)
net.train()
for epoch in range(epochs):
for str2var in train_batcher:
opt.zero_grad()
loss, argmax = net(str2var)
loss.backward()
opt.step()
train_batcher.state.loss = loss
train_batcher.state.targets = str2var['target']
train_batcher.state.argmax = argmax
net.eval()
for i, str2var in enumerate(dev_batcher):
t = str2var['target']
idx = str2var['index']
loss, argmax = net(str2var)
dev_batcher.state.loss = loss
dev_batcher.state.targets = str2var['target']
dev_batcher.state.argmax = argmax
def main():
if Config.process: preprocess(Config.dataset, delete_data=True)
input_keys = ['e1', 'rel', 'e2', 'e2_multi1', 'e2_multi2']
p = Pipeline(Config.dataset, keys=input_keys)
p.load_vocabs()
vocab = p.state['vocab']
num_entities = vocab['e1'].num_token
train_batcher = StreamBatcher(Config.dataset,
'train',
Config.batch_size,
randomize=True,
keys=input_keys)
dev_rank_batcher = StreamBatcher(Config.dataset,
'dev_ranking',
Config.batch_size,
randomize=False,
loader_threads=4,
keys=input_keys,
is_volatile=True)
test_rank_batcher = StreamBatcher(Config.dataset,
'test_ranking',
Config.batch_size,
randomize=False,
loader_threads=4,
keys=input_keys,
is_volatile=True)
# Load literals
numerical_literals = np.load(
f'data/{Config.dataset}/literals/numerical_literals.npy')
# Normalize literals
max_lit, min_lit = np.max(numerical_literals,
axis=0), np.min(numerical_literals, axis=0)
numerical_literals = (numerical_literals - min_lit) / (max_lit - min_lit +
1e-8)
# Load literal models
if Config.model_name is None:
model = DistMultLiteral(vocab['e1'].num_token, vocab['rel'].num_token,
numerical_literals)
elif Config.model_name == 'DistMultLiteral_highway':
model = DistMultLiteral_highway(vocab['e1'].num_token,
vocab['rel'].num_token,
numerical_literals)
elif Config.model_name == 'DistMultLiteral_gate':
model = DistMultLiteral_gate(vocab['e1'].num_token,
vocab['rel'].num_token,
numerical_literals)
elif Config.model_name == 'ComplEx':
model = ComplexLiteral(vocab['e1'].num_token, vocab['rel'].num_token,
numerical_literals)
elif Config.model_name == 'ConvE':
model = ConvELiteral(vocab['e1'].num_token, vocab['rel'].num_token,
numerical_literals)
elif Config.model_name == 'ConvEAlt':
model = ConvELiteralAlt(vocab['e1'].num_token, vocab['rel'].num_token,
numerical_literals)
elif Config.model_name == 'DistMultNN':
model = DistMultLiteralNN(vocab['e1'].num_token,
vocab['rel'].num_token, numerical_literals)
elif Config.model_name == 'DistMultNN2':
model = DistMultLiteralNN2(vocab['e1'].num_token,
vocab['rel'].num_token, numerical_literals)
else:
log.info('Unknown model: {0}', Config.model_name)
raise Exception("Unknown model!")
train_batcher.at_batch_prepared_observers.insert(
1, TargetIdx2MultiTarget(num_entities, 'e2_multi1',
'e2_multi1_binary'))
eta = ETAHook('train', print_every_x_batches=100)
train_batcher.subscribe_to_events(eta)
train_batcher.subscribe_to_start_of_epoch_event(eta)
train_batcher.subscribe_to_events(
LossHook('train', print_every_x_batches=100))
if Config.cuda:
model.cuda()
if load:
model_params = torch.load(model_path)
print(model)
total_param_size = []
params = [(key, value.size(), value.numel())
for key, value in model_params.items()]
for key, size, count in params:
total_param_size.append(count)
print(key, size, count)
print(np.array(total_param_size).sum())
model.load_state_dict(model_params)
model.eval()
ranking_and_hits(model, test_rank_batcher, vocab, 'test_evaluation')
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
else:
model.init()
total_param_size = []
params = [value.numel() for value in model.parameters()]
print(params)
print(np.sum(params))
opt = torch.optim.Adam(model.parameters(),
lr=Config.learning_rate,
weight_decay=Config.L2)
for epoch in range(epochs):
model.train()
for i, str2var in enumerate(train_batcher):
opt.zero_grad()
e1 = str2var['e1']
rel = str2var['rel']
e2_multi = str2var['e2_multi1_binary'].float()
# label smoothing
#e2_multi = ((1.0-Config.label_smoothing_epsilon)*e2_multi) + (1.0/e2_multi.size(1))
pred = model.forward(e1, rel)
loss = model.loss(pred, e2_multi)
loss.backward()
opt.step()
train_batcher.state.loss = loss
print('saving to {0}'.format(model_path))
torch.save(model.state_dict(), model_path)
model.eval()
ranking_and_hits(model, dev_rank_batcher, vocab, 'dev_evaluation')
if epoch % 3 == 0:
if epoch > 0:
ranking_and_hits(model, test_rank_batcher, vocab,
'test_evaluation')
