def beam_search(model: NMT, test_iter, beam_size: int,
max_decoding_time_step: int) -> List[List[Hypothesis]]:
""" Run beam search to construct hypotheses for a list of src-language sentences.
@param model (NMT): NMT Model
@param test_data_src (List[List[str]]): List of sentences (words) in source language, from test set.
@param beam_size (int): beam_size (# of hypotheses to hold for a translation at every step)
@param max_decoding_time_step (int): maximum sentence length that Beam search can produce
@returns hypotheses (List[List[Hypothesis]]): List of Hypothesis translations for every source sentence.
"""
was_training = model.training
model.eval()
hypotheses = []
with torch.no_grad():
for _, data in enumerate(test_iter):
print(data)
(src_sents, src_lengths), (_, _) = data.abc, data.d
example_hyps = model.beam_search(
src_sents,
src_lengths,
beam_size=beam_size,
max_decoding_time_step=max_decoding_time_step)
hypotheses.append(example_hyps)
if was_training: model.train(was_training)
return hypotheses
def beam_search(model: NMT, test_data_src: List[List[str]], beam_size: int,
max_decoding_time_step: int) -> List[List[Hypothesis]]:
""" Run beam search to construct hypotheses for a list of src-language sentences.
@param model (NMT): NMT Model
@param test_data_src (List[List[str]]): List of sentences (words) in source language, from test set.
@param beam_size (int): beam_size (# of hypotheses to hold for a translation at every step)
@param max_decoding_time_step (int): maximum sentence length that Beam search can produce
@returns hypotheses (List[List[Hypothesis]]): List of Hypothesis translations for every source sentence.
"""
was_training = model.training
model.eval()
hypotheses = []
with torch.no_grad():
for src_sent in tqdm(test_data_src, desc='Decoding', file=sys.stdout):
example_hyps = model.beam_search(
src_sent,
beam_size=beam_size,
max_decoding_time_step=max_decoding_time_step)
hypotheses.append(example_hyps)
if was_training: model.train(was_training)
return hypotheses
def beam_search(model: NMT, test_iterator: BucketIterator, beam_size: int,
max_decoding_time_step: int) -> List[List[Hypothesis]]:
""" Run beam search to construct hypotheses for a list of src-language sentences.
@param model (NMT): NMT Model
@param test_iterator BucketIterator: BucketIterator in source language, from test set.
@param beam_size (int): beam_size (# of hypotheses to hold for a translation at every step)
@param max_decoding_time_step (int): maximum sentence length that Beam search can produce
@returns hypotheses (List[List[Hypothesis]]): List of Hypothesis translations for every source sentence.
"""
was_training = model.training
model.eval()
hypotheses = []
with torch.no_grad():
# for src_sent in tqdm(test_data_src, desc='Decoding', file=sys.stdout):
for i, batch in enumerate(test_iterator):
src_sents, src_sents_lens = batch.src
src_sents = src_sents.permute(1, 0)
for j in range(len(src_sents_lens)):
src_sent = src_sents[j]
example_hyps = model.beam_search(
src_sent,
src_sents_lens[j],
beam_size=beam_size,
max_decoding_time_step=max_decoding_time_step)
hypotheses.append(example_hyps)
if was_training: model.train(was_training)
return hypotheses
def beam_search(model: NMT, test_data_src: List[List[str]], beam_size: int,
max_decoding_time_step: int) -> List[List[Hypothesis]]:
""" 对源句子列表使用beam search去构建假设.
@param model (NMT): NMT 模型
@param test_data_src (List[List[str]]): 源句子列表, 测试集中的.
@param beam_size (int): beam_size (每一步的候选数)
@param max_decoding_time_step (int): Beam search 能产生的最大句子长度
@returns hypotheses (List[List[Hypothesis]]): 每个源句子的beam_size个假设.
"""
was_training = model.training
model.eval()
hypotheses = [] # 所有句子的候选句列表
with torch.no_grad():
for src_sent in tqdm(test_data_src, desc='Decoding', file=sys.stdout):
example_hyps = model.beam_search(
src_sent,
beam_size=beam_size,
max_decoding_time_step=max_decoding_time_step)
hypotheses.append(example_hyps) # 把这句话的所有候选句加入列表
if was_training: model.train(was_training)
return hypotheses
def beam_search2(model1: NMT, model2: DPPNMT, test_data_src: List[List[str]],
beam_size: int, max_decoding_time_step: int,
test_data_tgt) -> List[List[Hypothesis]]:
""" Run beam search to construct hypotheses for a list of src-language sentences.
@param model (NMT): NMT Model
@param test_data_src (List[List[str]]): List of sentences (words) in source language, from test set.
@param beam_size (int): beam_size (# of hypotheses to hold for a translation at every step)
@param max_decoding_time_step (int): maximum sentence length that Beam search can produce
@returns hypotheses (List[List[Hypothesis]]): List of Hypothesis translations for every source sentence.
"""
model1.eval()
model2.eval()
i = 0
with torch.no_grad():
for src_sent in tqdm(test_data_src, desc='Decoding', file=sys.stdout):
hyp1 = model1.beam_search(
src_sent,
beam_size=beam_size,
max_decoding_time_step=max_decoding_time_step)
hyp2 = model2.beam_search(
src_sent,
beam_size=beam_size,
max_decoding_time_step=max_decoding_time_step)
ref = test_data_tgt[i][1:-1]
#print(ref, hyp1[0].value)
bleu_topk = sentence_bleu(ref, hyp1[0].value)
bleu_dpp = sentence_bleu(test_data_tgt[i], hyp2[0].value)
#print(bleu_topk, bleu_dpp)
if bleu_dpp > bleu_topk:
print(i)
print(" ".join(hyp1[0].value))
print(" ".join(hyp2[0].value))
print(" ".join(ref))
i += 1
def test(args):
test_data_src = read_corpus(args.test_src, source='src')
test_data_tgt = read_corpus(args.test_tgt, source='tgt')
test_data = list(zip(test_data_src, test_data_tgt))
if args.load_model:
print('load model from [%s]' % args.load_model)
params = torch.load(args.load_model,
map_location=lambda storage, loc: storage)
vocab = params['vocab']
saved_args = params['args']
state_dict = params['state_dict']
model = NMT(saved_args, vocab)
model.load_state_dict(state_dict)
else:
vocab = torch.load(args.vocab)
model = NMT(args, vocab)
model.eval()
if args.cuda:
# model = nn.DataParallel(model).cuda()
model = model.cuda()
hypotheses = decode(model, test_data)
top_hypotheses = [hyps[0] for hyps in hypotheses]
bleu_score = get_bleu([tgt for src, tgt in test_data], top_hypotheses)
word_acc = get_acc([tgt for src, tgt in test_data], top_hypotheses,
'word_acc')
sent_acc = get_acc([tgt for src, tgt in test_data], top_hypotheses,
'sent_acc')
print('Corpus Level BLEU: %f, word level acc: %f, sentence level acc: %f' %
(bleu_score, word_acc, sent_acc),
file=sys.stderr)
if args.save_to_file:
print('save decoding results to %s' % args.save_to_file)
with open(args.save_to_file, 'w') as f:
for hyps in hypotheses:
f.write(' '.join(hyps[0][1:-1]) + '\n')
if args.save_nbest:
nbest_file = args.save_to_file + '.nbest'
print('save nbest decoding results to %s' % nbest_file)
with open(nbest_file, 'w') as f:
for src_sent, tgt_sent, hyps in zip(test_data_src,
test_data_tgt, hypotheses):
print('Source: %s' % ' '.join(src_sent), file=f)
print('Target: %s' % ' '.join(tgt_sent), file=f)
print('Hypotheses:', file=f)
for i, hyp in enumerate(hyps, 1):
print('[%d] %s' % (i, ' '.join(hyp)), file=f)
print('*' * 30, file=f)
def sample(args):
train_data_src = read_corpus(args.train_src, source='src')
train_data_tgt = read_corpus(args.train_tgt, source='tgt')
train_data = zip(train_data_src, train_data_tgt)
if args.load_model:
print('load model from [%s]' % args.load_model)
params = torch.load(args.load_model,
map_location=lambda storage, loc: storage)
vocab = params['vocab']
opt = params['args']
state_dict = params['state_dict']
model = NMT(opt, vocab)
model.load_state_dict(state_dict)
else:
vocab = torch.load(args.vocab)
model = NMT(args, vocab)
model.eval()
if args.cuda:
# model = nn.DataParallel(model).cuda()
model = model.cuda()
print('begin sampling')
check_every = 10
train_iter = cum_samples = 0
train_time = time.time()
for src_sents, tgt_sents in data_iter(train_data,
batch_size=args.batch_size):
train_iter += 1
samples = model.sample(src_sents,
sample_size=args.sample_size,
to_word=True)
cum_samples += sum(len(sample) for sample in samples)
if train_iter % check_every == 0:
elapsed = time.time() - train_time
print('sampling speed: %d/s' % (cum_samples / elapsed))
cum_samples = 0
train_time = time.time()
for i, tgt_sent in enumerate(tgt_sents):
print('*' * 80)
print('target:' + ' '.join(tgt_sent))
tgt_samples = samples[i]
print('samples:')
for sid, sample in enumerate(tgt_samples, 1):
print('[%d] %s' % (sid, ' '.join(sample[1:-1])))
print('*' * 80)
def interactive(args):
assert args.load_model, 'You have to specify a pre-trained model'
print('load model from [%s]' % args.load_model)
params = torch.load(args.load_model,
map_location=lambda storage, loc: storage)
vocab = params['vocab']
saved_args = params['args']
state_dict = params['state_dict']
model = NMT(saved_args, vocab)
model.load_state_dict(state_dict)
model.eval()
if args.cuda:
# model = nn.DataParallel(model).cuda()
model = model.cuda()
while True:
src_sent = input('Source Sentence:')
src_sent = src_sent.strip().split(' ')
hyps = model.translate(src_sent)
for i, hyp in enumerate(hyps, 1):
print('Hypothesis #%d: %s' % (i, ' '.join(hyp)))
def beam_search(model: NMT, test_data_src: List[List[str]], beam_size: int, max_decoding_time_step: int)\
-> List[List[Hypothesis]]:
""" Run beam search to construct hypotheses for a list of src-language sentences.
:param NMT model: NMT Model
:param List[List[str]] test_data_src: List of sentences (words) in source language, from test set
:param int beam_size: beam_size (number of hypotheses to keep for a translation at every step)
:param int max_decoding_time_step: maximum sentence length that beam search can produce
:returns List[List[Hypothesis]] hypotheses: List of Hypothesis translations for every source sentence
"""
was_training = model.training
model.eval()
hypotheses = []
with torch.no_grad():
for src_sent in tqdm(test_data_src, desc='Decoding', file=sys.stdout):
example_hyps = model.beam_search(
src_sent,
beam_size=beam_size,
max_decoding_time_step=max_decoding_time_step)
hypotheses.append(example_hyps)
if was_training: model.train(was_training)
return hypotheses
def compute_lm_prob(args):
"""
given source-target sentence pairs, compute ppl and log-likelihood
"""
test_data_src = read_corpus(args.test_src, source='src')
test_data_tgt = read_corpus(args.test_tgt, source='tgt')
test_data = zip(test_data_src, test_data_tgt)
if args.load_model:
print('load model from [%s]' % args.load_model)
params = torch.load(args.load_model,
map_location=lambda storage, loc: storage)
vocab = params['vocab']
saved_args = params['args']
state_dict = params['state_dict']
model = NMT(saved_args, vocab)
model.load_state_dict(state_dict)
else:
vocab = torch.load(args.vocab)
model = NMT(args, vocab)
model.eval()
if args.cuda:
# model = nn.DataParallel(model).cuda()
model = model.cuda()
f = open(args.save_to_file, 'w')
for src_sent, tgt_sent in test_data:
src_sents = [src_sent]
tgt_sents = [tgt_sent]
batch_size = len(src_sents)
src_sents_len = [len(s) for s in src_sents]
pred_tgt_word_nums = [len(s[1:])
for s in tgt_sents] # omitting leading `<s>`
# (sent_len, batch_size)
src_sents_var = to_input_variable(src_sents,
model.vocab.src,
cuda=args.cuda,
is_test=True)
tgt_sents_var = to_input_variable(tgt_sents,
model.vocab.tgt,
cuda=args.cuda,
is_test=True)
# (tgt_sent_len, batch_size, tgt_vocab_size)
scores = model(src_sents_var, src_sents_len, tgt_sents_var[:-1])
# (tgt_sent_len * batch_size, tgt_vocab_size)
log_scores = F.log_softmax(scores.view(-1, scores.size(2)))
# remove leading <s> in tgt sent, which is not used as the target
# (batch_size * tgt_sent_len)
flattened_tgt_sents = tgt_sents_var[1:].view(-1)
# (batch_size * tgt_sent_len)
tgt_log_scores = torch.gather(
log_scores, 1, flattened_tgt_sents.unsqueeze(1)).squeeze(1)
# 0-index is the <pad> symbol
tgt_log_scores = tgt_log_scores * (
1. - torch.eq(flattened_tgt_sents, 0).float())
# (tgt_sent_len, batch_size)
tgt_log_scores = tgt_log_scores.view(-1, batch_size) # .permute(1, 0)
# (batch_size)
tgt_sent_scores = tgt_log_scores.sum(dim=0).squeeze()
tgt_sent_word_scores = [
tgt_sent_scores[i].item() / pred_tgt_word_nums[i]
for i in range(batch_size)
]
for src_sent, tgt_sent, score in zip(src_sents, tgt_sents,
tgt_sent_word_scores):
f.write('%s ||| %s ||| %f\n' %
(' '.join(src_sent), ' '.join(tgt_sent), score))
f.close()
def experiement(args: Dict, test_only, device):
""" Train and Test the NMT Model.
@param args (Dict): args from cmd line
"""
# train_data_src = read_corpus(args['--train-src'], source='src')
# train_data_tgt = read_corpus(args['--train-tgt'], source='tgt')
#
# dev_data_src = read_corpus(args['--dev-src'], source='src')
# dev_data_tgt = read_corpus(args['--dev-tgt'], source='tgt')
#
# train_data = list(zip(train_data_src, train_data_tgt))
# dev_data = list(zip(dev_data_src, dev_data_tgt))
train_batch_size = int(args['--batch-size'])
clip_grad = float(args['--clip-grad'])
valid_niter = int(args['--valid-niter'])
log_every = int(args['--log-every'])
model_save_path = args['--save-to']
use_pos_embed = False
if args['--use-pos-embed']:
use_pos_embed = True
use_copy = False
if args['--use-copy']:
use_copy = True
SRC, TRG, train_iterator, dev_iterator, test_iterator = load_data(
args['--train-data'], args['--dev-data'], args['--test-data'], device,
train_batch_size, (use_pos_embed or use_copy))
vocab = Vocab(SRC, TRG)
model = NMT(src_embed_size=int(args['--src-embed-size']),
dst_embed_size=int(args['--dst-embed-size']),
hidden_size=int(args['--hidden-size']),
dropout_rate=float(args['--dropout']),
vocab=vocab,
use_pos_embed=use_pos_embed,
use_copy=use_copy)
model.load_pretrained_embeddings(vocab)
# print("args: {}".format(args))
uniform_init = float(args['--uniform-init'])
if np.abs(uniform_init) > 0.:
print('uniformly initialize parameters [-%f, +%f]' %
(uniform_init, uniform_init),
file=sys.stderr)
for p in model.parameters():
p.data.uniform_(-uniform_init, uniform_init)
# def init_weights(m):
# for name, param in m.named_parameters():
# if 'weight' in name:
# nn.init.normal_(param.data, mean=0, std=0.01)
# else:
# nn.init.constant_(param.data, 0)
#
# model.apply(init_weights)
# vocab_mask = torch.ones(len(vocab.tgt))
# vocab_mask[vocab.tgt['<pad>']] = 0
print('use device: %s' % device, file=sys.stderr)
print(model)
para_count = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f'The model has {para_count:,} trainable parameters')
print("file path: {}".format(model_save_path))
if test_only:
model.eval()
decode(args, test_iterator, vocab, device)
exit(0)
# perform training
model.train()
model = model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=float(args['--lr']))
num_trial = 0
train_iter = patience = cum_loss = report_loss = cum_tgt_words = report_tgt_words = 0
cum_examples = report_examples = epoch = valid_num = 0
hist_valid_scores = []
train_time = begin_time = time.time()
print('begin Maximum Likelihood training')
while True:
epoch += 1
#perform training
model.train()
# for src_sents, tgt_sents in batch_iter(train_data, batch_size=train_batch_size, shuffle=True):
for i, batch in enumerate(train_iterator):
train_iter += 1
optimizer.zero_grad()
src_sents, src_sents_lens = batch.src
tgt_sents = batch.trg
batch_size = src_sents.shape[1]
example_losses = -model(src_sents, src_sents_lens,
tgt_sents) # (batch_size,)
batch_loss = example_losses.sum()
loss = batch_loss / batch_size
loss.backward()
# clip gradient
grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(),
clip_grad)
optimizer.step()
batch_losses_val = batch_loss.item()
report_loss += batch_losses_val
cum_loss += batch_losses_val
tgt_words_num_to_predict = sum(
len(s[1:]) for s in tgt_sents) # omitting leading `<s>`
report_tgt_words += tgt_words_num_to_predict
cum_tgt_words += tgt_words_num_to_predict
report_examples += batch_size
cum_examples += batch_size
# if train_iter % log_every == 0:
# print("")
print('epoch %d, iter %d, avg. loss %.2f, avg. ppl %.2f ' \
'cum. examples %d, speed %.2f words/sec, time elapsed %.2f sec' % (epoch, train_iter, report_loss / report_examples,
math.exp(report_loss / report_tgt_words),
cum_examples,
report_tgt_words / (time.time() - train_time),
time.time() - begin_time), file=sys.stderr)
train_time = time.time()
report_loss = report_tgt_words = report_examples = 0.
# perform validation
# model.eval()
# if train_iter % valid_niter == 0:
# print('epoch %d, iter %d, cum. loss %.2f, cum. ppl %.2f cum. examples %d' % (epoch, train_iter,
# cum_loss / cum_examples,
# np.exp(cum_loss / cum_tgt_words),
# cum_examples), file=sys.stderr)
cum_loss = cum_examples = cum_tgt_words = 0.
valid_num += 1
# print('begin validation ...', file=sys.stderr)
# compute dev. ppl and bleu
dev_ppl = evaluate_ppl(
model, dev_iterator,
batch_size=128) # dev batch size can be a bit larger
valid_metric = -dev_ppl
print('validation: iter %d, dev. ppl %f' % (train_iter, dev_ppl),
file=sys.stderr)
is_better = len(
hist_valid_scores) == 0 or valid_metric > max(hist_valid_scores)
hist_valid_scores.append(valid_metric)
if is_better:
patience = 0
# print('save currently the best model to [%s]' % model_save_path, file=sys.stderr)
model.save(model_save_path)
# also save the optimizers' state
torch.save(optimizer.state_dict(), model_save_path + '.optim')
elif patience < int(args['--patience']):
patience += 1
print('hit patience %d' % patience, file=sys.stderr)
if patience == int(args['--patience']):
num_trial += 1
print('hit #%d trial' % num_trial, file=sys.stderr)
if num_trial == int(args['--max-num-trial']):
print('early stop!', file=sys.stderr)
# exit(0)
break
# decay lr, and restore from previously best checkpoint
lr = optimizer.param_groups[0]['lr'] * float(
args['--lr-decay'])
print(
'load previously best model and decay learning rate to %f'
% lr,
file=sys.stderr)
# load model
params = torch.load(model_save_path,
map_location=lambda storage, loc: storage)
model.load_state_dict(params['state_dict'])
model = model.to(device)
print('restore parameters of the optimizers', file=sys.stderr)
optimizer.load_state_dict(
torch.load(model_save_path + '.optim'))
# set new lr
for param_group in optimizer.param_groups:
param_group['lr'] = lr
# reset patience
patience = 0
if epoch == int(args['--max-epoch']):
print('reached maximum number of epochs!', file=sys.stderr)
break
# perform testing
model.eval()
decode(args, test_iterator, vocab, device)
