def train(self):
wlog('Start training ... ')
assert wargs.sample_size < wargs.batch_size, 'Batch size < sample count'
# [low, high)
batch_count = len(self.train_data) if wargs.fine_tune is False else 0
batch_count_domain = len(self.train_data_domain)
batch_start_sample = tc.randperm(batch_count + batch_count_domain)[0]
wlog('Randomly select {} samples in the {}th/{} batch'.format(
wargs.sample_size, batch_start_sample,
batch_count + batch_count_domain))
bidx, eval_cnt = 0, [0]
wlog('Self-normalization alpha -> {}'.format(wargs.self_norm_alpha))
train_start = time.time()
wlog('')
wlog('#' * 120)
wlog('#' * 30, False)
wlog(' Start Training ', False)
wlog('#' * 30)
wlog('#' * 120)
#DANN
loss_domain = tc.nn.NLLLoss()
for epoch in range(wargs.start_epoch, wargs.max_epochs + 1):
epoch_start = time.time()
# train for one epoch on the training data
wlog('')
wlog('$' * 30, False)
wlog(' Epoch [{}/{}] '.format(epoch, wargs.max_epochs), False)
wlog('$' * 30)
if wargs.epoch_shuffle and epoch > wargs.epoch_shuffle_minibatch:
self.train_data.shuffle()
self.train_data_domain.shuffle()
# shuffle the original batch
shuffled_batch_idx = tc.randperm(batch_count + batch_count_domain)
sample_size = wargs.sample_size
epoch_loss, epoch_trg_words, epoch_num_correct = 0, 0, 0
show_loss, show_src_words, show_trg_words, show_correct_num = 0, 0, 0, 0
domain_loss = 0
sample_spend, eval_spend, epoch_bidx = 0, 0, 0
show_start = time.time()
for name, par in self.model.named_parameters():
if name.split('.')[0] != "domain_discriminator":
par.requires_grad = True
else:
par.requires_grad = False
for k in range(batch_count + batch_count_domain):
bidx += 1
epoch_bidx = k + 1
batch_idx = shuffled_batch_idx[
k] if epoch >= wargs.epoch_shuffle_minibatch else k
#p = float(k + epoch * (batch_count+batch_count_domain)) / wargs.max_epochs / (batch_count+batch_count_domain)
#alpha = 2. / (1. + np.exp(-10 * p)) - 1
if batch_idx < batch_count:
# (max_slen_batch, batch_size)
_, srcs, trgs, slens, srcs_m, trgs_m = self.train_data[
batch_idx]
self.model.zero_grad()
# (max_tlen_batch - 1, batch_size, out_size)
#forward compute DANN out of domain
decoder_outputs, domain_outputs = self.model(
srcs,
trgs[:-1],
srcs_m,
trgs_m[:-1],
'OUT',
alpha=wargs.alpha)
if len(decoder_outputs) == 2:
(decoder_outputs, _checks) = decoder_outputs
this_bnum = decoder_outputs.size(1)
#batch_loss, grad_output, batch_correct_num = memory_efficient(
# outputs, trgs[1:], trgs_m[1:], self.model.classifier)
#backward compute, now we have the grad
batch_loss, batch_correct_num, batch_log_norm = self.model.classifier.snip_back_prop(
decoder_outputs, trgs[1:], trgs_m[1:], wargs.snip_size)
#self.model.zero_grad()
#domain_outputs = self.model(srcs, None, srcs_m, None, 'OUT', alpha=alpha)
#if wargs.cross_entropy is True:
# domain_label = tc.zeros(len(domain_outputs))
# domain_label = domain_label.long()
# domain_label = domain_label.cuda()
#domainv_label = Variable(domain_label, volatile=False)
# domain_loss_a = loss_domain(tc.log(domain_outputs), domainv_label)
# domain_loss_a.backward(retain_graph = True if wargs.max_entropy is True else False)
if wargs.max_entropy is True:
#try to max this entropy
lam = wargs.alpha if epoch > 1 else 0
domain_loss_b = lam * tc.dot(tc.log(domain_outputs),
domain_outputs)
domain_loss_b.backward()
batch_src_words = srcs.data.ne(PAD).sum()
assert batch_src_words == slens.data.sum()
batch_trg_words = trgs[1:].data.ne(PAD).sum()
elif batch_idx >= batch_count:
#domain_loss_out.backward(retain_graph=True)
#DANN in domain compute
_, srcs_domain, trgs_domain, slens_domain, srcs_m_domain, trgs_m_domain = self.train_data_domain[
batch_idx - batch_count]
self.model.zero_grad()
decoder_outputs, domain_outputs = self.model(
srcs_domain,
trgs_domain[:-1],
srcs_m_domain,
trgs_m_domain[:-1],
'IN',
alpha=wargs.alpha)
if len(decoder_outputs) == 2:
(decoder_outputs, _checks) = decoder_outputs
this_bnum = decoder_outputs.size(1)
batch_loss, batch_correct_num, batch_log_norm = self.model.classifier.snip_back_prop(
decoder_outputs, trgs_domain[1:], trgs_m_domain[1:],
wargs.snip_size)
#domain_outputs = self.model(srcs_domain, None, srcs_m_domain, None, 'IN', alpha=alpha)
#if wargs.cross_entropy is True:
# domain_label = tc.ones(len(domain_outputs))
# domain_label = domain_label.long()
# domain_label = domain_label.cuda()
#domainv_label = Variable(domain_label, volatile=False)
# domain_loss_a = loss_domain(tc.log(domain_outputs), domainv_label)
# domain_loss_a.backward(retain_graph = True if wargs.max_entropy is True else False)
if wargs.max_entropy is True:
lam = wargs.alpha if epoch > 1 else 0
domain_loss_b = lam * tc.dot(tc.log(domain_outputs),
domain_outputs)
domain_loss_b.backward()
batch_src_words = srcs_domain.data.ne(PAD).sum()
assert batch_src_words == slens_domain.data.sum()
batch_trg_words = trgs_domain[1:].data.ne(PAD).sum()
_grad_nan = False
for n, p in self.model.named_parameters():
if p.grad is None:
debug('grad None | {}'.format(n))
continue
tmp_grad = p.grad.data.cpu().numpy()
if numpy.isnan(tmp_grad).any(
): # we check gradient here for vanishing Gradient
wlog("grad contains 'nan' | {}".format(n))
#wlog("gradient\n{}".format(tmp_grad))
_grad_nan = True
if n == 'decoder.l_f1_0.weight' or n == 's_init.weight' or n=='decoder.l_f1_1.weight' \
or n == 'decoder.l_conv.0.weight' or n == 'decoder.l_f2.weight':
debug('grad zeros |{:5} {}'.format(
str(not np.any(tmp_grad)), n))
if _grad_nan is True and wargs.dynamic_cyk_decoding is True:
for _i, items in enumerate(_checks):
wlog('step {} Variable----------------:'.format(_i))
#for item in items: wlog(item.cpu().data.numpy())
wlog('wen _check_tanh_sa ---------------')
wlog(items[0].cpu().data.numpy())
wlog('wen _check_a1_weight ---------------')
wlog(items[1].cpu().data.numpy())
wlog('wen _check_a1 ---------------')
wlog(items[2].cpu().data.numpy())
wlog('wen alpha_ij---------------')
wlog(items[3].cpu().data.numpy())
wlog('wen before_mask---------------')
wlog(items[4].cpu().data.numpy())
wlog('wen after_mask---------------')
wlog(items[5].cpu().data.numpy())
#outputs.backward(grad_output)
self.optim.step()
#del outputs, grad_output
show_loss += batch_loss
#domain_loss += domain_loss_a.data.clone()[0]
#domain_loss += domain_loss_out
show_correct_num += batch_correct_num
epoch_loss += batch_loss
epoch_num_correct += batch_correct_num
show_src_words += batch_src_words
show_trg_words += batch_trg_words
epoch_trg_words += batch_trg_words
batch_log_norm = tc.mean(tc.abs(batch_log_norm))
if epoch_bidx % wargs.display_freq == 0:
#print show_correct_num, show_loss, show_trg_words, show_loss/show_trg_words
ud = time.time() - show_start - sample_spend - eval_spend
wlog(
'Epo:{:>2}/{:>2} |[{:^5} {:^5} {:^5}k] |acc:{:5.2f}% |ppl:{:4.2f} '
'| |logZ|:{:.2f} '
'|stok/s:{:>4}/{:>2}={:>2} |ttok/s:{:>2} '
'|stok/sec:{:6.2f} |ttok/sec:{:6.2f} |elapsed:{:4.2f}/{:4.2f}m'
.format(epoch, wargs.max_epochs, epoch_bidx, batch_idx,
bidx / 1000,
(show_correct_num / show_trg_words) * 100,
math.exp(show_loss / show_trg_words),
batch_log_norm, batch_src_words, this_bnum,
int(batch_src_words / this_bnum),
int(batch_trg_words / this_bnum),
show_src_words / ud, show_trg_words / ud, ud,
(time.time() - train_start) / 60.))
show_loss, show_src_words, show_trg_words, show_correct_num = 0, 0, 0, 0
sample_spend, eval_spend = 0, 0
show_start = time.time()
if epoch_bidx % wargs.sampling_freq == 0:
sample_start = time.time()
self.model.eval()
#self.model.classifier.eval()
tor = Translator(self.model, self.sv, self.tv)
if batch_idx < batch_count:
# (max_len_batch, batch_size)
sample_src_tensor = srcs.t()[:sample_size]
sample_trg_tensor = trgs.t()[:sample_size]
tor.trans_samples(sample_src_tensor, sample_trg_tensor,
"OUT")
elif batch_idx >= batch_count:
sample_src_tensor = srcs_domain.t()[:sample_size]
sample_trg_tensor = trgs_domain.t()[:sample_size]
tor.trans_samples(sample_src_tensor, sample_trg_tensor,
"IN")
wlog('')
sample_spend = time.time() - sample_start
self.model.train()
# Just watch the translation of some source sentences in training data
if wargs.if_fixed_sampling and bidx == batch_start_sample:
# randomly select sample_size sample from current batch
rand_rows = np.random.choice(this_bnum,
sample_size,
replace=False)
sample_src_tensor = tc.Tensor(sample_size,
srcs.size(0)).long()
sample_src_tensor.fill_(PAD)
sample_trg_tensor = tc.Tensor(sample_size,
trgs.size(0)).long()
sample_trg_tensor.fill_(PAD)
for id in xrange(sample_size):
sample_src_tensor[id, :] = srcs.t()[rand_rows[id], :]
sample_trg_tensor[id, :] = trgs.t()[rand_rows[id], :]
if wargs.epoch_eval is not True and bidx > wargs.eval_valid_from and \
bidx % wargs.eval_valid_freq == 0:
eval_start = time.time()
eval_cnt[0] += 1
wlog('\nAmong epoch, batch [{}], [{}] eval save model ...'.
format(epoch_bidx, eval_cnt[0]))
self.mt_eval(epoch, epoch_bidx, "IN")
eval_spend = time.time() - eval_start
for name, par in self.model.named_parameters():
if name.split('.')[0] != "domain_discriminator":
par.requires_grad = False
else:
par.requires_grad = True
for k in range(batch_count + batch_count_domain):
epoch_bidx = k + 1
batch_idx = shuffled_batch_idx[
k] if epoch >= wargs.epoch_shuffle_minibatch else k
if batch_idx < batch_count:
_, srcs, trgs, slens, srcs_m, trgs_m = self.train_data[
batch_idx]
self.model.zero_grad()
# (max_tlen_batch - 1, batch_size, out_size)
#forward compute DANN out of domain
domain_outputs = self.model(srcs,
trgs[:-1],
srcs_m,
trgs_m[:-1],
'OUT',
alpha=wargs.alpha,
adv=True)
#if len(decoder_outputs) == 2: (decoder_outputs, _checks) = decoder_outputs
#this_bnum = decoder_outputs.size(1)
#batch_loss, grad_output, batch_correct_num = memory_efficient(
# outputs, trgs[1:], trgs_m[1:], self.model.classifier)
#backward compute, now we have the grad
#batch_loss, batch_correct_num, batch_log_norm = self.model.classifier.snip_back_prop(
# decoder_outputs, trgs[1:], trgs_m[1:], wargs.snip_size)
#self.model.zero_grad()
#domain_outputs = self.model(srcs, None, srcs_m, None, 'OUT', alpha=alpha)
if wargs.cross_entropy is True:
domain_label = tc.zeros(len(domain_outputs))
domain_label = domain_label.long()
domain_label = domain_label.cuda()
domainv_label = Variable(domain_label, volatile=False)
domain_loss_a = loss_domain(tc.log(domain_outputs),
domainv_label)
domain_loss_a.backward()
#if wargs.max_entropy is True:
#try to max this entropy
# domain_loss_b = -wargs.alpha*tc.dot(tc.log(domain_outputs), domain_outputs)
# domain_loss_b.backward()
batch_src_words = srcs.data.ne(PAD).sum()
assert batch_src_words == slens.data.sum()
batch_trg_words = trgs[1:].data.ne(PAD).sum()
elif batch_idx >= batch_count:
#domain_loss_out.backward(retain_graph=True)
#DANN in domain compute
_, srcs_domain, trgs_domain, slens_domain, srcs_m_domain, trgs_m_domain = self.train_data_domain[
batch_idx - batch_count]
self.model.zero_grad()
domain_outputs = self.model(srcs_domain,
trgs_domain[:-1],
srcs_m_domain,
trgs_m_domain[:-1],
'IN',
alpha=wargs.alpha,
adv=True)
#if len(decoder_outputs) == 2: (decoder_outputs, _checks) = decoder_outputs
#this_bnum = decoder_outputs.size(1)
#batch_loss, batch_correct_num, batch_log_norm = self.model.classifier.snip_back_prop(
# decoder_outputs, trgs_domain[1:], trgs_m_domain[1:], wargs.snip_size)
#domain_outputs = self.model(srcs_domain, None, srcs_m_domain, None, 'IN', alpha=alpha)
if wargs.cross_entropy is True:
domain_label = tc.ones(len(domain_outputs))
domain_label = domain_label.long()
domain_label = domain_label.cuda()
domainv_label = Variable(domain_label, volatile=False)
domain_loss_a = loss_domain(tc.log(domain_outputs),
domainv_label)
domain_loss_a.backward(
retain_graph=True
if wargs.max_entropy is True else False)
#if wargs.max_entropy is True:
# domain_loss_b = -wargs.alpha*tc.dot(tc.log(domain_outputs), domain_outputs)
# domain_loss_b.backward()
batch_src_words = srcs_domain.data.ne(PAD).sum()
assert batch_src_words == slens_domain.data.sum()
batch_trg_words = trgs_domain[1:].data.ne(PAD).sum()
self.optim.step()
show_loss += batch_loss
#domain_loss += domain_loss_a.data.clone()[0]
#domain_loss += domain_loss_out
show_correct_num += batch_correct_num
epoch_loss += batch_loss
epoch_num_correct += batch_correct_num
show_src_words += batch_src_words
show_trg_words += batch_trg_words
epoch_trg_words += batch_trg_words
#batch_log_norm = tc.mean(tc.abs(batch_log_norm))
if epoch_bidx % wargs.display_freq == 0:
#print show_correct_num, show_loss, show_trg_words, show_loss/show_trg_words
ud = time.time() - show_start - sample_spend - eval_spend
wlog(
'Epo:{:>2}/{:>2} |[{:^5} {:^5} {:^5}k] |acc:{:5.2f}% |ppl:{:4.2f} '
'| |logZ|:{:.2f} '
'|stok/s:{:>4}/{:>2}={:>2} |ttok/s:{:>2} '
'|stok/sec:{:6.2f} |ttok/sec:{:6.2f} |elapsed:{:4.2f}/{:4.2f}m'
.format(epoch, wargs.max_epochs, epoch_bidx, batch_idx,
bidx / 1000,
(show_correct_num / show_trg_words) * 100,
math.exp(show_loss / show_trg_words), 0,
batch_src_words, this_bnum,
int(batch_src_words / this_bnum),
int(batch_trg_words / this_bnum),
show_src_words / ud, show_trg_words / ud, ud,
(time.time() - train_start) / 60.))
show_loss, show_src_words, show_trg_words, show_correct_num = 0, 0, 0, 0
sample_spend, eval_spend = 0, 0
show_start = time.time()
if epoch_bidx % wargs.sampling_freq == 0:
sample_start = time.time()
self.model.eval()
#self.model.classifier.eval()
tor = Translator(self.model, self.sv, self.tv)
if batch_idx < batch_count:
# (max_len_batch, batch_size)
sample_src_tensor = srcs.t()[:sample_size]
sample_trg_tensor = trgs.t()[:sample_size]
tor.trans_samples(sample_src_tensor, sample_trg_tensor,
"OUT")
elif batch_idx >= batch_count:
sample_src_tensor = srcs_domain.t()[:sample_size]
sample_trg_tensor = trgs_domain.t()[:sample_size]
tor.trans_samples(sample_src_tensor, sample_trg_tensor,
"IN")
wlog('')
sample_spend = time.time() - sample_start
self.model.train()
# Just watch the translation of some source sentences in training data
if wargs.if_fixed_sampling and bidx == batch_start_sample:
# randomly select sample_size sample from current batch
rand_rows = np.random.choice(this_bnum,
sample_size,
replace=False)
sample_src_tensor = tc.Tensor(sample_size,
srcs.size(0)).long()
sample_src_tensor.fill_(PAD)
sample_trg_tensor = tc.Tensor(sample_size,
trgs.size(0)).long()
sample_trg_tensor.fill_(PAD)
for id in xrange(sample_size):
sample_src_tensor[id, :] = srcs.t()[rand_rows[id], :]
sample_trg_tensor[id, :] = trgs.t()[rand_rows[id], :]
avg_epoch_loss = epoch_loss / epoch_trg_words
avg_epoch_acc = epoch_num_correct / epoch_trg_words
wlog('\nEnd epoch [{}]'.format(epoch))
wlog('Train accuracy {:4.2f}%'.format(avg_epoch_acc * 100))
wlog('Average loss {:4.2f}'.format(avg_epoch_loss))
wlog('Train perplexity: {0:4.2f}'.format(math.exp(avg_epoch_loss)))
#wlog('Epoch domain loss is {:4.2f}'.format(float(domain_loss)))
wlog('End epoch, batch [{}], [{}] eval save model ...'.format(
epoch_bidx, eval_cnt[0]))
mteval_bleu = self.mt_eval(epoch, epoch_bidx, "IN")
self.optim.update_learning_rate(mteval_bleu, epoch)
epoch_time_consume = time.time() - epoch_start
wlog('Consuming: {:4.2f}s'.format(epoch_time_consume))
wlog('Train finished, comsuming {:6.2f} hours'.format(
(time.time() - train_start) / 3600))
wlog('Congratulations!')
# "问题 交换 了 意见 。"
#t = "( beijing , syndicated news ) the sino - us relation that was heated momentarily " \
# "by the us president bush 's visit to china is cooling down rapidly . china " \
# "confirmed yesterday that it has called off its naval fleet visit to the us " \
# "ports this year and refused to confirm whether the country 's vice president " \
# "hu jintao will visit the united states as planned ."
s = '当 林肯 去 新奥尔良 时 , 我 听到 密西 西比 河 的 歌声 。'
t = "When Lincoln goes to New Orleans, I hear Mississippi river's singing sound"
#s = '新奥尔良 是 爵士 音乐 的 发源 地 。'
#s = '新奥尔良 以 其 美食 而 闻名 。'
# = '休斯顿 是 仅 次于 新奥尔良 和 纽约 的 美国 第三 大 港 。'
s = [[src_vocab.key2idx[x] if x in src_vocab.key2idx else UNK for x in s.split(' ')]]
t = [[trg_vocab.key2idx[x] if x in trg_vocab.key2idx else UNK for x in t.split(' ')]]
tor.trans_samples(s, t)
sys.exit(0)
input_file = '{}{}.{}'.format(wargs.val_tst_dir, args.input_file, wargs.val_src_suffix)
input_abspath = os.path.realpath(input_file)
wlog('Translating test file {} ... '.format(input_abspath))
ref_file = '{}{}.{}'.format(wargs.val_tst_dir, args.input_file, wargs.val_ref_suffix)
test_src_tlst, _ = wrap_tst_data(input_abspath, src_vocab, char=wargs.src_char)
test_input_data = Input(test_src_tlst, None, batch_size=wargs.test_batch_size, batch_sort=False,
gpu_ids=args.gpu_ids)
batch_tst_data = None
if os.path.exists(ref_file) and False:
wlog('With force decoding test file {} ... to get alignments'.format(input_file))
wlog('\t\tRef file {}'.format(ref_file))
from tools.inputs_handler import wrap_data
def train(self):
wlog('Start training ... ')
assert wargs.sample_size < wargs.batch_size, 'Batch size < sample count'
# [low, high)
batch_count = len(self.train_data)
batch_start_sample = tc.randperm(batch_count)[-1]
wlog('Randomly select {} samples in the {}th/{} batch'.format(
wargs.sample_size, batch_start_sample, batch_count))
bidx, eval_cnt = 0, [0]
wlog('Self-normalization alpha -> {}'.format(wargs.self_norm_alpha))
tor_hook = Translator(self.model, self.sv, self.tv)
train_start = time.time()
wlog('\n' + '#' * 120 + '\n' + '#' * 30 + ' Start Training ' +
'#' * 30 + '\n' + '#' * 120)
batch_oracles, _checks = None, None
for epoch in range(wargs.start_epoch, wargs.max_epochs + 1):
epoch_start = time.time()
# train for one epoch on the training data
wlog('\n' + '$' * 30, 0)
wlog(' Epoch [{}/{}] '.format(epoch, wargs.max_epochs) + '$' * 30)
if wargs.epoch_shuffle and epoch > wargs.epoch_shuffle_minibatch:
self.train_data.shuffle()
# shuffle the original batch
shuffled_batch_idx = tc.randperm(batch_count)
sample_size = wargs.sample_size
epoch_loss, epoch_trg_words, epoch_num_correct, \
epoch_batch_logZ, epoch_n_sents = 0, 0, 0, 0, 0
show_loss, show_src_words, show_trg_words, show_correct_num, \
show_batch_logZ, show_n_sents = 0, 0, 0, 0, 0, 0
sample_spend, eval_spend, epoch_bidx = 0, 0, 0
show_start = time.time()
for k in range(batch_count):
bidx += 1
epoch_bidx = k + 1
batch_idx = shuffled_batch_idx[
k] if epoch >= wargs.epoch_shuffle_minibatch else k
# (max_slen_batch, batch_size)
#_, srcs, ttrgs_for_files, slens, srcs_m, trg_mask_for_files = self.train_data[batch_idx]
_, srcs, spos, ttrgs_for_files, ttpos_for_files, slens, srcs_m, trg_mask_for_files = self.train_data[
batch_idx]
trgs, tpos, trgs_m = ttrgs_for_files[0], ttpos_for_files[
0], trg_mask_for_files[0]
#self.model.zero_grad()
self.optim.zero_grad()
# (max_tlen_batch - 1, batch_size, out_size)
if wargs.model == 8:
gold, gold_mask = trgs[1:], trgs_m[1:]
# (L, B) -> (B, L)
T_srcs, T_spos, T_trgs, T_tpos = srcs.t(), spos.t(
), trgs.t(), tpos.t()
src, trg = (T_srcs, T_spos), (T_trgs, T_tpos)
gold, gold_mask = gold.t(), gold_mask.t()
#gold, gold_mask = trg[0][:, 1:], trgs_m.t()[:, 1:]
if gold.is_contiguous() is False: gold = gold.contiguous()
if gold_mask.is_contiguous() is False:
gold_mask = gold_mask.contiguous()
outputs = self.model(src, trg)
else:
gold, gold_mask = trgs[1:], trgs_m[1:]
outputs = self.model(srcs, trgs[:-1], srcs_m, trgs_m[:-1])
if len(outputs) == 2: (outputs, _checks) = outputs
if len(outputs) == 2: (outputs, attends) = outputs
this_bnum = outputs.size(1)
epoch_n_sents += this_bnum
show_n_sents += this_bnum
#batch_loss, batch_correct_num, batch_log_norm = self.classifier(outputs, trgs[1:], trgs_m[1:])
#batch_loss.div(this_bnum).backward()
#batch_loss = batch_loss.data[0]
#batch_correct_num = batch_correct_num.data[0]
batch_loss, batch_correct_num, batch_Z = self.classifier.snip_back_prop(
outputs, gold, gold_mask, wargs.snip_size)
self.optim.step()
grad_checker(self.model, _checks)
batch_src_words = srcs.data.ne(PAD).sum()
assert batch_src_words == slens.data.sum()
batch_trg_words = trgs[1:].data.ne(PAD).sum()
show_loss += batch_loss
show_correct_num += batch_correct_num
epoch_loss += batch_loss
epoch_num_correct += batch_correct_num
show_src_words += batch_src_words
show_trg_words += batch_trg_words
epoch_trg_words += batch_trg_words
show_batch_logZ += batch_Z
epoch_batch_logZ += batch_Z
if epoch_bidx % wargs.display_freq == 0:
#print show_correct_num, show_loss, show_trg_words, show_loss/show_trg_words
ud = time.time() - show_start - sample_spend - eval_spend
wlog(
'Epo:{:>2}/{:>2} |[{:^5} {:^5} {:^5}k] |acc:{:5.2f}% |ppl:{:4.2f} '
'||w-logZ|:{:.2f} ||s-logZ|:{:.2f} '
'|stok/s:{:>4}/{:>2}={:>2} |ttok/s:{:>2} '
'|stok/sec:{:6.2f} |ttok/sec:{:6.2f} |elapsed:{:4.2f}/{:4.2f}m'
.format(epoch, wargs.max_epochs, epoch_bidx, batch_idx,
bidx / 1000,
(show_correct_num / show_trg_words) * 100,
math.exp(show_loss / show_trg_words),
show_batch_logZ / show_trg_words,
show_batch_logZ / show_n_sents,
batch_src_words, this_bnum,
int(batch_src_words / this_bnum),
int(batch_trg_words / this_bnum),
show_src_words / ud, show_trg_words / ud, ud,
(time.time() - train_start) / 60.))
show_loss, show_src_words, show_trg_words, show_correct_num, \
show_batch_logZ, show_n_sents = 0, 0, 0, 0, 0, 0
sample_spend, eval_spend = 0, 0
show_start = time.time()
if epoch_bidx % wargs.sampling_freq == 0:
sample_start = time.time()
self.model.eval()
# (max_len_batch, batch_size)
sample_src_tensor = srcs.t()[:sample_size]
sample_trg_tensor = trgs.t()[:sample_size]
sample_src_tensor_pos = T_spos[:
sample_size] if wargs.model == 8 else None
tor_hook.trans_samples(sample_src_tensor,
sample_trg_tensor,
sample_src_tensor_pos)
wlog('')
sample_spend = time.time() - sample_start
self.model.train()
# Just watch the translation of some source sentences in training data
if wargs.if_fixed_sampling and bidx == batch_start_sample:
# randomly select sample_size sample from current batch
rand_rows = np.random.choice(this_bnum,
sample_size,
replace=False)
sample_src_tensor = tc.Tensor(sample_size,
srcs.size(0)).long()
sample_src_tensor.fill_(PAD)
sample_trg_tensor = tc.Tensor(sample_size,
trgs.size(0)).long()
sample_trg_tensor.fill_(PAD)
for id in xrange(sample_size):
sample_src_tensor[id, :] = srcs.t()[rand_rows[id], :]
sample_trg_tensor[id, :] = trgs.t()[rand_rows[id], :]
if wargs.epoch_eval is not True and bidx > wargs.eval_valid_from and \
bidx % wargs.eval_valid_freq == 0:
eval_start = time.time()
eval_cnt[0] += 1
wlog('\nAmong epoch, batch [{}], [{}] eval save model ...'.
format(epoch_bidx, eval_cnt[0]))
self.mt_eval(epoch, epoch_bidx)
eval_spend = time.time() - eval_start
avg_epoch_loss = epoch_loss / epoch_trg_words
avg_epoch_acc = epoch_num_correct / epoch_trg_words
wlog('\nEnd epoch [{}]'.format(epoch))
wlog('Train accuracy {:4.2f}%'.format(avg_epoch_acc * 100))
wlog('Average loss {:4.2f}'.format(avg_epoch_loss))
wlog('Train perplexity: {0:4.2f}'.format(math.exp(avg_epoch_loss)))
wlog('Train average |w-logZ|: {}/{}={} |s-logZ|: {}/{}={}'.format(
epoch_batch_logZ, epoch_trg_words,
epoch_batch_logZ / epoch_trg_words, epoch_batch_logZ,
epoch_n_sents, epoch_batch_logZ / epoch_n_sents))
wlog('End epoch, batch [{}], [{}] eval save model ...'.format(
epoch_bidx, eval_cnt[0]))
mteval_bleu = self.mt_eval(epoch, epoch_bidx)
self.optim.update_learning_rate(mteval_bleu, epoch)
epoch_time_consume = time.time() - epoch_start
wlog('Consuming: {:4.2f}s'.format(epoch_time_consume))
wlog('Finish training, comsuming {:6.2f} hours'.format(
(time.time() - train_start) / 3600))
wlog('Congratulations!')
class Trainer(object):
def __init__(self,
model_par,
train_data,
vocab_data,
optim,
lossCompute,
model,
valid_data=None,
tests_data=None):
self.model_par, self.model, self.lossCompute, self.optim = model_par, model, lossCompute, optim
self.sv, self.tv = vocab_data['src'].idx2key, vocab_data['trg'].idx2key
self.train_data, self.valid_data, self.tests_data = train_data, valid_data, tests_data
self.max_epochs, self.start_epoch = wargs.max_epochs, wargs.start_epoch
self.n_look = wargs.n_look
assert self.n_look <= wargs.batch_size, 'eyeball count > batch size'
self.n_batches = len(train_data) # [low, high)
self.look_xs, self.look_ys = None, None
if wargs.fix_looking is True:
rand_idxs = random.sample(range(train_data.n_sent), self.n_look)
wlog(
'randomly look {} samples frow the whole training data'.format(
self.n_look))
self.look_xs = [train_data.x_list[i][0] for i in rand_idxs]
self.look_ys = [train_data.y_list_files[i][0] for i in rand_idxs]
self.tor = Translator(model, self.sv, self.tv, gpu_ids=wargs.gpu_ids)
self.n_eval = 1
self.grad_accum_count = wargs.grad_accum_count
self.epoch_shuffle_train = wargs.epoch_shuffle_train
self.epoch_shuffle_batch = wargs.epoch_shuffle_batch
self.ss_cur_prob = wargs.ss_prob_begin
if wargs.ss_type is not None:
wlog(
'word-level optimizing bias between training and decoding ...')
if wargs.bleu_sampling is True:
wlog('sentence-level optimizing ...')
wlog('schedule sampling value {}'.format(self.ss_cur_prob))
if self.ss_cur_prob < 1. and wargs.bleu_sampling is True:
self.sampler = Nbs(self.model,
self.tv,
k=3,
noise=wargs.bleu_gumbel_noise,
batch_sample=True)
if self.grad_accum_count > 1:
assert (
wargs.chunk_size == 0
), 'to accumulate grads, disable target sequence truncating'
def accum_matrics(self, batch_size, xtoks, ytoks, nll, ok_ytoks, bow_loss):
self.look_sents += batch_size
self.e_sents += batch_size
self.look_nll += nll
self.look_bow_loss += bow_loss
self.look_ok_ytoks += ok_ytoks
self.e_nll += nll
self.e_ok_ytoks += ok_ytoks
self.look_xtoks += xtoks
self.look_ytoks += ytoks
self.e_ytoks += ytoks
def grad_accumulate(self, real_batches, e_idx, n_upds):
#if self.grad_accum_count > 1:
# self.model_par.zero_grad()
for batch in real_batches:
# (batch_size, max_slen_batch)
_, xs, y_for_files, bows, x_lens, xs_mask, y_mask_for_files, bows_mask = batch
_batch_size = xs.size(0)
ys, ys_mask = y_for_files[0], y_mask_for_files[0]
#wlog('x: {}, x_mask: {}, y: {}, y_mask: {}'.format(
# xs.size(), xs_mask.size(), ys.size(), ys_mask.size()))
if bows is not None:
bows, bows_mask = bows[0], bows_mask[0]
#wlog('bows: {}, bows_mask: {})'.format(bows.size(), bows_mask.size()))
_xtoks = xs.data.ne(PAD).sum().item()
assert _xtoks == x_lens.data.sum().item()
_ytoks = ys[:, 1:].data.ne(PAD).sum().item()
#if self.grad_accum_count == 1: self.model_par.zero_grad()
# exclude last target word from inputs
results = self.model_par(xs, ys[:, :-1], xs_mask, ys_mask[:, :-1],
self.ss_cur_prob)
logits, alphas, contexts = results['logit'], results[
'attend'], results['context']
# (batch_size, y_Lm1, out_size)
gold, gold_mask = ys[:, 1:].contiguous(), ys_mask[:,
1:].contiguous()
# 3. Compute loss in shards for memory efficiency.
_nll, _ok_ytoks, _bow_loss = self.lossCompute(
logits, e_idx, n_upds, gold, gold_mask, None, bows, bows_mask,
contexts)
self.accum_matrics(_batch_size, _xtoks, _ytoks, _nll, _ok_ytoks,
_bow_loss)
# 3. Update the parameters and statistics.
self.optim.step()
self.optim.optimizer.zero_grad()
#tc.cuda.empty_cache()
def look_samples(self, n_steps):
if n_steps % wargs.look_freq == 0:
look_start = time.time()
self.model_par.eval() # affect the dropout !!!
self.model.eval()
if self.look_xs is not None and self.look_ys is not None:
_xs, _ys = self.look_xs, self.look_ys
else:
rand_idxs = random.sample(range(self.train_data.n_sent),
self.n_look)
wlog('randomly look {} samples frow the whole training data'.
format(self.n_look))
_xs = [self.train_data.x_list[i][0] for i in rand_idxs]
_ys = [self.train_data.y_list_files[i][0] for i in rand_idxs]
self.tor.trans_samples(_xs, _ys)
wlog('')
self.look_spend = time.time() - look_start
self.model_par.train()
self.model.train()
def try_valid(self, e_idx, e_bidx, n_steps):
if n_steps > 150000: wargs.eval_valid_freq = 1000
if wargs.epoch_eval is not True and n_steps > wargs.eval_valid_from and \
n_steps % wargs.eval_valid_freq == 0:
eval_start = time.time()
wlog('\nAmong epoch, e_batch:{}, n_steps:{}, {}-th validation ...'.
format(e_bidx, n_steps, self.n_eval))
self.mt_eval(e_idx, e_bidx, n_steps)
self.eval_spend = time.time() - eval_start
def mt_eval(self, e_idx, e_bidx, n_steps):
state_dict = {
'model': self.model.state_dict(),
'epoch': e_idx,
'batch': e_bidx,
'steps': n_steps,
'optim': self.optim
}
if wargs.save_one_model:
model_file = '{}.pt'.format(wargs.model_prefix)
else:
model_file = '{}_e{}_upd{}.pt'.format(wargs.model_prefix, e_idx,
n_steps)
tc.save(state_dict, model_file)
wlog('Saving temporary model in {}'.format(model_file))
self.model_par.eval()
self.model.eval()
self.tor.trans_eval(self.valid_data, e_idx, e_bidx, n_steps,
model_file, self.tests_data)
self.model_par.train()
self.model.train()
self.n_eval += 1
def train(self):
wlog('start training ... ')
train_start = time.time()
wlog('\n' + '#' * 120 + '\n' + '#' * 30 + ' Start Training ' +
'#' * 30 + '\n' + '#' * 120)
batch_oracles, _checks, accum_batches, real_batches = None, None, 0, []
current_steps = self.optim.n_current_steps
self.model_par.train()
self.model.train()
show_start = time.time()
self.look_nll, self.look_ytoks, self.look_ok_ytoks, self.look_sents, self.look_bow_loss = 0, 0, 0, 0, 0
for e_idx in range(self.start_epoch, self.max_epochs + 1):
wlog('\n{} Epoch [{}/{}] {}'.format('$' * 30, e_idx,
self.max_epochs, '$' * 30))
if wargs.bow_loss is True:
wlog('bow: {}'.format(schedule_bow_lambda(e_idx, 5, 0.5, 0.2)))
# shuffle the training data for each epoch
if self.epoch_shuffle_train: self.train_data.shuffle()
self.e_nll, self.e_ytoks, self.e_ok_ytoks, self.e_sents = 0, 0, 0, 0
self.look_xtoks, self.look_spend, b_counter, self.eval_spend = 0, 0, 0, 0
epo_start = time.time()
if self.epoch_shuffle_batch:
shuffled_bidx = tc.randperm(self.n_batches)
#for bidx in range(self.n_batches):
bidx = 0
cond = True if wargs.lr_update_way != 'invsqrt' else self.optim.learning_rate > wargs.min_lr
while cond:
if self.train_data.eos() is True: break
if current_steps >= wargs.max_update:
wlog('Touch the max update {}'.format(wargs.max_update))
sys.exit(0)
b_counter += 1
e_bidx = shuffled_bidx[
bidx] if self.epoch_shuffle_batch else bidx
if wargs.ss_type is not None and self.ss_cur_prob < 1. and wargs.bleu_sampling:
batch_beam_trgs = self.sampler.beam_search_trans(
xs, xs_mask, ys_mask)
batch_beam_trgs = [
list(zip(*b)[0]) for b in batch_beam_trgs
]
#wlog(batch_beam_trgs)
batch_oracles = batch_search_oracle(
batch_beam_trgs, ys[1:], ys_mask[1:])
#wlog(batch_oracles)
batch_oracles = batch_oracles[:-1].cuda()
batch_oracles = self.model.decoder.trg_lookup_table(
batch_oracles)
batch = self.train_data[e_bidx]
real_batches.append(batch)
accum_batches += 1
if accum_batches == self.grad_accum_count:
self.grad_accumulate(real_batches, e_idx, current_steps)
current_steps = self.optim.n_current_steps
del real_batches
accum_batches, real_batches = 0, []
tc.cuda.empty_cache()
#grad_checker(self.model, _checks)
if current_steps % wargs.display_freq == 0:
#wlog('look_ok_ytoks:{}, look_nll:{}, look_ytoks:{}'.format(self.look_ok_ytoks, self.look_nll, self.look_ytoks))
ud = time.time(
) - show_start - self.look_spend - self.eval_spend
wlog(
'Epo:{:>2}/{:>2} |[{:^5}/{} {:^5}] |acc:{:5.2f}% |{:4.2f}/{:4.2f}=nll:{:4.2f} |bow:{:4.2f}'
' |w-ppl:{:4.2f} |x(y)/s:{:>4}({:>4})/{}={}({}) |x(y)/sec:{}({}) |lr:{:7.6f}'
' |{:4.2f}s/{:4.2f}m'.format(
e_idx, self.max_epochs, b_counter,
len(self.train_data), current_steps,
(self.look_ok_ytoks / self.look_ytoks) * 100,
self.look_nll, self.look_ytoks,
self.look_nll / self.look_ytoks,
self.look_bow_loss / self.look_ytoks,
math.exp(self.look_nll / self.look_ytoks),
self.look_xtoks, self.look_ytoks,
self.look_sents,
int(round(self.look_xtoks / self.look_sents)),
int(round(self.look_ytoks / self.look_sents)),
int(round(self.look_xtoks / ud)),
int(round(self.look_ytoks / ud)),
self.optim.learning_rate, ud,
(time.time() - train_start) / 60.))
self.look_nll, self.look_xtoks, self.look_ytoks, self.look_ok_ytoks, self.look_sents, self.look_bow_loss = 0, 0, 0, 0, 0, 0
self.look_spend, self.eval_spend = 0, 0
show_start = time.time()
self.look_samples(current_steps)
self.try_valid(e_idx, e_bidx, current_steps)
bidx += 1
avg_epo_acc, avg_epo_nll = self.e_ok_ytoks / self.e_ytoks, self.e_nll / self.e_ytoks
wlog('\nEnd epoch [{}]'.format(e_idx))
wlog('avg. w-acc: {:4.2f}%, w-nll: {:4.2f}, w-ppl: {:4.2f}'.format(
avg_epo_acc * 100, avg_epo_nll, math.exp(avg_epo_nll)))
if wargs.epoch_eval is True:
wlog(
'\nEnd epoch, e_batch:{}, n_steps:{}, {}-th validation ...'
.format(e_bidx, n_steps, self.n_eval))
self.mt_eval(e_idx, e_bidx, self.optim.n_current_steps)
# decay the probability value epslion of scheduled sampling per batch
if wargs.ss_type is not None:
self.ss_cur_prob = ss_prob_decay(e_idx) # start from 1.
epo_time_consume = time.time() - epo_start
wlog('Consuming: {:4.2f}s'.format(epo_time_consume))
wlog('Finish training, comsuming {:6.2f} hours'.format(
(time.time() - train_start) / 3600))
wlog('Congratulations!')
def train(self):
wlog('Start training ... ')
assert wargs.sample_size < wargs.batch_size, 'Batch size < sample count'
# [low, high)
batch_count = len(self.train_data)
batch_start_sample = tc.randperm(batch_count)[0]
wlog('Randomly select {} samples in the {}th/{} batch'.format(wargs.sample_size, batch_start_sample, batch_count))
bidx, eval_cnt, ss_eps_cur = 0, [0], wargs.ss_eps_begin
wlog('Self-normalization alpha -> {}'.format(wargs.self_norm_alpha))
train_start = time.time()
wlog('')
wlog('#' * 120)
wlog('#' * 30, False)
wlog(' Start Training ', False)
wlog('#' * 30)
wlog('#' * 120)
for epoch in range(wargs.start_epoch, wargs.max_epochs + 1):
epoch_start = time.time()
# train for one epoch on the training data
wlog('')
wlog('$' * 30, False)
wlog(' Epoch [{}/{}] '.format(epoch, wargs.max_epochs), False)
wlog('$' * 30)
wlog('Schedule sampling value {}'.format(ss_eps_cur))
if wargs.epoch_shuffle and epoch > wargs.epoch_shuffle_minibatch: self.train_data.shuffle()
# shuffle the original batch
shuffled_batch_idx = tc.randperm(batch_count)
sample_size = wargs.sample_size
epoch_loss, epoch_trg_words, epoch_num_correct = 0, 0, 0
show_loss, show_src_words, show_trg_words, show_correct_num = 0, 0, 0, 0
sample_spend, eval_spend, epoch_bidx = 0, 0, 0
show_start = time.time()
for k in range(batch_count):
bidx += 1
epoch_bidx = k + 1
batch_idx = shuffled_batch_idx[k] if epoch >= wargs.epoch_shuffle_minibatch else k
# (max_slen_batch, batch_size)
_, srcs, trgs, slens, srcs_m, trgs_m = self.train_data[batch_idx]
self.model.zero_grad()
# (max_tlen_batch - 1, batch_size, out_size)
outputs = self.model(srcs, trgs[:-1], srcs_m, trgs_m[:-1], ss_eps=ss_eps_cur)
if len(outputs) == 2: (outputs, _checks) = outputs
this_bnum = outputs.size(1)
#batch_loss, grad_output, batch_correct_num = memory_efficient(
# outputs, trgs[1:], trgs_m[1:], self.model.classifier)
batch_loss, batch_correct_num, batch_log_norm = self.model.decoder.classifier.snip_back_prop(
outputs, trgs[1:], trgs_m[1:], wargs.snip_size)
_grad_nan = False
for n, p in self.model.named_parameters():
if p.grad is None:
debug('grad None | {}'.format(n))
continue
tmp_grad = p.grad.data.cpu().numpy()
if numpy.isnan(tmp_grad).any(): # we check gradient here for vanishing Gradient
wlog("grad contains 'nan' | {}".format(n))
#wlog("gradient\n{}".format(tmp_grad))
_grad_nan = True
if n == 'decoder.l_f1_0.weight' or n == 's_init.weight' or n=='decoder.l_f1_1.weight' \
or n == 'decoder.l_conv.0.weight' or n == 'decoder.l_f2.weight':
debug('grad zeros |{:5} {}'.format(str(not np.any(tmp_grad)), n))
if _grad_nan is True and wargs.dynamic_cyk_decoding is True:
for _i, items in enumerate(_checks):
wlog('step {} Variable----------------:'.format(_i))
#for item in items: wlog(item.cpu().data.numpy())
wlog('wen _check_tanh_sa ---------------')
wlog(items[0].cpu().data.numpy())
wlog('wen _check_a1_weight ---------------')
wlog(items[1].cpu().data.numpy())
wlog('wen _check_a1 ---------------')
wlog(items[2].cpu().data.numpy())
wlog('wen alpha_ij---------------')
wlog(items[3].cpu().data.numpy())
wlog('wen before_mask---------------')
wlog(items[4].cpu().data.numpy())
wlog('wen after_mask---------------')
wlog(items[5].cpu().data.numpy())
#outputs.backward(grad_output)
self.optim.step()
#del outputs, grad_output
batch_src_words = srcs.data.ne(PAD).sum()
assert batch_src_words == slens.data.sum()
batch_trg_words = trgs[1:].data.ne(PAD).sum()
show_loss += batch_loss
show_correct_num += batch_correct_num
epoch_loss += batch_loss
epoch_num_correct += batch_correct_num
show_src_words += batch_src_words
show_trg_words += batch_trg_words
epoch_trg_words += batch_trg_words
batch_log_norm = tc.mean(tc.abs(batch_log_norm))
if epoch_bidx % wargs.display_freq == 0:
#print show_correct_num, show_loss, show_trg_words, show_loss/show_trg_words
ud = time.time() - show_start - sample_spend - eval_spend
wlog(
'Epo:{:>2}/{:>2} |[{:^5} {:^5} {:^5}k] |acc:{:5.2f}% |ppl:{:4.2f} '
'| |logZ|:{:.2f} '
'|stok/s:{:>4}/{:>2}={:>2} |ttok/s:{:>2} '
'|stok/sec:{:6.2f} |ttok/sec:{:6.2f} |elapsed:{:4.2f}/{:4.2f}m'.format(
epoch, wargs.max_epochs, epoch_bidx, batch_idx, bidx/1000,
(show_correct_num / show_trg_words) * 100,
math.exp(show_loss / show_trg_words), batch_log_norm,
batch_src_words, this_bnum, int(batch_src_words / this_bnum),
int(batch_trg_words / this_bnum),
show_src_words / ud, show_trg_words / ud, ud,
(time.time() - train_start) / 60.)
)
show_loss, show_src_words, show_trg_words, show_correct_num = 0, 0, 0, 0
sample_spend, eval_spend = 0, 0
show_start = time.time()
if epoch_bidx % wargs.sampling_freq == 0:
sample_start = time.time()
self.model.eval()
#self.model.classifier.eval()
tor = Translator(self.model, self.sv, self.tv)
# (max_len_batch, batch_size)
sample_src_tensor = srcs.t()[:sample_size]
sample_trg_tensor = trgs.t()[:sample_size]
tor.trans_samples(sample_src_tensor, sample_trg_tensor)
wlog('')
sample_spend = time.time() - sample_start
self.model.train()
# Just watch the translation of some source sentences in training data
if wargs.if_fixed_sampling and bidx == batch_start_sample:
# randomly select sample_size sample from current batch
rand_rows = np.random.choice(this_bnum, sample_size, replace=False)
sample_src_tensor = tc.Tensor(sample_size, srcs.size(0)).long()
sample_src_tensor.fill_(PAD)
sample_trg_tensor = tc.Tensor(sample_size, trgs.size(0)).long()
sample_trg_tensor.fill_(PAD)
for id in xrange(sample_size):
sample_src_tensor[id, :] = srcs.t()[rand_rows[id], :]
sample_trg_tensor[id, :] = trgs.t()[rand_rows[id], :]
if wargs.epoch_eval is not True and bidx > wargs.eval_valid_from and \
bidx % wargs.eval_valid_freq == 0:
eval_start = time.time()
eval_cnt[0] += 1
wlog('\nAmong epoch, batch [{}], [{}] eval save model ...'.format(
epoch_bidx, eval_cnt[0]))
self.mt_eval(epoch, epoch_bidx)
eval_spend = time.time() - eval_start
avg_epoch_loss = epoch_loss / epoch_trg_words
avg_epoch_acc = epoch_num_correct / epoch_trg_words
wlog('\nEnd epoch [{}]'.format(epoch))
wlog('Train accuracy {:4.2f}%'.format(avg_epoch_acc * 100))
wlog('Average loss {:4.2f}'.format(avg_epoch_loss))
wlog('Train perplexity: {0:4.2f}'.format(math.exp(avg_epoch_loss)))
wlog('End epoch, batch [{}], [{}] eval save model ...'.format(epoch_bidx, eval_cnt[0]))
mteval_bleu = self.mt_eval(epoch, epoch_bidx)
self.optim.update_learning_rate(mteval_bleu, epoch)
# decay the probability value epslion of scheduled sampling per batch
ss_eps_cur = schedule_sample_eps_decay(epoch) # start from 1
epoch_time_consume = time.time() - epoch_start
wlog('Consuming: {:4.2f}s'.format(epoch_time_consume))
wlog('Finish training, comsuming {:6.2f} hours'.format((time.time() - train_start) / 3600))
wlog('Congratulations!')