def perform_translation(input_file_path, translation_directory,
cloze_train_path, question_train_path,
fasttext_vectors_path, checkpoint_path):
params = get_params(
exp_name='translation',
dump_path=translation_directory,
cloze_train_path=cloze_train_path,
question_train_path=question_train_path,
cloze_test_path=input_file_path,
fasttext_vectors_path=fasttext_vectors_path,
checkpoint_path=checkpoint_path,
)
# check parameters
assert params.exp_name
check_all_data_params(params)
check_mt_model_params(params)
data = load_data(params, mono_only=True)
encoder, decoder, discriminator, lm = build_mt_model(params, data)
# initialize trainer / reload checkpoint / initialize evaluator
trainer = TrainerMT(encoder, decoder, discriminator, lm, data, params)
trainer.reload_checkpoint()
trainer.test_sharing() # check parameters sharing
evaluator = EvaluatorMT(trainer, data, params)
with torch.no_grad():
lang1, lang2 = 'cloze', 'question'
evaluator.encoder.eval()
evaluator.decoder.eval()
lang1_id = evaluator.params.lang2id[lang1]
lang2_id = evaluator.params.lang2id[lang2]
translations = []
dataset = evaluator.data['mono'][lang1]['test']
dataset.batch_size = params.batch_size
for i, (sent1, len1) in enumerate(
dataset.get_iterator(shuffle=False, group_by_size=False)()):
encoded = evaluator.encoder(sent1.cuda(), len1, lang1_id)
sent2_, len2_, _ = evaluator.decoder.generate(encoded, lang2_id)
lang1_text = convert_to_text(sent1, len1, evaluator.dico[lang1],
lang1_id, evaluator.params)
lang2_text = convert_to_text(sent2_, len2_, evaluator.dico[lang2],
lang2_id, evaluator.params)
translations += zip(lang1_text, lang2_text)
# export sentences to hypothesis file and restore BPE segmentation
out_name = os.path.join(translation_directory,
'output_translations.txt')
with open(out_name, 'w', encoding='utf-8') as f:
f.write('\n'.join(['\t'.join(st) for st in translations]) + '\n')
restore_segmentation(out_name)
return out_name
def load(params):
# check parameters
assert params.exp_name
check_all_data_params(params)
check_mt_model_params(params)
# initialize experiment / load data / build model
data = load_data(params)
encoder, decoder, discriminator, lm = build_mt_model(params, data)
# initialize trainer / reload checkpoint / initialize evaluator
trainer = TrainerMT(encoder, decoder, discriminator, lm, data, params)
trainer.reload_checkpoint()
trainer.test_sharing() # check parameters sharing
evaluator = EvaluatorMT(trainer, data, params)
return trainer, evaluator
def main(params):
# check parameters
assert params.exp_name
check_all_data_params(params)
check_mt_model_params(params)
# initialize experiment / load data / build model
logger = initialize_exp(params)
data = load_data(params)
encoder, decoder, discriminator, lm = build_mt_model(params, data)
# initialize trainer / reload checkpoint / initialize evaluator
trainer = TrainerMT(encoder, decoder, discriminator, lm, data, params)
trainer.reload_checkpoint()
trainer.test_sharing() # check parameters sharing
evaluator = EvaluatorMT(trainer, data, params)
# evaluation mode
if params.eval_only:
evaluator.run_all_evals(0)
exit()
# language model pretraining
if params.lm_before > 0:
logger.info("Pretraining language model for %i iterations ..." %
params.lm_before)
trainer.n_sentences = 0
for _ in range(params.lm_before):
for lang in params.langs:
trainer.lm_step(lang)
trainer.iter()
# define epoch size
if params.epoch_size == -1:
params.epoch_size = params.n_para
assert params.epoch_size > 0
# start training
for _ in range(trainer.epoch, params.max_epoch):
logger.info(
"====================== Starting epoch %i ... ======================"
% trainer.epoch)
trainer.n_sentences = 0
while trainer.n_sentences < params.epoch_size:
# discriminator training
for _ in range(params.n_dis):
trainer.discriminator_step()
# language model training
if params.lambda_lm > 0:
for _ in range(params.lm_after):
for lang in params.langs:
trainer.lm_step(lang)
# MT training (parallel data)
if params.lambda_xe_para > 0:
for lang1, lang2 in params.para_directions:
trainer.enc_dec_step(lang1, lang2, params.lambda_xe_para)
# MT training (back-parallel data)
if params.lambda_xe_back > 0:
for lang1, lang2 in params.back_directions:
trainer.enc_dec_step(lang1,
lang2,
params.lambda_xe_back,
back=True)
# autoencoder training (monolingual data)
if params.lambda_xe_mono > 0:
for lang in params.mono_directions:
trainer.enc_dec_step(lang, lang, params.lambda_xe_mono)
# AE - MT training (on the fly back-translation)
if params.lambda_xe_otfd > 0 or params.lambda_xe_otfa > 0:
# start on-the-fly batch generations
if not getattr(params, 'started_otf_batch_gen', False):
otf_iterator = trainer.otf_bt_gen_async()
params.started_otf_batch_gen = True
# update model parameters on subprocesses
if trainer.n_iter % params.otf_sync_params_every == 0:
trainer.otf_sync_params()
# get training batch from CPU
before_gen = time.time()
batches = next(otf_iterator)
trainer.gen_time += time.time() - before_gen
# training
for batch in batches:
lang1, lang2, lang3 = batch['lang1'], batch[
'lang2'], batch['lang3']
# 2-lang back-translation - autoencoding
if lang1 != lang2 == lang3:
trainer.otf_bt(batch, params.lambda_xe_otfa,
params.otf_backprop_temperature)
# 2-lang back-translation - parallel data
elif lang1 == lang3 != lang2:
trainer.otf_bt(batch, params.lambda_xe_otfd,
params.otf_backprop_temperature)
# 3-lang back-translation - parallel data
elif lang1 != lang2 and lang2 != lang3 and lang1 != lang3:
trainer.otf_bt(batch, params.lambda_xe_otfd,
params.otf_backprop_temperature)
trainer.iter()
# end of epoch
logger.info(
"====================== End of epoch %i ======================" %
trainer.epoch)
# evaluate discriminator / perplexity / BLEU
scores = evaluator.run_all_evals(trainer.epoch)
# print / JSON log
for k, v in scores.items():
logger.info('%s -> %.6f' % (k, v))
logger.info("__log__:%s" % json.dumps(scores))
# save best / save periodic / end epoch
trainer.save_best_model(scores)
trainer.save_periodic()
trainer.end_epoch(scores)
trainer.test_sharing()
# initialize trainer / reload checkpoint / initialize evaluator
encoder = nn.DataParallel(encoder,
device_ids=[0, 1, 2],
output_device=[0, 1, 2])
decoder = nn.DataParallel(decoder,
device_ids=[0, 1, 2],
output_device=[0, 1, 2])
if discriminator != None:
discriminator = nn.DataParallel(discriminator,
device_ids=[0, 1, 2],
output_device=[0, 1, 2])
if lm != None:
lm = nn.DataParallel(lm, device_ids=[0, 1, 2], output_device=[0, 1, 2])
trainer = TrainerMT(encoder, decoder, discriminator, lm, data, params)
trainer.reload_checkpoint()
trainer.test_sharing() # check parameters sharing
evaluator = EvaluatorMT(trainer, data, params)
# evaluation mode
if params.eval_only:
evaluator.run_all_evals(0)
exit()
# language model pretraining
if params.lm_before > 0:
logger.info("Pretraining language model for %i iterations ..." %
params.lm_before)
trainer.n_sentences = 0
for _ in range(params.lm_before):
for lang in params.langs:
def main(params):
# check parameters
assert params.exp_name
check_all_data_params(params)
check_mt_model_params(params)
# import pickle
# with open('params.pkl', 'wb') as f_in:
# pickle.dump(params, f_in)
# exit()
# initialize experiment / load data / build model
logger = initialize_exp(params)
data = load_data(params)
if params.model_path is None:
encoder, decoder, discriminator, lm = build_mt_model(
params, data, cuda=torch.cuda.is_available())
else:
model = torch.load(params.model_path)
encoder = model['enc']
decoder = model['dec']
discriminator = model['dis']
lm = model['lm']
# initialize trainer / reload checkpoint / initialize evaluator
trainer = TrainerMT(encoder, decoder, discriminator, lm, data, params)
trainer.reload_checkpoint()
trainer.test_sharing() # check parameters sharing
evaluator = EvaluatorMT(trainer, data, params)
# evaluation mode
if params.eval_only:
scores = evaluator.run_all_evals(0, params)
# print / JSON log
for k, v in scores.items():
logger.info('%s -> %.6f' % (k, v))
logger.info("__log__:%s" % json.dumps(scores))
exit()
# language model pretraining
if params.lm_before > 0:
logger.info("Pretraining language model for %i iterations ..." %
params.lm_before)
trainer.n_sentences = 0
for idx in range(params.lm_before):
for lang in params.mono_directions:
trainer.enc_dec_step(lang, lang, params.lambda_xe_mono)
# for lang in params.langs:
# trainer.lm_step(lang)
trainer.iter()
if (idx + 1) % 10000 == 0:
trainer.save_model('lmpre_%d' % idx)
# define epoch size
if params.epoch_size == -1:
params.epoch_size = params.n_para
assert params.epoch_size > 0
# start training
for epoch in range(trainer.epoch, params.max_epoch):
logger.info(
"====================== Starting epoch %i ... ======================"
% trainer.epoch)
trainer.n_sentences = 0
while trainer.n_sentences < params.epoch_size:
# discriminator training
for _ in range(params.n_dis):
trainer.discriminator_step()
# language model training
if params.lambda_lm > 0:
for _ in range(params.lm_after):
for lang in params.langs:
trainer.lm_step(lang)
# MT training (parallel data)
if params.lambda_xe_para > 0:
for lang1, lang2 in params.para_directions:
trainer.enc_dec_step(lang1, lang2, params.lambda_xe_para)
# MT training (back-parallel data)
if params.lambda_xe_back > 0:
for lang1, lang2 in params.back_directions:
trainer.enc_dec_step(lang1,
lang2,
params.lambda_xe_back,
back=True)
# autoencoder training (monolingual data)
if params.lambda_xe_mono > 0:
for lang in params.mono_directions:
trainer.enc_dec_step(lang, lang, params.lambda_xe_mono)
# AE - MT training (on the fly back-translation)
if params.lambda_xe_otfd > 0 or params.lambda_xe_otfa > 0:
# start on-the-fly batch generations
if not getattr(params, 'started_otf_batch_gen', False):
otf_iterator = trainer.otf_bt_gen_async()
params.started_otf_batch_gen = True
# update model parameters on subprocesses
if trainer.n_iter % params.otf_sync_params_every == 0:
trainer.otf_sync_params()
# get training batch from CPU
before_gen = time.time()
batches = next(otf_iterator)
trainer.gen_time += time.time() - before_gen
# training
for batch in batches:
lang1, lang2, lang3 = batch['lang1'], batch[
'lang2'], batch['lang3']
if params.reward_gamma_ap != 0 or params.reward_type_ar != 'None':
# 2-lang back-translation - autoencoding
if lang1 != lang2 == lang3:
trainer.otf_bt(batch, params.lambda_xe_otfa,
params.otf_backprop_temperature)
# 2-lang back-translation - parallel data
elif lang1 == lang3 != lang2:
if params.use_rl and epoch >= params.rl_start_epoch:
trainer.otf_bt_rl(
batch, params.lambda_xe_otfd,
params.otf_backprop_temperature,
params.reward_gamma_ap,
params.reward_gamma_ar,
params.reward_type_ar,
params.reward_thresh_ar,
params.reward_gamma_cv)
else:
trainer.otf_bt(batch, params.lambda_xe_otfd,
params.otf_backprop_temperature)
# 3-lang back-translation - parallel data
elif lang1 != lang2 and lang2 != lang3 and lang1 != lang3:
trainer.otf_bt(batch, params.lambda_xe_otfd,
params.otf_backprop_temperature)
else:
trainer.otf_bt(batch, params.lambda_xe_otfd,
params.otf_backprop_temperature)
trainer.iter()
# end of epoch
logger.info(
"====================== End of epoch %i ======================" %
trainer.epoch)
# evaluate discriminator / perplexity / BLEU
# scores=0
scores = evaluator.run_all_evals(trainer.epoch, params) #TODO
# print / JSON log
for k, v in scores.items():
logger.info('%s -> %.6f' % (k, v))
logger.info("__log__:%s" % json.dumps(scores))
# save best / save periodic / end epoch
# trainer.save_best_model(scores)
trainer.save_periodic()
trainer.end_epoch(scores)
trainer.test_sharing()
def main(params):
# check parameters
assert params.exp_name
check_all_data_params(params) # done
check_mt_model_params(params)
# initialize experiment / load data / build model
logger = initialize_exp(params)
data = load_data(params, mono_para=params.mono_para)
encoder, decoder_aux, decoder, discriminator, lm = build_mt_model_modified(
params, data)
#encoder, decoder1, decoder2, discriminator, lm = build_mt_model(params, data) # add new decoder
# initialize trainer / reload checkpoint / initialize evaluator
trainer = TrainerMT(encoder, decoder, decoder_aux, discriminator, lm, data,
params)
trainer.reload_checkpoint()
trainer.test_sharing() # check parameters sharing
evaluator = EvaluatorMT(trainer, data, params)
if params.eval_only:
evaluator.run_all_evals(0)
exit()
# language model pretraining
if params.lm_before > 0:
logger.info("Pretraining language model for %i iterations ..." %
params.lm_before)
trainer.n_sentences = 0
for _ in range(params.lm_before):
for lang in params.langs:
trainer.lm_step(lang)
trainer.iter()
# define epoch size
if params.epoch_size == -1:
params.epoch_size = params.n_para
assert params.epoch_size > 0
# start training
for _ in range(trainer.epoch, params.max_epoch):
logger.info(
"====================== Starting epoch %i ... ======================"
% trainer.epoch)
trainer.n_sentences = 0
while trainer.n_sentences < params.epoch_size:
# discriminator training : n_dis = 0 : we do not use dis
logger.info("num of sentences %i " % trainer.n_sentences)
for _ in range(params.n_dis):
trainer.discriminator_step()
# language model training : 0 : we do not train lm
if params.lambda_lm > 0:
for _ in range(params.lm_after):
for lang in params.langs:
trainer.lm_step(lang)
# autoencoder training (monolingual data)
if params.lambda_xe_mono > 0:
trainer.enc_dec_step(lang,
lang,
params.lambda_xe_mono,
mono=True)
# autoencoder training disfluent-fluent parallel monolingual dataset : Disfluent_mono -> Mono data
if params.lambda_xe_mono_para > 0:
for lang1, lang2 in params.mono_para_directions: # (en,en), (es,es)
if (lang1 in ['en']):
trainer.enc_dec_step(lang1,
lang2,
params.lambda_xe_mono_para,
mono_para=True)
# MT training (parallel data) : introduce new decoder here. Fisher Data
# E1->D2' and E2->D1'
if params.lambda_xe_para > 0:
for lang1, lang2 in params.para_directions: # (en,es), (es,en)
#logger.info(" para : %s %s" % (lang1, lang2))
#logger.info("training para")
trainer.enc_dec_step(lang1,
lang2,
params.lambda_xe_para,
para=True)
# MT training news-commentary data
# E1->D2'->D2 and E2->D1'->D1
if params.lambda_xe_para_aux > 0:
for lang1, lang2 in params.para_directions: # (en, es) , (es,en)
trainer.enc_dec_step(lang1,
lang2,
params.lambda_xe_para_aux,
aux=True)
# MT training (back-parallel data) : 0
if params.lambda_xe_back > 0:
for lang1, lang2 in params.back_directions:
trainer.enc_dec_step(lang1,
lang2,
params.lambda_xe_back,
back=True)
# AE - MT training (on the fly back-translation) : otfd in on (let's see what to do with it) set it t
if params.lambda_xe_otfd > 0 or params.lambda_xe_otfa > 0:
# start on-the-fly batch generations
if not getattr(params, 'started_otf_batch_gen', False):
otf_iterator = trainer.otf_bt_gen_async()
params.started_otf_batch_gen = True
# update model parameters on subprocesses
if trainer.n_iter % params.otf_sync_params_every == 0:
trainer.otf_sync_params()
# get training batch from CPU
before_gen = time.time()
batches = next(otf_iterator)
trainer.gen_time += time.time() - before_gen
# training
for batch in batches:
lang1, lang2, lang3 = batch['lang1'], batch[
'lang2'], batch['lang3']
# 2-lang back-translation - autoencoding
if lang1 != lang2 == lang3:
trainer.otf_bt(batch, params.lambda_xe_otfa,
params.otf_backprop_temperature)
# 2-lang back-translation - parallel data
elif lang1 == lang3 != lang2:
trainer.otf_bt(batch, params.lambda_xe_otfd,
params.otf_backprop_temperature)
# 3-lang back-translation - parallel data
elif lang1 != lang2 and lang2 != lang3 and lang1 != lang3:
trainer.otf_bt(batch, params.lambda_xe_otfd,
params.otf_backprop_temperature)
trainer.iter()
# end of epoch
logger.info(
"====================== End of epoch %i ======================" %
trainer.epoch)
# evaluate discriminator / perplexity / BLEU
scores = evaluator.run_all_evals(trainer.epoch)
# print / JSON log
for k, v in scores.items():
logger.info('%s -> %.6f' % (k, v))
logger.info("__log__:%s" % json.dumps(scores))
# save best / save periodic / end epoch
trainer.save_best_model(scores)
trainer.save_periodic()
trainer.end_epoch(scores)
trainer.test_sharing()