def main(params):
# initialize the multi-GPU / multi-node training
init_distributed_mode(params)
# initialize the experiment
logger = initialize_exp(params)
# initialize SLURM signal handler for time limit / pre-emption
init_signal_handler()
# load data
data = load_data(params)
# build model
if params.encoder_only:
model = build_model(params, data['dico'])
else:
encoder, decoder = build_model(params, data['dico'])
# float16
if params.fp16:
assert torch.backends.cudnn.enabled
if params.encoder_only:
model = network_to_half(model)
else:
encoder = network_to_half(encoder)
decoder = network_to_half(decoder)
# distributed
if params.multi_gpu:
logger.info("Using nn.parallel.DistributedDataParallel ...")
if params.fp16:
if params.encoder_only:
model = apex.parallel.DistributedDataParallel(
model, delay_allreduce=True)
else:
encoder = apex.parallel.DistributedDataParallel(
encoder, delay_allreduce=True)
decoder = apex.parallel.DistributedDataParallel(
decoder, delay_allreduce=True)
else:
if params.encoder_only:
model = nn.parallel.DistributedDataParallel(
model,
device_ids=[params.local_rank],
output_device=params.local_rank,
broadcast_buffers=True)
else:
encoder = nn.parallel.DistributedDataParallel(
encoder,
device_ids=[params.local_rank],
output_device=params.local_rank,
broadcast_buffers=True)
decoder = nn.parallel.DistributedDataParallel(
decoder,
device_ids=[params.local_rank],
output_device=params.local_rank,
broadcast_buffers=True)
# build trainer, reload potential checkpoints / build evaluator
if params.encoder_only:
trainer = SingleTrainer(model, data, params)
evaluator = SingleEvaluator(trainer, data, params)
else:
trainer = EncDecTrainer(encoder, decoder, data, params)
evaluator = EncDecEvaluator(trainer, data, params)
# evaluation
if params.eval_only:
scores = evaluator.run_all_evals(trainer)
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
logger.info("__log__:%s" % json.dumps(scores))
exit()
# set sampling probabilities for training
set_sampling_probs(data, params)
# language model training
for _ in range(params.max_epoch):
logger.info("============ Starting epoch %i ... ============" %
trainer.epoch)
trainer.n_sentences = 0
while trainer.n_sentences < trainer.epoch_size:
# CLM steps
for lang1, lang2 in shuf_order(params.clm_steps, params):
trainer.clm_step(lang1, lang2, params.lambda_clm)
# MLM steps (also includes TLM if lang2 is not None)
for lang1, lang2 in shuf_order(params.mlm_steps, params):
trainer.mlm_step(lang1, lang2, params.lambda_mlm)
# parallel classification steps
for lang1, lang2 in shuf_order(params.pc_steps, params):
trainer.pc_step(lang1, lang2, params.lambda_pc)
# denoising auto-encoder steps
for lang in shuf_order(params.ae_steps):
trainer.mt_step(lang, lang, params.lambda_ae)
# mass prediction steps
for lang in shuf_order(params.ms_steps):
p = np.random.random()
trainer.ms_step(lang, params.lambda_ms)
# machine translation steps
for lang1, lang2 in shuf_order(params.mt_steps, params):
trainer.mt_step(lang1, lang2, params.lambda_mt)
# back-translation steps
for lang1, lang2, lang3 in shuf_order(params.bt_steps):
trainer.bt_step(lang1, lang2, lang3, params.lambda_bt)
trainer.iter()
logger.info("============ End of epoch %i ============" %
trainer.epoch)
# evaluate perplexity
scores = evaluator.run_all_evals(trainer)
# print / JSON log
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
if params.is_master:
logger.info("__log__:%s" % json.dumps(scores))
# end of epoch
trainer.save_best_model(scores)
trainer.save_periodic()
trainer.end_epoch(scores)
def main(params):
# initialize the multi-GPU / multi-node training
init_distributed_mode(params)
# initialize the experiment
logger = initialize_exp(params)
# initialize SLURM signal handler for time limit / pre-emption
init_signal_handler()
# load data
data = load_data(params)
# build model
if params.encoder_only:
model = build_model(params, data['dico'])
else:
encoder, decoder = build_model(params, data['dico'])
# build trainer, reload potential checkpoints / build evaluator
if params.encoder_only:
trainer = SingleTrainer(model, data, params)
evaluator = SingleEvaluator(trainer, data, params)
else:
trainer = EncDecTrainer(encoder, decoder, data, params)
evaluator = EncDecEvaluator(trainer, data, params)
# evaluation
if params.eval_only:
scores = evaluator.run_all_evals(trainer)
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
logger.info("__log__:%s" % json.dumps(scores))
exit()
# print(params)
# input()
# set sampling probabilities for training
set_sampling_probs(data, params)
# language model training
for _ in range(params.max_epoch):
logger.info("============ Starting epoch %i ... ============" %
trainer.epoch)
trainer.n_sentences = 0
while trainer.n_sentences < trainer.epoch_size:
# CLM steps
for lang1, lang2 in shuf_order(params.clm_steps, params):
trainer.clm_step(lang1, lang2, params.lambda_clm)
# MLM steps (also includes TLM if lang2 is not None)
for lang1, lang2 in shuf_order(params.mlm_steps, params):
trainer.mlm_step(lang1, lang2, params.lambda_mlm)
# parallel classification steps
for lang1, lang2 in shuf_order(params.pc_steps, params):
trainer.pc_step(lang1, lang2, params.lambda_pc)
# denoising auto-encoder steps
for lang in shuf_order(params.ae_steps):
trainer.mt_step(lang, lang, params.lambda_ae)
# machine translation steps
for lang1, lang2 in shuf_order(params.mt_steps, params):
trainer.mt_step(lang1, lang2, params.lambda_mt)
# back-translation steps
for lang1, lang2, lang3 in shuf_order(params.bt_steps):
trainer.bt_step(lang1, lang2, lang3, params.lambda_bt)
trainer.iter()
logger.info("============ End of epoch %i ============" %
trainer.epoch)
# evaluate perplexity
scores = evaluator.run_all_evals(trainer)
# print / JSON log
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
if params.is_master:
logger.info("__log__:%s" % json.dumps(scores))
# end of epoch
trainer.save_best_model(scores)
trainer.save_periodic()
trainer.end_epoch(scores)
def main(params):
# initialize the multi-GPU / multi-node training
init_distributed_mode(params)
# initialize the experiment
logger = initialize_exp(params)
# initialize SLURM signal handler for time limit / pre-emption
init_signal_handler()
# load data
data = load_data(params)
# build model
# reload-model options are in here
if params.encoder_only:
model = build_model(params, data['dico'])
if params.use_adapters:
logger.info("Using adapters")
for param in model.named_parameters():
if param[0][:8] != "adapters":
param[1].requires_grad = False
for param_name, param in model.embeddings.named_parameters():
param.requires_grad = True
for param_name, param in model.position_embeddings.named_parameters(
):
param.requires_grad = True
for param_name, param in model.pred_layer.named_parameters():
param.requires_grad = True
for param in model.layer_norm_emb.parameters():
param.requires_grad = True
for param in model.named_parameters():
logger.info(param[0] + ' required grad = ' +
str(param[1].requires_grad))
else:
encoder, decoder = build_model(params, data['dico'])
# build trainer, reload potential checkpoints / build evaluator
if params.encoder_only:
trainer = SingleTrainer(model, data, params)
evaluator = SingleEvaluator(trainer, data, params)
logger.info("Number of trainable parameters (encoder): %i" % sum(
[p.numel()
for p in trainer.model.parameters() if p.requires_grad]))
else:
trainer = EncDecTrainer(encoder, decoder, data, params)
evaluator = EncDecEvaluator(trainer, data, params)
logger.info(
"Number of trainable parameters (encoder): %i" %
sum([p.numel() for p in encoder.parameters() if p.requires_grad]))
logger.info(
"Number of trainable parameters (decoder): %i" %
sum([p.numel() for p in decoder.parameters() if p.requires_grad]))
# evaluation
if params.eval_only:
scores = evaluator.run_all_evals(trainer)
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
logger.info("__log__:%s" % json.dumps(scores))
exit()
# set sampling probabilities for training
set_sampling_probs(data, params)
# language model training
for epoch in range(params.max_epoch):
logger.info("============ Starting epoch %i ... ============" %
trainer.epoch)
trainer.n_sentences = 0
while trainer.n_sentences < trainer.epoch_size:
# CLM steps
for lang1, lang2 in shuf_order(params.clm_steps, params):
trainer.clm_step(lang1, lang2, params.lambda_clm)
# MLM steps (also includes TLM if lang2 is not None)
for lang1, lang2 in shuf_order(params.mlm_steps, params):
trainer.mlm_step(lang1, lang2, params.lambda_mlm)
# parallel classification steps
for lang1, lang2 in shuf_order(params.pc_steps, params):
trainer.pc_step(lang1, lang2, params.lambda_pc)
# denoising auto-encoder
for lang in shuf_order(params.ae_steps):
trainer.mt_step(lang, lang, params.lambda_ae)
for lang1, lang2 in shuf_order(params.mt_steps, params):
trainer.mt_step(lang1, lang2, params.lambda_mt)
# back-translation
for lang1, lang2, lang3 in shuf_order(params.bt_steps):
trainer.bt_step(lang1, lang2, lang3, params.lambda_bt)
trainer.iter()
logger.info("============ End of epoch %i ============" %
trainer.epoch)
# evaluate perplexity
scores = evaluator.run_all_evals(trainer)
# print / JSON log
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
if params.is_master:
logger.info("__log__:%s" % json.dumps(scores))
# end of epoch
trainer.save_best_model(scores)
trainer.save_periodic()
trainer.end_epoch(scores)
def main(params):
# initialize the multi-GPU / multi-node training
init_distributed_mode(params)
# initialize the experiment
logger = initialize_exp(params)
# # initialize SLURM signal handler for time limit / pre-emption
# init_signal_handler()
# load data
data = load_data(params)
# build model
if params.encoder_only:
model = build_model(params, data['dico'])
else:
encoder, decoder = build_model(params, data['dico'])
# # float16
# if params.fp16:
# assert torch.backends.cudnn.enabled
# if params.encoder_only:
# model = network_to_half(model)
# else:
# encoder = network_to_half(encoder)
# decoder = network_to_half(decoder)
# # distributed
# if params.multi_gpu:
# logger.info("Using nn.parallel.DistributedDataParallel ...")
# if params.fp16:
# if params.encoder_only:
# model = apex.parallel.DistributedDataParallel(model, delay_allreduce=True)
# else:
# encoder = apex.parallel.DistributedDataParallel(encoder, delay_allreduce=True)
# decoder = apex.parallel.DistributedDataParallel(decoder, delay_allreduce=True)
# else:
# if params.encoder_only:
# model = nn.parallel.DistributedDataParallel(model, device_ids=[params.local_rank], output_device=params.local_rank, broadcast_buffers=True)
# else:
# encoder = nn.parallel.DistributedDataParallel(encoder, device_ids=[params.local_rank], output_device=params.local_rank, broadcast_buffers=True)
# decoder = nn.parallel.DistributedDataParallel(decoder, device_ids=[params.local_rank], output_device=params.local_rank, broadcast_buffers=True)
# build trainer, reload potential checkpoints / build evaluator
if params.encoder_only:
trainer = SingleTrainer(model, data, params)
evaluator = SingleEvaluator(trainer, data, params)
else:
trainer = EncDecTrainer(encoder, decoder, data, params)
evaluator = EncDecEvaluator(trainer, data, params)
# # evaluation
# if params.eval_only:
# scores = evaluator.run_all_evals(trainer)
# for k, v in scores.items():
# logger.info("%s -> %.6f" % (k, v))
# logger.info("__log__:%s" % json.dumps(scores))
# exit()
# set sampling probabilities for training
set_sampling_probs(data, params)
# language model training
for _ in range(params.max_epoch):
logger.info("============ Starting epoch %i ... ============" % trainer.epoch)
trainer.n_sentences = 0
trainer.n_images = 0
while trainer.n_sentences < trainer.epoch_size or trainer.n_images < trainer.epoch_size:
# CLM steps
for lang1, lang2 in shuf_order(params.clm_steps, params):
trainer.clm_step(lang1, lang2, params.lambda_clm)
# MLM steps (also includes TLM if lang2 is not None)
# shuf_order's result could be: ['fr', 'fr'] or ['en', 'fr'] or ['fr', 'en'] or ['en', 'en']
for lang1, lang2 in shuf_order(params.mlm_steps, params):
trainer.mlm_step(lang1, lang2, params.lambda_mlm)
# parallel classification steps
for lang1, lang2 in shuf_order(params.pc_steps, params):
trainer.pc_step(lang1, lang2, params.lambda_pc)
# Image-language pretraining steps
trainer.ipm_step("coco36", params.lambda_ipm)
# CMLM steps steps
for m1, m2 in shuf_order(params.cmlm_steps, params):
trainer.cmlm_step(m1, m2, params.lambda_cmlm)
# denoising auto-encoder steps
for lang in shuf_order(params.ae_steps):
trainer.mt_step(lang, lang, params.lambda_ae)
# machine translation steps
for lang1, lang2 in shuf_order(params.mt_steps, params):
trainer.mt_step(lang1, lang2, params.lambda_mt)
# back-translation steps
for lang1, lang2, lang3 in shuf_order(params.bt_steps):
trainer.bt_step(lang1, lang2, lang3, params.lambda_bt)
trainer.iter()
logger.info("============ End of epoch %i ============" % trainer.epoch)
def main(params):
# initialize the multi-GPU / multi-node training
init_distributed_mode(params)
# initialize the experiment
meta_params = copy.deepcopy(params).meta_params
params.meta_params = "..." # to long to be log
logger = initialize_exp(params)
params.meta_params = meta_params
# initialize SLURM signal handler for time limit / pre-emption
init_signal_handler()
# load data
data = load_data(params)
# todo : good params.n_words (We take the one from the first task have this parameter for the moment.)
"""
But we think that if all the task data are based on the same vocabulary, all these parameters will be the same,
and therefore no problem if we choose one at random.
"""
p = params.meta_params[data['key']]
# build model
if params.encoder_only:
model = build_model(params=p, dico=data['dico'])
else:
encoder, decoder = build_model(params=p, dico=data['dico'])
# todo : good pad_index and eos_index and ... (I'll take the one from the first task for the moment.)
"""
But we think that if all the task data are based on the same vocabulary, all these parameters will be the same,
and therefore no problem if we choose one at random.
"""
params.pad_index = p.pad_index
params.eos_index = p.eos_index
# build trainer, reload potential checkpoints / build evaluator
if params.encoder_only:
trainer = SingleTrainer(model, data, params)
evaluator = SingleEvaluator(trainer, data, params)
else:
trainer = EncDecTrainer(encoder, decoder, data, params)
evaluator = EncDecEvaluator(trainer, data, params)
# evaluation
if params.eval_only:
scores = evaluator.run_all_evals(trainer)
if not params.meta_learning:
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
else:
for lgs in params.meta_params.keys():
logger.info("============ task : %s " % lgs)
for k, v in scores[lgs].items():
if k != "epoch":
logger.info("%s -> %.6f" % (k, v))
logger.info("============ all")
for k, v in scores.items():
if not (k in (list(params.meta_params.keys()) + ['epoch'])):
logger.info("%s -> %.6f" % (k, v))
logger.info("__log__:%s" % json.dumps(scores))
exit()
# set sampling probabilities for training
set_sampling_probs(data, params)
# language model training
for _ in range(params.max_epoch):
logger.info("============ Starting epoch %i ... ============" %
trainer.epoch)
if not params.meta_learning:
trainer.n_sentences = 0
while trainer.n_sentences < trainer.epoch_size:
# CLM steps
for lang1, lang2 in shuf_order(params.clm_steps, params):
trainer.clm_step(lang1, lang2, params.lambda_clm)
# MLM steps (also includes TLM if lang2 is not None)
for lang1, lang2 in shuf_order(params.mlm_steps, params):
trainer.mlm_step(lang1, lang2, params.lambda_mlm)
# parallel classification steps
for lang1, lang2 in shuf_order(params.pc_steps, params):
trainer.pc_step(lang1, lang2, params.lambda_pc)
# denoising auto-encoder steps
for lang in shuf_order(params.ae_steps):
trainer.mt_step(lang, lang, params.lambda_ae)
# machine translation steps
for lang1, lang2 in shuf_order(params.mt_steps, params):
trainer.mt_step(lang1, lang2, params.lambda_mt)
# back-translation steps
for lang1, lang2, lang3 in shuf_order(params.bt_steps):
trainer.bt_step(lang1, lang2, lang3, params.lambda_bt)
trainer.iter()
else:
# our
trainer.n_sentences = {}
"""
Here we build language lists for each of our meta-taks. Indeed, for two language lists l1 and l2,
the objective will be done with l1[i] and l2[i] respectively, this for each index i of the two lists.
"""
lang1_dic, lang2_dic, lang3_dic = {}, {}, {}
"""
In the case of meta-learning, we have a (meta-)data dictionary for each (meta-)task,
so the keys are the languages conserved by the task.
"""
data_keys_dic = {}
# equivalent to "for task in list of task" in the original algorithm, except here we prepare all the tasks beforehand.
for lgs in params.meta_params.keys():
trainer.n_sentences[lgs] = 0
# CLM
try:
lang1_dic['clm_step']
except KeyError:
lang1_dic['clm_step'], lang2_dic[
'clm_step'], data_keys_dic['clm_step'] = [], [], []
for lang1, lang2 in shuf_order(
params.meta_params[lgs].clm_steps, params):
lang1_dic['clm_step'].append(lang1)
lang2_dic['clm_step'].append(lang2)
data_keys_dic['clm_step'].append(lgs)
# MLM
try:
lang1_dic['mlm_step']
except KeyError:
lang1_dic['mlm_step'], lang2_dic[
'mlm_step'], data_keys_dic['mlm_step'] = [], [], []
for lang1, lang2 in shuf_order(
params.meta_params[lgs].mlm_steps, params):
lang1_dic['mlm_step'].append(lang1)
lang2_dic['mlm_step'].append(lang2)
data_keys_dic['mlm_step'].append(lgs)
# parallel classification
try:
lang1_dic['pc_step']
except KeyError:
lang1_dic['pc_step'], lang2_dic['pc_step'], data_keys_dic[
'pc_step'] = [], [], []
for lang1, lang2 in shuf_order(
params.meta_params[lgs].pc_steps, params):
lang1_dic['pc_step'].append(lang1)
lang2_dic['pc_step'].append(lang2)
data_keys_dic['pc_step'].append(lgs)
# denoising auto-encoder
try:
lang1_dic['ae_step']
except KeyError:
lang1_dic['ae_step'], data_keys_dic['ae_step'] = [], []
for lang1 in shuf_order(params.meta_params[lgs].ae_steps):
lang1_dic['ae_step'].append(lang1)
data_keys_dic['ae_step'].append(lgs)
# machine translation
try:
lang1_dic['mt_step']
except KeyError:
lang1_dic['mt_step'], lang2_dic['mt_step'], data_keys_dic[
'mt_step'] = [], [], []
for lang1, lang2 in shuf_order(
params.meta_params[lgs].mt_steps, params):
lang1_dic['mt_step'].append(lang1)
lang2_dic['mt_step'].append(lang2)
data_keys_dic['mt_step'].append(lgs)
# back-translation
try:
lang1_dic['bt_step']
except KeyError:
lang1_dic['bt_step'], lang2_dic['bt_step'], lang3_dic[
'bt_step'], data_keys_dic['bt_step'] = [], [], [], []
for lang1, lang2, lang3 in shuf_order(
params.meta_params[lgs].bt_steps):
lang1_dic['bt_step'].append(lang1)
lang2_dic['bt_step'].append(lang2)
lang3_dic['bt_step'].append(lang3)
data_keys_dic['bt_step'].append(lgs)
flag = True
# equivalent to "while not done do" in the original algorithm
while flag:
# CLM steps
#print("clm_step", flag)
a = trainer.clm_step(lang1_dic['clm_step'],
lang2_dic['clm_step'], params.lambda_clm,
data_keys_dic['clm_step'])
#print("mlm_step", flag)
# MLM steps (also includes TLM if lang2 is not None)
b = trainer.mlm_step(lang1_dic['mlm_step'],
lang2_dic['mlm_step'], params.lambda_mlm,
data_keys_dic['mlm_step'])
# parallel classification steps
c = trainer.pc_step(lang1_dic['pc_step'], lang2_dic['pc_step'],
params.lambda_pc, data_keys_dic['pc_step'])
if isinstance(trainer, EncDecTrainer):
# denoising auto-encoder steps
d = trainer.mt_step(lang1_dic['ae_step'],
lang1_dic['ae_step'], params.lambda_ae,
data_keys_dic['ae_step'])
# machine translation steps
e = trainer.mt_step(lang1_dic['mt_step'],
lang2_dic['mt_step'], params.lambda_mt,
data_keys_dic['mt_step'])
# back-translation steps
f = trainer.bt_step(lang1_dic['bt_step'],
lang2_dic['bt_step'],
lang3_dic['bt_step'], params.lambda_bt,
data_keys_dic['bt_step'])
# do things better
if (not a) and (not b) and (not c) and (not d) and (
not e) and (not f):
flag = False # End of epoch
else:
flag = True
else:
# do things better
if (not a) and (not b) and (not c):
flag = False # End of epoch
else:
flag = True
trainer.iter()
logger.info("============ End of epoch %i ============" %
trainer.epoch)
# evaluate perplexity
scores = evaluator.run_all_evals(trainer)
# print / JSON log
if not params.meta_learning:
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
else:
for lgs in params.meta_params.keys():
logger.info("============ task : %s " % lgs)
for k, v in scores[lgs].items():
if k != "epoch":
logger.info("%s -> %.6f" % (k, v))
logger.info("============ all")
for k, v in scores.items():
if not (k in (list(params.meta_params.keys()) + ['epoch'])):
logger.info("%s -> %.6f" % (k, v))
if params.is_master:
logger.info("__log__:%s" % json.dumps(scores))
# end of epoch
trainer.save_best_model(scores)
trainer.save_periodic()
trainer.end_epoch(scores)
# our
logger.info("============ garbage collector collecting %d ..." %
gc.collect())
def main(params):
# initialize the multi-GPU / multi-node training
init_distributed_mode(params)
# initialize the experiment
logger = initialize_exp(params)
# initialize SLURM signal handler for time limit / pre-emption
init_signal_handler()
if params.other_seed > -1:
# deterministic
torch.manual_seed(params.other_seed)
torch.cuda.manual_seed(params.other_seed)
np.random.seed(params.other_seed)
random.seed(params.other_seed)
if params.iter_seed == -1:
# non-deterministic
params.iter_seed = None
# load data
data = load_data(params)
writer = SummaryWriter(params.dump_path + "/" + params.exp_name + "_log")
# build model
if params.encoder_only:
model = build_model(params, data['dico'])
else:
encoder, decoder = build_model(params, data['dico'])
# build trainer, reload potential checkpoints / build evaluator
if params.encoder_only:
trainer = SingleTrainer(model, data, params)
evaluator = SingleEvaluator(trainer, data, params)
else:
trainer = EncDecTrainer(encoder, decoder, data, params)
evaluator = EncDecEvaluator(trainer, data, params)
# evaluation
if params.eval_only:
scores = evaluator.run_all_evals(trainer)
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
logger.info("__log__:%s" % json.dumps(scores))
sys.exit()
# set sampling probabilities for training
set_sampling_probs(data, params)
_iter = 0
# dump initial weights
if params.save_initial:
trainer.save_checkpoint('initial', include_optimizers=False)
# language model training
for _ in range(params.max_epoch):
logger.info("============ Starting epoch %i ... ============" %
trainer.epoch)
trainer.n_sentences = 0
while trainer.n_sentences < trainer.epoch_size:
# MLM steps (also includes TLM if lang2 is not None)
for lang1, lang2 in shuf_order(params.mlm_steps, params):
if params.only_vlm:
# with visual features
trainer.vlm_step(lang1, lang2, params.lambda_mlm, _iter)
else:
trainer.mlm_step(lang1, lang2, params.lambda_mlm, _iter)
# parallel classification steps
for lang1, lang2 in shuf_order(params.pc_steps, params):
trainer.pc_step(lang1, lang2, params.lambda_pc)
# denoising auto-encoder steps
for lang in shuf_order(params.ae_steps):
trainer.mt_step(lang, lang, params.lambda_ae)
# back-translation steps
for lang1, lang2, lang3 in shuf_order(params.bt_steps):
trainer.bt_step(lang1, lang2, lang3, params.lambda_bt)
# machine translation steps
for lang1, lang2 in shuf_order(params.mt_steps, params):
trainer.mt_step(lang1, lang2, params.lambda_mt)
for lang1, lang2 in shuf_order(params.mmt_steps, params):
trainer.mmt_step(lang1, lang2, params.lambda_mt)
trainer.iter()
_iter += 1
logger.info("============ End of epoch %i ============" %
trainer.epoch)
# evaluate perplexity
scores = evaluator.run_all_evals(trainer)
# print / JSON log
for k, v in scores.items():
writer.add_scalar(k, v, _iter)
logger.info("%s -> %.6f" % (k, v))
if params.is_master:
logger.info("__log__:%s" % json.dumps(scores))
# end of epoch
trainer.save_best_model(scores)
trainer.save_periodic()
trainer.end_epoch(scores)
def main(params):
# initialize the multi-GPU / multi-node training
init_distributed_mode(params)
# initialize the experiment
logger = initialize_exp(params)
# initialize SLURM signal handler for time limit / pre-emption
init_signal_handler()
# load data
data = load_data(params)
if params.build_nmt_domain_feature:
import torch
from src.curriculum import build_nmt_domain_feature
dataset = data['para'][('de', 'en')]['train']
batches, indices = dataset.get_iterator(
shuffle=False,
group_by_size=params.group_by_size,
n_sentences=-1,
)
features = build_nmt_domain_feature(data, params, batches, dataset)
result = {'indices': indices, 'domain_feature': features}
torch.save(result, params.build_output_path)
return
if params.build_nlm_domain_feature:
import torch
from src.curriculum import build_nlm_domain_feature
dataset = data['para'][('de', 'en')]['train']
batches, indices = dataset.get_iterator(
shuffle=False,
group_by_size=params.group_by_size,
n_sentences=-1,
)
# dataset = data['mono_stream']['en']['train']
# indices = dataset.get_iterator(
# shuffle=False
# )
features, domain_score, sents = build_nlm_domain_feature(
data, params, batches, dataset)
result = {
'indices': indices,
'domain_feature': features,
'domain_score': domain_score,
'sents': sents
}
if params.build_output_path.endswith('pth'):
build_output_path = params.build_output_path
else:
build_output_path = f'{params.build_output_path}/final.pth'
torch.save(result, build_output_path)
return
# build model
if params.encoder_only:
model = build_model(params, data['dico'])
elif params.dual_encoder:
encoder1, encoder2 = build_model(params, data['dico'])
else:
encoder, decoder = build_model(params, data['dico'])
# build trainer, reload potential checkpoints / build evaluator
if params.encoder_only:
trainer = SingleTrainer(model, data, params)
evaluator = SingleEvaluator(trainer, data, params)
elif params.domains:
if params.curriculum_learning:
trainer = CurriculumTrainer(encoder, decoder, data, params)
elif params.dual_encoder:
trainer = DualEncoderTrainer(encoder1, encoder2, data, params)
else:
trainer = MultiDomainTrainer(encoder, decoder, data, params)
if params.local_adapt:
evaluator = MetaMultiDomainEvaluator(trainer, data, params)
else:
if params.curriculum_learning:
evaluator = EncDecEvaluator(trainer, data, params)
elif params.dual_encoder:
evaluator = DualEncoderEvaluator(trainer, data, params)
else:
evaluator = MultiDomainEvaluator(trainer, data, params)
else:
trainer = EncDecTrainer(encoder, decoder, data, params)
evaluator = EncDecEvaluator(trainer, data, params)
# evaluation
if params.eval_only:
scores = evaluator.run_all_evals(trainer)
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
logger.info("__log__:%s" % json.dumps(scores))
exit()
# set sampling probabilities for training
set_sampling_probs(data, params)
# language model training
for _ in range(params.max_epoch):
logger.info("============ Starting epoch %i ... ============" %
trainer.epoch)
trainer.n_sentences = 0
while trainer.n_sentences < trainer.epoch_size:
# CLM steps
for lang1, lang2 in shuf_order(params.clm_steps, params):
trainer.clm_step(lang1, lang2, params.lambda_clm)
# MLM steps (also includes TLM if lang2 is not None)
for lang1, lang2 in shuf_order(params.mlm_steps, params):
trainer.mlm_step(lang1, lang2, params.lambda_mlm)
# parallel classification steps
for lang1, lang2 in shuf_order(params.pc_steps, params):
trainer.pc_step(lang1, lang2, params.lambda_pc)
# denoising auto-encoder steps
for lang in shuf_order(params.ae_steps):
trainer.mt_step(lang, lang, params.lambda_ae)
# machine translation steps
for lang1, lang2 in shuf_order(params.mt_steps, params):
trainer.mt_step(lang1, lang2, params.lambda_mt)
# back-translation steps
for lang1, lang2, lang3 in shuf_order(params.bt_steps):
trainer.bt_step(lang1, lang2, lang3, params.lambda_bt)
trainer.iter()
if not params.dual_encoder and params.domains and params.domain_ratio_update_freq > 0 and trainer.n_total_iter % params.domain_ratio_update_freq == 0 and not params.sampling_uniform:
evaluator.update_language_sampler_multidomain()
evaluator.update_dataset_ratio(trainer)
if not params.dual_encoder and params.domain_reset_freq > 0 and trainer.n_total_iter % params.domain_reset_freq == 0:
evaluator.reset_dataset_ratio(trainer)
logger.info("============ End of epoch %i ============" %
trainer.epoch)
# evaluate perplexity
scores = evaluator.run_all_evals(trainer)
# print / JSON log
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
if params.is_master:
logger.info("__log__:%s" % json.dumps(scores))
# end of epoch
trainer.save_best_model(scores)
trainer.save_periodic()
trainer.end_epoch(scores)