def build_trainer(opt, device_id, model, fields, optim, model_saver=None):
"""
Simplify `Trainer` creation based on user `opt`s*
Args:
opt (:obj:`Namespace`): user options (usually from argument parsing)
model (:obj:`onmt.models.NMTModel`): the model to train
fields (dict): dict of fields
optim (:obj:`onmt.utils.Optimizer`): optimizer used during training
model_saver(:obj:`onmt.models.ModelSaverBase`): the utility object
used to save the model
"""
# 传词表是干嘛的？
train_loss = build_loss_compute(model, fields["tgt"].vocab, opt)
valid_loss = build_loss_compute(model,
fields["tgt"].vocab,
opt,
train=False)
# 截断后向传播？默认是0
trunc_size = opt.truncated_decoder # Badly named...
# 默认是2, 可以并行运行generator
shard_size = opt.max_generator_batches
# 梯度的normalization方法: tokens
norm_method = opt.normalization
# 一次更新梯度，batch_size * accum_count, 默认accum_count = 1
grad_accum_count = opt.accum_count
# 这里gpu的数量指定为分布式进程的数量
n_gpu = opt.world_size
if device_id >= 0:
gpu_rank = opt.gpu_ranks[device_id]
else:
gpu_rank = 0
n_gpu = 0
# 啥玩意
gpu_verbose_level = opt.gpu_verbose_level
report_manager = build_report_manager(opt)
trainer = Trainer(model,
train_loss,
valid_loss,
optim,
trunc_size,
shard_size,
norm_method,
grad_accum_count,
n_gpu,
gpu_rank,
gpu_verbose_level,
report_manager,
model_saver=model_saver)
return trainer
def build_trainer(opt, device_id, model, fields,
optim, model_saver=None):
"""
Simplify `Trainer` creation based on user `opt`s*
Args:
opt (:obj:`Namespace`): user options (usually from argument parsing)
model (:obj:`onmt.models.NMTModel`): the model to train
fields (dict): dict of fields
optim (:obj:`onmt.utils.Optimizer`): optimizer used during training
model_saver(:obj:`onmt.models.ModelSaverBase`): the utility object
used to save the model
"""
train_loss = [build_loss_compute(
model, fields["tgt"].vocab, opt), nn.NLLLoss(reduction='sum')]
valid_loss = build_loss_compute(
model, fields["tgt"].vocab, opt, train=False)
trunc_size = opt.truncated_decoder # Badly named...
shard_size = opt.max_generator_batches
norm_method = opt.normalization
grad_accum_count = opt.accum_count
n_gpu = opt.world_size
if device_id >= 0:
gpu_rank = opt.gpu_ranks[device_id]
else:
gpu_rank = 0
n_gpu = 0
gpu_verbose_level = opt.gpu_verbose_level
report_manager = build_report_manager(opt)
trainer = Trainer(model, train_loss, valid_loss, optim, trunc_size,
shard_size, norm_method,
grad_accum_count, n_gpu, gpu_rank,
gpu_verbose_level, report_manager,
model_saver=model_saver)
return trainer