def train(args, model, tokenizer, query_cache, passage_cache):
""" Train the model """
#if args.local_rank in [-1, 0]:
tb_writer = None
if is_first_worker():
tb_writer = SummaryWriter(log_dir=args.log_dir)
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
real_batch_size = args.train_batch_size * args.gradient_accumulation_steps * (
torch.distributed.get_world_size() if args.local_rank != -1 else 1)
# layerwise optimization for lamb
optimizer_grouped_parameters = []
for layer_name in [
"roberta.embeddings", "score_out", "downsample1", "downsample2",
"downsample3"
]:
layer = getattr_recursive(model, layer_name)
if layer is not None:
optimizer_grouped_parameters.append({"params": layer.parameters()})
if getattr_recursive(model, "roberta.encoder.layer") is not None:
for layer in model.roberta.encoder.layer:
optimizer_grouped_parameters.append({"params": layer.parameters()})
if len(optimizer_grouped_parameters) == 0:
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0
},
]
if args.optimizer.lower() == "lamb":
optimizer = Lamb(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
elif args.optimizer.lower() == "adamw":
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
else:
raise Exception(
"optimizer {0} not recognized! Can only be lamb or adamW".format(
args.optimizer))
# Check if saved optimizer or scheduler states exist
if os.path.isfile(
os.path.join(args.model_name_or_path,
"optimizer.pt")) and args.load_optimizer_scheduler:
# Load in optimizer and scheduler states
optimizer.load_state_dict(
torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True,
)
# Train!
logger.info("***** Running training *****")
#logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Max steps = %d", args.max_steps)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
global_step = 0
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
# set global_step to gobal_step of last saved checkpoint from model path
if "-" in args.model_name_or_path:
global_step = int(
args.model_name_or_path.split("-")[-1].split("/")[0])
else:
global_step = 0
logger.info(
" Continuing training from checkpoint, will skip to saved global_step"
)
logger.info(" Continuing training from global step %d", global_step)
tr_loss = 0.0
model.zero_grad()
model.train()
set_seed(args) # Added here for reproductibility
last_ann_no = -1
train_dataloader = None
train_dataloader_iter = None
dev_ndcg = 0
step = 0
if args.single_warmup:
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=args.max_steps)
while global_step < args.max_steps:
if step % args.gradient_accumulation_steps == 0 and global_step % args.logging_steps == 0:
# check if new ann training data is availabe
ann_no, ann_path, ndcg_json = get_latest_ann_data(args.ann_dir)
if ann_path is not None and ann_no != last_ann_no:
logger.info("Training on new add data at %s", ann_path)
with open(ann_path, 'r') as f:
ann_training_data = f.readlines()
dev_ndcg = ndcg_json['ndcg']
ann_checkpoint_path = ndcg_json['checkpoint']
ann_checkpoint_no = get_checkpoint_no(ann_checkpoint_path)
aligned_size = (len(ann_training_data) //
args.world_size) * args.world_size
ann_training_data = ann_training_data[:aligned_size]
logger.info("Total ann queries: %d", len(ann_training_data))
if args.triplet:
train_dataset = StreamingDataset(
ann_training_data,
GetTripletTrainingDataProcessingFn(
args, query_cache, passage_cache))
else:
train_dataset = StreamingDataset(
ann_training_data,
GetTrainingDataProcessingFn(args, query_cache,
passage_cache))
train_dataloader = DataLoader(train_dataset,
batch_size=args.train_batch_size)
train_dataloader_iter = iter(train_dataloader)
# re-warmup
if not args.single_warmup:
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=len(ann_training_data))
if args.local_rank != -1:
dist.barrier()
if is_first_worker():
# add ndcg at checkpoint step used instead of current step
tb_writer.add_scalar("dev_ndcg", dev_ndcg,
ann_checkpoint_no)
if last_ann_no != -1:
tb_writer.add_scalar("epoch", last_ann_no,
global_step - 1)
tb_writer.add_scalar("epoch", ann_no, global_step)
last_ann_no = ann_no
try:
batch = next(train_dataloader_iter)
except StopIteration:
logger.info("Finished iterating current dataset, begin reiterate")
train_dataloader_iter = iter(train_dataloader)
batch = next(train_dataloader_iter)
batch = tuple(t.to(args.device) for t in batch)
step += 1
if args.triplet:
inputs = {
"query_ids": batch[0].long(),
"attention_mask_q": batch[1].long(),
"input_ids_a": batch[3].long(),
"attention_mask_a": batch[4].long(),
"input_ids_b": batch[6].long(),
"attention_mask_b": batch[7].long()
}
else:
inputs = {
"input_ids_a": batch[0].long(),
"attention_mask_a": batch[1].long(),
"input_ids_b": batch[3].long(),
"attention_mask_b": batch[4].long(),
"labels": batch[6]
}
# sync gradients only at gradient accumulation step
if step % args.gradient_accumulation_steps == 0:
outputs = model(**inputs)
else:
with model.no_sync():
outputs = model(**inputs)
loss = outputs[
0] # model outputs are always tuple in transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
if step % args.gradient_accumulation_steps == 0:
loss.backward()
else:
with model.no_sync():
loss.backward()
tr_loss += loss.item()
if step % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer),
args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
if args.logging_steps > 0 and global_step % args.logging_steps == 0:
logs = {}
loss_scalar = tr_loss / args.logging_steps
learning_rate_scalar = scheduler.get_lr()[0]
logs["learning_rate"] = learning_rate_scalar
logs["loss"] = loss_scalar
tr_loss = 0
if is_first_worker():
for key, value in logs.items():
tb_writer.add_scalar(key, value, global_step)
logger.info(json.dumps({**logs, **{"step": global_step}}))
if is_first_worker(
) and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(args.output_dir,
"checkpoint-{}".format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
torch.save(optimizer.state_dict(),
os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(),
os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s",
output_dir)
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
tb_writer.close()
return global_step
def train(args, model, tokenizer, query_cache, passage_cache):
""" Train the model """
#if args.local_rank in [-1, 0]:
tb_writer = None
if is_first_worker():
tb_writer = SummaryWriter(log_dir=args.log_dir)
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
real_batch_size = args.train_batch_size * args.gradient_accumulation_steps * (
torch.distributed.get_world_size() if args.local_rank != -1 else 1)
# layerwise optimization for lamb
optimizer_grouped_parameters = []
no_decay = ["bias", "w", "b", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0
},
]
if args.optimizer.lower() == "lamb":
optimizer = Lamb(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
elif args.optimizer.lower() == "adamw":
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
else:
raise Exception(
"optimizer {0} not recognized! Can only be lamb or adamW".format(
args.optimizer))
# Check if saved optimizer or scheduler states exist
if os.path.isfile(
os.path.join(args.model_name_or_path,
"optimizer.pt")) and args.load_optimizer_scheduler:
# Load in optimizer and scheduler states
optimizer.load_state_dict(
torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True,
)
# Train!
logger.info("***** Running training *****")
#logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Max steps = %d", args.max_steps)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
global_step = 0
eval_cnt = 0
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
# set global_step to gobal_step of last saved checkpoint from model path
if "-" in args.model_name_or_path:
try:
global_step = int(
args.model_name_or_path.split("-")[-1].split("/")[0])
except:
global_step = 0
else:
global_step = 0
logger.info(
" Continuing training from checkpoint, will skip to saved global_step"
)
logger.info(" Continuing training from global step %d", global_step)
tr_loss = 0.0
model.zero_grad()
model.train()
set_seed(args) # Added here for reproductibility
last_ann_no = -1
train_dataloader = None
train_dataloader_iter = None
dev_ndcg = 0
step = 0
eval_path = os.path.join(args.data_dir, "eval_full.tsv")
eval_cfg = L1InputConfig("eval",
model,
None,
eval_path,
args.configObj.chunk_cfg,
qid=0,
docid=1,
query=4,
title=5,
anchor=6,
url=7,
click=8,
desc=9,
rating=10,
market=12,
lang=13)
def eval_fn(line, i):
return L1_process_fn(line, i, tokenizer, args, eval_cfg.map,
eval_cfg.chunk_cfg)
model.eval()
ideal_path = os.path.join(args.data_dir, "ideal_map_UN.tsv")
is_first_eval = (eval_cnt == 0)
args.global_step = global_step
# fidelity = eval_fidelity(args, model, eval_fn, eval_cfg.path, ideal_path, args.data_cache_dir, is_first_eval)
# eval_cnt+=1
# print("Fidelity: {0}".format(fidelity))
# if is_first_worker():
# tb_writer.add_scalar("fidelity", fidelity, global_step)
best_checkpoints = []
acc_accum = []
scheduler = None
while global_step < args.max_steps:
if step % args.gradient_accumulation_steps == 0 and global_step % args.logging_steps == 0:
# check if new ann training data is availabe
ann_no, ann_path, ndcg_json = get_latest_ann_data(args.ann_dir)
while ann_no == -1 or (ann_path is not None
and ann_no != last_ann_no):
try:
logger.info("Training on new add data at %s", ann_path)
with open(ann_path, 'r') as f:
ann_training_data = f.readlines()
dev_ndcg = ndcg_json['ndcg']
ann_checkpoint_path = ndcg_json['checkpoint']
ann_checkpoint_no = get_checkpoint_no(ann_checkpoint_path)
aligned_size = (len(ann_training_data) //
args.world_size) * args.world_size
ann_training_data = ann_training_data[:aligned_size]
logger.info("Total ann queries: %d",
len(ann_training_data))
if len(ann_training_data) == 0:
time.sleep(300)
continue
train_dataset = SimplifiedStreamingDataset(
ann_training_data,
args.configObj.process_fn(args, query_cache,
passage_cache))
train_dataloader = DataLoader(
train_dataset,
batch_size=args.train_batch_size,
num_workers=1)
train_dataloader_iter = iter(train_dataloader)
# re-warmup
if scheduler is None:
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=args.max_steps)
if args.local_rank != -1:
dist.barrier()
if is_first_worker():
# add ndcg at checkpoint step used instead of current step
# tb_writer.add_scalar("dev_ndcg", dev_ndcg, ann_checkpoint_no)
if last_ann_no != -1:
tb_writer.add_scalar("epoch", last_ann_no,
global_step - 1)
tb_writer.add_scalar("epoch", ann_no, global_step)
last_ann_no = ann_no
break
except:
if is_first_worker():
print("wait")
time.sleep(300)
ann_no, ann_path, ndcg_json = get_latest_ann_data(args.ann_dir)
try:
batch = next(train_dataloader_iter)
except StopIteration:
logger.info("Finished iterating current dataset, begin reiterate")
train_dataloader_iter = iter(train_dataloader)
batch = next(train_dataloader_iter)
batch = tuple(t.to(args.device).long() for t in batch)
step += 1
model.train()
# if args.triplet:
# inputs = {"query_ids": batch[0].long(), "attention_mask_q": batch[1].long(),
# "input_ids_a": batch[3].long(), "attention_mask_a": batch[4].long(),
# "input_ids_b": batch[6].long(), "attention_mask_b": batch[7].long()}
# else:
# inputs = {"input_ids_a": batch[0].long(), "attention_mask_a": batch[1].long(),
# "input_ids_b": batch[3].long(), "attention_mask_b": batch[4].long(),
# "labels": batch[6]}
# sync gradients only at gradient accumulation step
if step % args.gradient_accumulation_steps == 0:
outputs = model(*batch)
else:
with model.no_sync():
outputs = model(*batch)
loss = outputs[
0] # model outputs are always tuple in transformers (see doc)
acc = outputs[2]
acc_accum.append(acc.item())
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
if step % args.gradient_accumulation_steps == 0:
loss.backward()
else:
with model.no_sync():
loss.backward()
tr_loss += loss.item()
if step % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer),
args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
# print("w grad:", model.module.w.grad)
# print("w:", model.module.w)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
if args.logging_steps > 0 and global_step % args.logging_steps == 0:
logs = {}
if global_step % (args.logging_steps_per_eval *
args.logging_steps) == 0:
print("Train acc:", sum(acc_accum) * 1.0 / len(acc_accum))
acc_accum = []
model.eval()
is_first_eval = (eval_cnt == 0)
args.global_step = global_step
fidelity = eval_fidelity(args, model, eval_fn,
eval_cfg.path, ideal_path,
args.data_cache_dir,
is_first_eval)
eval_cnt += 1
if is_first_worker():
if len(best_checkpoints) < 10:
save_checkpoint(args, global_step, model,
tokenizer, optimizer, scheduler)
best_checkpoints.append((global_step, fidelity))
else:
worst_checkpoint = sorted(best_checkpoints,
key=lambda x: x[1])[0]
if fidelity > worst_checkpoint[1]:
save_checkpoint(args, global_step, model,
tokenizer, optimizer,
scheduler)
worst_cp_path = os.path.join(
args.output_dir, "checkpoint-{}".format(
str(worst_checkpoint[0])))
shutil.rmtree(worst_cp_path)
best_checkpoints.remove(worst_checkpoint)
best_checkpoints.append(
(global_step, fidelity))
else:
print("Fidelity not in top 10!")
assert len(best_checkpoints) == 10
save_checkpoint(args, -1, model, tokenizer)
print("Fidelity: {0}".format(fidelity))
logs["fidelity"] = fidelity
dist.barrier()
loss_scalar = tr_loss / args.logging_steps
learning_rate_scalar = scheduler.get_lr()[0]
logs["learning_rate"] = learning_rate_scalar
logs["loss"] = loss_scalar
tr_loss = 0
if is_first_worker():
for key, value in logs.items():
tb_writer.add_scalar(key, value, global_step)
logger.info(json.dumps({**logs, **{"step": global_step}}))
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
tb_writer.close()
return global_step
def train(args, model, tokenizer, train_dataloader):
""" Train the model """
#if args.local_rank in [-1, 0]:
tb_writer = None
if is_first_worker():
tb_writer = SummaryWriter(log_dir=args.log_dir)
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
real_batch_size = args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1)
if args.max_steps > 0:
t_total = args.max_steps
#args.num_train_epochs = args.max_steps // (args.expected_train_size // args.gradient_accumulation_steps) + 1
else:
t_total = args.expected_train_size // real_batch_size * args.num_train_epochs
# layerwise optimization for lamb
optimizer_grouped_parameters = []
layer_optim_params = set()
for layer_name in ["roberta.embeddings", "score_out", "downsample1", "downsample2", "downsample3", "embeddingHead"]:
layer = getattr_recursive(model, layer_name)
if layer is not None:
optimizer_grouped_parameters.append({"params": layer.parameters()})
for p in layer.parameters():
layer_optim_params.add(p)
if getattr_recursive(model, "roberta.encoder.layer") is not None:
for layer in model.roberta.encoder.layer:
optimizer_grouped_parameters.append({"params": layer.parameters()})
for p in layer.parameters():
layer_optim_params.add(p)
optimizer_grouped_parameters.append({"params": [p for p in model.parameters() if p not in layer_optim_params]})
if len(optimizer_grouped_parameters)==0:
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
]
if args.optimizer.lower()=="lamb":
optimizer = Lamb(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
elif args.optimizer.lower()=="adamw":
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
else:
raise Exception("optimizer {0} not recognized! Can only be lamb or adamW".format(args.optimizer))
if args.scheduler.lower()=="linear":
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
)
elif args.scheduler.lower()=="cosine":
scheduler = CosineAnnealingLR(optimizer, t_total, 1e-8)
else:
raise Exception("Scheduler {0} not recognized! Can only be linear or cosine".format(args.scheduler))
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt")
) and args.load_optimizer_scheduler:
# Load in optimizer and scheduler states
# if is_first_worker():
# op_state = torch.load(os.path.join(args.model_name_or_path, "optimizer.pt"))
# print([len(x['params']) for x in op_state['param_groups']])
# real_op_state = optimizer.state_dict()
# print([len(x['params']) for x in real_op_state['param_groups']])
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True,
)
# Train!
logger.info("***** Running training *****")
#logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size
* args.gradient_accumulation_steps
* (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
epochs_trained = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
# set global_step to gobal_step of last saved checkpoint from model path
try:
global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
epochs_trained = global_step // (args.expected_train_size // args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (args.expected_train_size // args.gradient_accumulation_steps)
logger.info(" Continuing training from checkpoint, will skip to saved global_step")
logger.info(" Continuing training from epoch %d", epochs_trained)
logger.info(" Continuing training from global step %d", global_step)
logger.info(" Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
except:
logger.info(" Start training from a pretrained model")
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(
epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0],
)
set_seed(args) # Added here for reproductibility
for m_epoch in train_iterator:
#epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
for step, batch in tqdm(enumerate(train_dataloader), desc="Iteration", disable=args.local_rank not in [-1, 0]):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
model.train()
batch = tuple(t.to(args.device).long() for t in batch)
if (step + 1) % args.gradient_accumulation_steps == 0:
outputs = model(*batch)
else:
with model.no_sync():
outputs = model(*batch)
loss = outputs[0] # model outputs are always tuple in transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
if (step + 1) % args.gradient_accumulation_steps == 0:
loss.backward()
else:
with model.no_sync():
loss.backward()
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
if is_first_worker() and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s", output_dir)
dist.barrier()
if args.logging_steps > 0 and global_step % args.logging_steps == 0:
logs = {}
if args.evaluate_during_training and global_step % (args.logging_steps_per_eval*args.logging_steps)==0:
model.eval()
reranking_mrr, full_ranking_mrr = passage_dist_eval(args, model, tokenizer)
if is_first_worker():
print("Reranking/Full ranking mrr: {0}/{1}".format(str(reranking_mrr), str(full_ranking_mrr)))
mrr_dict = {"reranking": float(reranking_mrr), "full_raking": float(full_ranking_mrr)}
tb_writer.add_scalars("mrr", mrr_dict, global_step)
print(args.output_dir)
loss_scalar = (tr_loss - logging_loss) / args.logging_steps
learning_rate_scalar = scheduler.get_lr()[0]
logs["learning_rate"] = learning_rate_scalar
logs["loss"] = loss_scalar
logging_loss = tr_loss
if is_first_worker():
for key, value in logs.items():
print(key, type(value))
tb_writer.add_scalar(key, value, global_step)
tb_writer.add_scalar("epoch", m_epoch, global_step)
print(json.dumps({**logs, **{"step": global_step}}))
dist.barrier()
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
tb_writer.close()
return global_step, tr_loss / global_step
def train(args, train_dataset, model_d, model_g, tokenizer):
""" Train the model """
if args.local_rank in [-1, 0]:
tb_writer = SummaryWriter()
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (
len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ["bias", "LayerNorm.weight"]
optimizer_d_grouped_parameters = [
{
"params": [
p for n, p in model_d.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model_d.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0
},
]
# optimizer_d = AdamW(optimizer_d_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
optimizer_d = Lamb(optimizer_d_grouped_parameters,
lr=args.learning_rate,
betas=(0.9, 0.999),
eps=1e-6)
scheduler_d = get_linear_schedule_with_warmup(
optimizer_d,
num_warmup_steps=args.warmup_steps,
num_training_steps=t_total)
optimizer_g_grouped_parameters = [
{
"params": [
p for n, p in model_g.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model_g.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0
},
]
# optimizer_g = AdamW(optimizer_g_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
optimizer_g = Lamb(optimizer_g_grouped_parameters,
lr=args.learning_rate,
betas=(0.9, 0.999),
eps=1e-6)
scheduler_g = get_linear_schedule_with_warmup(
optimizer_g,
num_warmup_steps=args.warmup_steps,
num_training_steps=t_total)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(
args.model_name_or_path, "optimizer_d.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler_d.pt")):
# Load in optimizer and scheduler states
optimizer_d.load_state_dict(
torch.load(os.path.join(args.model_name_or_path,
"optimizer_d.pt")))
scheduler_d.load_state_dict(
torch.load(os.path.join(args.model_name_or_path,
"scheduler_d.pt")))
if os.path.isfile(os.path.join(
args.model_name_or_path, "optimizer_g.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler_g.pt")):
# Load in optimizer and scheduler states
optimizer_g.load_state_dict(
torch.load(os.path.join(args.model_name_or_path,
"optimizer_g.pt")))
scheduler_g.load_state_dict(
torch.load(os.path.join(args.model_name_or_path,
"scheduler_g.pt")))
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model_d, optimizer_d = amp.initialize(model_d,
optimizer_d,
opt_level=args.fp16_opt_level)
model_g, optimizer_g = amp.initialize(model_g,
optimizer_g,
opt_level=args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model_d = torch.nn.DataParallel(model_d)
model_g = torch.nn.DataParallel(model_g)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model_d = torch.nn.parallel.DistributedDataParallel(
model_d,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True,
)
model_g = torch.nn.parallel.DistributedDataParallel(
model_g,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True,
)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
epochs_trained = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
# set global_step to gobal_step of last saved checkpoint from model path
global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
epochs_trained = global_step // (len(train_dataloader) //
args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (
len(train_dataloader) // args.gradient_accumulation_steps)
logger.info(
" Continuing training from checkpoint, will skip to saved global_step"
)
logger.info(" Continuing training from epoch %d", epochs_trained)
logger.info(" Continuing training from global step %d", global_step)
logger.info(" Will skip the first %d steps in the first epoch",
steps_trained_in_current_epoch)
model_to_resize_d = model_d.module if hasattr(
model_d,
"module") else model_d # Take care of distributed/parallel training
# model_to_resize_d.resize_token_embeddings(len(tokenizer))
model_to_resize_g = model_g.module if hasattr(
model_g,
"module") else model_g # Take care of distributed/parallel training
# model_to_resize_g.resize_token_embeddings(len(tokenizer))
# model_to_resize_d.bert.embeddings = model_to_resize_g.bert.embeddings
tr_loss, logging_loss = 0.0, 0.0
tr_loss_d, logging_loss_d = 0.0, 0.0
tr_loss_g, logging_loss_g = 0.0, 0.0
model_d.zero_grad()
model_g.zero_grad()
train_iterator = trange(
epochs_trained,
int(args.num_train_epochs),
desc="Epoch",
disable=args.local_rank not in [-1, 0],
)
set_seed(args) # Added here for reproductibility
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
model_d.train()
model_g.train()
# batch = tuple(t.to(args.device) for t in batch)
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"labels": batch[3]
}
if args.model_type != "distilbert":
inputs["token_type_ids"] = (
batch[2]
if args.model_type in ["bert", "xlnet", "albert"] else None
) # XLM, DistilBERT, RoBERTa, and XLM-RoBERTa don't use segment_ids
# outputs = model(**inputs)
# loss = outputs[0] # model outputs are always tuple in transformers (see doc)
masked_input_ids, mask_labels = mask_tokens(
inputs['input_ids'], tokenizer, args)
outputs_g = model_g(
input_ids=masked_input_ids.to(args.device),
masked_lm_labels=mask_labels.to(args.device),
attention_mask=inputs['attention_mask'].to(args.device),
token_type_ids=inputs['token_type_ids'].to(args.device))
masked_lm_loss, prediction_scores_g = outputs_g[0], outputs_g[1]
prediction_g = prediction_scores_g.max(dim=-1)[1].cpu()
acc_g = (prediction_g[mask_labels >= 0] == mask_labels[
mask_labels >= 0]).float().mean().item()
prediction_probs_g = F.softmax(prediction_scores_g, dim=-1).cpu()
bsz, seq_len, vocab_size = prediction_probs_g.size()
prediction_samples_g = torch.multinomial(prediction_probs_g.view(
-1, vocab_size),
num_samples=1)
prediction_samples_g = prediction_samples_g.view(bsz, seq_len)
input_ids_replace = inputs['input_ids'].clone()
input_ids_replace[mask_labels >= 0] = prediction_samples_g[
mask_labels >= 0]
labels_d = input_ids_replace.eq(inputs['input_ids']).long()
special_tokens_mask = [
tokenizer.get_special_tokens_mask(
val, already_has_special_tokens=True)
for val in inputs['input_ids'].tolist()
]
labels_d.masked_fill_(torch.tensor(special_tokens_mask,
dtype=torch.bool),
value=-100)
padding_mask = inputs['input_ids'].eq(tokenizer.pad_token_id)
labels_d.masked_fill_(padding_mask, value=-100)
labels_d_ones = labels_d[labels_d >= 0].float().mean().item()
acc_replace = 1 - ((labels_d == 0).sum().float() /
(mask_labels >= 0).sum().float()).item()
outputs_d = model_d(
input_ids=input_ids_replace.to(args.device),
attention_mask=inputs['attention_mask'].to(args.device),
token_type_ids=inputs['token_type_ids'].to(args.device),
labels=labels_d.to(args.device))
loss_d, prediction_scores_d = outputs_d[0], outputs_d[1]
prediction_d = prediction_scores_d.max(dim=-1)[1].cpu()
acc_d = (prediction_d[labels_d >= 0] == labels_d[labels_d >= 0]
).float().mean().item()
acc_d_0 = (prediction_d[labels_d == 0] == labels_d[labels_d == 0]
).float().mean().item()
acc_d_1 = (prediction_d[labels_d == 1] == labels_d[labels_d == 1]
).float().mean().item()
if args.n_gpu > 1:
loss_d = loss_d.mean(
) # mean() to average on multi-gpu parallel training
masked_lm_loss = masked_lm_loss.mean()
if args.gradient_accumulation_steps > 1:
loss_d = loss_d / args.gradient_accumulation_steps
masked_lm_loss = masked_lm_loss / args.gradient_accumulation_steps
lambd = 50
loss = loss_d * lambd + masked_lm_loss
if args.fp16:
loss_d = loss_d * lambd
with amp.scale_loss(loss_d, optimizer_d) as scaled_loss_d:
scaled_loss_d.backward()
with amp.scale_loss(masked_lm_loss,
optimizer_g) as scaled_loss_g:
scaled_loss_g.backward()
else:
loss.backward()
tr_loss += loss.item()
tr_loss_d += loss_d.item()
tr_loss_g += masked_lm_loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer_d), args.max_grad_norm)
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer_g), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model_d.parameters(),
args.max_grad_norm)
torch.nn.utils.clip_grad_norm_(model_g.parameters(),
args.max_grad_norm)
optimizer_d.step()
scheduler_d.step() # Update learning rate schedule
model_d.zero_grad()
optimizer_g.step()
scheduler_g.step() # Update learning rate schedule
model_g.zero_grad()
if args.local_rank in [
-1, 0
] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
logs = {}
# if (
# args.local_rank == -1 and args.evaluate_during_training
# ): # Only evaluate when single GPU otherwise metrics may not average well
# results = evaluate(args, model, tokenizer)
# for key, value in results.items():
# eval_key = "eval_{}".format(key)
# logs[eval_key] = value
loss_scalar = (tr_loss - logging_loss) / args.logging_steps
loss_scalar_d = (tr_loss_d -
logging_loss_d) / args.logging_steps
loss_scalar_g = (tr_loss_g -
logging_loss_g) / args.logging_steps
learning_rate_scalar_d = scheduler_d.get_lr()[0]
learning_rate_scalar_g = scheduler_g.get_lr()[0]
logs["learning_rate_d"] = learning_rate_scalar_d
logs["learning_rate_g"] = learning_rate_scalar_g
logs["loss"] = loss_scalar
logs["loss_d"] = loss_scalar_d
logs["loss_g"] = loss_scalar_g
logs["acc_repalce"] = acc_replace
logs["acc_d"] = acc_d
logs["acc_d_0"] = acc_d_0
logs["acc_d_1"] = acc_d_1
logs["acc_g"] = acc_g
logs["labels_d_ones"] = labels_d_ones
logs["masked_ratio"] = (mask_labels >= 0).float().sum(
).item() / (labels_d >= 0).sum().float().item()
logging_loss = tr_loss
logging_loss_d = tr_loss_d
logging_loss_g = tr_loss_g
for key, value in logs.items():
tb_writer.add_scalar(key, value, global_step)
print(json.dumps({**logs, **{"step": global_step}}))
# print(args.save_steps)
if args.local_rank in [
-1, 0
] and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(
args.output_dir, "checkpoint-{}".format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
output_dir_d = os.path.join(
output_dir, "checkpoint-d-{}".format(global_step))
output_dir_g = os.path.join(
output_dir, "checkpoint-g-{}".format(global_step))
if not os.path.exists(output_dir_d):
os.makedirs(output_dir_d)
if not os.path.exists(output_dir_g):
os.makedirs(output_dir_g)
model_to_save_d = (
model_d.module if hasattr(model_d, "module") else
model_d) # Take care of distributed/parallel training
model_to_save_g = (
model_g.module if hasattr(model_g, "module") else
model_g) # Take care of distributed/parallel training
model_to_save_d.save_pretrained(output_dir_d)
model_to_save_g.save_pretrained(output_dir_g)
tokenizer.save_pretrained(output_dir)
torch.save(args,
os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
torch.save(optimizer_d.state_dict(),
os.path.join(output_dir_d, "optimizer_d.pt"))
torch.save(scheduler_d.state_dict(),
os.path.join(output_dir_d, "scheduler_d.pt"))
torch.save(optimizer_g.state_dict(),
os.path.join(output_dir_d, "optimizer_g.pt"))
torch.save(scheduler_g.state_dict(),
os.path.join(output_dir_d, "scheduler_g.pt"))
logger.info("Saving optimizer and scheduler states to %s",
output_dir)
global_step += 1
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
if args.local_rank in [-1, 0]:
tb_writer.close()
return global_step, tr_loss / global_step
