def test_clip_grad_norm(max_norm, check_isfinite):
def gt_grad_global_norm(parameters):
ret = 0
for p in parameters:
grads = p.list_grad()
ret += (grads[0].asnumpy()**2).sum()
return np.sqrt(ret)
contexts = [mx.cpu(0), mx.cpu(1)]
net = mx.gluon.nn.HybridSequential()
# Create a network with 8 layers
for _ in range(8):
net.add(
mx.gluon.nn.Dense(1,
weight_initializer='ones',
bias_initializer='ones'))
net.initialize(ctx=contexts)
net.hybridize()
trainer = mx.gluon.Trainer(net.collect_params(),
'sgd',
update_on_kvstore=False)
for ctx in contexts:
with mx.autograd.record():
out = net(mx.np.ones((1, 1), ctx=ctx))
out.backward()
trainer.allreduce_grads()
# Cache the original gradient for checking
original_grad_l = [
p.list_grad()[0].asnumpy() for p in net.collect_params().values()
]
norm = grad_global_norm(net.collect_params().values())
gt_norm = gt_grad_global_norm(net.collect_params().values())
assert_almost_equal(norm, gt_norm, atol=1e-5)
with mx.cpu(2):
norm, ratio, is_finite = clip_grad_global_norm(
net.collect_params().values(), max_norm, check_isfinite)
assert_almost_equal(norm, gt_norm, atol=1e-5)
for p, orig_grad in zip(net.collect_params().values(), original_grad_l):
for ctx in contexts:
if max_norm > norm:
assert_almost_equal(p.grad(ctx).asnumpy(), orig_grad)
else:
ratio = max_norm / norm
assert_almost_equal(p.grad(ctx).asnumpy(), orig_grad * ratio)
def train(args):
_, num_parts, rank, local_rank, _, ctx_l = init_comm(
args.comm_backend, args.gpus)
if args.comm_backend == 'horovod':
logging_config(
args.save_dir,
name=f'train_transformer_rank{rank}_local{local_rank}_{num_parts}',
console=(rank == 0))
logging.info(args)
else:
logging_config(args.save_dir, name='train_transformer', console=True)
logging.info(args)
use_amp = args.fp16
if use_amp:
from mxnet import amp
src_tokenizer = create_tokenizer(args.src_tokenizer,
args.src_subword_model_path,
args.src_vocab_path)
tgt_tokenizer = create_tokenizer(args.tgt_tokenizer,
args.tgt_subword_model_path,
args.tgt_vocab_path)
base_tgt_tokenizer = MosesTokenizer(args.tgt_lang)
src_vocab = src_tokenizer.vocab
tgt_vocab = tgt_tokenizer.vocab
train_src_data, train_tgt_data = load_dataset_with_cache(
args.train_src_corpus,
args.train_tgt_corpus,
src_tokenizer,
tgt_tokenizer,
args.overwrite_cache,
local_rank,
max_src_length=args.max_src_length,
max_tgt_length=args.max_tgt_length,
pretokenized=not args.tokenize)
dev_src_data, dev_tgt_data = load_dataset_with_cache(
args.dev_src_corpus,
args.dev_tgt_corpus,
src_tokenizer,
tgt_tokenizer,
args.overwrite_cache,
local_rank,
pretokenized=not args.tokenize)
tgt_detok_sentences = []
tgt_raw_sentences = []
with open(args.dev_tgt_corpus, 'r') as in_f:
for line in in_f:
tgt_detok_sentences.append(
base_tgt_tokenizer.decode(
tgt_tokenizer.decode(line.split()).split()))
with open(args.dev_tgt_raw_corpus, 'r') as in_f:
for line in in_f:
tgt_raw_sentences.append(line.strip())
data_train = gluon.data.SimpleDataset([
(src_tokens, tgt_tokens, len(src_tokens), len(tgt_tokens), i)
for i, (src_tokens,
tgt_tokens) in enumerate(zip(train_src_data, train_tgt_data))
])
val_samples = [
(src_tokens, tgt_tokens, len(src_tokens), len(tgt_tokens), i)
for i, (src_tokens,
tgt_tokens) in enumerate(zip(dev_src_data, dev_tgt_data))
]
if args.comm_backend == 'horovod':
slice_begin = rank * (len(val_samples) // num_parts)
slice_end = min((rank + 1) * (len(val_samples) // num_parts),
len(val_samples))
data_val = gluon.data.SimpleDataset(val_samples[slice_begin:slice_end])
else:
data_val = gluon.data.SimpleDataset(val_samples)
# Construct the model + loss function
if args.cfg.endswith('.yml'):
cfg = TransformerModel.get_cfg().clone_merge(args.cfg)
else:
cfg = TransformerModel.get_cfg(args.cfg)
cfg.defrost()
cfg.MODEL.src_vocab_size = len(src_vocab)
cfg.MODEL.tgt_vocab_size = len(tgt_vocab)
cfg.MODEL.layout = 'TN'
cfg.freeze()
model = TransformerModel.from_cfg(cfg)
model.initialize(mx.init.Xavier(magnitude=args.magnitude), ctx=ctx_l)
model.hybridize()
for v in model.collect_params().values():
if v.grad_req != 'null':
v.grad_req = 'add'
# Do not apply weight decay to all the LayerNorm and bias
for _, v in model.collect_params('.*beta|.*gamma|.*bias').items():
v.wd_mult = 0.0
param_dict = deduplicate_param_dict(model.collect_params())
inference_model = TransformerInference(model=model)
inference_model.hybridize()
if local_rank == 0:
logging.info(model)
with open(os.path.join(args.save_dir, 'config.yml'), 'w') as cfg_f:
cfg_f.write(cfg.dump())
label_smooth_loss = LabelSmoothCrossEntropyLoss(
num_labels=len(tgt_vocab),
alpha=args.label_smooth_alpha,
from_logits=False)
label_smooth_loss.hybridize()
# Construct the beam search sampler
scorer = BeamSearchScorer(alpha=args.lp_alpha,
K=args.lp_k,
from_logits=False)
beam_search_sampler = BeamSearchSampler(beam_size=args.beam_size,
decoder=inference_model,
vocab_size=len(tgt_vocab),
eos_id=tgt_vocab.eos_id,
scorer=scorer,
stochastic=False,
max_length_a=args.max_length_a,
max_length_b=args.max_length_b)
logging.info(beam_search_sampler)
if args.comm_backend == 'horovod':
hvd.broadcast_parameters(param_dict, root_rank=0)
# Construct the trainer
if args.lr is None:
base_lr = 2.0 / math.sqrt(args.num_units) / math.sqrt(
args.warmup_steps)
else:
base_lr = args.lr
lr_scheduler = InverseSquareRootScheduler(
warmup_steps=args.warmup_steps,
base_lr=base_lr,
warmup_init_lr=args.warmup_init_lr)
optimizer_params = {
'learning_rate': args.lr,
'beta1': 0.9,
'beta2': 0.997,
'epsilon': 1e-9,
'lr_scheduler': lr_scheduler,
'wd': args.wd
}
user_provided_ptimizer_params = json.loads(args.optimizer_params)
optimizer_params.update(user_provided_ptimizer_params)
if args.fp16:
optimizer_params.update({'multi_precision': True})
if args.comm_backend == 'horovod':
trainer = hvd.DistributedTrainer(param_dict, args.optimizer,
optimizer_params)
else:
trainer = gluon.Trainer(param_dict,
args.optimizer,
optimizer_params,
update_on_kvstore=False)
# Load Data
if args.sampler == 'BoundedBudgetSampler':
train_batch_sampler = BoundedBudgetSampler(
lengths=[(ele[2], ele[3]) for ele in data_train],
max_num_tokens=args.max_num_tokens,
max_num_sentences=args.max_num_sentences,
shuffle=True,
seed=args.seed)
elif args.sampler == 'FixedBucketSampler':
if args.comm_backend == 'horovod':
raise NotImplementedError(
'FixedBucketSampler does not support horovod at present')
if args.bucket_scheme == 'constant':
bucket_scheme = ConstWidthBucket()
elif args.bucket_scheme == 'linear':
bucket_scheme = LinearWidthBucket()
elif args.bucket_scheme == 'exp':
bucket_scheme = ExpWidthBucket(bucket_len_step=1.2)
else:
raise NotImplementedError
# TODO(sxjscience) Support auto-bucket-size tuning
train_batch_sampler = FixedBucketSampler(lengths=[
(ele[2], ele[3]) for ele in data_train
],
batch_size=args.batch_size,
num_buckets=args.num_buckets,
ratio=args.bucket_ratio,
shuffle=True,
use_average_length=True,
bucket_scheme=bucket_scheme,
seed=args.seed)
else:
raise NotImplementedError
num_updates_per_epoch = int(
math.ceil(
len(train_batch_sampler) /
(num_parts * len(ctx_l) * args.num_accumulated)))
# Convert the batch sampler to multiple shards
if num_parts > 1:
train_batch_sampler = ShardedIterator(train_batch_sampler,
num_parts=num_parts,
part_index=rank,
even_size=True,
seed=args.seed + 1000 * rank)
logging.info(train_batch_sampler)
batchify_fn = bf.Tuple(bf.Pad(), bf.Pad(), bf.Stack(), bf.Stack(),
bf.Stack())
train_data_loader = gluon.data.DataLoader(
data_train,
batch_sampler=train_batch_sampler,
batchify_fn=batchify_fn,
num_workers=0)
val_data_loader = gluon.data.DataLoader(data_val,
batch_size=args.val_batch_size,
batchify_fn=batchify_fn,
num_workers=0,
shuffle=False)
params = [p for p in param_dict.values() if p.grad_req != 'null']
model_averager = AverageSGDTracker(param_dict)
log_start_time = time.time()
num_params, num_fixed_params = None, None
# TODO(sxjscience) Add a log metric class
log_avg_loss_l = [mx.np.array(0.0, ctx=ctx) for ctx in ctx_l]
# Maintain the denominator of the loss.
log_avg_loss_denom_l = [mx.np.array(0.0, ctx=ctx) for ctx in ctx_l]
log_wc_l = [mx.np.array(0, dtype=np.int64, ctx=ctx) for ctx in ctx_l]
log_tgt_wc_l = [mx.np.array(0, dtype=np.int64, ctx=ctx) for ctx in ctx_l]
log_avg_grad_norm = 0
log_iter_num = 0
if local_rank == 0:
writer = SummaryWriter(
logdir=os.path.join(args.save_dir, 'tensorboard'))
if use_amp:
amp.init_trainer(trainer)
train_multi_data_loader = grouper(repeat(train_data_loader), len(ctx_l))
# when args.epochs < 0, the model will keep training
if args.epochs < 0:
if args.max_update > 0:
total_train_iters = args.max_update
if args.num_averages > 0:
assert args.num_averages <= total_train_iters // args.save_iterval_update
avg_start_iter = (
total_train_iters // args.save_iterval_update -
args.num_averages) * args.save_iterval_update
else:
avg_start_iter = -1
else:
total_train_iters = np.inf
avg_start_iter = -1
else:
total_train_iters = args.epochs * num_updates_per_epoch
if args.num_averages > 0:
assert args.num_averages <= args.epochs
avg_start_iter = (args.epochs -
args.num_average) * num_updates_per_epoch
else:
avg_start_iter = -1
# Here, we are manually setting up the scale to 1.0 because
# in horovod, the scale can be the number of workers:
# See the code here: https://github.com/horovod/horovod/blob/125115583b7029196e2ec530decd4209459d5479/horovod/mxnet/__init__.py#L141
# Since we will need to use the dynamic scaling in amp, we will manually call amp.unscale().
# A scale that is larger than 1.0 can be problematic in this case.
trainer._scale = 1.0
if args.max_num_tokens > 0:
const_scale = args.max_num_tokens
else:
const_scale = 100
train_start_time = time.time()
for train_iter in range(total_train_iters):
model.zero_grad()
loss_denom_l = [mx.np.array(0.0, ctx=ctx) for ctx in ctx_l]
for i in range(args.num_accumulated):
loss_l = []
sample_data_l = next(train_multi_data_loader)
for j, (sample_data, ctx) in enumerate(zip(sample_data_l, ctx_l)):
src_token_ids, tgt_token_ids, src_valid_length,\
tgt_valid_length, sample_ids = sample_data
src_token_ids = src_token_ids.as_in_ctx(ctx)
tgt_token_ids = tgt_token_ids.as_in_ctx(ctx)
src_valid_length = src_valid_length.as_in_ctx(ctx)
tgt_valid_length = tgt_valid_length.as_in_ctx(ctx)
src_wc, tgt_wc, bs = src_valid_length.sum(), \
tgt_valid_length.sum(), src_token_ids.shape[0]
log_wc_l[j] += src_wc + tgt_wc
log_tgt_wc_l[j] += tgt_wc
token_count = (tgt_valid_length - 1).sum()
loss_denom_l[j] += token_count / const_scale
log_avg_loss_denom_l[j] += token_count / const_scale
with mx.autograd.record():
if model.layout == 'NT':
tgt_pred = model(src_token_ids, src_valid_length,
tgt_token_ids[:, :-1],
tgt_valid_length - 1)
tgt_labels = tgt_token_ids[:, 1:]
loss = label_smooth_loss(tgt_pred, tgt_labels)
loss = mx.npx.sequence_mask(
loss,
sequence_length=tgt_valid_length - 1,
use_sequence_length=True,
axis=1)
loss = loss.sum() / const_scale
loss_l.append(loss)
elif model.layout == 'TN':
tgt_pred = model(src_token_ids.T, src_valid_length,
tgt_token_ids.T[:-1, :],
tgt_valid_length - 1)
tgt_labels = tgt_token_ids.T[1:, :]
loss = label_smooth_loss(tgt_pred, tgt_labels)
loss = mx.npx.sequence_mask(
loss,
sequence_length=tgt_valid_length - 1,
use_sequence_length=True,
axis=0)
loss = loss.sum() / const_scale
loss_l.append(loss)
log_avg_loss_l[j] += loss
if use_amp:
with mx.autograd.record():
with amp.scale_loss(loss_l, trainer) as amp_loss_l:
for loss in amp_loss_l:
loss.backward()
else:
with mx.autograd.record():
for loss in loss_l:
loss.backward()
# Print the total number of parameters
if local_rank == 0 and num_params is None:
num_params, num_fixed_params = count_parameters(param_dict)
logging.info(
'Total Number of Parameters (not-fixed/fixed): {}/{}'.format(
num_params, num_fixed_params))
# All-Reduce the gradient
trainer.allreduce_grads()
if args.comm_backend == 'horovod':
# All-Reduce the loss denominator
assert len(loss_denom_l) == 1
loss_denom = hvd.allreduce(loss_denom_l[0],
average=False).asnumpy()
else:
loss_denom = sum([ele.asnumpy() for ele in loss_denom_l])
if use_amp:
# We need to first unscale the gradient and then perform allreduce.
grad_scale = trainer.amp_loss_scale * loss_denom
else:
grad_scale = loss_denom
if args.max_grad_norm is not None:
total_norm, ratio, is_finite\
= clip_grad_global_norm(params, args.max_grad_norm * grad_scale)
total_norm = total_norm / grad_scale
else:
total_norm = grad_global_norm(params)
total_norm = total_norm / grad_scale
log_avg_grad_norm += total_norm
log_iter_num += 1
trainer.update(loss_denom, ignore_stale_grad=True)
if avg_start_iter > 0 and train_iter >= avg_start_iter:
model_averager.step()
if ((train_iter + 1) % args.log_interval == 0
or train_iter + 1 == total_train_iters):
if args.comm_backend == 'horovod':
# Use allreduce to get the total number of tokens and loss
log_wc = hvd.allreduce(log_wc_l[0], average=False).asnumpy()
log_tgt_wc = hvd.allreduce(log_tgt_wc_l[0],
average=False).asnumpy()
log_avg_loss = hvd.allreduce(log_avg_loss_l[0] /
log_avg_loss_denom_l[0],
average=True)
log_avg_loss = log_avg_loss.asnumpy()
else:
log_wc = sum([ele.asnumpy() for ele in log_wc_l])
log_tgt_wc = sum([ele.asnumpy() for ele in log_tgt_wc_l])
log_avg_loss =\
sum([log_avg_loss_l[i].asnumpy() / log_avg_loss_denom_l[i].asnumpy()
for i in range(len(log_avg_loss_l))]) / len(log_avg_loss_l)
log_avg_grad_norm = log_avg_grad_norm / log_iter_num
log_end_time = time.time()
wps = log_wc / (log_end_time - log_start_time)
epoch_id = train_iter // num_updates_per_epoch
logging.info(
'[Epoch {} Iter {}/{}, Overall {}/{}] loss={:.4f}, ppl={:.4f}, '
'throughput={:.2f}K wps, total wc={:.2f}K, wpb={:.2f}K,'
' LR={}, gnorm={:.4f}, ETA={:.2f}h'.format(
epoch_id, train_iter % num_updates_per_epoch + 1,
num_updates_per_epoch,
train_iter + 1, total_train_iters, log_avg_loss,
np.exp(log_avg_loss), wps / 1000, log_wc / 1000,
log_tgt_wc / 1000 / log_iter_num, trainer.learning_rate,
log_avg_grad_norm,
(log_end_time - train_start_time) / (train_iter + 1) *
(total_train_iters - train_iter - 1) / 3600))
if local_rank == 0:
writer.add_scalar('throughput_wps', wps, train_iter)
writer.add_scalar('train_loss', log_avg_loss, train_iter)
writer.add_scalar('lr', trainer.learning_rate, train_iter)
writer.add_scalar('grad_norm', log_avg_grad_norm, train_iter)
# Reinitialize the log variables
log_start_time = time.time()
log_avg_loss_l = [mx.np.array(0.0, ctx=ctx) for ctx in ctx_l]
log_avg_loss_denom_l = [mx.np.array(0.0, ctx=ctx) for ctx in ctx_l]
log_avg_grad_norm = 0
log_iter_num = 0
log_wc_l = [
mx.np.array(0, dtype=np.int64, ctx=ctx) for ctx in ctx_l
]
log_tgt_wc_l = [
mx.np.array(0, dtype=np.int64, ctx=ctx) for ctx in ctx_l
]
if (args.max_update > 0 and (train_iter + 1) % args.save_interval_update == 0) \
or ((train_iter + 1) % num_updates_per_epoch == 0) \
or train_iter + 1 == total_train_iters:
epoch_id = (train_iter + 1) // num_updates_per_epoch
if local_rank == 0:
if args.max_update <= 0:
model.save_parameters(os.path.join(
args.save_dir, 'epoch{}.params'.format(epoch_id)),
deduplicate=True)
else:
model.save_parameters(os.path.join(
args.save_dir, 'iter{}.params'.format(train_iter + 1)),
deduplicate=True)
avg_val_loss, ntokens, pred_sentences, pred_lengths, sentence_ids\
= validation(model, val_data_loader, inference_model, beam_search_sampler,
tgt_tokenizer, ctx_l)
if args.comm_backend == 'horovod':
flatten_pred_sentences = np.concatenate(pred_sentences, axis=0)
all_val_loss = hvd.allgather(
mx.np.array([avg_val_loss * ntokens],
dtype=np.float32,
ctx=ctx_l[0]))
all_ntokens = hvd.allgather(
mx.np.array([ntokens], dtype=np.int64, ctx=ctx_l[0]))
flatten_pred_sentences = hvd.allgather(
mx.np.array(flatten_pred_sentences,
dtype=np.int32,
ctx=ctx_l[0]))
pred_lengths = hvd.allgather(
mx.np.array(pred_lengths, dtype=np.int64, ctx=ctx_l[0]))
sentence_ids = hvd.allgather(
mx.np.array(sentence_ids, dtype=np.int64, ctx=ctx_l[0]))
avg_val_loss = all_val_loss.asnumpy().sum(
) / all_ntokens.asnumpy().sum()
flatten_pred_sentences = flatten_pred_sentences.asnumpy()
pred_lengths = pred_lengths.asnumpy()
sentence_ids = sentence_ids.asnumpy()
pred_sentences = [None for _ in range(len(sentence_ids))]
ptr = 0
assert sentence_ids.min() == 0 and sentence_ids.max(
) == len(sentence_ids) - 1
for sentence_id, length in zip(sentence_ids, pred_lengths):
pred_sentences[sentence_id] = flatten_pred_sentences[ptr:(
ptr + length)]
ptr += length
if local_rank == 0:
# Perform detokenization
pred_sentences_bpe_decode = []
pred_sentences_raw = []
for sentence in pred_sentences:
bpe_decode_sentence = tgt_tokenizer.decode(
sentence.tolist())
raw_sentence = base_tgt_tokenizer.decode(
bpe_decode_sentence.split())
pred_sentences_bpe_decode.append(bpe_decode_sentence)
pred_sentences_raw.append(raw_sentence)
detok_sacrebleu_out = sacrebleu.corpus_bleu(
sys_stream=pred_sentences_bpe_decode,
ref_streams=[tgt_detok_sentences])
raw_sacrebleu_out = sacrebleu.corpus_bleu(
sys_stream=pred_sentences_raw,
ref_streams=[tgt_raw_sentences])
with open(
os.path.join(args.save_dir,
f'epoch{epoch_id}_dev_prediction.txt'),
'w') as of:
for line in pred_sentences_raw:
of.write(line + '\n')
logging.info(
'[Epoch {}][Iter {}/{}] validation loss/ppl={:.4f}/{:.4f}, '
'SacreBlEU={}, Detok SacreBLUE={}'.format(
epoch_id, train_iter, total_train_iters, avg_val_loss,
np.exp(avg_val_loss), raw_sacrebleu_out.score,
detok_sacrebleu_out.score))
writer.add_scalar('valid_loss', avg_val_loss, train_iter)
writer.add_scalar('valid_bleu', raw_sacrebleu_out.score,
train_iter)
if args.num_averages > 0:
model_averager.copy_back(
param_dict) # TODO(sxjscience) Rewrite using update
model.save_parameters(os.path.join(args.save_dir, 'average.params'),
deduplicate=True)
def train(args):
store, num_workers, rank, local_rank, is_master_node, ctx_l = init_comm(
args.comm_backend, args.gpus)
setup_logging(args, local_rank)
cfg, tokenizer, qa_net, use_segmentation = \
get_network(args.model_name, ctx_l,
args.classifier_dropout,
args.param_checkpoint,
args.backbone_path)
logging.info('Prepare training data')
train_features = get_squad_features(args, tokenizer, segment='train')
dataset_processor = SquadDatasetProcessor(
tokenizer=tokenizer,
doc_stride=args.doc_stride,
max_seq_length=args.max_seq_length,
max_query_length=args.max_query_length)
logging.info('Processing the Training data:')
train_dataset, num_answer_mismatch, num_unreliable \
= dataset_processor.get_train(train_features, skip_unreliable=True)
logging.info(
'Done! #Unreliable Span={} / #Mismatched Answer={} / #Total={}'.format(
num_unreliable, num_answer_mismatch, len(train_features)))
# Get dataset statistics
num_impossible = 0
for sample in train_dataset:
num_impossible += sample.is_impossible
logging.info('Before Chunking, #Train/Is Impossible = {}/{}'.format(
len(train_features),
sum([ele.is_impossible for ele in train_features])))
logging.info('After Chunking, #Train Sample/Is Impossible = {}/{}'.format(
len(train_dataset), num_impossible))
# Shuffle the dataset using a fixed seed across all workers
rs = np.random.RandomState(args.pre_shuffle_seed)
rs.shuffle(train_dataset)
sampler = SplitSampler(len(train_dataset),
num_parts=num_workers,
part_index=rank,
even_size=True)
train_dataloader = mx.gluon.data.DataLoader(
train_dataset,
batchify_fn=dataset_processor.BatchifyFunction,
batch_size=args.batch_size,
num_workers=0,
sampler=sampler)
if 'electra' in args.model_name:
# Froze parameters, does not work for albert model since parameters in all layers are shared
if args.untunable_depth > 0:
qa_net.backbone.frozen_params(args.untunable_depth)
if args.layerwise_decay > 0:
qa_net.backbone.apply_layerwise_decay(args.layerwise_decay)
logging.info('Creating distributed trainer...')
# Collect differentiable parameters
param_dict = qa_net.collect_params()
# Do not apply weight decay to all the LayerNorm and bias
for _, v in qa_net.collect_params('.*beta|.*gamma|.*bias').items():
v.wd_mult = 0.0
params = [p for p in param_dict.values() if p.grad_req != 'null']
# Set grad_req if gradient accumulation is required
num_accumulated = args.num_accumulated
if num_accumulated > 1:
logging.info(
'Using gradient accumulation. Effective global batch size = {}'.
format(num_accumulated * args.batch_size * len(ctx_l) *
num_workers))
for p in params:
p.grad_req = 'add'
# backend specific implementation
if args.comm_backend == 'horovod':
# Horovod: fetch and broadcast parameters
hvd.broadcast_parameters(param_dict, root_rank=0)
epoch_size = (len(train_dataloader) + len(ctx_l) - 1) // len(ctx_l)
if args.num_train_steps is not None:
num_train_steps = args.num_train_steps
else:
num_train_steps = int(args.epochs * epoch_size / args.num_accumulated)
if args.warmup_steps is not None:
warmup_steps = args.warmup_steps
else:
warmup_steps = int(num_train_steps * args.warmup_ratio)
assert warmup_steps is not None, 'Must specify either warmup_steps or warmup_ratio'
log_interval = args.log_interval
save_interval = args.save_interval if args.save_interval is not None\
else epoch_size // args.num_accumulated
logging.info(
'#Total Training Steps={}, Warmup={}, Save Interval={}'.format(
num_train_steps, warmup_steps, save_interval))
# set up optimization
lr_scheduler = PolyScheduler(max_update=num_train_steps,
base_lr=args.lr,
warmup_begin_lr=0,
pwr=1,
final_lr=0,
warmup_steps=warmup_steps,
warmup_mode='linear')
optimizer_params = {
'learning_rate': args.lr,
'wd': args.wd,
'lr_scheduler': lr_scheduler,
}
adam_betas = eval(args.adam_betas)
if args.optimizer == 'adamw':
optimizer_params.update({
'beta1': adam_betas[0],
'beta2': adam_betas[1],
'epsilon': args.adam_epsilon,
'correct_bias': False,
})
elif args.optimizer == 'adam':
optimizer_params.update({
'beta1': adam_betas[0],
'beta2': adam_betas[1],
'epsilon': args.adam_epsilon,
})
if args.comm_backend == 'horovod':
trainer = hvd.DistributedTrainer(param_dict, args.optimizer,
optimizer_params)
else:
trainer = mx.gluon.Trainer(param_dict,
args.optimizer,
optimizer_params,
update_on_kvstore=False)
log_span_loss = 0
log_answerable_loss = 0
log_total_loss = 0
log_sample_num = 0
global_tic = time.time()
tic = time.time()
for step_num, batch_data in enumerate(
grouper(repeat(train_dataloader),
len(ctx_l) * num_accumulated)):
for sample_l in grouper(batch_data, len(ctx_l)):
loss_l = []
span_loss_l = []
answerable_loss_l = []
for sample, ctx in zip(sample_l, ctx_l):
if sample is None:
continue
# Copy the data to device
tokens = sample.data.as_in_ctx(ctx)
log_sample_num += len(tokens)
segment_ids = sample.segment_ids.as_in_ctx(
ctx) if use_segmentation else None
valid_length = sample.valid_length.as_in_ctx(ctx)
p_mask = sample.masks.as_in_ctx(ctx)
gt_start = sample.gt_start.as_in_ctx(ctx).astype(np.int32)
gt_end = sample.gt_end.as_in_ctx(ctx).astype(np.int32)
is_impossible = sample.is_impossible.as_in_ctx(ctx).astype(
np.int32)
batch_idx = mx.np.arange(tokens.shape[0],
dtype=np.int32,
ctx=ctx)
p_mask = 1 - p_mask # In the network, we use 1 --> no_mask, 0 --> mask
with mx.autograd.record():
start_logits, end_logits, answerable_logits \
= qa_net(tokens, segment_ids, valid_length, p_mask, gt_start)
sel_start_logits = start_logits[batch_idx, gt_start]
sel_end_logits = end_logits[batch_idx, gt_end]
sel_answerable_logits = answerable_logits[batch_idx,
is_impossible]
span_loss = -0.5 * (sel_start_logits +
sel_end_logits).mean()
answerable_loss = -0.5 * sel_answerable_logits.mean()
loss = span_loss + answerable_loss
loss_l.append(loss)
span_loss_l.append(span_loss)
answerable_loss_l.append(answerable_loss)
for loss in loss_l:
loss.backward()
# All Reduce the Step Loss
log_span_loss += sum(
[ele.as_in_ctx(ctx_l[0]) for ele in span_loss_l]).asnumpy()
log_total_loss += sum([ele.as_in_ctx(ctx_l[0])
for ele in loss_l]).asnumpy()
log_answerable_loss += sum([
ele.as_in_ctx(ctx_l[0]) for ele in answerable_loss_l
]).asnumpy()
# update
trainer.allreduce_grads()
if args.max_grad_norm > 0:
total_norm, ratio, is_finite = clip_grad_global_norm(
params, args.max_grad_norm * num_workers)
else:
total_norm = grad_global_norm(params)
if args.comm_backend == 'horovod':
# Note that horovod.trainer._scale is default to num_workers,
# thus trainer.update(1) will scale the gradients by 1./num_workers
trainer.update(1, ignore_stale_grad=True)
else:
# gluon.trainer._scale is default to 1
trainer.update(num_workers, ignore_stale_grad=True)
total_norm = total_norm / num_workers
if args.num_accumulated > 1:
# set grad to zero for gradient accumulation
qa_net.zero_grad()
# saving
if local_rank == 0 and (step_num + 1) % save_interval == 0 or (
step_num + 1) >= num_train_steps:
version_prefix = 'squad' + args.version
ckpt_name = '{}_{}_{}.params'.format(args.model_name,
version_prefix,
(step_num + 1))
params_saved = os.path.join(args.output_dir, ckpt_name)
qa_net.save_parameters(params_saved)
ckpt_candidates = [
f for f in os.listdir(args.output_dir) if f.endswith('.params')
]
# keep last `max_saved_ckpt` checkpoints
if len(ckpt_candidates) > args.max_saved_ckpt:
ckpt_candidates.sort(key=lambda ele: (len(ele), ele))
os.remove(os.path.join(args.output_dir, ckpt_candidates[0]))
logging.info('Params saved in: {}'.format(params_saved))
# logging
if (step_num + 1) % log_interval == 0:
log_span_loss /= log_sample_num
log_answerable_loss /= log_sample_num
log_total_loss /= log_sample_num
toc = time.time()
logging.info(
'Step: {}/{}, Loss span/answer/total={:.4f}/{:.4f}/{:.4f},'
' LR={:.8f}, grad_norm={:.4f}. Time cost={:.2f}, Throughput={:.2f} samples/s'
' ETA={:.2f}h'.format(
(step_num + 1), num_train_steps, log_span_loss,
log_answerable_loss, log_total_loss, trainer.learning_rate,
total_norm, toc - tic, log_sample_num / (toc - tic),
(num_train_steps - (step_num + 1)) /
((step_num + 1) / (toc - global_tic)) / 3600))
tic = time.time()
log_span_loss = 0
log_answerable_loss = 0
log_total_loss = 0
log_sample_num = 0
num_samples_per_update = 0
if (step_num + 1) >= num_train_steps:
toc = time.time()
logging.info('Finish training step: {} within {} hours'.format(
step_num + 1, (toc - global_tic) / 3600))
break
return params_saved
def test_multi_head_rel_attn_score(num_heads, method, bidirectional, hybridize,
ctx):
batch_size = 6
query_length = 25
mem_length = 20
query_head_units = 7
# Initialize the attention cell with relative positional embedding
base_layout = 'NKT'
base_use_einsum = False
if method == 'shaw':
num_buckets = None
max_distance = 20
elif method == 't5':
num_buckets = 10
max_distance = 20
elif method == 'transformer_xl':
num_buckets = None
max_distance = None
else:
raise NotImplementedError
base_score_cell = RelAttentionScoreCell(query_units=num_heads *
query_head_units,
num_heads=num_heads,
dropout=0.0,
method=method,
num_buckets=num_buckets,
max_distance=max_distance,
layout=base_layout,
use_einsum=base_use_einsum)
base_score_cell.initialize()
if hybridize:
base_score_cell.hybridize()
# Generate the data
query = mx.np.random.normal(
0,
1, (batch_size, num_heads, query_length, query_head_units),
dtype=np.float32)
if method != 't5':
rel_score_grad = mx.np.random.normal(
0,
1, (batch_size, num_heads, query_length, mem_length),
dtype=np.float32)
else:
rel_score_grad = mx.np.random.normal(
0, 1, (num_heads, query_length, mem_length), dtype=np.float32)
query_positions = mx.np.arange(query_length, dtype=np.int32)
mem_positions = mx.np.arange(mem_length, dtype=np.int32)
rel_positions = mx.np.expand_dims(query_positions, axis=-1)\
- mx.np.expand_dims(mem_positions, axis=0)
mask = mx.np.random.randint(0,
2, (batch_size, query_length, mem_length),
dtype=np.int32)
query.attach_grad()
with mx.autograd.record():
rel_score = base_score_cell(rel_positions, query)
rel_score.backward(rel_score_grad)
original_rel_score = rel_score.asnumpy()
original_grad_norm = grad_global_norm(
base_score_cell.collect_params().values())
original_query_grad_norm = np.linalg.norm(query.grad.asnumpy())
assert original_grad_norm > 0
# 1. Test for permutation equivariant
# We can permutate the query, rel_positions and the rel_score_grad and the result should
# always be the same.
query_perm = mx.np.array(np.random.permutation(query_length),
dtype=np.int32)
mem_perm = mx.np.array(np.random.permutation(mem_length, ), dtype=np.int32)
query.grad[:] = 0
with mx.autograd.record():
rel_score = base_score_cell(rel_positions[query_perm, :][:, mem_perm],
query[:, :, query_perm, :])
if method != 't5':
rel_score.backward(rel_score_grad[:, :, query_perm, :][:, :, :,
mem_perm])
else:
rel_score.backward(rel_score_grad[:, query_perm, :][:, :,
mem_perm])
permutated_out = rel_score.asnumpy()
permutated_grad_norm = grad_global_norm(
base_score_cell.collect_params().values())
permutated_query_grad_norm = np.linalg.norm(query.grad.asnumpy())
if method != 't5':
assert_allclose(
original_rel_score[:, :,
query_perm.asnumpy(), :][:, :, :,
mem_perm.asnumpy()],
permutated_out, 1E-4, 1E-4)
else:
assert_allclose(
original_rel_score[:, query_perm.asnumpy(), :][:, :,
mem_perm.asnumpy()],
permutated_out, 1E-4, 1E-4)
assert_allclose(permutated_grad_norm, original_grad_norm, 1E-4, 1E-4)
assert_allclose(permutated_query_grad_norm, original_query_grad_norm, 1E-4,
1E-4)
# 2. Test for different layout + use/not use einsum
for layout in ['NKT', 'NTK', 'TNK']:
for use_einsum in [False, True]:
if layout == base_layout and use_einsum == base_use_einsum:
continue
score_cell = RelAttentionScoreCell(query_units=num_heads *
query_head_units,
num_heads=num_heads,
dropout=0.0,
method=method,
num_buckets=num_buckets,
max_distance=max_distance,
layout=layout,
use_einsum=use_einsum)
score_cell.initialize()
if hybridize:
score_cell.hybridize()
score_cell.load_dict({
name: param.data()
for name, param in base_score_cell.collect_params().items()
})
query.attach_grad()
query.grad[:] = 0
with mx.autograd.record():
if layout == 'NKT':
rel_score = score_cell(rel_positions, query)
rel_score.backward(rel_score_grad)
elif layout == 'NTK':
rel_score = score_cell(rel_positions,
query.transpose((0, 2, 1, 3)))
rel_score.backward(rel_score_grad)
elif layout == 'TNK':
rel_score = score_cell(rel_positions,
query.transpose((2, 0, 1, 3)))
rel_score.backward(rel_score_grad)
else:
raise NotImplementedError
assert_allclose(rel_score.asnumpy(), original_rel_score, 1E-5,
1E-5)
layout_query_grad_norm = np.linalg.norm(query.grad.asnumpy())
assert_allclose(layout_query_grad_norm, original_query_grad_norm,
1E-5, 1E-5)
