def __init__(self, model_params, **kwargs):
"""Create a Trainer, and give it the parameters needed to instantiate the model
:param model_params: The model parameters
:param kwargs: See below
:Keyword Arguments:
* *nsteps* (`int`) -- If we should report every n-steps, this should be passed
* *ema_decay* (`float`) -- If we are doing an exponential moving average, what decay to us4e
* *clip* (`int`) -- If we are doing gradient clipping, what value to use
* *optim* (`str`) -- The name of the optimizer we are using
* *lr* (`float`) -- The learning rate we are using
* *mom* (`float`) -- If we are using SGD, what value to use for momentum
* *beta1* (`float`) -- Adam-specific hyper-param, defaults to `0.9`
* *beta2* (`float`) -- Adam-specific hyper-param, defaults to `0.999`
* *epsilon* (`float`) -- Adam-specific hyper-param, defaults to `1e-8
"""
super().__init__()
if type(model_params) is dict:
self.model = create_model_for('classify', **model_params)
else:
self.model = model_params
self.optimizer = EagerOptimizer(loss, **kwargs)
self.nsteps = kwargs.get('nsteps', six.MAXSIZE)
self._checkpoint = tf.train.Checkpoint(
optimizer=self.optimizer.optimizer, model=self.model)
checkpoint_dir = kwargs.get(
'checkpoint_dir', '{}-{}'.format("./tf-classify", os.getpid()))
log.warning(f"Checkpoint dir {checkpoint_dir}")
self.checkpoint_manager = tf.train.CheckpointManager(
self._checkpoint, directory=checkpoint_dir, max_to_keep=5)
strategy_type = kwargs.get('strategy_type', 'mirror')
self.eval_device = kwargs.get('eval_device', '/device:GPU:0')
if strategy_type == 'tpu':
self.eval_device = '/device:CPU:0'
gpus = 0
else:
gpus = int(kwargs.get('gpus', 1))
endpoint = kwargs.get('endpoint')
self.strategy = create_distribute_strategy(strategy_type, gpus,
endpoint)
def __init__(self, model_params, **kwargs):
super().__init__()
if type(model_params) is dict:
self.model = create_model_for('lm', **model_params)
else:
self.model = model_params
loss_fn = loss_with_state if self.model.requires_state else loss_without_state
self.optimizer = EagerOptimizer(loss_fn, **kwargs)
self.nsteps = kwargs.get('nsteps', 500)
self._checkpoint = tf.train.Checkpoint(
optimizer=self.optimizer.optimizer, model=self.model)
checkpoint_dir = '{}-{}'.format("./tf-lm", os.getpid())
self.checkpoint_manager = tf.train.CheckpointManager(
self._checkpoint, directory=checkpoint_dir, max_to_keep=5)
strategy_type = kwargs.get('strategy_type', 'mirror')
gpus = int(kwargs.get('gpus', 1))
endpoint = kwargs.get('endpoint')
self.strategy = create_distribute_strategy(strategy_type, gpus,
endpoint)
def __init__(self, model_params, **kwargs):
super().__init__()
if type(model_params) is dict:
self.model = create_model_for('seq2seq', **model_params)
else:
self.model = model_params
self.tgt_rlut = kwargs['tgt_rlut']
self.optimizer = EagerOptimizer(loss, **kwargs)
self.nsteps = kwargs.get('nsteps', 500)
self._checkpoint = tf.train.Checkpoint(
optimizer=self.optimizer.optimizer, model=self.model)
checkpoint_dir = '{}-{}'.format("./tf-seq2seq", os.getpid())
self.checkpoint_manager = tf.train.CheckpointManager(
self._checkpoint, directory=checkpoint_dir, max_to_keep=5)
strategy_type = kwargs.get('strategy_type', 'mirror')
gpus = int(kwargs.get('gpus', 1))
endpoint = kwargs.get('endpoint')
self.strategy = create_distribute_strategy(strategy_type, gpus,
endpoint)
self.bleu_n_grams = int(kwargs.get("bleu_n_grams", 4))
def main():
parser = ArgumentParser()
parser.add_argument("--basedir", type=str)
parser.add_argument("--train_dir",
type=str,
required=True,
help='Training directory')
parser.add_argument("--valid_dir",
type=str,
required=True,
help='Validation directory')
parser.add_argument(
"--train_md",
type=str,
help="Training metadata YAML, defaults to `{train_dir}/md.yml`")
parser.add_argument(
"--valid_md",
type=str,
help="Validation metadata YAML, defaults to `{valid_dir}/md.yml`")
parser.add_argument("--label_file",
type=str,
help="JSON file mapping labels to integers",
default="labels.json")
parser.add_argument("--dataset_key",
default="tlm",
help="dataset key for basedir")
parser.add_argument(
"--embed_type",
type=str,
default='default',
choices=["default", "positional", "learned-positional"],
help="register label of the embeddings")
parser.add_argument("--d_model",
type=int,
default=512,
help="Model dimension (and embedding dsz)")
parser.add_argument("--d_ff", type=int, default=2048, help="FFN dimension")
parser.add_argument("--num_heads",
type=int,
default=8,
help="Number of heads")
parser.add_argument("--num_layers",
type=int,
default=8,
help="Number of layers")
parser.add_argument("--num_train_workers",
type=int,
default=4,
help="Number train workers")
parser.add_argument("--distribute",
type=str,
default="mirror",
choices=["mirror", "tpu", "nccl"])
parser.add_argument("--tpu_ep",
type=str,
help="The TPU endpoint if using `distribute=tpu`")
parser.add_argument("--nctx",
type=int,
default=256,
help="Max input length")
parser.add_argument("--file_type",
default='tfrecord',
choices=['json', 'tfrecord'],
help="Glob pattern for data")
parser.add_argument("--batch_size",
type=int,
default=256,
help="Batch Size")
parser.add_argument("--subword_model_file",
type=str,
help="The BPE model file",
required=True)
parser.add_argument("--subword_vocab_file",
type=str,
help="The BPE subword vocab",
required=True)
parser.add_argument("--dropout", type=float, default=0.1, help="Dropout")
parser.add_argument("--ffn_pdrop",
type=float,
default=0.0,
help="Dropout in the dense stack")
parser.add_argument("--layer_drop",
type=float,
default=0.0,
help="LayerDrop to apply")
parser.add_argument("--optim",
default="adamw",
type=str,
help="Optimizer to use (defaults to adamw)")
parser.add_argument("--lr",
type=float,
default=4.0e-4,
help="Learning rate")
parser.add_argument("--clip",
type=float,
default=1.0,
help="Clipping gradient norm")
parser.add_argument("--weight_decay",
type=float,
default=1.0e-2,
help="Weight decay")
parser.add_argument("--epochs",
type=int,
default=32,
help="Num training epochs")
parser.add_argument(
"--restart",
type=str2bool,
help="Option allows you to restart from a previous checkpoint")
parser.add_argument("--warmup_steps",
type=int,
default=10000,
help="Num warmup steps")
parser.add_argument("--saves_per_epoch",
type=int,
default=10,
help="The number of checkpoints to save per epoch")
parser.add_argument(
'--rpr_k',
help=
'Relative attention positional sizes pass 0 if you dont want relative attention',
type=int,
default=[8],
nargs='+')
parser.add_argument(
'--rpr_value_on',
type=str2bool,
default=True,
help=
"In relative attention, whether add positional correction to values in addition to the "
"correction to attention matrix")
parser.add_argument('--windowed_ra',
type=str2bool,
default=False,
help="whether prevent attention beyond rpr_k")
parser.add_argument("--strategy",
help="Training strategy, defaults to `mirror`",
choices=["mirror"])
parser.add_argument("--npz",
help="Should we write out NPZ files?",
type=str2bool,
default=False)
parser.add_argument("--tb",
help="Turn on tensorboard?",
type=str2bool,
default=False)
parser.add_argument(
"--convert_only",
help="Should we just convert this file to NPZ and exit?",
type=str2bool,
default=False)
args = parser.parse_args()
SET_TRAIN_FLAG(True)
if args.convert_only:
args.restart = True
if args.basedir is None:
args.basedir = f'lm-{args.dataset_key}-bpe-{os.getpid()}'
logging.basicConfig(level=logging.INFO)
logger.info(f"Writing results to {args.basedir}")
if args.tb:
logdir = f"logs/scalars/{os.getpid()}"
file_writer = tf.summary.create_file_writer(logdir + "/metrics")
file_writer.set_as_default()
logger.info(f"Set up tensorboard logdir {logdir}")
strategy = create_distribute_strategy(args.distribute, args.tpu_ep)
num_replicas = strategy.num_replicas_in_sync
logger.info(f"Using {num_replicas} replicas in this job.")
vectorizer = BPEVectorizer1D(model_file=args.subword_model_file,
vocab_file=args.subword_vocab_file,
mxlen=args.nctx)
vocab = {'x': vectorizer.vocab}
preproc_data = baseline.embeddings.load_embeddings(
'x',
dsz=args.d_model,
known_vocab=vocab['x'],
preserve_vocab_indices=True,
embed_type=args.embed_type)
vocabs = preproc_data['vocab']
train_md = args.train_md if args.train_md else os.path.join(
args.train_dir, 'md.yml')
num_train_samples = get_num_samples(train_md)
valid_md = args.valid_md if args.valid_md else os.path.join(
args.valid_dir, 'md.yml')
num_valid_samples = get_num_samples(valid_md)
labels = read_json_tf(args.label_file)
num_labels = len(labels)
def dataset_train_fn(input_context):
global_batchsz = args.batch_size
base_batchsz = input_context.get_per_replica_batch_size(global_batchsz)
ds = get_dataset(args.train_dir, args.file_type,
args.num_train_workers).batch(base_batchsz)
return ds.shard(input_context.num_input_pipelines,
input_context.input_pipeline_id)
train_loader = strategy.experimental_distribute_datasets_from_function(
dataset_train_fn)
def dataset_test_fn(input_context):
global_batchsz = args.batch_size
base_batchsz = input_context.get_per_replica_batch_size(global_batchsz)
ds = get_dataset(args.valid_dir,
args.file_type,
args.num_train_workers,
shuffle=False).batch(base_batchsz)
return ds.shard(input_context.num_input_pipelines,
input_context.input_pipeline_id)
valid_loader = strategy.experimental_distribute_datasets_from_function(
dataset_test_fn)
os.makedirs(args.basedir, exist_ok=True)
# We want to make sure to save our input vocab into the basedir for reuse later
write_json(vocabs, os.path.join(args.basedir, 'vocabs.json'))
embeddings = {'x': preproc_data['embeddings']}
logger.info("Loaded embeddings")
logger.info("Loaded datasets")
logger.info("Using embedding type [%s]", args.embed_type)
if len(args.rpr_k) == 0 or args.rpr_k[0] < 1:
args.rpr_k = None
elif len(args.rpr_k) == 1:
args.rpr_k = args.rpr_k[0]
model = TransformerTagger(num_labels, embeddings, **vars(args))
logger.info("Loaded model and loss")
steps_per_epoch = num_train_samples // args.batch_size
steps_per_valid_epoch = num_valid_samples // args.batch_size
update_on = steps_per_epoch // args.saves_per_epoch
report_on = max(10, update_on) // 10
logger.info(
f"Steps per epoch: {steps_per_epoch}. Saving checkpoint every {update_on} steps."
)
lr_decay = CosineDecaySchedulerTensorFlow(steps_per_epoch * args.epochs,
lr=args.lr)
linear_warmup = WarmupLinearSchedulerTensorFlow(args.warmup_steps,
lr=args.lr)
lr_sched = CompositeLRSchedulerTensorFlow(linear_warmup, lr_decay)
optimizer = EagerOptimizer(loss_function,
optim=args.optim,
lr_function=lr_sched,
weight_decay=args.weight_decay,
clip=args.clip,
lr=args.lr)
checkpoint = tf.train.Checkpoint(optimizer=optimizer.optimizer,
model=model)
checkpoint_manager = tf.train.CheckpointManager(checkpoint,
directory=args.basedir,
max_to_keep=5)
start_epoch = 0
if args.restart:
# The global step gets automatically updated here
# so we dont have to worry about our LR regimen
checkpoint.restore(checkpoint_manager.latest_checkpoint)
current_step = optimizer.global_step
start_epoch = current_step // steps_per_epoch
def _replicated_train_step(inputs):
"""This runs on a single replica"""
x, y = inputs
per_replica_loss = optimizer.update(model, {'x': x}, y, num_replicas)
return per_replica_loss
@tf.function
def _distributed_train_step(inputs: Tuple[tf.Tensor, tf.Tensor]):
"""Runs across multiple replicas and aggregates the results.
:param inputs:
:return:
"""
per_replica_loss = strategy.run(_replicated_train_step,
args=(inputs, ))
return strategy.reduce(tf.distribute.ReduceOp.SUM,
per_replica_loss,
axis=None)
def _replicated_test_step(inputs):
"""This runs on a single replica"""
x, y = inputs
per_replica_loss = loss_function(model, {'x': x}, y) / num_replicas
return per_replica_loss
@tf.function
def _distributed_test_step(inputs: Tuple[tf.Tensor, tf.Tensor]):
"""Runs across multiple replicas and aggregates the results.
:param inputs:
:return:
"""
per_replica_loss = strategy.run(_replicated_test_step, args=(inputs, ))
return strategy.reduce(tf.distribute.ReduceOp.SUM,
per_replica_loss,
axis=None)
timer = Timer()
with strategy.scope():
for epoch in range(start_epoch, args.epochs):
SET_TRAIN_FLAG(True)
logger.info('Starting epoch %d', epoch + 1)
avg_loss = Average('average_train_loss')
metrics = {}
timer.start()
train_iter = iter(train_loader)
for i in range(steps_per_epoch):
try:
loss = _distributed_train_step(next(train_iter))
avg_loss.update(loss.numpy().item())
tf.summary.scalar("train_loss",
data=loss,
step=optimizer.global_step)
except Exception as e:
logger.error(
f"Exception at training step {i+1}/{steps_per_epoch}. Skipping"
)
pass
if args.convert_only:
logger.warning(
"Convert only flag specified. Stopping after one step"
)
steps = optimizer.global_step.numpy()
npz_checkpoint = os.path.join(
args.basedir, f'checkpoint-step-{steps}.npz')
save_tlm_output_npz(model, npz_checkpoint)
return
steps = optimizer.global_step.numpy()
if (steps + 1) % report_on == 0:
logger.info(avg_loss)
if (steps + 1) % update_on == 0:
elapsed = timer.elapsed(True)
logger.info('elapsed time this epoch %d min', elapsed)
logger.info('elapsed step time %f steps/min', i / elapsed)
checkpoint_manager.save()
if args.npz:
npz_checkpoint = os.path.join(
args.basedir, f'checkpoint-step-{steps}.npz')
save_tlm_output_npz(model, npz_checkpoint)
# How much time elapsed in minutes
elapsed = timer.elapsed(True)
train_token_loss = avg_loss.avg
# This is the average training token-level loss across all machines
# This is the token-level training perplexity
train_token_ppl = math.exp(train_token_loss)
metrics['train_elapsed_min'] = elapsed
metrics['average_train_loss'] = train_token_loss
metrics['train_ppl'] = train_token_ppl
metrics['lr'] = float(
lr_sched(tf.cast(optimizer.global_step,
tf.float32)).numpy().item())
avg_valid_loss = Average('average_valid_loss')
timer.start()
SET_TRAIN_FLAG(False)
valid_iter = iter(valid_loader)
for i in range(steps_per_valid_epoch):
try:
valid_loss = _distributed_test_step(next(valid_iter))
tf.summary.scalar('valid_loss',
data=valid_loss,
step=optimizer.global_step)
avg_valid_loss.update(valid_loss.numpy().item())
except Exception as e:
logger.error(
f"Exception at validation step {i+1}/{steps_per_valid_epoch}. Skipping"
)
pass
valid_token_loss = avg_valid_loss.avg
valid_token_ppl = math.exp(valid_token_loss)
elapsed = timer.elapsed(True)
metrics['valid_elapsed_min'] = elapsed
metrics['average_valid_loss'] = valid_token_loss
metrics['average_valid_word_ppl'] = valid_token_ppl
logger.info(json.dumps(metrics, indent=4))
def train():
parser = ArgumentParser()
parser.add_argument("--basedir", type=str)
parser.add_argument("--train_dir",
type=str,
required=True,
help='Training directory')
parser.add_argument("--valid_dir",
type=str,
required=True,
help='Validation directory')
parser.add_argument(
"--train_md",
type=str,
help="Training metadata YAML, defaults to `{train_dir}/md.yml`")
parser.add_argument(
"--valid_md",
type=str,
help="Validation metadata YAML, defaults to `{valid_dir}/md.yml`")
parser.add_argument("--dataset_key",
default="tlm",
help="dataset key for basedir")
parser.add_argument(
"--embed_type",
type=str,
default='default',
choices=["default", "positional", "learned-positional"],
help="register label of the embeddings")
parser.add_argument("--gen_d_model",
type=int,
default=256,
help="Model dimension (and embedding dsz)")
parser.add_argument("--gen_d_ff",
type=int,
default=1024,
help="FFN dimension")
parser.add_argument(
"--gen_d_k",
type=int,
default=None,
help="Dimension per head. Use if num_heads=1 to reduce dims")
parser.add_argument("--gen_num_heads",
type=int,
default=8,
help="Number of heads")
parser.add_argument("--gen_num_layers",
type=int,
default=8,
help="Number of layers")
parser.add_argument(
'--gen_rpr_k',
help=
'Relative attention positional sizes pass 0 if you dont want relative attention',
type=int,
default=[8],
nargs='+')
parser.add_argument('--windowed_ra',
type=str2bool,
default=False,
help="whether prevent attention beyond rpr_k")
parser.add_argument("--gen_loss_scale",
type=float,
default=50.0,
help="Scaling for loss function")
parser.add_argument("--gen_dropout",
type=float,
default=0.1,
help="Dropout")
parser.add_argument(
'--discrim_rpr_k',
help=
'Relative attention positional sizes pass 0 if you dont want relative attention',
type=int,
default=[8],
nargs='+')
parser.add_argument("--discrim_d_model",
type=int,
default=512,
help="Model dimension (and embedding dsz)")
parser.add_argument("--discrim_d_ff",
type=int,
default=2048,
help="FFN dimension")
parser.add_argument(
"--discrim_d_k",
type=int,
default=None,
help="Dimension per head. Use if num_heads=1 to reduce dims")
parser.add_argument("--discrim_num_heads",
type=int,
default=8,
help="Number of heads")
parser.add_argument("--discrim_num_layers",
type=int,
default=8,
help="Number of layers")
parser.add_argument("--discrim_dropout",
type=float,
default=0.1,
help="Dropout")
parser.add_argument("--num_train_workers",
type=int,
default=4,
help="Number train workers")
parser.add_argument("--distribute",
type=str,
default="mirror",
choices=["mirror", "tpu", "nccl"])
parser.add_argument("--tpu_ep",
type=str,
help="The TPU endpoint if using `distribute=tpu`")
parser.add_argument("--nctx",
type=int,
default=256,
help="Max input length")
parser.add_argument("--file_type",
default='tfrecord',
choices=['json', 'tfrecord'],
help="Glob pattern for data")
parser.add_argument("--batch_size",
type=int,
default=256,
help="Batch Size")
parser.add_argument("--subword_model_file",
type=str,
help="The BPE model file",
required=True)
parser.add_argument("--subword_vocab_file",
type=str,
help="The BPE subword vocab",
required=True)
parser.add_argument("--optim",
default="adam",
type=str,
help="Optimizer to use (defaults to adam)")
parser.add_argument("--lr",
type=float,
default=4.0e-4,
help="Learning rate")
parser.add_argument("--clip",
type=float,
default=1.0,
help="Clipping gradient norm")
parser.add_argument("--weight_decay",
type=float,
default=1.0e-2,
help="Weight decay")
parser.add_argument("--epochs",
type=int,
default=32,
help="Num training epochs")
parser.add_argument(
"--restart",
type=str2bool,
help="Option allows you to restart from a previous checkpoint")
parser.add_argument("--warmup_steps",
type=int,
default=10000,
help="Num warmup steps")
parser.add_argument("--causal",
type=str2bool,
default=False,
help="Use CLM (causal) instead of MLM")
parser.add_argument("--saves_per_epoch",
type=int,
default=10,
help="The number of checkpoints to save per epoch")
parser.add_argument("--strategy",
help="Training strategy, defaults to `mirror`",
choices=["mirror"])
parser.add_argument("--npz",
help="Should we write out NPZ files?",
type=str2bool,
default=False)
parser.add_argument("--tb",
help="Turn on tensorboard?",
type=str2bool,
default=False)
parser.add_argument(
"--convert_only",
help="Should we just convert this file to NPZ and exit?",
type=str2bool,
default=False)
args = parser.parse_args()
SET_TRAIN_FLAG(True)
if args.convert_only:
args.restart = True
args.npz = True
if args.basedir is None:
args.basedir = f'discrim-{args.dataset_key}-bpe-{os.getpid()}'
logging.basicConfig(level=logging.INFO)
logger.info(f"Writing results to {args.basedir}")
if args.tb:
logdir = f"logs/scalars/{os.getpid()}"
file_writer = tf.summary.create_file_writer(logdir + "/metrics")
file_writer.set_as_default()
logger.info(f"Set up tensorboard logdir {logdir}")
strategy = create_distribute_strategy(args.distribute, args.tpu_ep)
num_replicas = strategy.num_replicas_in_sync
logger.info(f"Using {num_replicas} replicas in this job.")
vectorizer = BPEVectorizer1D(model_file=args.subword_model_file,
vocab_file=args.subword_vocab_file,
mxlen=args.nctx)
vocab = {'x': vectorizer.vocab}
gen_preproc_data = baseline.embeddings.load_embeddings(
'x',
dsz=args.gen_d_model,
known_vocab=vocab['x'],
preserve_vocab_indices=True,
embed_type=args.embed_type)
vocabs = gen_preproc_data['vocab']
discrim_preproc_data = baseline.embeddings.load_embeddings(
'x',
dsz=args.discrim_d_model,
known_vocab=vocab['x'],
preserve_vocab_indices=True,
embed_type=args.embed_type)
def dataset_train_fn(input_context):
batch_size = input_context.get_per_replica_batch_size(args.batch_size)
ds = get_dataset(args.train_dir, args.file_type,
args.num_train_workers).batch(batch_size)
return ds.shard(input_context.num_input_pipelines,
input_context.input_pipeline_id)
train_loader = strategy.experimental_distribute_datasets_from_function(
dataset_train_fn)
def dataset_test_fn(input_context):
batch_size = input_context.get_per_replica_batch_size(args.batch_size)
ds = get_dataset(args.valid_dir,
args.file_type,
args.num_train_workers,
shuffle=False).batch(batch_size)
return ds.shard(input_context.num_input_pipelines,
input_context.input_pipeline_id)
valid_loader = strategy.experimental_distribute_datasets_from_function(
dataset_test_fn)
train_md = args.train_md if args.train_md else os.path.join(
args.train_dir, 'md.yml')
num_train_samples = get_num_samples(train_md)
valid_md = args.valid_md if args.valid_md else os.path.join(
args.valid_dir, 'md.yml')
num_valid_samples = get_num_samples(valid_md)
os.makedirs(args.basedir, exist_ok=True)
# We want to make sure to save our input vocab into the basedir for reuse later
write_json(vocabs, os.path.join(args.basedir, 'vocabs.json'))
gen_embeddings = {'x': gen_preproc_data['embeddings']}
discrim_embeddings = {'x': discrim_preproc_data['embeddings']}
logger.info("Loaded embeddings")
logger.info("Loaded datasets")
logger.info("Using embedding type [%s]", args.embed_type)
if len(args.gen_rpr_k) == 0 or args.gen_rpr_k[0] < 1:
gen_rpr_k = None
elif len(args.gen_rpr_k) == 1:
gen_rpr_k = args.gen_rpr_k[0]
else:
gen_rpr_k = args.gen_rpr_k
if len(args.discrim_rpr_k) == 0 or args.discrim_rpr_k[0] < 1:
discrim_rpr_k = None
elif len(args.gen_rpr_k) == 1:
discrim_rpr_k = args.discrim_rpr_k[0]
else:
discrim_rpr_k = args.discrim_rpr_k
gen_model = TransformerMaskedLanguageModel.create(
gen_embeddings,
hsz=args.gen_d_model,
d_ff=args.gen_d_ff,
tie_weights=True,
dropout=args.gen_dropout,
gpu=False,
num_heads=args.gen_num_heads,
layers=args.gen_num_layers,
rpr_k=gen_rpr_k,
d_k=args.gen_d_k,
windowed_ra=args.windowed_ra,
src_keys=['x'],
tgt_key='x')
discrim_model = TransformerDiscriminator(discrim_embeddings,
d_model=args.discrim_d_model,
d_ff=args.discrim_d_ff,
dropout=args.discrim_dropout,
num_heads=args.discrim_num_heads,
layers=args.discrim_num_layers,
rpr_k=discrim_rpr_k,
d_k=args.discrim_d_k)
logger.info("Loaded model and loss")
steps_per_epoch = num_train_samples // args.batch_size
steps_per_valid_epoch = num_valid_samples // args.batch_size
update_on = steps_per_epoch // args.saves_per_epoch
report_on = max(10, update_on) // 10
logger.info(
f"Steps per epoch: {steps_per_epoch}. Saving checkpoint every {update_on} steps."
)
lr_decay = CosineDecaySchedulerTensorFlow(steps_per_epoch * args.epochs,
lr=args.lr)
linear_warmup = WarmupLinearSchedulerTensorFlow(args.warmup_steps,
lr=args.lr)
lr_sched = CompositeLRSchedulerTensorFlow(linear_warmup, lr_decay)
mask_value = vocabs.get("[MASK]", vocabs.get("<MASK>", -1))
if mask_value == -1:
logger.error("We could not find a suitable masking token in the vocab")
return
optimizer, clip = create_keras_optimizer(**vars(args))
discrim_checkpoint = tf.train.Checkpoint(optimizer=optimizer,
model=discrim_model)
discrim_checkpoint_manager = tf.train.CheckpointManager(
discrim_checkpoint,
directory=os.path.join(args.basedir, 'discrim'),
max_to_keep=5)
gen_checkpoint = tf.train.Checkpoint(optimizer=optimizer,
model=discrim_model)
gen_checkpoint_manager = tf.train.CheckpointManager(gen_checkpoint,
directory=os.path.join(
args.basedir,
'gen'),
max_to_keep=5)
mask_value = vocabs.get("[MASK]", vocabs.get("<MASK>", -1))
if mask_value == -1:
logger.error("We could not find a suitable masking token in the vocab")
return
if args.restart:
# The global step gets automatically updated here
# so we dont have to worry about our LR regimen
gen_checkpoint.restore(gen_checkpoint_manager.latest_checkpoint)
discrim_checkpoint.restore(
discrim_checkpoint_manager.latest_checkpoint)
def _replicated_train_step(inputs):
"""This runs on a single replica"""
noised_x, labels = inputs
with tf.GradientTape() as tape:
gen_loss_step, discrim_loss_step, acc = gen_vs_discrim(
noised_x, labels, gen_model, discrim_model, mask_value)
loss_value = (args.gen_loss_scale * gen_loss_step +
discrim_loss_step) / num_replicas
grads = tape.gradient(
loss_value,
gen_model.trainable_variables + discrim_model.trainable_variables)
grads, _ = tf.clip_by_global_norm(grads, clip)
optimizer.apply_gradients(
zip(
grads, gen_model.trainable_variables +
discrim_model.trainable_variables))
return loss_value, gen_loss_step, discrim_loss_step, acc
@tf.function
def _distributed_train_step(inputs: Tuple[tf.Tensor, tf.Tensor]):
"""Runs across multiple replicas and aggregates the results.
:param inputs:
:return:
"""
loss, gen_loss, discrim_loss, acc = strategy.run(
_replicated_train_step, args=(inputs, ))
sum_loss = strategy.reduce(tf.distribute.ReduceOp.SUM, loss, axis=None)
sum_gen_loss = strategy.reduce(tf.distribute.ReduceOp.SUM,
gen_loss,
axis=None)
sum_discrim_loss = strategy.reduce(tf.distribute.ReduceOp.SUM,
discrim_loss,
axis=None)
sum_acc = strategy.reduce(tf.distribute.ReduceOp.SUM, acc, axis=None)
return sum_loss, sum_gen_loss, sum_discrim_loss, sum_acc
def _replicated_test_step(inputs):
"""This runs on a single replica"""
noised_x, labels = inputs
gen_loss_step, discrim_loss_step, acc = gen_vs_discrim(
noised_x, labels, gen_model, discrim_model, mask_value)
loss_value = (args.gen_loss_scale * gen_loss_step +
discrim_loss_step) / num_replicas
return loss_value, gen_loss_step, discrim_loss_step, acc
@tf.function
def _distributed_test_step(inputs: Tuple[tf.Tensor, tf.Tensor]):
"""Runs across multiple replicas and aggregates the results.
:param inputs:
:return:
"""
loss, gen_loss, discrim_loss, acc = strategy.run(_replicated_test_step,
args=(inputs, ))
sum_loss = strategy.reduce(tf.distribute.ReduceOp.SUM, loss, axis=None)
sum_gen_loss = strategy.reduce(tf.distribute.ReduceOp.SUM,
gen_loss,
axis=None)
sum_discrim_loss = strategy.reduce(tf.distribute.ReduceOp.SUM,
discrim_loss,
axis=None)
sum_acc = strategy.reduce(tf.distribute.ReduceOp.SUM, acc, axis=None)
return sum_loss, sum_gen_loss, sum_discrim_loss, sum_acc
# This is the training loop
start_epoch = 0
timer = Timer()
with strategy.scope():
for epoch in range(start_epoch, args.epochs):
SET_TRAIN_FLAG(True)
logger.info('Starting epoch %d', epoch + 1)
avg_loss = Average('average_train_loss')
avg_gen_loss = Average('average_gen_loss')
avg_discrim_loss = Average('average_discrim_loss')
avg_acc = Average('average_train_acc')
metrics = {}
timer.start()
train_iter = iter(train_loader)
for i in range(steps_per_epoch):
loss, gen_loss, discrim_loss, acc = _distributed_train_step(
next(train_iter))
avg_loss.update(loss.numpy().item())
avg_gen_loss.update(gen_loss.numpy().item())
avg_discrim_loss.update(discrim_loss.numpy().item())
avg_acc.update(acc.numpy().item())
tf.summary.scalar("train_loss",
data=loss,
step=optimizer.iterations)
tf.summary.scalar("train_gen_loss",
data=gen_loss,
step=optimizer.iterations)
tf.summary.scalar("train_discrim_loss",
data=discrim_loss,
step=optimizer.iterations)
tf.summary.scalar("train_acc",
data=acc,
step=optimizer.iterations)
if args.convert_only:
logger.warning(
"Convert only flag specified. Stopping after one step"
)
steps = optimizer.iterations.numpy()
npz_checkpoint = os.path.join(args.basedir,
f'discrim-step-{steps}.npz')
save_tlm_npz(discrim_model, npz_checkpoint)
npz_checkpoint = os.path.join(args.basedir,
f'gen-step-{steps}.npz')
save_tlm_npz(gen_model, npz_checkpoint)
return
if (i + 1) % report_on == 0:
logging.info(avg_loss)
logging.info(avg_gen_loss)
logging.info(avg_discrim_loss)
logging.info(avg_acc)
if (i + 1) % update_on == 0:
elapsed = timer.elapsed(True)
logging.info('elapsed time this epoch %d min', elapsed)
logging.info('elapsed step time %f steps/min', i / elapsed)
gen_checkpoint_manager.save()
discrim_checkpoint_manager.save()
if args.npz:
steps = optimizer.iterations.numpy()
npz_checkpoint = os.path.join(
args.basedir, f'discrim-step-{steps}.npz')
save_tlm_npz(discrim_model, npz_checkpoint)
npz_checkpoint = os.path.join(args.basedir,
f'gen-step-{steps}.npz')
save_tlm_npz(gen_model, npz_checkpoint)
# This is the average training token-level loss across all machines
# This is the token-level training perplexity
metrics['train_elapsed_min'] = timer.elapsed(True)
metrics['average_train_loss'] = avg_loss.avg
metrics['average_gen_loss'] = avg_gen_loss.avg
metrics['average_discrim_loss'] = avg_discrim_loss.avg
metrics['average_train_acc'] = avg_acc.avg
metrics['lr'] = float(
lr_sched(tf.cast(optimizer.global_step,
tf.float32)).numpy().item())
avg_valid_loss = Average('average_valid_loss')
avg_valid_gen_loss = Average('average_valid_gen_loss')
avg_valid_discrim_loss = Average('average_valid_discrim_loss')
avg_valid_acc = Average('average_valid_acc')
timer.start()
SET_TRAIN_FLAG(False)
valid_iter = iter(valid_loader)
for i in range(steps_per_valid_epoch):
valid_loss, valid_gen_loss, valid_discrim_loss, valid_acc = _distributed_test_step(
next(valid_iter))
tf.summary.scalar('valid_loss',
data=valid_loss,
step=optimizer.iterations)
tf.summary.scalar('valid_gen_loss',
data=valid_gen_loss,
step=optimizer.iterations)
tf.summary.scalar('valid_discrim_loss',
data=valid_discrim_loss,
step=optimizer.iterations)
tf.summary.scalar('valid_acc',
data=valid_acc,
step=optimizer.iterations)
avg_valid_loss.update(valid_loss.numpy().item())
avg_valid_gen_loss.update(valid_gen_loss.numpy().item())
avg_valid_discrim_loss.update(
valid_discrim_loss.numpy().item())
avg_valid_acc.update(valid_acc.numpy().item())
metrics['valid_elapsed_min'] = timer.elapsed(True)
metrics['average_valid_loss'] = avg_valid_loss.avg
metrics['average_valid_gen_loss'] = avg_valid_gen_loss.avg
metrics['average_valid_discrim_loss'] = avg_valid_discrim_loss.avg
metrics['average_valid_acc'] = avg_valid_acc.avg
logger.info(json.dumps(metrics, indent=4))
def main():
parser = ArgumentParser()
parser.add_argument("--basedir", type=str)
parser.add_argument("--train_dir", type=str, required=True, help='Training directory')
parser.add_argument("--valid_dir", type=str, required=True, help='Validation directory')
parser.add_argument("--train_md", type=str, help="Training metadata YAML, defaults to `{train_dir}/md.yml`")
parser.add_argument("--valid_md", type=str, help="Validation metadata YAML, defaults to `{valid_dir}/md.yml`")
parser.add_argument("--dataset_key", default="tlm",
help="dataset key for basedir")
parser.add_argument("--embed_type", type=str, default='default',
choices=["default", "positional", "learned-positional"],
help="register label of the embeddings")
parser.add_argument("--d_model", type=int, default=512, help="Model dimension (and embedding dsz)")
parser.add_argument("--d_ff", type=int, default=2048, help="FFN dimension")
parser.add_argument("--d_k", type=int, default=None, help="Dimension per head. Use if num_heads=1 to reduce dims")
parser.add_argument("--num_heads", type=int, default=8, help="Number of heads")
parser.add_argument("--num_layers", type=int, default=8, help="Number of layers")
parser.add_argument("--num_train_workers", type=int, default=4, help="Number train workers")
parser.add_argument("--distribute", type=str, default="mirror", choices=["mirror", "tpu", "nccl"])
parser.add_argument("--tpu_ep", type=str, help="The TPU endpoint if using `distribute=tpu`")
parser.add_argument("--nctx", type=int, default=256, help="Max input length (x)")
parser.add_argument("--file_type", default='tfrecord', choices=['json', 'jsonl', 'tfrecord'], help="Glob pattern for data")
parser.add_argument("--batch_size", type=int, default=256, help="Batch Size")
parser.add_argument("--subword_model_file", type=str, help="The BPE model file", required=False)
parser.add_argument("--subword_vocab_file", type=str, help="The BPE subword vocab", required=True)
parser.add_argument("--subword_type", type=str, choices=["bpe", "wordpiece", "sentencepiece"], default="bpe")
parser.add_argument("--dropout", type=float, default=0.1, help="Dropout")
parser.add_argument("--layer_drop", type=float, default=0.0, help="LayerDrop to apply")
parser.add_argument("--ff_pdrop", type=float, default=0.1, help="Dropout in the dense stack")
parser.add_argument("--optim", default="adamw", type=str, help="Optimizer to use (defaults to adamw)")
parser.add_argument("--lr", type=float, default=4.0e-4, help="Learning rate")
parser.add_argument("--clip", type=float, default=1.0, help="Clipping gradient norm")
parser.add_argument("--weight_decay", type=float, default=1.0e-2, help="Weight decay")
parser.add_argument("--epochs", type=int, default=32, help="Num training epochs")
parser.add_argument("--restart", type=str2bool, help="Option allows you to restart from a previous checkpoint")
parser.add_argument("--warmup_steps", type=int, default=10000, help="Num warmup steps")
parser.add_argument("--saves_per_epoch", type=int, default=10, help="The number of checkpoints to save per epoch")
parser.add_argument('--rpr_k',
help='Relative attention positional sizes pass 0 if you dont want relative attention',
type=int, default=[8], nargs='+')
parser.add_argument('--ra_type', type=str, help="Specify a relative attention type")
parser.add_argument("--reduction_d_k", type=int, default=64, help="Dimensions of Key and Query in the single headed"
"reduction layers")
parser.add_argument("--reduction_type", type=str, default="2ha",
help="If using a dual encoder, specifies the reduction type")
parser.add_argument("--stacking_layers", type=int, nargs='+', default=[])
parser.add_argument("--loss", type=str, default='symmetric',
choices=['contrastive', 'symmetric'])
parser.add_argument("--learn_temp", type=str2bool, default=True,
help="If 'constrastive' or 'symmetric' loss, should we learn the temperature scaling")
parser.add_argument("--init_temp", type=float,
help="Initialize the temperature for 'contrastive' or 'symmetric' loss")
parser.add_argument("--npz", help="Should we write out NPZ files?", type=str2bool, default=False)
parser.add_argument("--tb", help="Turn on tensorboard?", type=str2bool, default=False)
parser.add_argument("--convert_only", help="Should we just convert this file to NPZ and exit?", type=str2bool, default=False)
parser.add_argument("--extra_tokens", help="What extra tokens should we use", nargs="+", default=["[CLS]", "[MASK]"])
args = parser.parse_args()
if args.tpu_ep is not None and args.file_type != 'tfrecord':
raise Exception("For TPUs, TFRecord format is required!")
SET_TRAIN_FLAG(True)
if args.convert_only:
args.restart = True
if args.basedir is None:
args.basedir = 'paired-{}-bpe-{}'.format(args.dataset_key, os.getpid())
logging.basicConfig(level=logging.INFO)
logger.info(f"Writing results to {args.basedir}")
if args.tb:
logdir = f"logs/scalars/{os.getpid()}"
file_writer = tf.summary.create_file_writer(logdir + "/metrics")
file_writer.set_as_default()
logger.info(f"Set up tensorboard logdir {logdir}")
strategy = create_distribute_strategy(args.distribute, args.tpu_ep, len(get_env_gpus(None)))
num_replicas = strategy.num_replicas_in_sync
logger.info(f"Using {num_replicas} replicas in this job.")
Vec1D = get_subword_vec1d(args.subword_type)
vectorizer = Vec1D(model_file=args.subword_model_file,
vocab_file=args.subword_vocab_file,
mxlen=args.nctx,
extra_tokens=args.extra_tokens)
preproc_data = baseline.embeddings.load_embeddings('x', dsz=args.d_model, known_vocab=vectorizer.vocab,
preserve_vocab_indices=True,
embed_type=args.embed_type)
vocabs = preproc_data['vocab']
train_md = args.train_md if args.train_md else os.path.join(args.train_dir, 'md.yml')
num_train_samples = get_num_samples(train_md)
valid_md = args.valid_md if args.valid_md else os.path.join(args.valid_dir, 'md.yml')
num_valid_samples = get_num_samples(valid_md)
is_curriculum = True if isinstance(num_train_samples, Mapping) else False
def dataset_train_fn(input_context):
global_batchsz = args.batch_size
base_batchsz = input_context.get_per_replica_batch_size(global_batchsz)
ds = None
num_shards = input_context.num_input_pipelines
index = input_context.input_pipeline_id
if is_curriculum:
for sub in num_train_samples.keys():
train_curr_dir = os.path.join(args.train_dir, str(sub))
batchsz_scale_factor = args.nctx // sub
this_batchsz = base_batchsz * batchsz_scale_factor
curr_ds = get_dataset(train_curr_dir, args.file_type, args.num_train_workers, num_shards, index).batch(this_batchsz, drop_remainder=True)
if ds is None:
ds = curr_ds
else:
ds = ds.concatenate(curr_ds)
else:
ds = get_dataset(args.train_dir, args.file_type, args.num_train_workers, num_shards, index).batch(base_batchsz)
return ds
train_loader = strategy.experimental_distribute_datasets_from_function(dataset_train_fn)
def dataset_test_fn(input_context):
global_batchsz = args.batch_size
base_batchsz = input_context.get_per_replica_batch_size(global_batchsz)
ds = None
num_shards = input_context.num_input_pipelines
index = input_context.input_pipeline_id
if is_curriculum:
for sub in num_valid_samples.keys():
valid_curr_dir = os.path.join(args.valid_dir, str(sub))
batchsz_scale_factor = args.nctx // sub
this_batchsz = base_batchsz * batchsz_scale_factor
curr_ds = get_dataset(valid_curr_dir, args.file_type, args.num_train_workers, num_shards, index, shuffle=False).batch(
this_batchsz, drop_remainder=True)
if ds is None:
ds = curr_ds
else:
ds = ds.concatenate(curr_ds)
else:
ds = get_dataset(args.valid_dir, args.file_type, args.num_train_workers, num_shards, index, shuffle=False).batch(base_batchsz)
return ds
valid_loader = strategy.experimental_distribute_datasets_from_function(dataset_test_fn)
os.makedirs(args.basedir, exist_ok=True)
# We want to make sure to save our input vocab into the basedir for reuse later
write_json(vocabs, os.path.join(args.basedir, 'vocabs.json'))
embeddings = preproc_data['embeddings']
logger.info("Loaded embeddings")
logger.info("Loaded datasets")
logger.info("Using embedding type [%s]", args.embed_type)
if len(args.rpr_k) == 0 or args.rpr_k[0] < 1:
rpr_k = None
elif len(args.rpr_k) == 1:
rpr_k = args.rpr_k[0]
else:
rpr_k = args.rpr_k
logger.info("Creating dual encoder")
model = PairedModel(embeddings, args.d_model, args.d_ff, args.dropout, args.num_heads, args.num_layers, rpr_k=rpr_k,
d_k=args.d_k, reduction_d_k=args.reduction_d_k, stacking_layers=args.stacking_layers,
ffn_pdrop=args.ff_pdrop, reduction_type=args.reduction_type, freeze_encoders=False,
ra_type=args.ra_type)
loss_function = model.create_loss(loss_type=args.loss, init_temp=args.init_temp, learn_temp=args.learn_temp)
logger.info("Loaded model and loss")
if is_curriculum:
steps_per_epoch = 0
steps_per_valid_epoch = 0
for k, v in num_train_samples.items():
steps_per_epoch += int(num_train_samples[k] // (args.batch_size * (args.nctx / k)))
for k, v in num_valid_samples.items():
steps_per_valid_epoch += int(num_valid_samples[k] // (args.batch_size * (args.nctx / k)))
else:
steps_per_epoch = num_train_samples // args.batch_size
steps_per_valid_epoch = num_valid_samples // args.batch_size
update_on = steps_per_epoch // args.saves_per_epoch
report_on = max(10, update_on) // 10
logger.info(f"Steps per epoch: {steps_per_epoch}. Saving checkpoint every {update_on} steps.")
lr_decay = CosineDecaySchedulerTensorFlow(steps_per_epoch * args.epochs, lr=args.lr)
linear_warmup = WarmupLinearSchedulerTensorFlow(args.warmup_steps, lr=args.lr)
lr_sched = CompositeLRSchedulerTensorFlow(linear_warmup, lr_decay)
optimizer = EagerOptimizer(loss_function, optim=args.optim, lr_function=lr_sched, weight_decay=args.weight_decay,
clip=args.clip, lr=args.lr)
checkpoint = tf.train.Checkpoint(optimizer=optimizer.optimizer, model=model)
checkpoint_manager = tf.train.CheckpointManager(checkpoint,
directory=args.basedir,
max_to_keep=5)
if args.restart:
# The global step gets automatically updated here
# so we dont have to worry about our LR regimen
checkpoint.restore(checkpoint_manager.latest_checkpoint)
def _replicated_train_step(inputs):
"""This runs on a single replica"""
x, y = inputs
per_replica_loss = optimizer.update(model, x, y, num_replicas)
return per_replica_loss
@tf.function
def _distributed_train_step(inputs: Tuple[tf.Tensor, tf.Tensor]):
"""Runs across multiple replicas and aggregates the results.
:param inputs:
:return:
"""
per_replica_loss = strategy.run(_replicated_train_step, args=(inputs,))
return strategy.reduce(tf.distribute.ReduceOp.SUM, per_replica_loss, axis=None)
def _replicated_test_step(inputs):
"""This runs on a single replica"""
x, y = inputs
per_replica_loss = loss_function(model, x, y) / num_replicas
return per_replica_loss
@tf.function
def _distributed_test_step(inputs: Tuple[tf.Tensor, tf.Tensor]):
"""Runs across multiple replicas and aggregates the results.
:param inputs:
:return:
"""
per_replica_loss = strategy.run(_replicated_test_step, args=(inputs,))
return strategy.reduce(tf.distribute.ReduceOp.SUM, per_replica_loss, axis=None)
# This is the training loop
start_epoch = 0
timer = Timer()
with strategy.scope():
for epoch in range(start_epoch, args.epochs):
SET_TRAIN_FLAG(True)
logger.info('Starting epoch %d', epoch + 1)
avg_loss = Average('average_train_loss')
metrics = {}
timer.start()
train_iter = iter(train_loader)
for i in range(steps_per_epoch):
loss = _distributed_train_step(next(train_iter))
avg_loss.update(loss.numpy().item())
tf.summary.scalar("train_loss", data=loss, step=optimizer.global_step)
if args.convert_only:
logger.warning("Convert only flag specified. Stopping after one step")
steps = optimizer.global_step.numpy()
npz_checkpoint = os.path.join(args.basedir, f'checkpoint-step-{steps}.npz')
save_transformer_de_npz(model, npz_checkpoint)
return
if (i + 1) % report_on == 0:
logging.info(avg_loss)
if (i + 1) % update_on == 0:
elapsed = timer.elapsed(True)
logging.info('elapsed time this epoch %d min', elapsed)
logging.info('elapsed step time %f steps/min', i / elapsed)
checkpoint_manager.save()
if args.npz:
steps = optimizer.global_step.numpy()
npz_checkpoint = os.path.join(args.basedir, f'checkpoint-step-{steps}.npz')
save_transformer_de_npz(model, npz_checkpoint)
# How much time elapsed in minutes
train_token_loss = avg_loss.avg
# This is the average training token-level loss across all machines
# This is the token-level training perplexity
train_token_ppl = math.exp(train_token_loss)
metrics['train_elapsed_min'] = timer.elapsed(True)
metrics['average_train_loss'] = train_token_loss
metrics['train_ppl'] = train_token_ppl
metrics['lr'] = float(lr_sched(tf.cast(optimizer.global_step, tf.float32)).numpy().item())
avg_valid_loss = Average('average_valid_loss')
timer.start()
SET_TRAIN_FLAG(False)
valid_iter = iter(valid_loader)
for i in range(steps_per_valid_epoch):
valid_loss = _distributed_test_step(next(valid_iter))
tf.summary.scalar('valid_loss', data=valid_loss, step=optimizer.global_step)
avg_valid_loss.update(valid_loss.numpy().item())
valid_token_loss = avg_valid_loss.avg
valid_token_ppl = math.exp(valid_token_loss)
metrics['valid_elapsed_min'] = timer.elapsed(True)
metrics['average_valid_loss'] = valid_token_loss
metrics['average_valid_word_ppl'] = valid_token_ppl
logger.info(json.dumps(metrics, indent=4))
def main():
parser = ArgumentParser()
parser.add_argument("--basedir", type=str)
parser.add_argument("--train_dir", type=str, required=True, help='Training directory')
parser.add_argument("--valid_dir", type=str, required=True, help='Validation directory')
parser.add_argument("--train_md", type=str, help="Training metadata YAML, defaults to `{train_dir}/md.yml`")
parser.add_argument("--valid_md", type=str, help="Validation metadata YAML, defaults to `{valid_dir}/md.yml`")
parser.add_argument("--dataset_key", default="tlm",
help="dataset key for basedir")
parser.add_argument("--embed_type", type=str, default='default',
choices=["default", "positional", "learned-positional"],
help="register label of the embeddings")
parser.add_argument("--d_model", type=int, default=512, help="Model dimension (and embedding dsz)")
parser.add_argument("--d_ff", type=int, default=2048, help="FFN dimension")
parser.add_argument("--d_k", type=int, default=None, help="Dimension per head. Use if num_heads=1 to reduce dims")
parser.add_argument("--num_heads", type=int, default=8, help="Number of heads")
parser.add_argument("--num_layers", type=int, default=8, help="Number of layers")
parser.add_argument("--num_train_workers", type=int, default=4, help="Number train workers")
parser.add_argument("--distribute", type=str, default="mirror", choices=["mirror", "tpu", "nccl"])
parser.add_argument("--tpu_ep", type=str, help="The TPU endpoint if using `distribute=tpu`")
parser.add_argument("--nctx", type=int, default=512, help="Max input length")
parser.add_argument("--file_type", default='tfrecord', choices=['json', 'tfrecord'], help="Glob pattern for data")
parser.add_argument("--batch_size", type=int, default=256, help="Batch Size")
parser.add_argument("--subword_model_file", type=str, help="The BPE model file", required=False)
parser.add_argument("--subword_vocab_file", type=str, help="The BPE subword vocab", required=False)
parser.add_argument("--subword_type", type=str, choices=["bpe", "wordpiece", "sentencepiece"], default="bpe")
parser.add_argument("--dropout", type=float, default=0.1, help="Dropout")
parser.add_argument("--ffn_pdrop", type=float, default=0.0, help="Dropout in the dense stack")
parser.add_argument("--layer_drop", type=float, default=0.0, help="LayerDrop to apply")
parser.add_argument("--optim", default="adamw", type=str, help="Optimizer to use (defaults to adamw)")
parser.add_argument("--lr", type=float, default=4.0e-4, help="Learning rate")
parser.add_argument("--clip", type=float, default=1.0, help="Clipping gradient norm")
parser.add_argument("--weight_decay", type=float, default=1.0e-2, help="Weight decay")
parser.add_argument("--epochs", type=int, default=32, help="Num training epochs")
parser.add_argument("--restart", type=str2bool, help="Option allows you to restart from a previous checkpoint")
parser.add_argument("--warmup_steps", type=int, default=10000, help="Num warmup steps")
parser.add_argument("--causal", type=str2bool, default=False, help="Use CLM (causal) instead of MLM")
parser.add_argument("--mlp", type=str2bool, default=False, help="Use Gated MLP")
parser.add_argument("--saves_per_epoch", type=int, default=10, help="The number of checkpoints to save per epoch")
parser.add_argument('--rpr_k',
help='Relative attention positional sizes pass 0 if you dont want relative attention',
type=int, default=[8], nargs='+')
parser.add_argument('--rpr_value_on', type=str2bool, default=True,
help="In relative attention, whether add positional correction to values in addition to the "
"correction to attention matrix")
parser.add_argument('--ra_type', type=str, help="Specify a relative attention type")
parser.add_argument('--windowed_ra', type=str2bool, default=False, help="whether prevent attention beyond rpr_k")
parser.add_argument("--strategy", help="Training strategy, defaults to `mirror`", choices=["mirror"])
parser.add_argument("--npz", help="Should we write out NPZ files?", type=str2bool, default=False)
parser.add_argument("--tb", help="Turn on tensorboard?", type=str2bool, default=False)
parser.add_argument("--convert_only", help="Should we just convert this file to NPZ and exit?", type=str2bool, default=False)
parser.add_argument("--extra_tokens", help="What extra tokens should we use", nargs="+", default=["[CLS]", "[MASK]"])
parser.add_argument("--eps", help="Epsilon", default=1e-6, type=float)
parser.add_argument("--beta2", help="Epsilon", default=0.98, type=float)
parser.add_argument("--grad_accum", help="Number of iterations to accum grads", default=1, type=int)
parser.add_argument("--transformer_type", help="Transformer layer type")
args = parser.parse_args()
SET_TRAIN_FLAG(True)
if args.convert_only:
args.restart = True
if args.basedir is None:
args.basedir = f'lm-{args.dataset_key}-bpe-{os.getpid()}'
logging.basicConfig(level=logging.INFO)
logger.info(f"Writing results to {args.basedir}")
if args.tb:
logdir = f"{args.basedir}/scalars/{os.getpid()}"
file_writer = tf.summary.create_file_writer(logdir + "/metrics")
file_writer.set_as_default()
logger.info(f"Set up tensorboard logdir {logdir}")
strategy = create_distribute_strategy(args.distribute, args.tpu_ep)
num_replicas = strategy.num_replicas_in_sync
logger.info(f"Using {num_replicas} replicas in this job.")
Vec1D = get_subword_vec1d(args.subword_type)
vectorizer = Vec1D(model_file=args.subword_model_file,
vocab_file=args.subword_vocab_file,
mxlen=args.nctx,
extra_tokens=args.extra_tokens)
vocab = {'x': vectorizer.vocab}
preproc_data = baseline.embeddings.load_embeddings('x', dsz=args.d_model, known_vocab=vocab['x'],
preserve_vocab_indices=True,
embed_type=args.embed_type)
vocabs = preproc_data['vocab']
train_md = args.train_md if args.train_md else os.path.join(args.train_dir, 'md.yml')
num_train_samples = get_num_samples(train_md)
valid_md = args.valid_md if args.valid_md else os.path.join(args.valid_dir, 'md.yml')
num_valid_samples = get_num_samples(valid_md)
is_curriculum = True if isinstance(num_train_samples, Mapping) else False
def dataset_train_fn(input_context):
global_batchsz = args.batch_size
base_batchsz = input_context.get_per_replica_batch_size(global_batchsz)
ds = None
num_shards = input_context.num_input_pipelines
index = input_context.input_pipeline_id
if is_curriculum:
for sub in num_train_samples.keys():
train_curr_dir = os.path.join(args.train_dir, str(sub))
batchsz_scale_factor = args.nctx // sub
this_batchsz = base_batchsz * batchsz_scale_factor
curr_ds = get_dataset(train_curr_dir, args.file_type, args.num_train_workers, num_shards, index, causal=args.causal).batch(this_batchsz, drop_remainder=True)
if ds is None:
ds = curr_ds
else:
ds = ds.concatenate(curr_ds)
else:
ds = get_dataset(args.train_dir, args.file_type, args.num_train_workers, num_shards, index, causal=args.causal).batch(base_batchsz)
return ds
train_loader = strategy.experimental_distribute_datasets_from_function(dataset_train_fn)
def dataset_test_fn(input_context):
global_batchsz = args.batch_size
base_batchsz = input_context.get_per_replica_batch_size(global_batchsz)
num_shards = input_context.num_input_pipelines
index = input_context.input_pipeline_id
ds = None
if is_curriculum:
for sub in num_valid_samples.keys():
valid_curr_dir = os.path.join(args.valid_dir, str(sub))
batchsz_scale_factor = args.nctx // sub
this_batchsz = base_batchsz * batchsz_scale_factor
curr_ds = get_dataset(valid_curr_dir, args.file_type, args.num_train_workers, num_shards, index, causal=args.causal).batch(
this_batchsz, drop_remainder=True)
if ds is None:
ds = curr_ds
else:
ds = ds.concatenate(curr_ds)
else:
ds = get_dataset(args.valid_dir, args.file_type, args.num_train_workers, num_shards, index, shuffle=False, causal=args.causal).batch(base_batchsz)
return ds
valid_loader = strategy.experimental_distribute_datasets_from_function(dataset_test_fn)
os.makedirs(args.basedir, exist_ok=True)
# We want to make sure to save our input vocab into the basedir for reuse later
write_json(vocabs, os.path.join(args.basedir, 'vocabs.json'))
embeddings = {'x': preproc_data['embeddings']}
logger.info("Loaded embeddings")
logger.info("Loaded datasets")
logger.info("Using embedding type [%s]", args.embed_type)
model = create_model(args, embeddings)
if isinstance(model, GatedMLPLanguageModel) and is_curriculum:
raise Exception("Variable tensor lengths not currently supported for gMLP")
logger.info("Loaded model and loss")
eff_batch_size = args.batch_size * args.grad_accum
logger.info(f"eff batch size: {eff_batch_size}, {args.batch_size}(b) x {args.grad_accum}(ga)")
if is_curriculum:
steps_per_epoch = 0
steps_per_valid_epoch = 0
for k, v in num_train_samples.items():
steps_per_epoch += int(num_train_samples[k] // (eff_batch_size * (args.nctx / k)))
for k, v in num_valid_samples.items():
steps_per_valid_epoch += int(num_valid_samples[k] // (args.batch_size * (args.nctx / k)))
else:
steps_per_epoch = num_train_samples // eff_batch_size
steps_per_valid_epoch = num_valid_samples // args.batch_size
update_on = steps_per_epoch // args.saves_per_epoch
report_on = max(10, update_on) // 10
logger.info(f"Steps per epoch: {steps_per_epoch}. Saving checkpoint every {update_on} steps.")
lr_decay = CosineDecaySchedulerTensorFlow(steps_per_epoch * args.epochs, lr=args.lr)
linear_warmup = WarmupLinearSchedulerTensorFlow(args.warmup_steps, lr=args.lr)
lr_sched = CompositeLRSchedulerTensorFlow(linear_warmup, lr_decay)
optimizer = EagerOptimizer(loss_function, optim=args.optim, lr_function=lr_sched, weight_decay=args.weight_decay, clip=args.clip,
lr=args.lr, epsilon=args.eps, beta2=args.beta2)
checkpoint = tf.train.Checkpoint(optimizer=optimizer.optimizer, model=model)
checkpoint_manager = tf.train.CheckpointManager(checkpoint,
directory=args.basedir,
max_to_keep=5)
grad_accum = GradientAccumulator()
start_epoch = 0
if args.restart:
# The global step gets automatically updated here
# so we dont have to worry about our LR regimen
checkpoint.restore(checkpoint_manager.latest_checkpoint)
current_step = optimizer.global_step
start_epoch = current_step // steps_per_epoch
def _replicated_forward_step(inputs):
"""This runs on a single replica"""
x, y = inputs
per_replica_grads, per_replica_loss = optimizer.get_grads_and_loss(model, {'x': x}, y, num_replicas * args.grad_accum)
grad_accum(per_replica_grads)
return per_replica_loss
def _replicated_optz_step():
optimizer.apply_grads(model, grad_accum.gradients)
@tf.function
def _distributed_optz_step():
strategy.run(_replicated_optz_step)
@tf.function
def _distributed_forward_step(inputs: Tuple[tf.Tensor, tf.Tensor]):
"""Runs across multiple replicas and aggregates the results.
:param inputs:
:return:
"""
per_replica_loss = strategy.run(_replicated_forward_step, args=(inputs,))
return strategy.reduce(tf.distribute.ReduceOp.SUM, per_replica_loss, axis=None)
def _replicated_test_step(inputs):
"""This runs on a single replica"""
x, y = inputs
per_replica_loss = loss_function(model, {'x': x}, y) / num_replicas
return per_replica_loss
@tf.function
def _distributed_test_step(inputs: Tuple[tf.Tensor, tf.Tensor]):
"""Runs across multiple replicas and aggregates the results.
:param inputs:
:return:
"""
per_replica_loss = strategy.run(_replicated_test_step, args=(inputs,))
return strategy.reduce(tf.distribute.ReduceOp.SUM, per_replica_loss, axis=None)
timer = Timer()
with strategy.scope():
for epoch in range(start_epoch, args.epochs):
timer.start()
SET_TRAIN_FLAG(True)
logger.info('Starting epoch %d', epoch + 1)
avg_loss = Average('average_train_loss')
metrics = {}
train_iter = iter(train_loader)
step_loss = 0
iterations = steps_per_epoch * args.batch_size
for i in range(iterations):
try:
loss = _distributed_forward_step(next(train_iter))
step_loss += loss
if (i + 1) % args.grad_accum == 0:
# This does a gradient update
_distributed_optz_step()
# Now reset the gradient accumulator
grad_accum.reset()
# Now update the loss info
tf.summary.scalar("train_loss", data=step_loss, step=optimizer.global_step)
avg_loss.update(step_loss.numpy().item())
# Now reset the loss
step_loss = 0
steps = optimizer.global_step.numpy()
if (steps + 1) % report_on == 0:
logger.info(avg_loss)
if (steps + 1) % update_on == 0:
elapsed = timer.elapsed(True)
logger.info('elapsed time this epoch %d min', elapsed)
logger.info('elapsed step time %f steps/min', i/elapsed)
checkpoint_manager.save()
if args.npz:
npz_checkpoint = os.path.join(args.basedir, f'checkpoint-step-{steps}.npz')
save_tlm_npz(model, npz_checkpoint)
except Exception as e:
logger.error(e)
logger.error(f"Exception at training iter {i+1}/{iterations}. Skipping")
pass
if args.convert_only:
logger.warning("Convert only flag specified. Stopping after one step")
steps = optimizer.global_step.numpy()
npz_checkpoint = os.path.join(args.basedir, f'checkpoint-step-{steps}.npz')
save_tlm_npz(model, npz_checkpoint)
return
# How much time elapsed in minutes
train_token_loss = avg_loss.avg
# This is the average training token-level loss across all machines
# This is the token-level training perplexity
train_token_ppl = math.exp(train_token_loss)
metrics['train_elapsed_min'] = timer.elapsed(True)
metrics['average_train_loss'] = train_token_loss
metrics['train_ppl'] = train_token_ppl
metrics['lr'] = float(lr_sched(tf.cast(optimizer.global_step, tf.float32)).numpy().item())
avg_valid_loss = Average('average_valid_loss')
timer.start()
SET_TRAIN_FLAG(False)
valid_iter = iter(valid_loader)
for i in range(steps_per_valid_epoch):
try:
valid_loss = _distributed_test_step(next(valid_iter))
tf.summary.scalar('valid_loss', data=valid_loss, step=optimizer.global_step)
avg_valid_loss.update(valid_loss.numpy().item())
except Exception as e:
logger.error(f"Exception at validation step {i+1}/{steps_per_valid_epoch}. Skipping")
pass
valid_token_loss = avg_valid_loss.avg
valid_token_ppl = math.exp(valid_token_loss)
metrics['valid_elapsed_min'] = timer.elapsed(True)
metrics['average_valid_loss'] = valid_token_loss
metrics['average_valid_word_ppl'] = valid_token_ppl
logger.info(json.dumps(metrics, indent=4))
