def create_model(embeddings, **kwargs):
unif = kwargs['unif']
if 'sess' not in kwargs:
kwargs['sess'] = tf.Session()
weight_initializer = tf.random_uniform_initializer(-unif, unif)
with tf.variable_scope('Model', initializer=weight_initializer):
lm = create_lang_model(BASELINE_LM_MODELS, embeddings, **kwargs)
return lm
def create_model(embeddings, **kwargs):
lm = create_lang_model(BASELINE_LM_MODELS, embeddings, **kwargs)
return lm
def run(basedir=None,
train_file=None,
valid_file=None,
dataset_key='tlm',
embed_type='default',
d_model=512,
d_ff=2048,
d_k=None,
num_heads=8,
num_layers=8,
num_train_workers=4,
nctx=256,
file_type='json',
batch_size=256,
subword_model_file=None,
subword_vocab_file=None,
dropout=0.1,
ffn_pdrop=0.0,
layer_drop=0.0,
lr_scheduler='cosine',
lr_decay_steps=None,
lr_decay_rate=None,
lr_alpha=0.0,
optim='adamw',
lr=4.0e-4,
clip=1.0,
weight_decay=1.0e-2,
epochs=32,
restart_from=None,
restart_tt=None,
warmup_steps=10000,
saves_per_epoch=10,
mlm=True,
preprocessed=True,
rpr_k=[8],
rpr_value_on=False,
windowed_ra=False,
device="cuda",
distributed=False,
local_rank=-1,
extra_tokens=["[CLS]", "[MASK]"],
do_early_stopping=False,
model_type='transformer-mlm',
modules=[],
ra_type=None,
transformer_type=None,
**kwargs):
if basedir is None:
basedir = 'lm-{}-bpe-{}'.format(dataset_key, os.getpid())
logging.basicConfig(
level=logging.INFO if local_rank in [-1, 0] else logging.WARN)
for module in modules:
import_user_module(module)
num_gpus = get_num_gpus_multiworker()
distributed = distributed or num_gpus > 1
logger.info(f"Using {num_gpus} GPUs in this job.")
do_on_demand_masking = mlm and not preprocessed
if do_on_demand_masking:
logger.info(f"On-demand masking is turned on")
if distributed:
device, updated_local_rank = init_distributed(local_rank)
local_rank = updated_local_rank
if file_type == 'tfrecord':
reader_type = 'tfrecord'
elif preprocessed:
reader_type = 'preprocessed'
else:
reader_type = 'lang'
reader = MultiFileDatasetReader(src_nctx=nctx,
model_file=subword_model_file,
vocab_file=subword_vocab_file,
file_type=file_type,
reader_type=reader_type,
record_keys=['x', 'y'] if mlm else ['x'],
extra_tokens=extra_tokens)
# This looks a bit funny but the streaming reader ignores our vocab and gives us the one from the subword_model
# However, we do need to get counts from our dataset for validation so we can calculate the perplexity
vocab = reader.build_vocab([valid_file])
# If we are not using chars, then use 'x' for both input and output
preproc_data = baseline.embeddings.load_embeddings(
'x',
dsz=d_model,
known_vocab=vocab['x'],
preserve_vocab_indices=True,
embed_type=embed_type)
vocabs = preproc_data['vocab']
os.makedirs(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(basedir, 'vocabs.json'))
embeddings = {'x': preproc_data['embeddings']}
logger.info("Loaded embeddings")
train_set = reader.load(train_file, vocabs)
valid_set = reader.load(valid_file,
vocabs,
distribute=False,
shuffle=False)
train_loader = DataLoader(train_set,
batch_size=batch_size,
num_workers=num_train_workers)
valid_loader = DataLoader(valid_set, batch_size=batch_size)
logger.info("Loaded datasets")
logger.info("Using embedding type [%s]", embed_type)
if 'mlm' in model_type:
mask_from = vocabs
vocab_size = len(mask_from)
mask_value = mask_from.get("[MASK]")
if mask_value == -1:
logger.error(
"We could not find a suitable masking token in the vocab")
return
if len(rpr_k) == 0 or rpr_k[0] < 1:
rpr_k = None
elif len(rpr_k) == 1:
rpr_k = None if rpr_k[0] == 0 else rpr_k[0]
if ra_type != None and ra_type != 'shaw' and rpr_k is not None:
print(
f"Relative attention mismatch. You requested {ra_type} with rpr set. Setting it to 0"
)
rpr_k = None
model = create_lang_model(
embeddings,
hsz=d_model,
nctx=nctx, # Only for gMLP
d_ff=d_ff,
tie_weights=True,
dropout=dropout,
gpu=False,
num_heads=num_heads,
layers=num_layers,
rpr_k=rpr_k,
d_k=d_k,
ffn_pdrop=ffn_pdrop,
windowed_ra=windowed_ra,
rpr_value_on=rpr_value_on,
layer_drop=layer_drop,
model_type=model_type,
ra_type=ra_type,
transformer_type=transformer_type,
src_keys=['x'],
tgt_key='x')
model.to(device)
loss_function = model.create_loss()
loss_function.to(device)
logger.info("Loaded model and loss")
steps_per_epoch = len(train_loader) // num_gpus
update_on = steps_per_epoch // saves_per_epoch
report_on = max(10, update_on) // 10
logger.info(
f"Steps per epoch per GPU: {steps_per_epoch}. Saving checkpoint every {update_on} steps."
)
lr_decay = get_lr_decay(lr_scheduler,
lr,
steps_per_epoch,
epochs,
logger,
decay_steps=lr_decay_steps,
decay_rate=lr_decay_rate,
alpha=lr_alpha)
linear_warmup = WarmupLinearSchedulerPyTorch(warmup_steps, lr=lr)
lr_sched = CompositeLRScheduler(linear_warmup, lr_decay, lr=lr)
global_step = 0
start_epoch = 0
if restart_from:
if restart_from.endswith('npz'):
load_tlm_npz(model, restart_from)
else:
model.load_state_dict(torch.load(restart_from))
vec = restart_from.split("-")
if restart_tt:
tick_type = restart_tt
else:
tick_type = vec[-2]
step_num = int(vec[-1].split(".")[0])
if tick_type == 'epoch':
start_epoch = step_num
global_step = start_epoch * steps_per_epoch
elif tick_type == 'step':
start_epoch = step_num // steps_per_epoch
global_step = step_num
else:
logger.warning(
f"The previous tick was {step_num} but command-line specifies to ignore, setting to 0"
)
logger.info(
"Restarting from a previous checkpoint %s.\n\tStarting at global_step=%d, epoch=%d",
restart_from, global_step, start_epoch + 1)
optimizer = OptimizerManager(model,
global_step,
optim=optim,
lr=lr,
lr_function=lr_sched,
weight_decay=weight_decay)
logger.info("Model has {:,} parameters".format(
sum(p.numel() for p in model.parameters() if p.requires_grad)))
# Prepare model for distributed training if needed
if distributed:
# This program assume pure data parallelism, each model is on a single gpu
# If we wanted to support model and data parallelism we would need to update
# the selection of gpus based on rank, it would need to select multiple ids
# based on rank, here we select only a single gpu and use it for input and
# output.
model = DistributedDataParallel(model,
device_ids=[device],
output_device=device,
find_unused_parameters=True)
logger.info("Model located on %s", device)
model_base = os.path.join(basedir, 'checkpoint')
steps = global_step
best_valid_loss = np.inf
timer = Timer()
for epoch in range(start_epoch, epochs):
avg_loss = Average('average_train_loss')
metrics = {}
optimizer.zero_grad()
timer.start()
model.train()
train_itr = iter(train_loader)
for i in range(steps_per_epoch):
batch = next(train_itr)
steps += 1
x, y = batch
inputs = x.to(device)
labels = y.to(device)
if do_on_demand_masking:
inputs, labels, _ = on_demand_mlm_masking(
inputs, labels, mask_value, vocab_size)
inputs = {'x': inputs}
labels = labels.contiguous()
logits = model(inputs, None)[0].contiguous()
if mlm:
loss = loss_function(logits, labels)
else:
shift_logits = logits[:, -1]
shift_labels = labels[:, 1:]
loss = loss_function(shift_logits, shift_labels)
loss.backward()
avg_loss.update(loss.item())
torch.nn.utils.clip_grad_norm_(model.parameters(), clip)
optimizer.step()
optimizer.zero_grad()
if (i + 1) % report_on == 0:
logging.info(avg_loss)
if (i + 1) % update_on == 0 and local_rank < 1:
elapsed = timer.elapsed(True)
logging.info('elapsed time this epoch %d min', elapsed)
logging.info('elapsed step time %f steps/min', i / elapsed)
logging.info('LR: %f', optimizer.current_lr)
if not do_early_stopping:
save_checkpoint(model, model_base, steps, tick_type='step')
else:
valid_token_loss = validate(model, loss_function,
valid_loader, avg_loss, timer,
metrics, do_on_demand_masking,
mlm, mask_value, vocab_size,
device)
if valid_token_loss < best_valid_loss:
best_valid_loss = valid_token_loss
logger.info(
f"New best valid loss: {best_valid_loss}. Saving checkpoint..."
)
save_checkpoint(model,
model_base,
steps,
tick_type='step')
model.train()
if not do_early_stopping:
_ = validate(model, loss_function, valid_loader, avg_loss, timer,
metrics, do_on_demand_masking, mlm, mask_value,
vocab_size, device)
save_checkpoint(model, model_base, epoch, tick_type='epoch')