def main(args, init_distributed=False):
utils.import_user_module(args)
utils.handle_save_path(args)
assert args.max_tokens is not None or args.max_sentences is not None, \
'Must specify batch size either with --max-tokens or --max-sentences'
# Initialize CUDA and distributed training
#if torch.cuda.is_available() and not args.cpu:
# torch.cuda.set_device(args.device_id)
torch.manual_seed(args.seed)
if init_distributed:
args.distributed_rank = distributed_utils.distributed_init(args)
if distributed_utils.is_master(args):
checkpoint_utils.verify_checkpoint_directory(args.save_dir)
# Print args
print(f"| Configs: {args}")
# Setup task, e.g., translation, language modeling, etc.
task = tasks.setup_task(args)
# Load valid dataset (we load training data below, based on the latest checkpoint)
for valid_sub_split in args.valid_subset.split(','):
task.load_dataset(valid_sub_split, combine=False, epoch=0)
# Build model and criterion
model = task.build_model(args)
criterion = task.build_criterion(args)
print(
f"| Model: {args.arch} \n| Criterion: {criterion.__class__.__name__}")
# Log architecture
if args.train_subtransformer:
print(" \n\n\t\tWARNING!!! Training one single SubTransformer\n\n")
print(
f"| SubTransformer Arch: {utils.get_subtransformer_config(args)} \n"
)
else:
print(" \n\n\t\tWARNING!!! Training SuperTransformer\n\n")
print(f"| SuperTransformer Arch: {model} \n")
# Log model size
if args.train_subtransformer:
print(
f"| SubTransformer size (without embedding weights): {model.get_sampled_params_numel(utils.get_subtransformer_config(args))}"
)
embed_size = args.decoder_embed_dim_subtransformer * len(task.tgt_dict)
print(f"| Embedding layer size: {embed_size} \n")
else:
model_s = 0
# if use model.state_dict, then will add 2 more parameters, they are encoder.version and decoder.version. Should not count them
for name, param in model.named_parameters():
if 'embed' not in name:
model_s += param.numel()
print(
f"| SuperTransofmer model size (without embedding weights): {model_s}"
)
print(
f"| Embedding layer size: {sum(p.numel() for p in model.parameters() if p.requires_grad) - model_s} \n"
)
# specify the length of the dummy input for profile
# for iwslt, the average length is 23, for wmt, that is 30
dummy_sentence_length_dict = {'iwslt': 23, 'wmt': 30}
if 'iwslt' in args.arch:
dummy_sentence_length = dummy_sentence_length_dict['iwslt']
elif 'wmt' in args.arch:
dummy_sentence_length = dummy_sentence_length_dict['wmt']
else:
raise NotImplementedError
dummy_src_tokens = [2] + [7] * (dummy_sentence_length - 1)
dummy_prev = [7] * (dummy_sentence_length - 1) + [2]
# profile the overall FLOPs number
if args.profile_flops:
import torchprofile
config_subtransformer = utils.get_subtransformer_config(args)
model.set_sample_config(config_subtransformer)
model.profile(mode=True)
macs = torchprofile.profile_macs(model,
args=(torch.tensor([dummy_src_tokens],
dtype=torch.long),
torch.tensor([30]),
torch.tensor([dummy_prev],
dtype=torch.long)))
model.profile(mode=False)
last_layer_macs = config_subtransformer['decoder'][
'decoder_embed_dim'] * dummy_sentence_length * len(task.tgt_dict)
print(f"| Total FLOPs: {macs * 2}")
print(f"| Last layer FLOPs: {last_layer_macs * 2}")
print(
f"| Total FLOPs without last layer: {(macs - last_layer_macs) * 2} \n"
)
exit(0)
with torch.autograd.set_detect_anomaly(True):
# Build trainer
trainer = Trainer(args, task, model, criterion)
print(f"| Training on {args.distributed_world_size} GPUs")
# print(f"| Max tokens per GPU = {args.max_tokens} and max sentences per GPU = {args.max_sentences} \n")
print(
f"| Max tokens per GPU = {args.max_tokens} and max sentences per GPU = {None} \n"
)
# Measure model latency, the program will exit after profiling latency
if args.latcpu or args.latgpu:
utils.measure_latency(args, model, dummy_src_tokens, dummy_prev)
exit(0)
# Load the latest checkpoint if one is available and restore the corresponding train iterator
extra_state, epoch_itr = checkpoint_utils.load_checkpoint(args, trainer)
# Evaluate the SubTransformer
if args.validate_subtransformer:
config = utils.get_subtransformer_config(args)
trainer.set_sample_config(config)
valid_loss = validate(args, trainer, task, epoch_itr, ['valid'],
'SubTransformer')
print(f"| SubTransformer validation loss:{valid_loss}")
# Loop boundaries
max_epoch = args.max_epoch or math.inf
max_update = args.max_update or math.inf
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
valid_subsets = args.valid_subset.split(',')
represent_configs = utils.get_represent_configs(args)
# Main training loop
while lr > args.stop_min_lr and epoch_itr.epoch < max_epoch and trainer.get_num_updates(
) < max_update:
# train for one epoch
train(args, trainer, task, epoch_itr)
if not args.disable_validation and epoch_itr.epoch % args.validate_interval == 0:
for k, v in represent_configs.items():
trainer.set_sample_config(config=v)
valid_losses = validate(args,
trainer,
task,
epoch_itr,
valid_subsets,
sampled_arch_name=k)
else:
valid_losses = [None]
# update the best loss and get current lr; the real lr scheduling is done in trainer.train_step()
lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0])
# save checkpoint epoch level
if epoch_itr.epoch % args.save_interval == 0:
checkpoint_utils.save_checkpoint(args, trainer, epoch_itr,
valid_losses[0])
train_meter.stop()
print('| Done training in {:.1f} seconds'.format(train_meter.sum))
def main(args):
utils.import_user_module(args)
utils.handle_save_path(args)
assert args.max_tokens is not None or args.max_sentences is not None, \
'Must specify batch size either with --max-tokens or --max-sentences'
# Initialize CUDA and distributed training
if torch.cuda.is_available() and not args.cpu:
torch.cuda.set_device(args.device_id)
torch.manual_seed(args.seed)
# Print args
print(args)
# Setup task, e.g., translation, language modeling, etc.
task = tasks.setup_task(args)
# Load valid dataset (we load training data below, based on the latest checkpoint)
for valid_sub_split in args.valid_subset.split(','):
task.load_dataset(valid_sub_split, combine=False, epoch=0)
# Build model and criterion
model = task.build_model(args)
criterion = task.build_criterion(args)
print(model)
# Build trainer
trainer = Trainer(args, task, model, criterion)
# Load the latest checkpoint if one is available and restore the corresponding train iterator
args.train_subset = 'valid' # no need to train, so just set a small subset to save loading time
extra_state, epoch_itr = checkpoint_utils.load_checkpoint(args, trainer)
# run evolutionary search to find the model with lowest loss and satisfies the latency requirement
evolver = Evolution(args, trainer, task, epoch_itr)
best_config = evolver.run_evo_search()
with open(args.write_config_path, 'w') as fid:
encoder_layer_num = best_config['encoder']['encoder_layer_num']
decoder_layer_num = best_config['decoder']['decoder_layer_num']
fid.write(
f"encoder-embed-dim-subtransformer: {best_config['encoder']['encoder_embed_dim']}\n"
)
fid.write(
f"decoder-embed-dim-subtransformer: {best_config['decoder']['decoder_embed_dim']}\n\n"
)
fid.write(
f"encoder-ffn-embed-dim-all-subtransformer: {best_config['encoder']['encoder_ffn_embed_dim'][:encoder_layer_num]}\n"
)
fid.write(
f"decoder-ffn-embed-dim-all-subtransformer: {best_config['decoder']['decoder_ffn_embed_dim'][:decoder_layer_num]}\n\n"
)
fid.write(
f"encoder-layer-num-subtransformer: {best_config['encoder']['encoder_layer_num']}\n"
)
fid.write(
f"decoder-layer-num-subtransformer: {best_config['decoder']['decoder_layer_num']}\n\n"
)
fid.write(
f"encoder-self-attention-heads-all-subtransformer: {best_config['encoder']['encoder_self_attention_heads'][:encoder_layer_num]}\n"
)
fid.write(
f"decoder-self-attention-heads-all-subtransformer: {best_config['decoder']['decoder_self_attention_heads'][:decoder_layer_num]}\n"
)
fid.write(
f"decoder-ende-attention-heads-all-subtransformer: {best_config['decoder']['decoder_ende_attention_heads'][:decoder_layer_num]}\n\n"
)
fid.write(
f"decoder-arbitrary-ende-attn-all-subtransformer: {best_config['decoder']['decoder_arbitrary_ende_attn'][:decoder_layer_num]}\n\n"
)
def main(args, init_distributed=False):
utils.import_user_module(args)
utils.handle_save_path(args)
assert args.max_tokens is not None or args.max_sentences is not None, \
'Must specify batch size either with --max-tokens or --max-sentences'
# Initialize CUDA and distributed training
if torch.cuda.is_available() and not args.cpu:
torch.cuda.set_device(args.device_id)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if init_distributed:
args.distributed_rank = distributed_utils.distributed_init(args)
if distributed_utils.is_master(args):
checkpoint_utils.verify_checkpoint_directory(args.save_dir)
# Print args
print(args)
# Setup task, e.g., translation, language modeling, etc.
task = tasks.setup_task(args)
# Load valid dataset (we load training data below, based on the latest checkpoint)
for valid_sub_split in args.valid_subset.split(','):
task.load_dataset(valid_sub_split, combine=False, epoch=0)
# Build model and criterion
model = task.build_model(args)
criterion = task.build_criterion(args)
print(model)
print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__))
print('| num. model params: {} (num. trained: {})'.format(
sum(p.numel() for p in model.parameters()),
sum(p.numel() for p in model.parameters() if p.requires_grad),
))
# Build trainer
trainer = Trainer(args, task, model, criterion)
print('| training on {} GPUs'.format(args.distributed_world_size))
print('| max tokens per GPU = {} and max sentences per GPU = {}'.format(
args.max_tokens,
args.max_sentences,
))
# Load the latest checkpoint if one is available and restore the
# corresponding train iterator
extra_state, epoch_itr = checkpoint_utils.load_checkpoint(args, trainer)
# Train until the learning rate gets too small
max_epoch = args.max_epoch or math.inf
max_update = args.max_update or math.inf
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
valid_subsets = args.valid_subset.split(',')
while (
lr > args.min_lr
and (
epoch_itr.epoch < max_epoch
# allow resuming training from the final checkpoint
or epoch_itr._next_epoch_itr is not None
)
and trainer.get_num_updates() < max_update
):
# train for one epoch
train(args, trainer, task, epoch_itr)
if not args.disable_validation and epoch_itr.epoch % args.validate_interval == 0:
valid_losses = validate(args, trainer, task, epoch_itr, valid_subsets)
else:
valid_losses = [None]
# only use first validation loss to update the learning rate
lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0])
# save checkpoint
if epoch_itr.epoch % args.save_interval == 0:
checkpoint_utils.save_checkpoint(args, trainer, epoch_itr, valid_losses[0])
# early stop
if should_stop_early(args, valid_losses[0]):
print('| Early stop since valid performance hasn\'t improved for last {} runs'.format(args.patience))
break
reload_dataset = ':' in getattr(args, 'data', '')
# sharded data: get train iterator for next epoch
epoch_itr = trainer.get_train_iterator(epoch_itr.epoch, load_dataset=reload_dataset)
train_meter.stop()
print('| done training in {:.1f} seconds'.format(train_meter.sum))