def main(args):
print(args)
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
torch.cuda.set_device(args.device_id)
torch.manual_seed(args.seed)
# Load dataset
splits = ['train', 'valid']
if data.has_binary_files(args.data, splits):
dataset = data.load_dataset(
args.data, splits, args.source_lang, args.target_lang)
else:
dataset = data.load_raw_text_dataset(
args.data, splits, args.source_lang, args.target_lang)
if args.source_lang is None or args.target_lang is None:
# record inferred languages in args, so that it's saved in checkpoints
args.source_lang, args.target_lang = dataset.src, dataset.dst
print('| [{}] dictionary: {} types'.format(dataset.src, len(dataset.src_dict)))
print('| [{}] dictionary: {} types'.format(dataset.dst, len(dataset.dst_dict)))
for split in splits:
print('| {} {} {} examples'.format(args.data, split, len(dataset.splits[split])))
# Build model and criterion
model = models.build_model(args, dataset.src_dict, dataset.dst_dict)
criterion = criterions.build_criterion(args, dataset.src_dict, dataset.dst_dict)
print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__))
print('| num. model params: {}'.format(sum(p.data.numel() for p in model.parameters())))
# Build trainer
trainer = Trainer(args, 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
os.makedirs(args.save_dir, exist_ok=True)
checkpoint_path = os.path.join(args.save_dir, args.restore_file)
extra_state = trainer.load_checkpoint(checkpoint_path)
if extra_state is not None:
epoch = extra_state['epoch']
batch_offset = extra_state['batch_offset']
print('| loaded checkpoint {} (epoch {})'.format(checkpoint_path, epoch))
if batch_offset == 0:
trainer.lr_step(epoch)
epoch += 1
else:
epoch, batch_offset = 1, 0
# Train until the learning rate gets too small
max_epoch = args.max_epoch or math.inf
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
while lr > args.min_lr and epoch <= max_epoch:
# train for one epoch
train(args, trainer, dataset, epoch, batch_offset)
# evaluate on validate set
for k, subset in enumerate(args.valid_subset.split(',')):
val_loss = validate(args, trainer, dataset, subset, epoch)
if k == 0:
# only use first validation loss to update the learning schedule
lr = trainer.lr_step(epoch, val_loss)
# save checkpoint
if not args.no_save:
save_checkpoint(trainer, args, epoch, 0, val_loss)
epoch += 1
batch_offset = 0
train_meter.stop()
print('| done training in {:.1f} seconds'.format(train_meter.sum))
def main(args):
print(args)
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
torch.cuda.set_device(args.device_id)
torch.manual_seed(args.seed)
# Load dataset
splits = ['train', 'valid']
if data.has_binary_files(args.data, splits):
dataset = data.load_dataset(args.data, splits, args.source_lang,
args.target_lang)
else:
dataset = data.load_raw_text_dataset(args.data, splits,
args.source_lang,
args.target_lang)
if args.source_lang is None or args.target_lang is None:
# record inferred languages in args, so that it's saved in checkpoints
args.source_lang, args.target_lang = dataset.src, dataset.dst
print('| [{}] dictionary: {} types'.format(dataset.src,
len(dataset.src_dict)))
print('| [{}] dictionary: {} types'.format(dataset.dst,
len(dataset.dst_dict)))
for split in splits:
print('| {} {} {} examples'.format(args.data, split,
len(dataset.splits[split])))
# Build model and criterion
model = models.build_model(args, dataset.src_dict, dataset.dst_dict)
if 0.0 < args.rank_scale < 1.0:
patch_transformer(args, model)
if args.wd2fd:
no_decay, skiplist = [
'fc1', 'fc2', 'embed_tokens', 'embed_positions', 'out_embed'
], []
else:
no_decay, skiplist = [], [
'fc1', 'fc2', 'embed_tokens', 'embed_positions', 'out_embed'
]
else:
no_decay, skiplist = [], []
spectral_init(args, model)
criterion = criterions.build_criterion(args, dataset.src_dict,
dataset.dst_dict)
print('| model {}, criterion {}'.format(args.arch,
criterion.__class__.__name__))
print('| num. model params: {}'.format(
sum(p.data.numel() for p in model.parameters())))
# Build trainer
no_decay, skiplist = [], []
if args.wd2fd_quekey:
no_decay.extend(['_query.weight', '_key.weight'])
else:
skiplist.append('quekey')
if args.wd2fd_outval:
no_decay.extend(['_value.weight', 'output_perform.weight'])
else:
skiplist.append('outval')
trainer = Trainer(args,
model,
criterion,
skiplist=skiplist,
no_decay=no_decay)
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
os.makedirs(args.save_dir, exist_ok=True)
checkpoint_path = os.path.join(args.save_dir, args.restore_file)
extra_state = trainer.load_checkpoint(checkpoint_path)
if extra_state is not None:
epoch = extra_state['epoch']
batch_offset = extra_state['batch_offset']
print('| loaded checkpoint {} (epoch {})'.format(
checkpoint_path, epoch))
if batch_offset == 0:
trainer.lr_step(epoch)
epoch += 1
else:
epoch, batch_offset = 1, 0
if args.distributed_rank <= 0:
writer = SummaryWriter(args.save_dir)
with open(os.path.join(args.save_dir, 'args.json'), 'w') as f:
json.dump(vars(args), f, indent=4)
else:
writer = SummaryWriter(
os.path.join(args.save_dir, str(args.distributed_rank)))
# 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()
while lr > args.min_lr and epoch <= max_epoch:
if args.distributed_rank <= 0:
writer.add_scalar('hyper/lr', lr, epoch)
for form in ['QueKey', 'OutVal']:
frobnorm, nucnorm, bound, nonorth = [], [], [], []
for module in model.modules():
if hasattr(module, form.lower()):
U, VT = getattr(module, form.lower()).get_UVT()
for u, vt in zip(U, VT):
frobnorm.append(frobenius_norm(u, vt))
nucnorm.append(
torch.norm(torch.matmul(u, vt), 'nuc'))
bound.append(
(u.pow(2).sum() + vt.pow(2).sum()) / 2.)
nonorth.append(sum(non_orthogonality(u, vt)) / 2.)
writer.add_scalar('FrobNorm/' + form,
sum(frobnorm) / len(frobnorm), epoch)
writer.add_scalar('NucNorm/' + form,
sum(nucnorm) / len(nucnorm), epoch)
writer.add_scalar('NucNorm/' + form + '-Bound',
sum(bound) / len(bound), epoch)
writer.add_scalar('NonOrth/' + form,
sum(nonorth) / len(nonorth), epoch)
frobnorm, nucnorm, bound, nonorth = [], [], [], []
for name, module in model.named_modules():
if not any(
block in name for block in
['embed', '_query', '_key', '_value', 'output_perform']):
if hasattr(module,
'frobgrad') and not hasattr(module, 'get_UVT'):
U, VT = module.U.data, module.VT.data
frobnorm.append(frobenius_norm(U, VT))
nucnorm.append(torch.norm(torch.matmul(U, VT), 'nuc'))
nonorth.append(sum(non_orthogonality(U, VT)) / 2.)
bound.append((U.pow(2).sum() + VT.pow(2).sum()) / 2.)
elif hasattr(module, 'weight'):
frobnorm.append(torch.norm(module.weight.data))
nucnorm.append(torch.norm(module.weight.data, 'nuc'))
writer.add_scalar('FrobNorm/Linear',
sum(frobnorm) / len(frobnorm), epoch)
writer.add_scalar('NucNorm/Linear',
sum(nucnorm) / len(nucnorm), epoch)
if nonorth:
writer.add_scalar('NucNorm/Linear-Bound',
sum(bound) / len(bound), epoch)
writer.add_scalar('NonOrth/Linear',
sum(nonorth) / len(nonorth), epoch)
# train for one epoch
train(args, trainer, dataset, epoch, batch_offset)
# evaluate on validate set
if epoch % args.validate_interval == 0:
for k, subset in enumerate(args.valid_subset.split(',')):
val_loss = validate(args, trainer, dataset, subset, epoch)
if k == 0:
# only use first validation loss to update the learning schedule
lr = trainer.lr_step(epoch, val_loss)
# save checkpoint
if not args.no_save:
save_checkpoint(trainer, args, epoch, 0, val_loss)
for k in ['loss', 'nll_loss']:
writer.add_scalar('valid/' + k,
trainer.meters['valid_' + k].avg, epoch)
writer.add_scalar('train/' + k,
trainer.meters['train_' + k].avg, epoch)
else:
lr = trainer.lr_step(epoch)
epoch += 1
batch_offset = 0
if trainer.get_num_updates() >= max_update:
break
train_meter.stop()
print('| done training in {:.1f} seconds'.format(train_meter.sum))
writer.flush()
newpar = sum(p.numel() for p in model.parameters())
if 0.0 < args.rank_scale < 1.0:
args.rank_scale = 1.0
origpar = sum(p.numel() for p in models.build_model(
args, dataset.src_dict, dataset.dst_dict).parameters())
else:
origpar = newpar
if args.distributed_rank <= 0:
with open(os.path.join(args.save_dir, 'results.json'), 'w') as f:
json.dump(
{
'final validation loss':
trainer.meters['valid_nll_loss'].avg,
'original parameter count': origpar,
'compressed parameter count': newpar,
'compression ratio': newpar / origpar
},
f,
indent=4)
def main():
parser = options.get_parser('Trainer')
dataset_args = options.add_dataset_args(parser)
dataset_args.add_argument('--max-tokens',
default=6000,
type=int,
metavar='N',
help='maximum number of tokens in a batch')
dataset_args.add_argument('--max-sentences',
type=int,
metavar='N',
help='maximum number of sentences in a batch')
dataset_args.add_argument(
'--train-subset',
default='train',
metavar='SPLIT',
choices=['train', 'valid', 'test'],
help='data subset to use for training (train, valid, test)')
dataset_args.add_argument(
'--valid-subset',
default='valid',
metavar='SPLIT',
help='comma separated list ofdata subsets '
' to use for validation (train, valid, valid1,test, test1)')
options.add_optimization_args(parser)
options.add_checkpoint_args(parser)
options.add_model_args(parser)
args = utils.parse_args_and_arch(parser)
if args.no_progress_bar and args.log_format is None:
args.log_format = 'simple'
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
torch.manual_seed(args.seed)
# Load dataset
splits = ['train', 'valid']
if data.has_binary_files(args.data, splits):
dataset = data.load_dataset(args.data, splits, args.source_lang,
args.target_lang)
else:
dataset = data.load_raw_text_dataset(args.data, splits,
args.source_lang,
args.target_lang)
if args.source_lang is None or args.target_lang is None:
# record inferred languages in args, so that it's saved in checkpoints
args.source_lang, args.target_lang = dataset.src, dataset.dst
print(args)
print('| [{}] dictionary: {} types'.format(dataset.src,
len(dataset.src_dict)))
print('| [{}] dictionary: {} types'.format(dataset.dst,
len(dataset.dst_dict)))
for split in splits:
print('| {} {} {} examples'.format(args.data, split,
len(dataset.splits[split])))
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
num_gpus = torch.cuda.device_count()
print(
'| using {} GPUs (with max tokens per GPU = {} and max sentences per GPU = {})'
.format(num_gpus, args.max_tokens, args.max_sentences))
# Build model and criterion
model = utils.build_model(args, dataset.src_dict, dataset.dst_dict)
criterion = utils.build_criterion(args, dataset.src_dict, dataset.dst_dict)
print('| model {}, criterion {}'.format(args.arch,
criterion.__class__.__name__))
# The max number of positions can be different for train and valid
# e.g., RNNs may support more positions at test time than seen in training
max_positions_train = (args.max_source_positions,
args.max_target_positions)
max_positions_valid = (min(args.max_source_positions,
model.max_encoder_positions()),
min(args.max_target_positions,
model.max_decoder_positions()))
# Start multiprocessing
trainer = MultiprocessingTrainer(args, model, criterion)
# Load the latest checkpoint if one is available
checkpoint_path = os.path.join(args.save_dir, args.restore_file)
extra_state = trainer.load_checkpoint(checkpoint_path)
if extra_state is not None:
epoch = extra_state['epoch']
batch_offset = extra_state['batch_offset']
print('| loaded checkpoint {} (epoch {})'.format(
checkpoint_path, epoch))
if batch_offset == 0:
epoch += 1
else:
epoch, batch_offset = 1, 0
# Train until the learning rate gets too small
val_loss = None
max_epoch = args.max_epoch or math.inf
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
while lr > args.min_lr and epoch <= max_epoch:
# train for one epoch
train(args, epoch, batch_offset, trainer, dataset, max_positions_train,
num_gpus)
# evaluate on validate set
for k, subset in enumerate(args.valid_subset.split(',')):
val_loss = validate(args, epoch, trainer, dataset,
max_positions_valid, subset, num_gpus)
if k == 0:
if not args.no_save:
# save checkpoint
save_checkpoint(trainer, args, epoch, 0, val_loss)
# only use first validation loss to update the learning schedule
lr = trainer.lr_step(val_loss, epoch)
epoch += 1
batch_offset = 0
train_meter.stop()
print('| done training in {:.1f} seconds'.format(train_meter.sum))
# Stop multiprocessing
trainer.stop()
def main():
parser = options.get_parser('Trainer')
dataset_args = options.add_dataset_args(parser)
dataset_args.add_argument('--max-tokens', default=6000, type=int, metavar='N',
help='maximum number of tokens in a batch')
dataset_args.add_argument('--max-sentences', type=int, metavar='N',
help='maximum number of sentences in a batch')
dataset_args.add_argument('--train-subset', default='train', metavar='SPLIT',
choices=['train', 'valid', 'test'],
help='data subset to use for training (train, valid, test)')
dataset_args.add_argument('--valid-subset', default='valid', metavar='SPLIT',
help='comma separated list of data subsets '
' to use for validation (train, valid, valid1,test, test1)')
options.add_optimization_args(parser)
options.add_checkpoint_args(parser)
options.add_model_args(parser)
args = utils.parse_args_and_arch(parser)
if args.no_progress_bar and args.log_format is None:
args.log_format = 'simple'
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
torch.manual_seed(args.seed)
# Load dataset
splits = ['train', 'valid']
if data.has_binary_files(args.data, splits):
dataset = data.load_dataset(args.data, splits, args.source_lang, args.target_lang)
else:
dataset = data.load_raw_text_dataset(args.data, splits, args.source_lang, args.target_lang)
if args.source_lang is None or args.target_lang is None:
# record inferred languages in args, so that it's saved in checkpoints
args.source_lang, args.target_lang = dataset.src, dataset.dst
print(args)
print('| [{}] dictionary: {} types'.format(dataset.src, len(dataset.src_dict)))
print('| [{}] dictionary: {} types'.format(dataset.dst, len(dataset.dst_dict)))
for split in splits:
print('| {} {} {} examples'.format(args.data, split, len(dataset.splits[split])))
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
num_gpus = torch.cuda.device_count()
print('| using {} GPUs (with max tokens per GPU = {} and max sentences per GPU = {})'.format(
num_gpus, args.max_tokens, args.max_sentences))
# Build model and criterion
model = utils.build_model(args, dataset.src_dict, dataset.dst_dict)
criterion = utils.build_criterion(args, dataset.src_dict, dataset.dst_dict)
print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__))
# The max number of positions can be different for train and valid
# e.g., RNNs may support more positions at test time than seen in training
max_positions_train = (args.max_source_positions, args.max_target_positions)
max_positions_valid = (
min(args.max_source_positions, model.max_encoder_positions()),
min(args.max_target_positions, model.max_decoder_positions())
)
# Start multiprocessing
trainer = MultiprocessingTrainer(args, model, criterion)
# Load the latest checkpoint if one is available
checkpoint_path = os.path.join(args.save_dir, args.restore_file)
extra_state = trainer.load_checkpoint(checkpoint_path)
if extra_state is not None:
epoch = extra_state['epoch']
batch_offset = extra_state['batch_offset']
print('| loaded checkpoint {} (epoch {})'.format(checkpoint_path, epoch))
if batch_offset == 0:
epoch += 1
else:
epoch, batch_offset = 1, 0
# Train until the learning rate gets too small
val_loss = None
max_epoch = args.max_epoch or math.inf
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
while lr > args.min_lr and epoch <= max_epoch:
# train for one epoch
train(args, epoch, batch_offset, trainer, dataset, max_positions_train, num_gpus)
# evaluate on validate set
for k, subset in enumerate(args.valid_subset.split(',')):
val_loss = validate(args, epoch, trainer, dataset, max_positions_valid, subset, num_gpus)
if k == 0:
if not args.no_save:
# save checkpoint
save_checkpoint(trainer, args, epoch, 0, val_loss)
# only use first validation loss to update the learning schedule
lr = trainer.lr_step(val_loss, epoch)
epoch += 1
batch_offset = 0
train_meter.stop()
print('| done training in {:.1f} seconds'.format(train_meter.sum))
# Stop multiprocessing
trainer.stop()