def generate_samplers(self):
# Generate train samplers
train_samplers = []
for i in range(self.num_proc):
print("Building training sampler for proc %d" % i)
t1 = time.time()
# for each GPU, allocate num_proc // num_GPU processes
train_sampler_head = dgl.contrib.sampling.EdgeSampler(
self.train_data,
seed_edges=F.tensor(self.train_data.edge_parts[i]),
batch_size=self.batch_size,
neg_sample_size=self.neg_sample_size,
chunk_size=self.neg_sample_size,
negative_mode='head',
num_workers=self.num_workers,
shuffle=True,
exclude_positive=False,
return_false_neg=False,
)
train_sampler_tail = dgl.contrib.sampling.EdgeSampler(
self.train_data,
seed_edges=F.tensor(self.train_data.edge_parts[i]),
batch_size=self.batch_size,
neg_sample_size=self.neg_sample_size,
chunk_size=self.neg_sample_size,
negative_mode='tail',
num_workers=self.num_workers,
shuffle=True,
exclude_positive=False,
return_false_neg=False,
)
print(train_sampler_head)
print(train_sampler_tail)
train_samplers.append(
NewBidirectionalOneShotIterator(
dataloader_head=train_sampler_head,
dataloader_tail=train_sampler_tail,
neg_chunk_size=self.neg_sample_size,
neg_sample_size=self.neg_sample_size,
is_chunked=True,
num_nodes=self.n_entities,
has_edge_importance=self.has_edge_importance,
))
print("Training sampler for proc {} created, it takes {} seconds".
format(i,
time.time() - t1))
def run(args, logger):
init_time_start = time.time()
# load dataset and samplers
dataset = get_dataset(args.data_path, args.dataset, args.format,
args.data_files)
if args.neg_sample_size_eval < 0:
args.neg_sample_size_eval = dataset.n_entities
args.batch_size = get_compatible_batch_size(args.batch_size,
args.neg_sample_size)
args.batch_size_eval = get_compatible_batch_size(args.batch_size_eval,
args.neg_sample_size_eval)
args.eval_filter = not args.no_eval_filter
if args.neg_deg_sample_eval:
assert not args.eval_filter, "if negative sampling based on degree, we can't filter positive edges."
train_data = TrainDataset(dataset, args, ranks=args.num_proc)
# if there is no cross partition relaiton, we fall back to strict_rel_part
args.strict_rel_part = args.mix_cpu_gpu and (train_data.cross_part
== False)
args.soft_rel_part = args.mix_cpu_gpu and args.soft_rel_part and train_data.cross_part
args.num_workers = 8 # fix num_worker to 8
if args.num_proc > 1:
train_samplers = []
for i in range(args.num_proc):
train_sampler_head = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
args.neg_sample_size,
mode='head',
num_workers=args.num_workers,
shuffle=True,
exclude_positive=False,
rank=i)
train_sampler_tail = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
args.neg_sample_size,
mode='tail',
num_workers=args.num_workers,
shuffle=True,
exclude_positive=False,
rank=i)
train_samplers.append(
NewBidirectionalOneShotIterator(train_sampler_head,
train_sampler_tail,
args.neg_sample_size,
args.neg_sample_size, True,
dataset.n_entities))
train_sampler = NewBidirectionalOneShotIterator(
train_sampler_head, train_sampler_tail, args.neg_sample_size,
args.neg_sample_size, True, dataset.n_entities)
else: # This is used for debug
train_sampler_head = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
args.neg_sample_size,
mode='head',
num_workers=args.num_workers,
shuffle=True,
exclude_positive=False)
train_sampler_tail = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
args.neg_sample_size,
mode='tail',
num_workers=args.num_workers,
shuffle=True,
exclude_positive=False)
train_sampler = NewBidirectionalOneShotIterator(
train_sampler_head, train_sampler_tail, args.neg_sample_size,
args.neg_sample_size, True, dataset.n_entities)
if args.valid or args.test:
if len(args.gpu) > 1:
args.num_test_proc = args.num_proc if args.num_proc < len(
args.gpu) else len(args.gpu)
else:
args.num_test_proc = args.num_proc
eval_dataset = EvalDataset(dataset, args)
if args.valid:
if args.num_proc > 1:
valid_sampler_heads = []
valid_sampler_tails = []
for i in range(args.num_proc):
valid_sampler_head = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_eval,
args.neg_sample_size_eval,
args.eval_filter,
mode='chunk-head',
num_workers=args.num_workers,
rank=i,
ranks=args.num_proc)
valid_sampler_tail = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_eval,
args.neg_sample_size_eval,
args.eval_filter,
mode='chunk-tail',
num_workers=args.num_workers,
rank=i,
ranks=args.num_proc)
valid_sampler_heads.append(valid_sampler_head)
valid_sampler_tails.append(valid_sampler_tail)
else: # This is used for debug
valid_sampler_head = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_eval,
args.neg_sample_size_eval,
args.eval_filter,
mode='chunk-head',
num_workers=args.num_workers,
rank=0,
ranks=1)
valid_sampler_tail = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_eval,
args.neg_sample_size_eval,
args.eval_filter,
mode='chunk-tail',
num_workers=args.num_workers,
rank=0,
ranks=1)
if args.test:
if args.num_test_proc > 1:
test_sampler_tails = []
test_sampler_heads = []
for i in range(args.num_test_proc):
test_sampler_head = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_eval,
args.neg_sample_size_eval,
args.eval_filter,
mode='chunk-head',
num_workers=args.num_workers,
rank=i,
ranks=args.num_test_proc)
test_sampler_tail = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_eval,
args.neg_sample_size_eval,
args.eval_filter,
mode='chunk-tail',
num_workers=args.num_workers,
rank=i,
ranks=args.num_test_proc)
test_sampler_heads.append(test_sampler_head)
test_sampler_tails.append(test_sampler_tail)
else:
test_sampler_head = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_eval,
args.neg_sample_size_eval,
args.eval_filter,
mode='chunk-head',
num_workers=args.num_workers,
rank=0,
ranks=1)
test_sampler_tail = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_eval,
args.neg_sample_size_eval,
args.eval_filter,
mode='chunk-tail',
num_workers=args.num_workers,
rank=0,
ranks=1)
# load model
model = load_model(logger, args, dataset.n_entities, dataset.n_relations)
if args.num_proc > 1 or args.async_update:
model.share_memory()
# We need to free all memory referenced by dataset.
eval_dataset = None
dataset = None
print('Total initialize time {:.3f} seconds'.format(time.time() -
init_time_start))
# train
start = time.time()
rel_parts = train_data.rel_parts if args.strict_rel_part or args.soft_rel_part else None
cross_rels = train_data.cross_rels if args.soft_rel_part else None
if args.num_proc > 1:
procs = []
barrier = mp.Barrier(args.num_proc)
for i in range(args.num_proc):
valid_sampler = [valid_sampler_heads[i], valid_sampler_tails[i]
] if args.valid else None
proc = mp.Process(target=train_mp,
args=(args, model, train_samplers[i],
valid_sampler, i, rel_parts, cross_rels,
barrier))
procs.append(proc)
proc.start()
for proc in procs:
proc.join()
else:
valid_samplers = [valid_sampler_head, valid_sampler_tail
] if args.valid else None
train(args, model, train_sampler, valid_samplers, rel_parts=rel_parts)
print('training takes {} seconds'.format(time.time() - start))
if args.save_emb is not None:
if not os.path.exists(args.save_emb):
os.mkdir(args.save_emb)
model.save_emb(args.save_emb, args.dataset)
# We need to save the model configurations as well.
conf_file = os.path.join(args.save_emb, 'config.json')
with open(conf_file, 'w') as outfile:
json.dump(
{
'dataset': args.dataset,
'model': args.model_name,
'emb_size': args.hidden_dim,
'max_train_step': args.max_step,
'batch_size': args.batch_size,
'neg_sample_size': args.neg_sample_size,
'lr': args.lr,
'gamma': args.gamma,
'double_ent': args.double_ent,
'double_rel': args.double_rel,
'neg_adversarial_sampling': args.neg_adversarial_sampling,
'adversarial_temperature': args.adversarial_temperature,
'regularization_coef': args.regularization_coef,
'regularization_norm': args.regularization_norm
},
outfile,
indent=4)
# test
if args.test:
start = time.time()
if args.num_test_proc > 1:
queue = mp.Queue(args.num_test_proc)
procs = []
for i in range(args.num_test_proc):
proc = mp.Process(target=test_mp,
args=(args, model, [
test_sampler_heads[i],
test_sampler_tails[i]
], i, 'Test', queue))
procs.append(proc)
proc.start()
total_metrics = {}
metrics = {}
logs = []
for i in range(args.num_test_proc):
log = queue.get()
logs = logs + log
for metric in logs[0].keys():
metrics[metric] = sum([log[metric]
for log in logs]) / len(logs)
for k, v in metrics.items():
print('Test average {} : {}'.format(k, v))
for proc in procs:
proc.join()
else:
test(args, model, [test_sampler_head, test_sampler_tail])
print('testing takes {:.3f} seconds'.format(time.time() - start))
def run(args, logger):
# load dataset and samplers
dataset = get_dataset(args.data_path, args.dataset, args.format)
n_entities = dataset.n_entities
n_relations = dataset.n_relations
if args.neg_sample_size_test < 0:
args.neg_sample_size_test = n_entities
train_data = TrainDataset(dataset, args, ranks=args.num_proc)
if args.num_proc > 1:
train_samplers = []
for i in range(args.num_proc):
train_sampler_head = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
mode='PBG-head',
num_workers=args.num_worker,
shuffle=True,
exclude_positive=True,
rank=i)
train_sampler_tail = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
mode='PBG-tail',
num_workers=args.num_worker,
shuffle=True,
exclude_positive=True,
rank=i)
train_samplers.append(
NewBidirectionalOneShotIterator(train_sampler_head,
train_sampler_tail, True,
n_entities))
else:
train_sampler_head = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
mode='PBG-head',
num_workers=args.num_worker,
shuffle=True,
exclude_positive=True)
train_sampler_tail = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
mode='PBG-tail',
num_workers=args.num_worker,
shuffle=True,
exclude_positive=True)
train_sampler = NewBidirectionalOneShotIterator(
train_sampler_head, train_sampler_tail, True, n_entities)
if args.valid or args.test:
eval_dataset = EvalDataset(dataset, args)
if args.valid:
# Here we want to use the regualr negative sampler because we need to ensure that
# all positive edges are excluded.
if args.num_proc > 1:
valid_sampler_heads = []
valid_sampler_tails = []
for i in range(args.num_proc):
valid_sampler_head = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_valid,
mode='PBG-head',
num_workers=args.num_worker,
rank=i,
ranks=args.num_proc)
valid_sampler_tail = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_valid,
mode='PBG-tail',
num_workers=args.num_worker,
rank=i,
ranks=args.num_proc)
valid_sampler_heads.append(valid_sampler_head)
valid_sampler_tails.append(valid_sampler_tail)
else:
valid_sampler_head = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_valid,
mode='PBG-head',
num_workers=args.num_worker,
rank=0,
ranks=1)
valid_sampler_tail = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_valid,
mode='PBG-tail',
num_workers=args.num_worker,
rank=0,
ranks=1)
if args.test:
# Here we want to use the regualr negative sampler because we need to ensure that
# all positive edges are excluded.
if args.num_proc > 1:
test_sampler_tails = []
test_sampler_heads = []
for i in range(args.num_proc):
test_sampler_head = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_test,
mode='PBG-head',
num_workers=args.num_worker,
rank=i,
ranks=args.num_proc)
test_sampler_tail = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_test,
mode='PBG-tail',
num_workers=args.num_worker,
rank=i,
ranks=args.num_proc)
test_sampler_heads.append(test_sampler_head)
test_sampler_tails.append(test_sampler_tail)
else:
test_sampler_head = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_test,
mode='PBG-head',
num_workers=args.num_worker,
rank=0,
ranks=1)
test_sampler_tail = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_test,
mode='PBG-tail',
num_workers=args.num_worker,
rank=0,
ranks=1)
# We need to free all memory referenced by dataset.
eval_dataset = None
dataset = None
# load model
model = load_model(logger, args, n_entities, n_relations)
if args.num_proc > 1:
model.share_memory()
# train
start = time.time()
if args.num_proc > 1:
procs = []
for i in range(args.num_proc):
valid_samplers = [valid_sampler_heads[i], valid_sampler_tails[i]
] if args.valid else None
proc = mp.Process(target=train,
args=(args, model, train_samplers[i],
valid_samplers))
procs.append(proc)
proc.start()
for proc in procs:
proc.join()
else:
valid_samplers = [valid_sampler_head, valid_sampler_tail
] if args.valid else None
train(args, model, train_sampler, valid_samplers)
print('training takes {} seconds'.format(time.time() - start))
if args.save_emb is not None:
if not os.path.exists(args.save_emb):
os.mkdir(args.save_emb)
model.save_emb(args.save_emb, args.dataset)
# test
if args.test:
if args.num_proc > 1:
procs = []
for i in range(args.num_proc):
proc = mp.Process(target=test,
args=(args, model, [
test_sampler_heads[i],
test_sampler_tails[i]
]))
procs.append(proc)
proc.start()
for proc in procs:
proc.join()
else:
test(args, model, [test_sampler_head, test_sampler_tail])
def run(args, logger):
# load dataset and samplers
dataset = get_dataset(args.data_path, args.dataset, args.format)
n_entities = dataset.n_entities
n_relations = dataset.n_relations
if args.neg_sample_size_test < 0:
args.neg_sample_size_test = n_entities
args.eval_filter = not args.no_eval_filter
if args.neg_deg_sample_eval:
assert not args.eval_filter, "if negative sampling based on degree, we can't filter positive edges."
# When we generate a batch of negative edges from a set of positive edges,
# we first divide the positive edges into chunks and corrupt the edges in a chunk
# together. By default, the chunk size is equal to the negative sample size.
# Usually, this works well. But we also allow users to specify the chunk size themselves.
if args.neg_chunk_size < 0:
args.neg_chunk_size = args.neg_sample_size
if args.neg_chunk_size_valid < 0:
args.neg_chunk_size_valid = args.neg_sample_size_valid
if args.neg_chunk_size_test < 0:
args.neg_chunk_size_test = args.neg_sample_size_test
train_data = TrainDataset(dataset, args, ranks=args.num_proc)
if args.num_proc > 1:
train_samplers = []
for i in range(args.num_proc):
train_sampler_head = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
args.neg_chunk_size,
mode='chunk-head',
num_workers=args.num_worker,
shuffle=True,
exclude_positive=True,
rank=i)
train_sampler_tail = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
args.neg_chunk_size,
mode='chunk-tail',
num_workers=args.num_worker,
shuffle=True,
exclude_positive=True,
rank=i)
train_samplers.append(
NewBidirectionalOneShotIterator(train_sampler_head,
train_sampler_tail,
args.neg_chunk_size, True,
n_entities))
else:
train_sampler_head = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
args.neg_chunk_size,
mode='chunk-head',
num_workers=args.num_worker,
shuffle=True,
exclude_positive=True)
train_sampler_tail = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
args.neg_chunk_size,
mode='chunk-tail',
num_workers=args.num_worker,
shuffle=True,
exclude_positive=True)
train_sampler = NewBidirectionalOneShotIterator(
train_sampler_head, train_sampler_tail, args.neg_chunk_size, True,
n_entities)
# for multiprocessing evaluation, we don't need to sample multiple batches at a time
# in each process.
num_workers = args.num_worker
if args.num_proc > 1:
num_workers = 1
if args.valid or args.test:
eval_dataset = EvalDataset(dataset, args)
if args.valid:
# Here we want to use the regualr negative sampler because we need to ensure that
# all positive edges are excluded.
if args.num_proc > 1:
valid_sampler_heads = []
valid_sampler_tails = []
for i in range(args.num_proc):
valid_sampler_head = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_valid,
args.neg_chunk_size_valid,
args.eval_filter,
mode='chunk-head',
num_workers=num_workers,
rank=i,
ranks=args.num_proc)
valid_sampler_tail = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_valid,
args.neg_chunk_size_valid,
args.eval_filter,
mode='chunk-tail',
num_workers=num_workers,
rank=i,
ranks=args.num_proc)
valid_sampler_heads.append(valid_sampler_head)
valid_sampler_tails.append(valid_sampler_tail)
else:
valid_sampler_head = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_valid,
args.neg_chunk_size_valid,
args.eval_filter,
mode='chunk-head',
num_workers=num_workers,
rank=0,
ranks=1)
valid_sampler_tail = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_valid,
args.neg_chunk_size_valid,
args.eval_filter,
mode='chunk-tail',
num_workers=num_workers,
rank=0,
ranks=1)
if args.test:
# Here we want to use the regualr negative sampler because we need to ensure that
# all positive edges are excluded.
if args.num_proc > 1:
test_sampler_tails = []
test_sampler_heads = []
for i in range(args.num_proc):
test_sampler_head = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_test,
args.neg_chunk_size_test,
args.eval_filter,
mode='chunk-head',
num_workers=num_workers,
rank=i,
ranks=args.num_proc)
test_sampler_tail = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_test,
args.neg_chunk_size_test,
args.eval_filter,
mode='chunk-tail',
num_workers=num_workers,
rank=i,
ranks=args.num_proc)
test_sampler_heads.append(test_sampler_head)
test_sampler_tails.append(test_sampler_tail)
else:
test_sampler_head = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_test,
args.neg_chunk_size_test,
args.eval_filter,
mode='chunk-head',
num_workers=num_workers,
rank=0,
ranks=1)
test_sampler_tail = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_test,
args.neg_chunk_size_test,
args.eval_filter,
mode='chunk-tail',
num_workers=num_workers,
rank=0,
ranks=1)
# We need to free all memory referenced by dataset.
eval_dataset = None
dataset = None
# load model
model = load_model(logger, args, n_entities, n_relations)
if args.num_proc > 1:
model.share_memory()
# train
start = time.time()
if args.num_proc > 1:
procs = []
for i in range(args.num_proc):
rel_parts = train_data.rel_parts if args.rel_part else None
valid_samplers = [valid_sampler_heads[i], valid_sampler_tails[i]
] if args.valid else None
proc = mp.Process(target=train,
args=(args, model, train_samplers[i], i,
rel_parts, valid_samplers))
procs.append(proc)
proc.start()
for proc in procs:
proc.join()
else:
valid_samplers = [valid_sampler_head, valid_sampler_tail
] if args.valid else None
train(args, model, train_sampler, valid_samplers)
print('training takes {} seconds'.format(time.time() - start))
if args.save_emb is not None:
if not os.path.exists(args.save_emb):
os.mkdir(args.save_emb)
model.save_emb(args.save_emb, args.dataset)
# test
if args.test:
start = time.time()
if args.num_proc > 1:
queue = mp.Queue(args.num_proc)
procs = []
for i in range(args.num_proc):
proc = mp.Process(target=test,
args=(args, model, [
test_sampler_heads[i],
test_sampler_tails[i]
], i, 'Test', queue))
procs.append(proc)
proc.start()
total_metrics = {}
for i in range(args.num_proc):
metrics = queue.get()
for k, v in metrics.items():
if i == 0:
total_metrics[k] = v / args.num_proc
else:
total_metrics[k] += v / args.num_proc
for k, v in metrics.items():
print('Test average {} at [{}/{}]: {}'.format(
k, args.step, args.max_step, v))
for proc in procs:
proc.join()
else:
test(args, model, [test_sampler_head, test_sampler_tail])
print('test:', time.time() - start)
def start_worker(args, logger):
"""Start kvclient for training
"""
init_time_start = time.time()
time.sleep(WAIT_TIME) # wait for launch script
server_namebook = dgl.contrib.read_ip_config(filename=args.ip_config)
args.machine_id = get_local_machine_id(server_namebook)
dataset, entity_partition_book, local2global = get_partition_dataset(
args.data_path, args.dataset, args.format, args.machine_id)
n_entities = dataset.n_entities
n_relations = dataset.n_relations
print('Partition %d n_entities: %d' % (args.machine_id, n_entities))
print("Partition %d n_relations: %d" % (args.machine_id, n_relations))
entity_partition_book = F.tensor(entity_partition_book)
relation_partition_book = get_long_tail_partition(dataset.n_relations,
args.total_machine)
relation_partition_book = F.tensor(relation_partition_book)
local2global = F.tensor(local2global)
relation_partition_book.share_memory_()
entity_partition_book.share_memory_()
local2global.share_memory_()
train_data = TrainDataset(dataset, args, ranks=args.num_client)
# if there is no cross partition relaiton, we fall back to strict_rel_part
args.strict_rel_part = args.mix_cpu_gpu and (train_data.cross_part
== False)
args.soft_rel_part = args.mix_cpu_gpu and args.soft_rel_part and train_data.cross_part
if args.neg_sample_size_eval < 0:
args.neg_sample_size_eval = dataset.n_entities
args.batch_size = get_compatible_batch_size(args.batch_size,
args.neg_sample_size)
args.batch_size_eval = get_compatible_batch_size(args.batch_size_eval,
args.neg_sample_size_eval)
args.num_workers = 8 # fix num_workers to 8
train_samplers = []
for i in range(args.num_client):
train_sampler_head = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
args.neg_sample_size,
mode='head',
num_workers=args.num_workers,
shuffle=True,
exclude_positive=False,
rank=i)
train_sampler_tail = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
args.neg_sample_size,
mode='tail',
num_workers=args.num_workers,
shuffle=True,
exclude_positive=False,
rank=i)
train_samplers.append(
NewBidirectionalOneShotIterator(train_sampler_head,
train_sampler_tail,
args.neg_sample_size,
args.neg_sample_size, True,
n_entities))
dataset = None
model = load_model(logger, args, n_entities, n_relations)
model.share_memory()
print('Total initialize time {:.3f} seconds'.format(time.time() -
init_time_start))
rel_parts = train_data.rel_parts if args.strict_rel_part or args.soft_rel_part else None
cross_rels = train_data.cross_rels if args.soft_rel_part else None
procs = []
barrier = mp.Barrier(args.num_client)
for i in range(args.num_client):
proc = mp.Process(target=dist_train_test,
args=(args, model, train_samplers[i],
entity_partition_book, relation_partition_book,
local2global, i, rel_parts, cross_rels,
barrier))
procs.append(proc)
proc.start()
for proc in procs:
proc.join()
def main():
args = ArgParser().parse_args()
prepare_save_path(args)
args.neg_sample_size_eval = 1000
set_global_seed(args.seed)
init_time_start = time.time()
dataset = get_dataset(args, args.data_path, args.dataset, args.format,
args.delimiter, args.data_files,
args.has_edge_importance)
args.batch_size = get_compatible_batch_size(args.batch_size,
args.neg_sample_size)
args.batch_size_eval = get_compatible_batch_size(args.batch_size_eval,
args.neg_sample_size_eval)
#print(args)
set_logger(args)
print("To build training dataset")
t1 = time.time()
train_data = TrainDataset(dataset,
args,
has_importance=args.has_edge_importance)
print("Training dataset built, it takes %s seconds" % (time.time() - t1))
args.num_workers = 8 # fix num_worker to 8
print("Building training sampler")
t1 = time.time()
train_sampler_head = train_data.create_sampler(
batch_size=args.batch_size,
num_workers=args.num_workers,
neg_sample_size=args.neg_sample_size,
neg_mode='head')
train_sampler_tail = train_data.create_sampler(
batch_size=args.batch_size,
num_workers=args.num_workers,
neg_sample_size=args.neg_sample_size,
neg_mode='tail')
train_sampler = NewBidirectionalOneShotIterator(train_sampler_head,
train_sampler_tail)
print("Training sampler created, it takes %s seconds" % (time.time() - t1))
if args.valid or args.test:
if len(args.gpu) > 1:
args.num_test_proc = args.num_proc if args.num_proc < len(
args.gpu) else len(args.gpu)
else:
args.num_test_proc = args.num_proc
print("To create eval_dataset")
t1 = time.time()
eval_dataset = EvalDataset(dataset, args)
print("eval_dataset created, it takes %d seconds" % (time.time() - t1))
if args.valid:
if args.num_proc > 1:
valid_samplers = []
for i in range(args.num_proc):
print("creating valid sampler for proc %d" % i)
t1 = time.time()
valid_sampler_tail = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
mode='tail',
num_workers=args.num_workers,
rank=i,
ranks=args.num_proc)
valid_samplers.append(valid_sampler_tail)
print(
"Valid sampler for proc %d created, it takes %s seconds" %
(i, time.time() - t1))
else:
valid_sampler_tail = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
mode='tail',
num_workers=args.num_workers,
rank=0,
ranks=1)
valid_samplers = [valid_sampler_tail]
for arg in vars(args):
logging.info('{:20}:{}'.format(arg, getattr(args, arg)))
print("To create model")
t1 = time.time()
model = BaseKEModel(args=args,
n_entities=dataset.n_entities,
n_relations=dataset.n_relations,
model_name=args.model_name,
hidden_size=args.hidden_dim,
entity_feat_dim=dataset.entity_feat.shape[1],
relation_feat_dim=dataset.relation_feat.shape[1],
gamma=args.gamma,
double_entity_emb=args.double_ent,
relation_times=args.ote_size,
scale_type=args.scale_type)
model.entity_feat = dataset.entity_feat
model.relation_feat = dataset.relation_feat
print(len(model.parameters()))
if args.cpu_emb:
print("using cpu emb\n" * 5)
else:
print("using gpu emb\n" * 5)
optimizer = paddle.optimizer.Adam(learning_rate=args.mlp_lr,
parameters=model.parameters())
lr_tensor = paddle.to_tensor(args.lr)
global_step = 0
tic_train = time.time()
log = {}
log["loss"] = 0.0
log["regularization"] = 0.0
for step in range(0, args.max_step):
pos_triples, neg_triples, ids, neg_head = next(train_sampler)
loss = model.forward(pos_triples, neg_triples, ids, neg_head)
log["loss"] = loss.numpy()[0]
if args.regularization_coef > 0.0 and args.regularization_norm > 0:
coef, nm = args.regularization_coef, args.regularization_norm
reg = coef * norm(model.entity_embedding.curr_emb, nm)
log['regularization'] = reg.numpy()[0]
loss = loss + reg
loss.backward()
optimizer.step()
if args.cpu_emb:
model.entity_embedding.step(lr_tensor)
optimizer.clear_grad()
if (step + 1) % args.log_interval == 0:
speed = args.log_interval / (time.time() - tic_train)
logging.info(
"step: %d, train loss: %.5f, regularization: %.4e, speed: %.2f steps/s"
% (step, log["loss"], log["regularization"], speed))
log["loss"] = 0.0
tic_train = time.time()
if args.valid and (
step + 1
) % args.eval_interval == 0 and step > 1 and valid_samplers is not None:
print("Valid begin")
valid_start = time.time()
valid_input_dict = test(args,
model,
valid_samplers,
step,
rank=0,
mode='Valid')
paddle.save(
valid_input_dict,
os.path.join(args.save_path, "valid_{}.pkl".format(step)))
# Save the model for the inference
if (step + 1) % args.save_step == 0:
print("The step:{}, save model path:{}".format(
step + 1, args.save_path))
model.save_model()
print("Save model done.")
def run(args, logger):
train_time_start = time.time()
# load dataset and samplers
dataset = get_dataset(args.data_path, args.dataset, args.format,
args.data_files)
n_entities = dataset.n_entities
n_relations = dataset.n_relations
if args.neg_sample_size_test < 0:
args.neg_sample_size_test = n_entities
args.eval_filter = not args.no_eval_filter
if args.neg_deg_sample_eval:
assert not args.eval_filter, "if negative sampling based on degree, we can't filter positive edges."
# When we generate a batch of negative edges from a set of positive edges,
# we first divide the positive edges into chunks and corrupt the edges in a chunk
# together. By default, the chunk size is equal to the negative sample size.
# Usually, this works well. But we also allow users to specify the chunk size themselves.
if args.neg_chunk_size < 0:
args.neg_chunk_size = args.neg_sample_size
if args.neg_chunk_size_valid < 0:
args.neg_chunk_size_valid = args.neg_sample_size_valid
if args.neg_chunk_size_test < 0:
args.neg_chunk_size_test = args.neg_sample_size_test
num_workers = args.num_worker
train_data = TrainDataset(dataset, args, ranks=args.num_proc)
# if there is no cross partition relaiton, we fall back to strict_rel_part
args.strict_rel_part = args.mix_cpu_gpu and (train_data.cross_part
== False)
args.soft_rel_part = args.mix_cpu_gpu and args.soft_rel_part and train_data.cross_part
# Automatically set number of OMP threads for each process if it is not provided
# The value for GPU is evaluated in AWS p3.16xlarge
# The value for CPU is evaluated in AWS x1.32xlarge
if args.nomp_thread_per_process == -1:
if len(args.gpu) > 0:
# GPU training
args.num_thread = 4
else:
# CPU training
args.num_thread = 1
else:
args.num_thread = args.nomp_thread_per_process
if args.num_proc > 1:
train_samplers = []
for i in range(args.num_proc):
train_sampler_head = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
args.neg_chunk_size,
mode='head',
num_workers=num_workers,
shuffle=True,
exclude_positive=False,
rank=i)
train_sampler_tail = train_data.create_sampler(
args.batch_size,
args.neg_sample_size,
args.neg_chunk_size,
mode='tail',
num_workers=num_workers,
shuffle=True,
exclude_positive=False,
rank=i)
train_samplers.append(
NewBidirectionalOneShotIterator(train_sampler_head,
train_sampler_tail,
args.neg_chunk_size,
args.neg_sample_size, True,
n_entities))
else:
train_sampler_head = train_data.create_sampler(args.batch_size,
args.neg_sample_size,
args.neg_chunk_size,
mode='head',
num_workers=num_workers,
shuffle=True,
exclude_positive=False)
train_sampler_tail = train_data.create_sampler(args.batch_size,
args.neg_sample_size,
args.neg_chunk_size,
mode='tail',
num_workers=num_workers,
shuffle=True,
exclude_positive=False)
train_sampler = NewBidirectionalOneShotIterator(
train_sampler_head, train_sampler_tail, args.neg_chunk_size,
args.neg_sample_size, True, n_entities)
# for multiprocessing evaluation, we don't need to sample multiple batches at a time
# in each process.
if args.num_proc > 1:
num_workers = 1
if args.valid or args.test:
if len(args.gpu) > 1:
args.num_test_proc = args.num_proc if args.num_proc < len(
args.gpu) else len(args.gpu)
else:
args.num_test_proc = args.num_proc
eval_dataset = EvalDataset(dataset, args)
if args.valid:
# Here we want to use the regualr negative sampler because we need to ensure that
# all positive edges are excluded.
if args.num_proc > 1:
valid_sampler_heads = []
valid_sampler_tails = []
for i in range(args.num_proc):
valid_sampler_head = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_valid,
args.neg_chunk_size_valid,
args.eval_filter,
mode='chunk-head',
num_workers=num_workers,
rank=i,
ranks=args.num_proc)
valid_sampler_tail = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_valid,
args.neg_chunk_size_valid,
args.eval_filter,
mode='chunk-tail',
num_workers=num_workers,
rank=i,
ranks=args.num_proc)
valid_sampler_heads.append(valid_sampler_head)
valid_sampler_tails.append(valid_sampler_tail)
else:
valid_sampler_head = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_valid,
args.neg_chunk_size_valid,
args.eval_filter,
mode='chunk-head',
num_workers=num_workers,
rank=0,
ranks=1)
valid_sampler_tail = eval_dataset.create_sampler(
'valid',
args.batch_size_eval,
args.neg_sample_size_valid,
args.neg_chunk_size_valid,
args.eval_filter,
mode='chunk-tail',
num_workers=num_workers,
rank=0,
ranks=1)
if args.test:
# Here we want to use the regualr negative sampler because we need to ensure that
# all positive edges are excluded.
# We use a maximum of num_gpu in test stage to save GPU memory.
if args.num_test_proc > 1:
test_sampler_tails = []
test_sampler_heads = []
for i in range(args.num_test_proc):
test_sampler_head = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_test,
args.neg_chunk_size_test,
args.eval_filter,
mode='chunk-head',
num_workers=num_workers,
rank=i,
ranks=args.num_test_proc)
test_sampler_tail = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_test,
args.neg_chunk_size_test,
args.eval_filter,
mode='chunk-tail',
num_workers=num_workers,
rank=i,
ranks=args.num_test_proc)
test_sampler_heads.append(test_sampler_head)
test_sampler_tails.append(test_sampler_tail)
else:
test_sampler_head = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_test,
args.neg_chunk_size_test,
args.eval_filter,
mode='chunk-head',
num_workers=num_workers,
rank=0,
ranks=1)
test_sampler_tail = eval_dataset.create_sampler(
'test',
args.batch_size_eval,
args.neg_sample_size_test,
args.neg_chunk_size_test,
args.eval_filter,
mode='chunk-tail',
num_workers=num_workers,
rank=0,
ranks=1)
# We need to free all memory referenced by dataset.
eval_dataset = None
dataset = None
# load model
model = load_model(logger, args, n_entities, n_relations)
if args.num_proc > 1 or args.async_update:
model.share_memory()
print('Total data loading time {:.3f} seconds'.format(time.time() -
train_time_start))
# train
start = time.time()
rel_parts = train_data.rel_parts if args.strict_rel_part or args.soft_rel_part else None
cross_rels = train_data.cross_rels if args.soft_rel_part else None
if args.num_proc > 1:
procs = []
barrier = mp.Barrier(args.num_proc)
for i in range(args.num_proc):
valid_sampler = [valid_sampler_heads[i], valid_sampler_tails[i]
] if args.valid else None
proc = mp.Process(target=train_mp,
args=(args, model, train_samplers[i],
valid_sampler, i, rel_parts, cross_rels,
barrier))
procs.append(proc)
proc.start()
for proc in procs:
proc.join()
else:
valid_samplers = [valid_sampler_head, valid_sampler_tail
] if args.valid else None
train(args, model, train_sampler, valid_samplers, rel_parts=rel_parts)
print('training takes {} seconds'.format(time.time() - start))
if args.save_emb is not None:
if not os.path.exists(args.save_emb):
os.mkdir(args.save_emb)
model.save_emb(args.save_emb, args.dataset)
# We need to save the model configurations as well.
conf_file = os.path.join(args.save_emb, 'config.json')
with open(conf_file, 'w') as outfile:
json.dump(
{
'dataset': args.dataset,
'model': args.model_name,
'emb_size': args.hidden_dim,
'max_train_step': args.max_step,
'batch_size': args.batch_size,
'neg_sample_size': args.neg_sample_size,
'lr': args.lr,
'gamma': args.gamma,
'double_ent': args.double_ent,
'double_rel': args.double_rel,
'neg_adversarial_sampling': args.neg_adversarial_sampling,
'adversarial_temperature': args.adversarial_temperature,
'regularization_coef': args.regularization_coef,
'regularization_norm': args.regularization_norm
},
outfile,
indent=4)
# test
if args.test:
start = time.time()
if args.num_test_proc > 1:
queue = mp.Queue(args.num_test_proc)
procs = []
for i in range(args.num_test_proc):
proc = mp.Process(target=test_mp,
args=(args, model, [
test_sampler_heads[i],
test_sampler_tails[i]
], i, 'Test', queue))
procs.append(proc)
proc.start()
total_metrics = {}
metrics = {}
logs = []
for i in range(args.num_test_proc):
log = queue.get()
logs = logs + log
for metric in logs[0].keys():
metrics[metric] = sum([log[metric]
for log in logs]) / len(logs)
for k, v in metrics.items():
print('Test average {} at [{}/{}]: {}'.format(
k, args.step, args.max_step, v))
for proc in procs:
proc.join()
else:
test(args, model, [test_sampler_head, test_sampler_tail])
print('test:', time.time() - start)
def start_worker(args, logger):
"""Start kvclient for training
"""
train_time_start = time.time()
server_namebook = dgl.contrib.read_ip_config(filename=args.ip_config)
args.machine_id = get_local_machine_id(server_namebook)
dataset, entity_partition_book, local2global = get_partition_dataset(
args.data_path,
args.dataset,
args.format,
args.machine_id)
n_entities = dataset.n_entities
n_relations = dataset.n_relations
print('Partition %d n_entities: %d' % (args.machine_id, n_entities))
print("Partition %d n_relations: %d" % (args.machine_id, n_relations))
entity_partition_book = F.tensor(entity_partition_book)
relation_partition_book = get_long_tail_partition(dataset.n_relations, args.total_machine)
relation_partition_book = F.tensor(relation_partition_book)
local2global = F.tensor(local2global)
relation_partition_book.share_memory_()
entity_partition_book.share_memory_()
local2global.share_memory_()
model = load_model(logger, args, n_entities, n_relations)
model.share_memory()
# When we generate a batch of negative edges from a set of positive edges,
# we first divide the positive edges into chunks and corrupt the edges in a chunk
# together. By default, the chunk size is equal to the negative sample size.
# Usually, this works well. But we also allow users to specify the chunk size themselves.
if args.neg_chunk_size < 0:
args.neg_chunk_size = args.neg_sample_size
num_workers = NUM_WORKER
train_data = TrainDataset(dataset, args, ranks=args.num_client)
train_samplers = []
for i in range(args.num_client):
train_sampler_head = train_data.create_sampler(args.batch_size,
args.neg_sample_size,
args.neg_chunk_size,
mode='head',
num_workers=num_workers,
shuffle=True,
exclude_positive=False,
rank=i)
train_sampler_tail = train_data.create_sampler(args.batch_size,
args.neg_sample_size,
args.neg_chunk_size,
mode='tail',
num_workers=num_workers,
shuffle=True,
exclude_positive=False,
rank=i)
train_samplers.append(NewBidirectionalOneShotIterator(train_sampler_head, train_sampler_tail,
args.neg_chunk_size, args.neg_sample_size,
True, n_entities))
dataset = None
print('Total data loading time {:.3f} seconds'.format(time.time() - train_time_start))
rel_parts = train_data.rel_parts if args.strict_rel_part or args.soft_rel_part else None
cross_rels = train_data.cross_rels if args.soft_rel_part else None
args.num_thread = NUM_THREAD
procs = []
barrier = mp.Barrier(args.num_client)
for i in range(args.num_client):
proc = mp.Process(target=dist_train_test, args=(args,
model,
train_samplers[i],
entity_partition_book,
relation_partition_book,
local2global,
i,
rel_parts,
cross_rels,
barrier))
procs.append(proc)
proc.start()
for proc in procs:
proc.join()
