def test(args, model, device, test_loader):
model.eval()
test_loss = 0
correct = 0
with torch.no_grad():
for batch_idx, (data, target) in enumerate(test_loader):
# SM Distributed: Moves input tensors to the GPU ID used by the current process
# based on the set_device call.
data, target = data.to(device), target.to(device)
# Since test_step returns scalars instead of tensors,
# test_step decorated with smp.step will return lists instead of StepOutput objects.
loss_batch, correct_batch = test_step(model, data, target)
test_loss += sum(loss_batch)
correct += sum(correct_batch)
if args.num_batches and batch_idx + 1 == args.num_batches:
break
test_loss /= len(test_loader.dataset)
if smp.mp_rank() == 0:
print("\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n".
format(
test_loss,
correct,
len(test_loader.dataset),
100.0 * correct / len(test_loader.dataset),
))
return test_loss
def setup_training(args):
assert torch.cuda.is_available()
if args.smp > 0:
# Initialize SMP. The configuration is obtained from the parameters passed to
# the Sagemaker PyTorch estimator.
smp.init()
# SMP: Set the device to the GPU ID used by the current process.
# Input tensors should be transferred to this device.
torch.cuda.set_device(smp.local_rank())
device = torch.device("cuda", smp.local_rank())
args.n_gpu = 1
# if args.local_rank == -1:
# device = torch.device("cuda")
# args.n_gpu = torch.cuda.device_count()
# args.allreduce_post_accumulation = False
# args.allreduce_post_accumulation_fp16 = False
# else:
# torch.cuda.set_device(args.local_rank)
# device = torch.device("cuda", args.local_rank)
# # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
# torch.distributed.init_process_group(backend='nccl', init_method='env://')
# args.n_gpu = 1
if args.gradient_accumulation_steps == 1:
args.allreduce_post_accumulation = False
args.allreduce_post_accumulation_fp16 = False
print(
"device: {} n_gpu: {}, mp_rank: {}, rank: {}, distributed training: {}, 16-bits training: {}"
.format(device, args.n_gpu, smp.mp_rank(), smp.rank(),
bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if args.train_batch_size % args.gradient_accumulation_steps != 0:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, batch size {} should be divisible"
.format(args.gradient_accumulation_steps, args.train_batch_size))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
if (not args.resume_from_checkpoint and os.path.exists(args.output_dir) and
(os.listdir(args.output_dir)
and any([i.startswith("ckpt")
for i in os.listdir(args.output_dir)]))):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if (not args.resume_from_checkpoint
or not os.path.exists(args.output_dir)) and is_main_process():
os.makedirs(args.output_dir, exist_ok=True)
return device, args
def is_world_process_zero(self) -> bool:
"""
Whether or not this process is the global main process (when training in a distributed fashion on several
machines, this is only going to be :obj:`True` for one process).
"""
if self.is_model_parallel_enabled:
return smp.rank() == 0 and smp.local_rank() == 0 and smp.mp_rank() == 0 and smp.dp_rank() == 0
else:
return super().is_world_process_zero()
def dist_setting(args):
# args.data_parallel = False
print("args.data_parallel : {}".format(args.data_parallel))
print("args.model_parallel : {}".format(args.model_parallel))
print("args.apex : {}".format(args.apex))
args.world_size = 1
args.host_num = args.hosts.index(args.current_host)
if args.data_parallel:
args.world_size = sdp.get_world_size()
args.rank = sdp.get_rank() # total rank in all hosts
args.local_rank = sdp.get_local_rank() # rank per host
elif args.model_parallel:
args.world_size = smp.size()
args.local_rank = smp.local_rank() # rank per host
args.rank = smp.rank()
args.dp_size = smp.dp_size()
args.dp_rank = smp.dp_rank()
print(
"smp.rank() : {}, smp.size() : {}, smp.mp_rank() : {}, smp.local_size() : {}, smp.get_mp_group() : {}, smp.get_dp_group() : {}, smp.local_rank() : {}, smp.dp_size() : {}, smp.dp_rank() : {}"
.format(smp.rank(), smp.size(), smp.mp_rank(), smp.local_size(),
smp.get_mp_group(), smp.get_dp_group(), smp.local_rank(),
smp.dp_size(), smp.dp_rank()))
else:
args.world_size = len(args.hosts) * args.num_gpus
if args.local_rank is not None:
args.rank = args.num_gpus * args.host_num + \
args.local_rank # total rank in all hosts
dist.init_process_group(backend=args.backend,
rank=args.rank,
world_size=args.world_size)
logger.info(
'Initialized the distributed environment: \'{}\' backend on {} nodes. '
.format(args.backend, dist.get_world_size()) +
'Current host rank is {}. Number of gpus: {}'.format(
dist.get_rank(), args.num_gpus))
print("**** [dist_setting] args.rank : {}".format(args.rank))
print("args.world_size : {}".format(args.world_size))
print("Use GPU: {} for training".format(args.local_rank))
args.lr = args.lr * float(args.world_size)
args.batch_size //= args.world_size // args.num_gpus
args.batch_size = max(args.batch_size, 1)
return args
