def main():
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size',
type=int,
default=64,
metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size',
type=int,
default=1000,
metavar='N',
help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs',
type=int,
default=14,
metavar='N',
help='number of epochs to train (default: 14)')
parser.add_argument('--lr',
type=float,
default=1.0,
metavar='LR',
help='learning rate (default: 1.0)')
parser.add_argument('--gamma',
type=float,
default=0.7,
metavar='M',
help='Learning rate step gamma (default: 0.7)')
parser.add_argument('--seed',
type=int,
default=1,
metavar='S',
help='random seed (default: 1)')
parser.add_argument(
'--log-interval',
type=int,
default=10,
metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--save-model',
action='store_true',
default=False,
help='For Saving the current Model')
parser.add_argument('--verbose',
action='store_true',
default=False,
help='For displaying SM Data Parallel-specific logs')
parser.add_argument('--data-path',
type=str,
default='/tmp/data',
help='Path for downloading '
'the MNIST dataset')
args = parser.parse_args()
args.world_size = dist.get_world_size()
args.rank = rank = dist.get_rank()
args.local_rank = local_rank = dist.get_local_rank()
args.lr = 1.0
args.batch_size //= args.world_size // 8
args.batch_size = max(args.batch_size, 1)
data_path = args.data_path
if args.verbose:
print('Hello from rank', rank, 'of local_rank', local_rank,
'in world size of', args.world_size)
if not torch.cuda.is_available():
raise Exception(
"Must run SM Distributed DataParallel MNIST example on CUDA-capable devices."
)
torch.manual_seed(args.seed)
device = torch.device("cuda")
if local_rank == 0:
train_dataset = datasets.MNIST(data_path,
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ),
(0.3081, ))
]))
else:
time.sleep(8)
train_dataset = datasets.MNIST(data_path,
train=True,
download=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ),
(0.3081, ))
]))
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset, num_replicas=args.world_size, rank=rank)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=0,
pin_memory=True,
sampler=train_sampler)
if rank == 0:
test_loader = torch.utils.data.DataLoader(
datasets.MNIST(data_path,
train=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])),
batch_size=args.test_batch_size,
shuffle=True)
model = DDP(Net().to(device))
torch.cuda.set_device(local_rank)
model.cuda(local_rank)
optimizer = optim.Adadelta(model.parameters(), lr=args.lr)
scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
for epoch in range(1, args.epochs + 1):
train(args, model, device, train_loader, optimizer, epoch)
if rank == 0:
test(model, device, test_loader)
scheduler.step()
if args.save_model:
torch.save(model.state_dict(), "mnist_cnn.pt")
def train_model(args):
from torchvision import datasets, models
from tqdm import tqdm
imgs, trn_df, vld_df = _get_images(args.train_dir,
args.num_folds,
args.vld_fold_idx,
data_type='train')
trn_loader, vld_loader = _get_data_loader(imgs, trn_df, vld_df)
logger.info("=== Getting Pre-trained model ===")
model = models.resnet18(pretrained=True)
last_hidden_units = model.fc.in_features
model.fc = torch.nn.Linear(last_hidden_units, 186)
# len_buffer = len(list(module.buffers()))
# logger.info("=== Buffer ===")
# print(f"len_buffer={len_buffer}")
# print(list(model.buffers()))
# SDP: Pin each GPU to a single process
# Use SMDataParallel PyTorch DDP for efficient distributed training
model = DDP(model.to(args.device), broadcast_buffers=False)
# SDP: Pin each GPU to a single SDP process.
torch.cuda.set_device(args.local_rank)
model.cuda(args.local_rank)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
loss_fn = nn.CrossEntropyLoss()
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode='max',
verbose=True,
patience=5,
factor=0.5)
best_score = -1
training_stats = []
logger.info("=== Start Training ===")
for epoch_id in range(args.num_epochs):
################################################################################
# ==> Training phase
################################################################################
trn_loss = []
model.train()
# Measure how long the training epoch takes.
t0 = time.time()
running_loss = 0.0
for batch_id, (inputs, targets) in enumerate((trn_loader)):
inputs = inputs.cuda()
targets = targets.cuda()
targets_gra = targets[:, 0]
targets_vow = targets[:, 1]
targets_con = targets[:, 2]
# 50%의 확률로 원본 데이터 그대로 사용
if np.random.rand() < 0.5:
logits = model(inputs)
grapheme = logits[:, :168]
vowel = logits[:, 168:179]
cons = logits[:, 179:]
loss1 = loss_fn(grapheme, targets_gra)
loss2 = loss_fn(vowel, targets_vow)
loss3 = loss_fn(cons, targets_con)
else:
lam = np.random.beta(1.0, 1.0)
rand_index = torch.randperm(inputs.size()[0])
shuffled_targets_gra = targets_gra[rand_index]
shuffled_targets_vow = targets_vow[rand_index]
shuffled_targets_con = targets_con[rand_index]
bbx1, bby1, bbx2, bby2 = _rand_bbox(inputs.size(), lam)
inputs[:, :, bbx1:bbx2,
bby1:bby2] = inputs[rand_index, :, bbx1:bbx2, bby1:bby2]
# 픽셀 비율과 정확히 일치하도록 lambda 파라메터 조정
lam = 1 - ((bbx2 - bbx1) * (bby2 - bby1) /
(inputs.size()[-1] * inputs.size()[-2]))
logits = model(inputs)
grapheme = logits[:, :168]
vowel = logits[:, 168:179]
cons = logits[:, 179:]
loss1 = loss_fn(grapheme, targets_gra) * lam + loss_fn(
grapheme, shuffled_targets_gra) * (1. - lam)
loss2 = loss_fn(vowel, targets_vow) * lam + loss_fn(
vowel, shuffled_targets_vow) * (1. - lam)
loss3 = loss_fn(cons, targets_con) * lam + loss_fn(
cons, shuffled_targets_con) * (1. - lam)
loss = 0.5 * loss1 + 0.25 * loss2 + 0.25 * loss3
trn_loss.append(loss.item())
running_loss += loss.item()
loss.backward()
optimizer.step()
optimizer.zero_grad()
# Printing vital information
if (batch_id + 1) % (args.log_interval) == 0:
s = f'[Epoch {epoch_id} Batch {batch_id+1}/{len(trn_loader)}] ' \
f'loss: {running_loss / args.log_interval:.4f}'
print(s)
running_loss = 0
# Measure how long this epoch took.
trn_time = _format_time(time.time() - t0)
if args.rank == 0:
################################################################################
# ==> Validation phase
################################################################################
val_loss = []
val_true = []
val_pred = []
model.eval()
# === Validation phase ===
logger.info('=== Start Validation ===')
with torch.no_grad():
for inputs, targets in vld_loader:
inputs = inputs.cuda()
targets = targets.cuda()
logits = model(inputs)
grapheme = logits[:, :168]
vowel = logits[:, 168:179]
cons = logits[:, 179:]
loss= 0.5* loss_fn(grapheme, targets[:,0]) + 0.25*loss_fn(vowel, targets[:,1]) + \
0.25*loss_fn(vowel, targets[:,2])
val_loss.append(loss.item())
grapheme = grapheme.cpu().argmax(dim=1).data.numpy()
vowel = vowel.cpu().argmax(dim=1).data.numpy()
cons = cons.cpu().argmax(dim=1).data.numpy()
val_true.append(targets.cpu().numpy())
val_pred.append(np.stack([grapheme, vowel, cons], axis=1))
val_true = np.concatenate(val_true)
val_pred = np.concatenate(val_pred)
val_loss = np.mean(val_loss)
trn_loss = np.mean(trn_loss)
score_g = recall_score(val_true[:, 0],
val_pred[:, 0],
average='macro')
score_v = recall_score(val_true[:, 1],
val_pred[:, 1],
average='macro')
score_c = recall_score(val_true[:, 2],
val_pred[:, 2],
average='macro')
final_score = np.average([score_g, score_v, score_c],
weights=[2, 1, 1])
# Printing vital information
s = f'[Epoch {epoch_id}] ' \
f'trn_loss: {trn_loss:.4f}, vld_loss: {val_loss:.4f}, score: {final_score:.4f}, ' \
f'score_each: [{score_g:.4f}, {score_v:.4f}, {score_c:.4f}]'
print(s)
################################################################################
# ==> Save checkpoint and training stats
################################################################################
if final_score > best_score:
best_score = final_score
state_dict = model.cpu().state_dict()
model = model.cuda()
torch.save(state_dict, os.path.join(args.model_dir,
'model.pth'))
# Record all statistics from this epoch
training_stats.append({
'epoch': epoch_id + 1,
'trn_loss': trn_loss,
'trn_time': trn_time,
'val_loss': val_loss,
'score': final_score,
'score_g': score_g,
'score_v': score_v,
'score_c': score_c
})
# === Save Model Parameters ===
logger.info("Model successfully saved at: {}".format(
args.model_dir))
def main():
# Training settings
parser = argparse.ArgumentParser(description="PyTorch MNIST Example")
parser.add_argument(
"--batch-size",
type=int,
default=64,
metavar="N",
help="input batch size for training (default: 64)",
)
parser.add_argument(
"--test-batch-size",
type=int,
default=1000,
metavar="N",
help="input batch size for testing (default: 1000)",
)
parser.add_argument(
"--epochs",
type=int,
default=14,
metavar="N",
help="number of epochs to train (default: 14)",
)
parser.add_argument("--lr",
type=float,
default=1.0,
metavar="LR",
help="learning rate (default: 1.0)")
parser.add_argument(
"--gamma",
type=float,
default=0.7,
metavar="M",
help="Learning rate step gamma (default: 0.7)",
)
parser.add_argument("--seed",
type=int,
default=1,
metavar="S",
help="random seed (default: 1)")
parser.add_argument(
"--log-interval",
type=int,
default=10,
metavar="N",
help="how many batches to wait before logging training status",
)
parser.add_argument("--save-model",
action="store_true",
default=False,
help="For Saving the current Model")
parser.add_argument(
"--verbose",
action="store_true",
default=False,
help="For displaying smdistributed.dataparallel-specific logs",
)
parser.add_argument(
"--data-path",
type=str,
default="/tmp/data",
help="Path for downloading "
"the MNIST dataset",
)
args = parser.parse_args()
args.world_size = dist.get_world_size()
args.rank = rank = dist.get_rank()
args.local_rank = local_rank = dist.get_local_rank()
args.lr = 1.0
args.batch_size //= args.world_size // 8
args.batch_size = max(args.batch_size, 1)
data_path = args.data_path
if args.verbose:
print(
"Hello from rank",
rank,
"of local_rank",
local_rank,
"in world size of",
args.world_size,
)
if not torch.cuda.is_available():
raise CUDANotFoundException(
"Must run smdistributed.dataparallel MNIST example on CUDA-capable devices."
)
torch.manual_seed(args.seed)
device = torch.device("cuda")
# select a single rank per node to download data
is_first_local_rank = local_rank == 0
if is_first_local_rank:
train_dataset = datasets.MNIST(
data_path,
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
]),
)
dist.barrier() # prevent other ranks from accessing the data early
if not is_first_local_rank:
train_dataset = datasets.MNIST(
data_path,
train=True,
download=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
]),
)
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset, num_replicas=args.world_size, rank=rank)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=0,
pin_memory=True,
sampler=train_sampler,
)
if rank == 0:
test_loader = torch.utils.data.DataLoader(
datasets.MNIST(
data_path,
train=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
]),
),
batch_size=args.test_batch_size,
shuffle=True,
)
model = DDP(Net().to(device))
torch.cuda.set_device(local_rank)
model.cuda(local_rank)
optimizer = optim.Adadelta(model.parameters(), lr=args.lr)
scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
for epoch in range(1, args.epochs + 1):
train(args, model, device, train_loader, optimizer, epoch)
if rank == 0:
test(model, device, test_loader)
scheduler.step()
if args.save_model:
torch.save(model.state_dict(), "mnist_cnn.pt")
def sdp_init(model, optimizer, args):
model = DDP(model.to(args.device), broadcast_buffers=False)
# model = DDP(model, device_ids=[args.rank], broadcast_buffers=False)
model.cuda(args.local_rank)
return model, optimizer, args
def train(cfg, args):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
if use_amp:
# Initialize mixed-precision training
use_mixed_precision = cfg.DTYPE == "float16"
amp_opt_level = 'O1' if use_mixed_precision else 'O0'
model, optimizer = amp.initialize(model, optimizer, opt_level=amp_opt_level)
if args.distributed:
# if use_apex_ddp:
# model = DDP(model, delay_allreduce=True)
# else:
# SMDataParallel: Wrap the PyTorch model with SMDataParallel’s DDP
model = DDP(model, device_ids=[dist.get_local_rank()], broadcast_buffers=False)
#model = DDP(model)
print("model parameter size: ", sum(p.numel() for p in model.parameters() if p.requires_grad))
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
# SMDataParallel: Save model on master node.
save_to_disk = dist.get_rank() == 0
checkpointer = DetectronCheckpointer(
cfg, model, optimizer, scheduler, output_dir, save_to_disk
)
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
data_loader, iters_per_epoch = make_data_loader(
cfg,
is_train=True,
is_distributed=args.distributed,
start_iter=arguments["iteration"],
data_dir = args.data_dir
)
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
# set the callback function to evaluate and potentially
# early exit each epoch
if cfg.PER_EPOCH_EVAL:
per_iter_callback_fn = functools.partial(
mlperf_test_early_exit,
iters_per_epoch=iters_per_epoch,
tester=functools.partial(test, cfg=cfg),
model=model,
distributed=args.distributed,
min_bbox_map=cfg.MIN_BBOX_MAP,
min_segm_map=cfg.MIN_MASK_MAP)
else:
per_iter_callback_fn = None
do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
use_amp,
cfg,
per_iter_end_callback_fn=per_iter_callback_fn,
)
return model