class ShufflenetConfig:
BATCH_SIZE = 128
LEARNING_RATE = 0.0625
MOMENTUM = 0.9
WEIGHT_DECAY = (lambda self, n, p: 4e-5
if n.find("weight") >= 0 and len(p.shape) > 1 else 0)
EPOCHS = 240
SCHEDULER = "Linear"
COLOR_JITTOR = T.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4)
def build_dataset(args):
train_dataset = data.dataset.ImageNet(args.data, train=True)
train_sampler = data.Infinite(
data.RandomSampler(train_dataset,
batch_size=args.batch_size,
drop_last=True))
train_dataloader = data.DataLoader(
train_dataset,
sampler=train_sampler,
transform=T.Compose([ # Baseline Augmentation for small models
T.RandomResizedCrop(224),
T.RandomHorizontalFlip(),
T.Normalize(mean=[103.530, 116.280, 123.675],
std=[57.375, 57.120, 58.395]), # BGR
T.ToMode("CHW"),
]) if args.arch in ("resnet18", "resnet34") else T.Compose(
[ # Facebook Augmentation for large models
T.RandomResizedCrop(224),
T.RandomHorizontalFlip(),
T.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4),
T.Normalize(mean=[103.530, 116.280, 123.675],
std=[57.375, 57.120, 58.395]), # BGR
T.ToMode("CHW"),
]),
num_workers=args.workers,
)
valid_dataset = data.dataset.ImageNet(args.data, train=False)
valid_sampler = data.SequentialSampler(valid_dataset,
batch_size=100,
drop_last=False)
valid_dataloader = data.DataLoader(
valid_dataset,
sampler=valid_sampler,
transform=T.Compose([
T.Resize(256),
T.CenterCrop(224),
T.Normalize(mean=[103.530, 116.280, 123.675],
std=[57.375, 57.120, 58.395]), # BGR
T.ToMode("CHW"),
]),
num_workers=args.workers,
)
return train_dataloader, valid_dataloader
def worker(rank, world_size, args):
if world_size > 1:
# Initialize distributed process group
logger.info("init distributed process group {} / {}".format(
rank, world_size))
dist.init_process_group(
master_ip="localhost",
master_port=23456,
world_size=world_size,
rank=rank,
dev=rank,
)
save_dir = os.path.join(args.save, args.arch)
model = getattr(M, args.arch)()
optimizer = optim.SGD(
get_parameters(model),
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay,
)
# Define train and valid graph
@jit.trace(symbolic=True)
def train_func(image, label):
model.train()
logits = model(image)
loss = F.cross_entropy_with_softmax(logits, label, label_smooth=0.1)
acc1, acc5 = F.accuracy(logits, label, (1, 5))
optimizer.backward(loss) # compute gradients
if dist.is_distributed(): # all_reduce_mean
loss = dist.all_reduce_sum(loss,
"train_loss") / dist.get_world_size()
acc1 = dist.all_reduce_sum(acc1,
"train_acc1") / dist.get_world_size()
acc5 = dist.all_reduce_sum(acc5,
"train_acc5") / dist.get_world_size()
return loss, acc1, acc5
@jit.trace(symbolic=True)
def valid_func(image, label):
model.eval()
logits = model(image)
loss = F.cross_entropy_with_softmax(logits, label, label_smooth=0.1)
acc1, acc5 = F.accuracy(logits, label, (1, 5))
if dist.is_distributed(): # all_reduce_mean
loss = dist.all_reduce_sum(loss,
"valid_loss") / dist.get_world_size()
acc1 = dist.all_reduce_sum(acc1,
"valid_acc1") / dist.get_world_size()
acc5 = dist.all_reduce_sum(acc5,
"valid_acc5") / dist.get_world_size()
return loss, acc1, acc5
# Build train and valid datasets
logger.info("preparing dataset..")
train_dataset = data.dataset.ImageNet(args.data, train=True)
train_sampler = data.RandomSampler(train_dataset,
batch_size=args.batch_size,
drop_last=True)
train_queue = data.DataLoader(
train_dataset,
sampler=train_sampler,
transform=T.Compose([
T.RandomResizedCrop(224),
T.RandomHorizontalFlip(),
T.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4),
T.Normalize(mean=[103.530, 116.280, 123.675],
std=[57.375, 57.120, 58.395]), # BGR
T.ToMode("CHW"),
]),
num_workers=args.workers,
)
train_queue = infinite_iter(train_queue)
valid_dataset = data.dataset.ImageNet(args.data, train=False)
valid_sampler = data.SequentialSampler(valid_dataset,
batch_size=100,
drop_last=False)
valid_queue = data.DataLoader(
valid_dataset,
sampler=valid_sampler,
transform=T.Compose([
T.Resize(256),
T.CenterCrop(224),
T.Normalize(mean=[103.530, 116.280, 123.675],
std=[57.375, 57.120, 58.395]), # BGR
T.ToMode("CHW"),
]),
num_workers=args.workers,
)
# Start training
objs = AverageMeter("Loss")
top1 = AverageMeter("Acc@1")
top5 = AverageMeter("Acc@5")
total_time = AverageMeter("Time")
t = time.time()
for step in range(0, args.steps + 1250 + 1):
# Linear learning rate decay
decay = 1.0
decay = 1 - float(step) / args.steps if step < args.steps else 0.0
for param_group in optimizer.param_groups:
param_group["lr"] = args.learning_rate * decay
image, label = next(train_queue)
image = image.astype("float32")
label = label.astype("int32")
n = image.shape[0]
optimizer.zero_grad()
loss, acc1, acc5 = train_func(image, label)
optimizer.step()
top1.update(100 * acc1.numpy()[0], n)
top5.update(100 * acc5.numpy()[0], n)
objs.update(loss.numpy()[0], n)
total_time.update(time.time() - t)
t = time.time()
if step % args.report_freq == 0 and rank == 0:
logger.info(
"TRAIN %06d %f %s %s %s %s",
step,
args.learning_rate * decay,
objs,
top1,
top5,
total_time,
)
objs.reset()
top1.reset()
top5.reset()
total_time.reset()
if step % 10000 == 0 and rank == 0:
logger.info("SAVING %06d", step)
mge.save(
model.state_dict(),
os.path.join(save_dir, "checkpoint-{:06d}.pkl".format(step)),
)
if step % 10000 == 0 and step != 0:
_, valid_acc, valid_acc5 = infer(valid_func, valid_queue, args)
logger.info("TEST %06d %f, %f", step, valid_acc, valid_acc5)
mge.save(model.state_dict(),
os.path.join(save_dir, "checkpoint-{:06d}.pkl".format(step)))
_, valid_acc, valid_acc5 = infer(valid_func, valid_queue, args)
logger.info("TEST %06d %f, %f", step, valid_acc, valid_acc5)
def worker(rank, world_size, args):
# pylint: disable=too-many-statements
mge.set_log_file(os.path.join(args.save, args.arch, "log.txt"))
if world_size > 1:
# Initialize distributed process group
logger.info("init distributed process group {} / {}".format(
rank, world_size))
dist.init_process_group(
master_ip="localhost",
master_port=23456,
world_size=world_size,
rank=rank,
dev=rank,
)
save_dir = os.path.join(args.save, args.arch)
model = getattr(M, args.arch)()
step_start = 0
if args.model:
logger.info("load weights from %s", args.model)
model.load_state_dict(mge.load(args.model))
step_start = int(args.model.split("-")[1].split(".")[0])
optimizer = optim.SGD(
get_parameters(model),
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay,
)
# Define train and valid graph
@jit.trace(symbolic=True)
def train_func(image, label):
model.train()
logits = model(image)
loss = F.cross_entropy_with_softmax(logits, label, label_smooth=0.)
acc1, acc5 = F.accuracy(logits, label, (1, 5))
optimizer.backward(loss) # compute gradients
if dist.is_distributed(): # all_reduce_mean
loss = dist.all_reduce_sum(loss,
"train_loss") / dist.get_world_size()
acc1 = dist.all_reduce_sum(acc1,
"train_acc1") / dist.get_world_size()
acc5 = dist.all_reduce_sum(acc5,
"train_acc5") / dist.get_world_size()
return loss, acc1, acc5
@jit.trace(symbolic=True)
def valid_func(image, label):
model.eval()
logits = model(image)
loss = F.cross_entropy_with_softmax(logits, label, label_smooth=0.)
acc1, acc5 = F.accuracy(logits, label, (1, 5))
if dist.is_distributed(): # all_reduce_mean
loss = dist.all_reduce_sum(loss,
"valid_loss") / dist.get_world_size()
acc1 = dist.all_reduce_sum(acc1,
"valid_acc1") / dist.get_world_size()
acc5 = dist.all_reduce_sum(acc5,
"valid_acc5") / dist.get_world_size()
return loss, acc1, acc5
# Build train and valid datasets
logger.info("preparing dataset..")
train_dataset = data.dataset.ImageNet(args.data, train=True)
train_sampler = data.Infinite(
data.RandomSampler(train_dataset,
batch_size=args.batch_size,
drop_last=True))
train_queue = data.DataLoader(
train_dataset,
sampler=train_sampler,
transform=T.Compose([
T.RandomResizedCrop(224),
T.RandomHorizontalFlip(),
T.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4),
T.ToMode("CHW"),
]),
num_workers=args.workers,
)
valid_dataset = data.dataset.ImageNet(args.data, train=False)
valid_sampler = data.SequentialSampler(valid_dataset,
batch_size=100,
drop_last=False)
valid_queue = data.DataLoader(
valid_dataset,
sampler=valid_sampler,
transform=T.Compose([
T.Resize(256),
T.CenterCrop(224),
T.ToMode("CHW"),
]),
num_workers=args.workers,
)
# Start training
objs = AverageMeter("Loss")
top1 = AverageMeter("Acc@1")
top5 = AverageMeter("Acc@5")
total_time = AverageMeter("Time")
t = time.time()
for step in range(step_start, args.steps + 1):
# Linear learning rate decay
decay = 1.0
decay = 1 - float(step) / args.steps if step < args.steps else 0
for param_group in optimizer.param_groups:
param_group["lr"] = args.learning_rate * decay
image, label = next(train_queue)
time_data = time.time() - t
image = image.astype("float32")
label = label.astype("int32")
n = image.shape[0]
optimizer.zero_grad()
loss, acc1, acc5 = train_func(image, label)
optimizer.step()
top1.update(100 * acc1.numpy()[0], n)
top5.update(100 * acc5.numpy()[0], n)
objs.update(loss.numpy()[0], n)
total_time.update(time.time() - t)
time_iter = time.time() - t
t = time.time()
if step % args.report_freq == 0 and rank == 0:
logger.info(
"TRAIN Iter %06d: lr = %f,\tloss = %f,\twc_loss = 1,\tTop-1 err = %f,\tTop-5 err = %f,\tdata_time = %f,\ttrain_time = %f,\tremain_hours=%f",
step,
args.learning_rate * decay,
float(objs.__str__().split()[1]),
1 - float(top1.__str__().split()[1]) / 100,
1 - float(top5.__str__().split()[1]) / 100,
time_data,
time_iter - time_data,
time_iter * (args.steps - step) / 3600,
)
objs.reset()
top1.reset()
top5.reset()
total_time.reset()
if step % 10000 == 0 and rank == 0 and step != 0:
logger.info("SAVING %06d", step)
mge.save(
model.state_dict(),
os.path.join(save_dir, "checkpoint-{:06d}.pkl".format(step)),
)
if step % 50000 == 0 and step != 0:
_, valid_acc, valid_acc5 = infer(valid_func, valid_queue, args)
logger.info(
"TEST Iter %06d: loss = %f,\tTop-1 err = %f,\tTop-5 err = %f",
step, _, 1 - valid_acc / 100, 1 - valid_acc5 / 100)
mge.save(model.state_dict(),
os.path.join(save_dir, "checkpoint-{:06d}.pkl".format(step)))
_, valid_acc, valid_acc5 = infer(valid_func, valid_queue, args)
logger.info("TEST Iter %06d: loss=%f,\tTop-1 err = %f,\tTop-5 err = %f",
step, _, 1 - valid_acc / 100, 1 - valid_acc5 / 100)
def worker(rank, world_size, args):
if world_size > 1:
# Initialize distributed process group
logger.info("init distributed process group {} / {}".format(
rank, world_size))
dist.init_process_group(
master_ip="localhost",
master_port=23456,
world_size=world_size,
rank=rank,
dev=rank,
)
save_dir = os.path.join(args.save, args.arch)
model = getattr(M, args.arch)()
optimizer = optim.SGD(
model.parameters(requires_grad=True),
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay,
)
scheduler = optim.MultiStepLR(optimizer, [30, 60, 80])
# Define train and valid graph
@jit.trace(symbolic=True)
def train_func(image, label):
model.train()
logits = model(image)
loss = F.cross_entropy_with_softmax(logits, label)
acc1, acc5 = F.accuracy(logits, label, (1, 5))
optimizer.backward(loss) # compute gradients
if dist.is_distributed(): # all_reduce_mean
loss = dist.all_reduce_sum(loss,
"train_loss") / dist.get_world_size()
acc1 = dist.all_reduce_sum(acc1,
"train_acc1") / dist.get_world_size()
acc5 = dist.all_reduce_sum(acc5,
"train_acc5") / dist.get_world_size()
return loss, acc1, acc5
@jit.trace(symbolic=True)
def valid_func(image, label):
model.eval()
logits = model(image)
loss = F.cross_entropy_with_softmax(logits, label)
acc1, acc5 = F.accuracy(logits, label, (1, 5))
if dist.is_distributed(): # all_reduce_mean
loss = dist.all_reduce_sum(loss,
"valid_loss") / dist.get_world_size()
acc1 = dist.all_reduce_sum(acc1,
"valid_acc1") / dist.get_world_size()
acc5 = dist.all_reduce_sum(acc5,
"valid_acc5") / dist.get_world_size()
return loss, acc1, acc5
# Build train and valid datasets
logger.info("preparing dataset..")
train_dataset = data.dataset.ImageNet(args.data, train=True)
train_sampler = data.RandomSampler(train_dataset,
batch_size=args.batch_size,
drop_last=True)
train_queue = data.DataLoader(
train_dataset,
sampler=train_sampler,
transform=T.Compose([ # Baseline Augmentation for small models
T.RandomResizedCrop(224),
T.RandomHorizontalFlip(),
T.Normalize(mean=[103.530, 116.280, 123.675],
std=[57.375, 57.120, 58.395]), # BGR
T.ToMode("CHW"),
]) if args.arch in ("resnet18", "resnet34") else T.Compose(
[ # Facebook Augmentation for large models
T.RandomResizedCrop(224),
T.RandomHorizontalFlip(),
T.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4),
T.Lighting(0.1),
T.Normalize(mean=[103.530, 116.280, 123.675],
std=[57.375, 57.120, 58.395]), # BGR
T.ToMode("CHW"),
]),
num_workers=args.workers,
)
valid_dataset = data.dataset.ImageNet(args.data, train=False)
valid_sampler = data.SequentialSampler(valid_dataset,
batch_size=100,
drop_last=False)
valid_queue = data.DataLoader(
valid_dataset,
sampler=valid_sampler,
transform=T.Compose([
T.Resize(256),
T.CenterCrop(224),
T.Normalize(mean=[103.530, 116.280, 123.675],
std=[57.375, 57.120, 58.395]), # BGR
T.ToMode("CHW"),
]),
num_workers=args.workers,
)
# Start training
top1_acc = 0
for epoch in range(0, args.epochs):
logger.info("Epoch %d LR %.3e", epoch, scheduler.get_lr()[0])
_, train_acc, train_acc5 = train(train_func,
train_queue,
optimizer,
args,
epoch=epoch)
logger.info("Epoch %d Train %.3f / %.3f", epoch, train_acc, train_acc5)
_, valid_acc, valid_acc5 = infer(valid_func,
valid_queue,
args,
epoch=epoch)
logger.info("Epoch %d Valid %.3f / %.3f", epoch, valid_acc, valid_acc5)
scheduler.step()
if rank == 0: # save checkpoint
mge.save(
{
"epoch": epoch + 1,
"state_dict": model.state_dict(),
"accuracy": valid_acc,
},
os.path.join(save_dir, "checkpoint.pkl"),
)
if valid_acc > top1_acc:
top1_acc = valid_acc
shutil.copy(
os.path.join(save_dir, "checkpoint.pkl"),
os.path.join(save_dir, "model_best.pkl"),
)