def save_checkpoint(self, out_dir, create_symlink=True):
"""Save the checkpoint.
Args:
out_dir (str): The directory that checkpoints are saved.
save_optimizer (bool, optional): Whether to save the optimizer to
the checkpoint. Defaults to True.
create_symlink (bool, optional): Whether to create a symlink
"latest.pth" to point to the latest checkpoint.
Defaults to True.
"""
filename_tmpl = "epoch_{}"
filename = filename_tmpl.format(self.epoch + 1)
filepath = osp.join(out_dir, filename)
self.logger.info('save checkpoint to {}'.format(filepath))
mkdir_or_exist(filepath)
if isinstance(self.model.optimizers, dict):
for key in self.model.optimizers.keys():
submodule = getattr(self.model, key, None)
assert submodule is not None, "model should have submodule {}".format(
key)
assert isinstance(
submodule, Module
), "submodule should be instance of megengine.module.Module"
mge.save(submodule.state_dict(),
osp.join(filepath, key + module_ckpt_suffix))
mge.save(self.model.optimizers[key].state_dict(),
osp.join(filepath, key + optim_ckpt_suffix))
else:
raise TypeError(
" the type of optimizers should be dict for save_checkpoint")
if create_symlink:
pass
def test_load_quantized():
from megengine.core.tensor import dtype
data_shape = (2, 28)
data = tensor(np.random.random(data_shape), dtype="float32")
data = data.astype(dtype.qint8(0.1))
mlp = MLP()
quantize_qat(mlp)
quantize(mlp)
mlp.dense0.weight = Parameter(
mlp.dense0.weight.astype(dtype.qint8(0.001)).numpy())
mlp.dense1.weight = Parameter(
mlp.dense1.weight.astype(dtype.qint8(0.0002)).numpy())
mlp.eval()
pred0 = mlp(data)
with BytesIO() as fout:
mge.save(mlp.state_dict(), fout)
fout.seek(0)
checkpoint = mge.load(fout)
# change mlp weight.
mlp.dense0.weight = Parameter(
mlp.dense0.weight.astype(dtype.qint8(0.00001)).numpy())
mlp.dense1.weight = Parameter(
mlp.dense1.weight.astype(dtype.qint8(0.2)).numpy())
mlp.load_state_dict(checkpoint)
pred1 = mlp(data)
np.testing.assert_allclose(pred0.astype("float32").numpy(),
pred1.astype("float32").numpy(),
atol=5e-6)
def train(dataloader, args):
writer = SummaryWriter("runs")
net = Net()
net.train()
optimizer = SGD(net.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.wd)
gm = GradManager().attach(net.parameters())
epoch_length = len(dataloader)
for epoch in range(args.epoch):
for step, (batch_data, batch_label) in enumerate(dataloader):
batch_label = batch_label.astype(np.int32)
data, label = mge.tensor(batch_data), mge.tensor(batch_label)
with gm:
pred = net(data)
loss = F.loss.cross_entropy(pred, label)
gm.backward(loss)
optimizer.step().clear_grad()
if step % 50 == 0:
print("epoch:{}, iter:{}, loss:{}".format(epoch + 1, step, float(loss))) # noqa
writer.add_scalar("loss", float(loss), epoch * epoch_length + step)
if (epoch + 1) % 5 == 0:
mge.save(
net.state_dict(), os.path.join(gettempdir(), f"mnist_net_e{epoch + 1}.pkl"),
) # noqa
def update_model(model_path):
"""
Update the dumped model with test cases for new reference values.
The model with pre-trained weights is trained for one iter with the test data attached.
The loss and updated net state dict is dumped.
.. code-block:: python
from test_correctness import update_model
update_model('mnist_model_with_test.mge') # for gpu
update_model('mnist_model_with_test_cpu.mge') # for cpu
"""
net = MnistNet(has_bn=True)
checkpoint = mge.load(model_path)
net.load_state_dict(checkpoint["net_init"])
lr = checkpoint["sgd_lr"]
opt = SGD(net.parameters(), lr=lr)
data = tensor(dtype=np.float32)
label = tensor(dtype=np.int32)
data.set_value(checkpoint["data"])
label.set_value(checkpoint["label"])
opt.zero_grad()
loss = train(data, label, net=net, opt=opt)
opt.step()
xpu_name = get_xpu_name()
checkpoint.update(
{"net_updated": net.state_dict(), "loss": loss.numpy(), "xpu": xpu_name}
)
mge.save(checkpoint, model_path)
def test_load_quantized():
data_shape = (2, 28)
data = tensor(np.random.random(data_shape), dtype="float32")
data = data.astype(mgb.dtype.qint8(0.1))
mlp = MLP()
quantize_qat(mlp)
quantize(mlp)
mlp.dense0.weight = Parameter(
mlp.dense0.weight.astype(mgb.dtype.qint8(0.001)).numpy())
mlp.dense1.weight = Parameter(
mlp.dense1.weight.astype(mgb.dtype.qint8(0.0002)).numpy())
mlp.eval()
pred0 = mlp(data)
with BytesIO() as fout:
mge.save(mlp.state_dict(), fout)
fout.seek(0)
checkpoint = mge.load(fout)
# change mlp weight.
mlp.dense0.weight = Parameter(
mlp.dense0.weight.astype(mgb.dtype.qint8(0.00001)).numpy())
mlp.dense1.weight = Parameter(
mlp.dense1.weight.astype(mgb.dtype.qint8(0.2)).numpy())
mlp.load_state_dict(checkpoint)
pred1 = mlp(data)
assertTensorClose(pred0.astype("float32").numpy(),
pred1.astype("float32").numpy(),
max_err=5e-6)
def worker(args):
current_network = import_from_file(args.file)
model = current_network.Net(current_network.Cfg())
model.train()
if dist.get_rank() == 0:
logger.info(get_config_info(model.cfg))
logger.info(repr(model))
params_with_grad = []
for name, param in model.named_parameters():
if "bottom_up.conv1" in name and model.cfg.backbone_freeze_at >= 1:
continue
if "bottom_up.layer1" in name and model.cfg.backbone_freeze_at >= 2:
continue
params_with_grad.append(param)
opt = SGD(
params_with_grad,
lr=model.cfg.basic_lr * args.batch_size,
momentum=model.cfg.momentum,
weight_decay=model.cfg.weight_decay * dist.get_world_size(),
)
gm = GradManager()
if dist.get_world_size() > 1:
gm.attach(params_with_grad,
callbacks=[dist.make_allreduce_cb("SUM", dist.WORLD)])
else:
gm.attach(params_with_grad)
if args.weight_file is not None:
# model.backbone.bottom_up.load_state_dict(weights, strict=False)
logger.info("Loading Base-Pretrain weights...")
weights = mge.load(args.weight_file)
weight_new = {k: v for k, v in weights.items() if 'pred_' not in k}
model.load_state_dict(weight_new, strict=False)
if dist.get_world_size() > 1:
dist.bcast_list_(model.parameters(), dist.WORLD) # sync parameters
if dist.get_rank() == 0:
logger.info("Prepare dataset")
train_loader = iter(
build_dataloader(args.batch_size, args.dataset_dir, model.cfg))
for epoch in range(model.cfg.max_epoch):
train_one_epoch(model, train_loader, opt, gm, epoch, args)
if dist.get_rank() == 0:
save_path = "logs/{}/epoch_{}.pkl".format(
os.path.basename(args.file).split(".")[0], epoch)
mge.save(
{
"epoch": epoch,
"state_dict": model.state_dict()
},
save_path,
)
logger.info("dump weights to %s", save_path)
def worker(rank, world_size, args):
if world_size > 1:
dist.init_process_group(
master_ip="localhost",
master_port=23456,
world_size=world_size,
rank=rank,
dev=rank,
)
logger.info("Init process group for gpu%d done", rank)
sys.path.insert(0, os.path.dirname(args.file))
current_network = importlib.import_module(
os.path.basename(args.file).split(".")[0])
model = current_network.Net(current_network.Cfg(),
batch_size=args.batch_size)
params = model.parameters(requires_grad=True)
model.train()
if rank == 0:
logger.info(get_config_info(model.cfg))
opt = optim.SGD(
params,
lr=model.cfg.basic_lr * world_size * model.batch_size,
momentum=model.cfg.momentum,
weight_decay=model.cfg.weight_decay,
)
if args.weight_file is not None:
weights = mge.load(args.weight_file)
model.backbone.bottom_up.load_state_dict(weights)
logger.info("Prepare dataset")
loader = build_dataloader(model.batch_size, args.dataset_dir, model.cfg)
train_loader = iter(loader["train"])
for epoch_id in range(model.cfg.max_epoch):
for param_group in opt.param_groups:
param_group["lr"] = (model.cfg.basic_lr * world_size *
model.batch_size *
(model.cfg.lr_decay_rate**bisect.bisect_right(
model.cfg.lr_decay_sates, epoch_id)))
tot_steps = model.cfg.nr_images_epoch // (model.batch_size *
world_size)
train_one_epoch(model, train_loader, opt, tot_steps, rank, epoch_id,
world_size)
if rank == 0:
save_path = "log-of-{}/epoch_{}.pkl".format(
os.path.basename(args.file).split(".")[0], epoch_id)
mge.save(
{
"epoch": epoch_id,
"state_dict": model.state_dict()
},
save_path,
)
logger.info("dump weights to %s", save_path)
def worker(rank, gpu_num, args):
# using sublinear
os.environ[
"MGB_COMP_GRAPH_OPT"] = "enable_sublinear_memory_opt=1;seq_opt.enable_seq_comp_node_opt=0"
os.environ["MGB_SUBLINEAR_MEMORY_GENETIC_NR_ITER"] = '50'
if gpu_num > 1:
dist.init_process_group(
master_ip="localhost",
master_port=23456,
world_size=gpu_num,
rank=rank,
dev=rank,
)
logger.info("Init process group for gpu%d done", rank)
model = network.Network()
params = model.parameters(requires_grad=True)
model.train()
opt = optim.SGD(
params,
lr=cfg.basic_lr * gpu_num * cfg.batch_per_gpu,
momentum=cfg.momentum,
weight_decay=cfg.weight_decay,
)
if cfg.pretrain_weight is not None:
weights = mge.load(cfg.pretrain_weight)
del weights['fc.weight']
del weights['fc.bias']
model.resnet50.load_state_dict(weights)
logger.info("Prepare dataset")
train_loader = dataset.train_dataset(rank)
logger.info("Training...")
for epoch_id in range(cfg.max_epoch):
for param_group in opt.param_groups:
param_group["lr"] = (cfg.basic_lr * gpu_num * cfg.batch_per_gpu *
(cfg.lr_decay_rate**bisect.bisect_right(
cfg.lr_decay_sates, epoch_id)))
max_steps = cfg.nr_images_epoch // (cfg.batch_per_gpu * gpu_num)
train_one_epoch(model, train_loader, opt, max_steps, rank, epoch_id,
gpu_num)
if rank == 0:
save_path = os.path.join(cfg.model_dir,
'epoch_{}.pkl'.format(epoch_id + 1))
mge.save(
{
"epoch": epoch_id,
"state_dict": model.state_dict()
},
save_path,
)
logger.info("dump weights to %s", save_path)
def test_tensor_serialization():
def tensor_eq(a, b):
assert a.dtype == b.dtype
assert a.device == b.device
np.testing.assert_equal(a.numpy(), b.numpy())
with TemporaryFile() as f:
data = np.random.randint(low=0, high=7, size=[233])
a = Tensor(data, device="xpux", dtype=np.int32)
pickle.dump(a, f)
f.seek(0)
b = pickle.load(f)
np.testing.assert_equal(a.numpy(), b.numpy())
with TemporaryFile() as f:
a = Parameter(np.random.random(size=(233, 2)).astype(np.float32))
pickle.dump(a, f)
f.seek(0)
b = pickle.load(f)
assert isinstance(b, Parameter)
np.testing.assert_equal(a.numpy(), b.numpy())
with TemporaryFile() as f:
a = Tensor(np.random.random(size=(2, 233)).astype(np.float32))
pickle.dump(a, f)
f.seek(0)
b = pickle.load(f)
assert type(b) is Tensor
np.testing.assert_equal(a.numpy(), b.numpy())
with TemporaryFile() as f:
a = Tensor(np.random.random(size=(2, 233)).astype(np.float32))
mge.save(a, f)
f.seek(0)
b = mge.load(f, map_location="cpux")
assert type(b) is Tensor
assert "cpu" in str(b.device)
np.testing.assert_equal(a.numpy(), b.numpy())
with TemporaryFile() as f:
if mge.is_cuda_available():
device_org = mge.get_default_device()
mge.set_default_device("gpu0")
a = Tensor(np.random.random(size=(2, 233)).astype(np.float32))
mge.save(a, f)
f.seek(0)
mge.set_default_device("cpux")
b = mge.load(f, map_location={"gpu0": "cpu0"})
assert type(b) is Tensor
assert "cpu0" in str(b.device)
np.testing.assert_equal(a.numpy(), b.numpy())
mge.set_default_device(device_org)
def test_save_load():
net = Simple()
optim = optimizer.SGD(net.parameters(), lr=1.0, momentum=0.9)
optim.clear_grad()
gm = ad.GradManager().attach(net.parameters())
data = tensor([2.34])
with gm:
loss = net(data)
gm.backward(loss)
optim.step()
model_name = "simple.pkl"
print("save to {}".format(model_name))
mge.save(
{
"name": "simple",
"state_dict": net.state_dict(),
"opt_state": optim.state_dict(),
},
model_name,
)
# Load param to cpu
checkpoint = mge.load(model_name, map_location="cpu0")
device_save = mge.get_default_device()
mge.set_default_device("cpu0")
net = Simple()
net.load_state_dict(checkpoint["state_dict"])
optim = optimizer.SGD(net.parameters(), lr=1.0, momentum=0.9)
optim.load_state_dict(checkpoint["opt_state"])
print("load done")
os.remove("simple.pkl")
with gm:
loss = net([1.23])
gm.backward(loss)
optim.step()
# Restore device
mge.set_default_device(device_save)
def _check_module(build_func: Callable):
net = build_func()
net.eval()
buffer = io.BytesIO()
mge.save(net.state_dict(), buffer)
buffer.seek(0)
inp = Tensor(np.random.random(size=(5, 3, 32, 32)))
traced_net = trace_module(build_func(), inp)
traced_net.load_state_dict(mge.load(buffer))
_check_param(net, traced_net)
buffer.seek(0)
traced_net = trace_module(build_func(), inp).flatten()
traced_net.load_state_dict(mge.load(buffer))
_check_param(net, traced_net)
def test_optimizer_serialization():
data, data_shape, label, label_shape = get_input()
mlp = MLP()
opt = SGD(mlp.parameters(), lr=0.01, momentum=0.9)
slots = TensorDict()
for param in mlp.parameters():
slots[param] = np.zeros(param.shape).astype(np.float32)
pred = mlp(data)
loss = F.square_loss(pred, label.reshape(-1, 1))
opt.zero_grad()
opt.backward(loss)
opt.step()
for param in mlp.parameters():
slot = slots[param]
slot *= 0.9
slot -= param.grad.numpy() * 0.01
with BytesIO() as fout:
save(opt.state_dict(), fout)
fout.seek(0)
state_dict = load(fout)
opt1 = SGD(mlp.parameters(), lr=0.02, momentum=0.8)
opt1.load_state_dict(state_dict)
data.set_value(np.random.random(data_shape).astype(np.float32))
label.set_value(np.random.randint(0, 10, label_shape))
pred = mlp(data)
loss = F.square_loss(pred, label.reshape(-1, 1))
opt1.zero_grad()
opt1.backward(loss)
orig_params = TensorDict()
for param in mlp.parameters():
orig_params[param] = np.copy(param.numpy())
opt1.step()
for param in mlp.parameters():
orig_param = orig_params[param]
slot = slots[param]
slot *= 0.9
slot -= param.grad.numpy() * 0.01
assertTensorClose(param.numpy(), orig_param + slot)
def test_state_dict():
data_shape = (2, 28)
data = tensor(np.random.random(data_shape))
mlp = MLP()
pred0 = mlp(data)
with BytesIO() as fout:
mge.save(mlp.state_dict(), fout)
fout.seek(0)
state_dict = mge.load(fout)
state_dict["extra"] = None
mlp1 = MLP()
mlp1.load_state_dict(state_dict, strict=False)
pred1 = mlp1(data)
np.testing.assert_allclose(pred0.numpy(), pred1.numpy(), atol=5e-6)
with pytest.raises(KeyError):
mlp1.load_state_dict(state_dict)
del state_dict["extra"]
del state_dict["dense0.bias"]
with pytest.raises(KeyError):
mlp1.load_state_dict(state_dict)
def test_shared_param_1d():
net = Simple2()
assert net.conv0.weight is net.conv1.weight
data = tensor(np.random.random((1, 1, 8)).astype(np.float32))
np.testing.assert_allclose(
net.conv0(data).numpy(),
net.conv1(data).numpy())
with BytesIO() as f:
mge.save(net, f)
f.seek(0)
net1 = mge.load(f)
assert net1.conv0.weight is net1.conv1.weight
np.testing.assert_allclose(
net1.conv0(data).numpy(),
net1.conv1(data).numpy())
with BytesIO() as f:
mge.save(net.conv0, f)
f.seek(0)
conv0 = mge.load(f)
with BytesIO() as f:
mge.save(net.conv1, f)
f.seek(0)
conv1 = mge.load(f)
assert conv0.weight is not conv1.weight
np.testing.assert_allclose(conv0(data).numpy(), conv1(data).numpy())
def test_pickle_module():
data_shape = (2, 28)
data = tensor(np.random.random(data_shape))
mlp = MLP()
# pickle before forward
with BytesIO() as fout:
mge.save(mlp, fout)
fout.seek(0)
mlp1 = mge.load(fout)
pred0 = mlp1(data)
pred1 = mlp(data)
# pickle after forward
with BytesIO() as fout:
mge.save(mlp, fout)
fout.seek(0)
mlp1 = mge.load(fout)
pred2 = mlp1(data)
np.testing.assert_allclose(pred0.numpy(), pred1.numpy(), atol=5e-6)
np.testing.assert_allclose(pred0.numpy(), pred2.numpy(), atol=5e-6)
def save_checkpoint(self,
directory,
global_step,
optimizer=None,
name=None):
r"""
Saves checkpoint at a certain global step during training. Optimizer state
is also saved together.
Args:
directory (str): Path to save checkpoint to.
global_step (int): The global step variable during training.
optimizer (Optimizer): Optimizer state to be saved concurrently.
name (str): The name to save the checkpoint file as.
Returns:
None
"""
# Create directory to save to
if not os.path.exists(directory):
os.makedirs(directory)
# Build checkpoint dict to save.
ckpt_dict = {
'model_state_dict':
self.state_dict(),
'optimizer_state_dict':
optimizer.state_dict() if optimizer is not None else None,
'global_step':
global_step
}
# Save the file with specific name
if name is None:
name = "{}_{}_steps.pth".format(
os.path.basename(directory), # netD or netG
global_step)
megengine.save(ckpt_dict, os.path.join(directory, name))
def map_weights(weight_file, output_file):
torch_weights = numpy_weights(weight_file)
new_dict = OrderedDict()
for k, v in torch_weights.items():
if "num_batches_tracked" in k:
print("drop: {}".format(k))
continue
if k.endswith("bias"):
print("bias key: {}".format(k))
v = v.reshape(1, -1, 1, 1)
new_dict[k] = v
elif "dconv" in k and "conv.weight" in k:
print("depthwise conv key: {}".format(k))
cout, cin, k1, k2 = v.shape
v = v.reshape(cout, 1, cin, k1, k2)
new_dict[k] = v
else:
new_dict[k] = v
mge.save(new_dict, output_file)
print("save weights to {}".format(output_file))
def test_pickle_module():
data_shape = (2, 28)
data = tensor()
data.set_value(np.random.random(data_shape))
mlp = MLP()
# pickle before forward
with BytesIO() as fout:
mge.save(mlp, fout)
fout.seek(0)
mlp1 = mge.load(fout)
pred0 = mlp1(data)
pred1 = mlp(data)
# pickle after forward
with BytesIO() as fout:
mge.save(mlp, fout)
fout.seek(0)
mlp1 = mge.load(fout)
pred2 = mlp1(data)
assertTensorClose(pred0.numpy(), pred1.numpy(), max_err=5e-6)
assertTensorClose(pred0.numpy(), pred2.numpy(), max_err=5e-6)
def worker(rank, world_size, args):
# pylint: disable=too-many-statements
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 + "." + args.mode)
if not os.path.exists(save_dir):
os.makedirs(save_dir, exist_ok=True)
mge.set_log_file(os.path.join(save_dir, "log.txt"))
model = models.__dict__[args.arch]()
cfg = config.get_finetune_config(args.arch)
cfg.LEARNING_RATE *= world_size # scale learning rate in distributed training
total_batch_size = cfg.BATCH_SIZE * world_size
steps_per_epoch = 1280000 // total_batch_size
total_steps = steps_per_epoch * cfg.EPOCHS
if args.mode != "normal":
Q.quantize_qat(model, Q.ema_fakequant_qconfig)
if args.checkpoint:
logger.info("Load pretrained weights from %s", args.checkpoint)
ckpt = mge.load(args.checkpoint)
ckpt = ckpt["state_dict"] if "state_dict" in ckpt else ckpt
model.load_state_dict(ckpt, strict=False)
if args.mode == "quantized":
raise ValueError("mode = quantized only used during inference")
Q.quantize(model)
optimizer = optim.SGD(
get_parameters(model, cfg),
lr=cfg.LEARNING_RATE,
momentum=cfg.MOMENTUM,
)
# 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.Infinite(
data.RandomSampler(train_dataset,
batch_size=cfg.BATCH_SIZE,
drop_last=True))
train_queue = data.DataLoader(
train_dataset,
sampler=train_sampler,
transform=T.Compose([
T.RandomResizedCrop(224),
T.RandomHorizontalFlip(),
cfg.COLOR_JITTOR,
T.Normalize(mean=128),
T.ToMode("CHW"),
]),
num_workers=args.workers,
)
train_queue = 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=128),
T.ToMode("CHW"),
]),
num_workers=args.workers,
)
def adjust_learning_rate(step, epoch):
learning_rate = cfg.LEARNING_RATE
if cfg.SCHEDULER == "Linear":
learning_rate *= 1 - float(step) / total_steps
elif cfg.SCHEDULER == "Multistep":
learning_rate *= cfg.SCHEDULER_GAMMA**bisect.bisect_right(
cfg.SCHEDULER_STEPS, epoch)
else:
raise ValueError(cfg.SCHEDULER)
for param_group in optimizer.param_groups:
param_group["lr"] = learning_rate
return learning_rate
# 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, total_steps):
# Linear learning rate decay
epoch = step // steps_per_epoch
learning_rate = adjust_learning_rate(step, epoch)
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 e%d %06d %f %s %s %s %s", epoch, step,
learning_rate, 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(
{
"step": step,
"state_dict": model.state_dict()
},
os.path.join(save_dir, "checkpoint.pkl"),
)
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({
"step": step,
"state_dict": model.state_dict()
}, os.path.join(save_dir, "checkpoint-final.pkl"))
_, 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(world_size, args):
# pylint: disable=too-many-statements
rank = dist.get_rank()
if world_size > 1:
logger.info("init distributed process group {} / {}".format(
rank, world_size))
save_dir = os.path.join(args.save, args.arch + "." + args.mode)
if not os.path.exists(save_dir):
os.makedirs(save_dir, exist_ok=True)
mge.set_log_file(os.path.join(save_dir, "log.txt"))
model = models.__dict__[args.arch]()
cfg = config.get_config(args.arch)
cfg.LEARNING_RATE *= world_size # scale learning rate in distributed training
total_batch_size = cfg.BATCH_SIZE * world_size
steps_per_epoch = 1280000 // total_batch_size
total_steps = steps_per_epoch * cfg.EPOCHS
if args.mode != "normal":
quantize_qat(model, qconfig=Q.ema_fakequant_qconfig)
if world_size > 1:
# Sync parameters
dist.bcast_list_(model.parameters(), dist.WORLD)
# Autodiff gradient manager
gm = autodiff.GradManager().attach(
model.parameters(),
callbacks=dist.make_allreduce_cb("MEAN") if world_size > 1 else None,
)
optimizer = optim.SGD(
get_parameters(model, cfg),
lr=cfg.LEARNING_RATE,
momentum=cfg.MOMENTUM,
)
# Define train and valid graph
def train_func(image, label):
with gm:
model.train()
logits = model(image)
loss = F.loss.cross_entropy(logits, label, label_smooth=0.1)
acc1, acc5 = F.topk_accuracy(logits, label, (1, 5))
gm.backward(loss)
optimizer.step().clear_grad()
return loss, acc1, acc5
def valid_func(image, label):
model.eval()
logits = model(image)
loss = F.loss.cross_entropy(logits, label, label_smooth=0.1)
acc1, acc5 = F.topk_accuracy(logits, label, (1, 5))
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=cfg.BATCH_SIZE,
drop_last=True))
train_queue = data.DataLoader(
train_dataset,
sampler=train_sampler,
transform=T.Compose([
T.RandomResizedCrop(224),
T.RandomHorizontalFlip(),
cfg.COLOR_JITTOR,
T.Normalize(mean=128),
T.ToMode("CHW"),
]),
num_workers=args.workers,
)
train_queue = 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=128),
T.ToMode("CHW")
]),
num_workers=args.workers,
)
def adjust_learning_rate(step, epoch):
learning_rate = cfg.LEARNING_RATE
if cfg.SCHEDULER == "Linear":
learning_rate *= 1 - float(step) / total_steps
elif cfg.SCHEDULER == "Multistep":
learning_rate *= cfg.SCHEDULER_GAMMA**bisect.bisect_right(
cfg.SCHEDULER_STEPS, epoch)
else:
raise ValueError(cfg.SCHEDULER)
for param_group in optimizer.param_groups:
param_group["lr"] = learning_rate
return learning_rate
# 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, total_steps):
# Linear learning rate decay
epoch = step // steps_per_epoch
learning_rate = adjust_learning_rate(step, epoch)
image, label = next(train_queue)
image = mge.tensor(image, dtype="float32")
label = mge.tensor(label, dtype="int32")
n = image.shape[0]
loss, acc1, acc5 = train_func(image, label)
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 e%d %06d %f %s %s %s %s",
epoch,
step,
learning_rate,
objs,
top1,
top5,
total_time,
)
objs.reset()
top1.reset()
top5.reset()
total_time.reset()
if step != 0 and step % 10000 == 0 and rank == 0:
logger.info("SAVING %06d", step)
mge.save(
{
"step": step,
"state_dict": model.state_dict()
},
os.path.join(save_dir, "checkpoint.pkl"),
)
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(
{
"step": step,
"state_dict": model.state_dict()
},
os.path.join(save_dir, "checkpoint-final.pkl"),
)
_, valid_acc, valid_acc5 = infer(valid_func, valid_queue, args)
logger.info("TEST %06d %f, %f", step, valid_acc, valid_acc5)
x = self.t1(x, pos)
logit = x
#logit = self.fc(x)
return logit
def forward(self, x, pos):
x = self.get_logit(x, pos)
flatten = x.reshape(-1, E)
return flatten
if __name__ == '__main__':
trans = ReverseString()
mge.save(trans.state_dict(), 'test.mge')
from dataset import get_dataloader, make_string_from_tensor
data_loader = get_dataloader(instance_num=1024)
inp_sequence = mge.tensor(dtype='float32')
position = mge.tensor(dtype='float32')
trans.eval()
for step, batch in enumerate(data_loader):
inp, label, pos, mask = batch
inp_sequence.set_value(inp)
position.set_value(pos)
out = trans.get_logit(inp_sequence, position)
inp = make_string_from_tensor(inp)
print(inp)
print(out)
print(out.shape)
def worker(rank, world_size, ngpus_per_node, args):
# pylint: disable=too-many-statements
if rank == 0:
os.makedirs(os.path.join(args.save, args.arch), exist_ok=True)
megengine.logger.set_log_file(
os.path.join(args.save, args.arch, "log.txt"))
# init process group
if world_size > 1:
dist.init_process_group(
master_ip=args.dist_addr,
port=args.dist_port,
world_size=world_size,
rank=rank,
device=rank % ngpus_per_node,
backend="nccl",
)
logging.info("init process group rank %d / %d", dist.get_rank(),
dist.get_world_size())
# build dataset
train_dataloader, valid_dataloader = build_dataset(args)
train_queue = iter(train_dataloader) # infinite
steps_per_epoch = 1280000 // (world_size * args.batch_size)
# build model
model = resnet_model.__dict__[args.arch]()
# Sync parameters
if world_size > 1:
dist.bcast_list_(model.parameters(), dist.WORLD)
# Autodiff gradient manager
gm = autodiff.GradManager().attach(
model.parameters(),
callbacks=dist.make_allreduce_cb("SUM") if world_size > 1 else None,
)
# Optimizer
opt = optim.SGD(
model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay *
world_size, # scale weight decay in "SUM" mode
)
# train and valid func
def train_step(image, label):
with gm:
logits = model(image)
loss = F.nn.cross_entropy(logits, label)
acc1, acc5 = F.topk_accuracy(logits, label, topk=(1, 5))
gm.backward(loss)
opt.step().clear_grad()
return loss, acc1, acc5
def valid_step(image, label):
logits = model(image)
loss = F.nn.cross_entropy(logits, label)
acc1, acc5 = F.topk_accuracy(logits, label, topk=(1, 5))
# calculate mean values
if world_size > 1:
loss = F.distributed.all_reduce_sum(loss) / world_size
acc1 = F.distributed.all_reduce_sum(acc1) / world_size
acc5 = F.distributed.all_reduce_sum(acc5) / world_size
return loss, acc1, acc5
# multi-step learning rate scheduler with warmup
def adjust_learning_rate(step):
lr = args.lr * 0.1**bisect.bisect_right(
[30 * steps_per_epoch, 60 * steps_per_epoch, 80 * steps_per_epoch],
step)
if step < 5 * steps_per_epoch: # warmup
lr = args.lr * (step / (5 * steps_per_epoch))
for param_group in opt.param_groups:
param_group["lr"] = lr
return lr
# start training
objs = AverageMeter("Loss")
top1 = AverageMeter("Acc@1")
top5 = AverageMeter("Acc@5")
clck = AverageMeter("Time")
for step in range(0, args.epochs * steps_per_epoch):
lr = adjust_learning_rate(step)
t = time.time()
image, label = next(train_queue)
image = megengine.tensor(image, dtype="float32")
label = megengine.tensor(label, dtype="int32")
loss, acc1, acc5 = train_step(image, label)
objs.update(loss.item())
top1.update(100 * acc1.item())
top5.update(100 * acc5.item())
clck.update(time.time() - t)
if step % args.print_freq == 0 and dist.get_rank() == 0:
logging.info(
"Epoch %d Step %d, LR %.4f, %s %s %s %s",
step // steps_per_epoch,
step,
lr,
objs,
top1,
top5,
clck,
)
objs.reset()
top1.reset()
top5.reset()
clck.reset()
if (step + 1) % steps_per_epoch == 0:
model.eval()
_, valid_acc1, valid_acc5 = valid(valid_step, valid_dataloader,
args)
model.train()
logging.info(
"Epoch %d Test Acc@1 %.3f, Acc@5 %.3f",
(step + 1) // steps_per_epoch,
valid_acc1,
valid_acc5,
)
megengine.save(
{
"epoch": (step + 1) // steps_per_epoch,
"state_dict": model.state_dict(),
},
os.path.join(args.save, args.arch, "checkpoint.pkl"),
)
def worker(args):
current_network = import_from_file(args.file)
model = current_network.Net(current_network.Cfg())
model.train()
if dist.get_rank() == 0:
logger.info(get_config_info(model.cfg))
logger.info(repr(model))
backbone_params = []
head_params = []
for name, param in model.named_parameters():
if "backbone" in name:
backbone_params.append(param)
else:
head_params.append(param)
opt = SGD(
[
{
"params": backbone_params,
"lr": model.cfg.learning_rate * 0.1
},
{
"params": head_params
},
],
lr=model.cfg.learning_rate,
momentum=model.cfg.momentum,
weight_decay=model.cfg.weight_decay * dist.get_world_size(),
)
gm = GradManager()
if dist.get_world_size() > 1:
gm.attach(model.parameters(),
callbacks=[dist.make_allreduce_cb("SUM", dist.WORLD)])
else:
gm.attach(model.parameters())
cur_epoch = 0
if args.resume is not None:
pretrained = mge.load(args.resume)
cur_epoch = pretrained["epoch"] + 1
model.load_state_dict(pretrained["state_dict"])
opt.load_state_dict(pretrained["opt"])
if dist.get_rank() == 0:
logger.info("load success: epoch %d", cur_epoch)
if dist.get_world_size() > 1:
dist.bcast_list_(model.parameters(), dist.WORLD) # sync parameters
if dist.get_rank() == 0:
logger.info("Prepare dataset")
train_loader = iter(
build_dataloader(model.cfg.batch_size, args.dataset_dir, model.cfg))
for epoch in range(cur_epoch, model.cfg.max_epoch):
train_one_epoch(model, train_loader, opt, gm, epoch)
if dist.get_rank() == 0:
save_path = "log-of-{}/epoch_{}.pkl".format(
os.path.basename(args.file).split(".")[0], epoch)
mge.save(
{
"epoch": epoch,
"state_dict": model.state_dict(),
"opt": opt.state_dict()
}, save_path)
logger.info("dump weights to %s", save_path)
total_steps += 1
result = {
"train_loss": sum_loss / total_steps,
"train_accuracy": sum_accuracy / total_examples,
}
logger.info("***** Train results *****")
for key in sorted(result.keys()):
logger.info("%s = %s", key, str(result[key]))
if __name__ == "__main__":
bert, config, vocab_file = create_hub_bert(args.pretrained_bert,
pretrained=True)
args.vocab_file = vocab_file
model = BertForSequenceClassification(config, num_labels=2, bert=bert)
mrpc_dataset = MRPCDataset(args)
optimizer = optim.Adam(
model.parameters(requires_grad=True),
lr=args.learning_rate,
)
mrpc_dataset = MRPCDataset(args)
train_dataloader, train_size = mrpc_dataset.get_train_dataloader()
eval_dataloader, eval_size = mrpc_dataset.get_eval_dataloader()
for epoch in range(args.num_train_epochs):
logger.info("***** Epoch {} *****".format(epoch + 1))
train(train_dataloader, model, optimizer)
mge.save(model.state_dict(), args.save_model_path)
eval(eval_dataloader, model)
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"),
)
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(master_ip, port, world_size, rank, configs):
if world_size > 1:
dist.init_process_group(
master_ip=master_ip,
port=port,
world_size=world_size,
rank=rank,
device=rank,
)
logger.info("init process group for gpu{} done".format(rank))
# set up logger
os.makedirs(configs["base_dir"], exist_ok=True)
worklog_path = os.path.join(configs["base_dir"], "worklog.txt")
mge.set_log_file(worklog_path)
# prepare model-related components
model = FaceRecognitionModel(configs)
# prepare data-related components
preprocess = T.Compose([T.Normalize(mean=127.5, std=128), T.ToMode("CHW")])
augment = T.Compose([T.RandomHorizontalFlip()])
train_dataset = get_train_dataset(configs["dataset"],
dataset_dir=configs["dataset_dir"])
train_sampler = data.RandomSampler(train_dataset,
batch_size=configs["batch_size"],
drop_last=True)
train_queue = data.DataLoader(train_dataset,
sampler=train_sampler,
transform=T.Compose([augment, preprocess]))
# prepare optimize-related components
configs["learning_rate"] = configs["learning_rate"] * dist.get_world_size()
if dist.get_world_size() > 1:
dist.bcast_list_(model.parameters())
gm = ad.GradManager().attach(
model.parameters(), callbacks=[dist.make_allreduce_cb("mean")])
else:
gm = ad.GradManager().attach(model.parameters())
opt = optim.SGD(
model.parameters(),
lr=configs["learning_rate"],
momentum=configs["momentum"],
weight_decay=configs["weight_decay"],
)
# try to load checkpoint
model, start_epoch = try_load_latest_checkpoint(model, configs["base_dir"])
# do training
def train_one_epoch():
def train_func(images, labels):
opt.clear_grad()
with gm:
loss, accuracy, _ = model(images, labels)
gm.backward(loss)
if dist.is_distributed():
# all_reduce_mean
loss = dist.functional.all_reduce_sum(
loss) / dist.get_world_size()
accuracy = dist.functional.all_reduce_sum(
accuracy) / dist.get_world_size()
opt.step()
return loss, accuracy
model.train()
average_loss = AverageMeter("loss")
average_accuracy = AverageMeter("accuracy")
data_time = AverageMeter("data_time")
train_time = AverageMeter("train_time")
total_step = len(train_queue)
data_iter = iter(train_queue)
for step in range(total_step):
# get next batch of data
data_tic = time.time()
images, labels = next(data_iter)
data_toc = time.time()
# forward pass & backward pass
train_tic = time.time()
images = mge.tensor(images, dtype="float32")
labels = mge.tensor(labels, dtype="int32")
loss, accuracy = train_func(images, labels)
train_toc = time.time()
# do the statistics and logging
n = images.shape[0]
average_loss.update(loss.item(), n)
average_accuracy.update(accuracy.item() * 100, n)
data_time.update(data_toc - data_tic)
train_time.update(train_toc - train_tic)
if step % configs["log_interval"] == 0 and dist.get_rank() == 0:
logger.info(
"epoch: %d, step: %d, %s, %s, %s, %s",
epoch,
step,
average_loss,
average_accuracy,
data_time,
train_time,
)
for epoch in range(start_epoch, configs["num_epoch"]):
adjust_learning_rate(opt, epoch, configs)
train_one_epoch()
if dist.get_rank() == 0:
checkpoint_path = os.path.join(configs["base_dir"],
f"epoch-{epoch+1}-checkpoint.pkl")
mge.save(
{
"epoch": epoch + 1,
"state_dict": model.state_dict()
},
checkpoint_path,
)
sta = time.time()
for step, (inputs_batched, labels_batched) in enumerate(dataloader):
optimizer.zero_grad()
labels_batched = np.squeeze(labels_batched, -1).astype(np.int32)
logits, loss = train_func(inputs_batched,
labels_batched,
optimizer=optimizer,
net=unet)
optimizer.step()
total_loss += loss.numpy().item()
predicted = np.argmax(logits, axis=1)
correct += ((predicted == labels_batched).sum().item() / (256 * 256.))
total += labels_batched.shape[0]
print("==>>epoch: {:0>3}, loss: {:.4f}, acc: {:.4f}, time: {:.4f}".format(
epoch, total_loss / len(dataloader), correct / total,
(time.time() - sta)))
epoch_loss = total_loss / len(dataloader)
if epoch_loss < loss_src:
print("model saved")
loss_src = epoch_loss
mge.save(unet.state_dict(), path)
print("-" * 50)
print("-use time :", time.time() - star)
def worker(master_ip, port, 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=master_ip,
port=port,
world_size=world_size,
rank=rank,
device=rank,
)
model_name = "{}_{}x{}".format(args.arch, cfg.input_shape[0], cfg.input_shape[1])
save_dir = os.path.join(args.save, model_name)
model = getattr(kpm, args.arch)()
model.train()
start_epoch = 0
if args.resume is not None:
file = mge.load(args.resume)
model.load_state_dict(file["state_dict"])
start_epoch = file["epoch"]
optimizer = optim.Adam(
model.parameters(), lr=cfg.initial_lr, weight_decay=cfg.weight_decay
)
gm = GradManager()
if dist.get_world_size() > 1:
gm.attach(
model.parameters(), callbacks=[dist.make_allreduce_cb("SUM", dist.WORLD)],
)
else:
gm.attach(model.parameters())
if dist.get_world_size() > 1:
dist.bcast_list_(model.parameters(), dist.WORLD) # sync parameters
# Build train datasets
logger.info("preparing dataset..")
ann_file = os.path.join(
cfg.data_root, "annotations", "person_keypoints_train2017.json"
)
train_dataset = COCOJoints(
cfg.data_root,
ann_file,
image_set="train2017",
order=("image", "keypoints", "boxes", "info"),
)
logger.info("Num of Samples: {}".format(len(train_dataset)))
train_sampler = data.RandomSampler(
train_dataset, batch_size=cfg.batch_size, drop_last=True
)
transforms = [
T.Normalize(mean=cfg.img_mean, std=cfg.img_std),
RandomHorizontalFlip(0.5, keypoint_flip_order=cfg.keypoint_flip_order)
]
if cfg.half_body_transform:
transforms.append(
HalfBodyTransform(
cfg.upper_body_ids, cfg.lower_body_ids, cfg.prob_half_body
)
)
if cfg.extend_boxes:
transforms.append(
ExtendBoxes(cfg.x_ext, cfg.y_ext, cfg.input_shape[1] / cfg.input_shape[0])
)
transforms += [
RandomBoxAffine(
degrees=cfg.rotate_range,
scale=cfg.scale_range,
output_shape=cfg.input_shape,
rotate_prob=cfg.rotation_prob,
scale_prob=cfg.scale_prob,
)
]
transforms += [T.ToMode()]
train_queue = data.DataLoader(
train_dataset,
sampler=train_sampler,
num_workers=args.workers,
transform=T.Compose(transforms=transforms, order=train_dataset.order,),
collator=HeatmapCollator(
cfg.input_shape,
cfg.output_shape,
cfg.keypoint_num,
cfg.heat_thr,
cfg.heat_kernels if args.multi_scale_supervision else cfg.heat_kernels[-1:],
cfg.heat_range,
),
)
# Start training
for epoch in range(start_epoch, cfg.epochs):
loss = train(model, train_queue, optimizer, gm, epoch=epoch)
logger.info("Epoch %d Train %.6f ", epoch, loss)
if rank == 0 and epoch % cfg.save_freq == 0: # save checkpoint
mge.save(
{"epoch": epoch + 1, "state_dict": model.state_dict()},
os.path.join(save_dir, "epoch_{}.pkl".format(epoch)),
)