Skip to content

add fp16 and loss_scaler #292

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ShawnXuan merged 1 commit into from
Jan 25, 2022
Merged

add fp16 and loss_scaler #292

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions RecommenderSystems/dlrm/config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,6 +93,12 @@ def str_list(x):
parser.add_argument(
"--value_memory_kind", type=str_list, default="device,host"
)
parser.add_argument(
"--use_fp16", action="store_true", help="Run model with amp"
)
parser.add_argument(
"--loss_scale_policy", type=str, default="static", help="static or dynamic"
)
parser.add_argument(
"--test_name", type=str, default="noname_test"
)
Expand Down
9 changes: 7 additions & 2 deletions RecommenderSystems/dlrm/graph.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,10 +2,12 @@


class DLRMValGraph(flow.nn.Graph):
def __init__(self, wdl_module, dataloader):
def __init__(self, wdl_module, dataloader, use_fp16=False):
super(DLRMValGraph, self).__init__()
self.module = wdl_module
self.dataloader = dataloader
if use_fp16:
self.config.enable_amp(True)

def build(self):
(
Expand All @@ -22,12 +24,15 @@ def build(self):


class DLRMTrainGraph(flow.nn.Graph):
def __init__(self, wdl_module, dataloader, bce_loss, optimizer, lr_scheduler=None):
def __init__(self, wdl_module, dataloader, bce_loss, optimizer, lr_scheduler=None, grad_scaler=None, use_fp16=False):
super(DLRMTrainGraph, self).__init__()
self.module = wdl_module
self.dataloader = dataloader
self.bce_loss = bce_loss
self.add_optimizer(optimizer, lr_sch=lr_scheduler)
if use_fp16:
self.config.enable_amp(True)
self.set_grad_scaler(grad_scaler)

def build(self):
(
Expand Down
13 changes: 11 additions & 2 deletions RecommenderSystems/dlrm/train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,15 +57,24 @@ def __init__(self):
self.dlrm_module.parameters(), lr=args.learning_rate
)
self.lr_scheduler = make_lr_scheduler(args, self.opt)
if args.loss_scale_policy == "static":
self.grad_scaler = flow.amp.StaticGradScaler(1024)
else:
self.grad_scaler = flow.amp.GradScaler(
init_scale=1073741824,
growth_factor=2.0,
backoff_factor=0.5,
growth_interval=2000,
)

self.loss = flow.nn.BCELoss(reduction="none").to("cuda")
if self.execution_mode == "graph":
self.eval_graph = DLRMValGraph(
self.dlrm_module, self.val_dataloader
self.dlrm_module, self.val_dataloader, args.use_fp16
)
self.train_graph = DLRMTrainGraph(
self.dlrm_module, self.train_dataloader, self.loss, self.opt,
self.lr_scheduler
self.lr_scheduler, self.grad_scaler, args.use_fp16
)

def init_model(self):
Expand Down