support FP16 gradient compression

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/70

DDP supports an fp16_compress_hook which compresses the gradient to FP16 before communication. This can result in a significant speed up.

Add one argument `_C.MODEL.DDP_FP16_GRAD_COMPRESS` to trigger it.

Reviewed By: zhanghang1989

Differential Revision: D28467701

fbshipit-source-id: 3c80865222f48eb8fe6947ea972448c445ee3ef3

d2go/utils/get_default_cfg.py

@@ -59,6 +59,8 @@ def get_default_cfg(_C):
 
     # Set find_unused_parameters for DistributedDataParallel.
     _C.MODEL.DDP_FIND_UNUSED_PARAMETERS = False
+    # Set FP16 gradient compression for DistributedDataParallel.
+    _C.MODEL.DDP_FP16_GRAD_COMPRESS = False
 
     # Set default optimizer
     _C.SOLVER.OPTIMIZER = "sgd"

tools/train_net.py

@@ -16,7 +16,7 @@ from d2go.setup import (
     setup_after_launch,
 )
 from d2go.utils.misc import print_metrics_table, dump_trained_model_configs
-from torch.nn.parallel import DistributedDataParallel
+from detectron2.engine.defaults import create_ddp_model
 
 
 logger = logging.getLogger("d2go.tools.train_net")
@@ -53,9 +53,9 @@ def main(
             "metrics": metrics,
         }
 
-    if comm.get_world_size() > 1:
-        model = DistributedDataParallel(
+    model = create_ddp_model(
         model,
+        fp16_compression=cfg.MODEL.DDP_FP16_GRAD_COMPRESS,
         device_ids=None if cfg.MODEL.DEVICE == "cpu" else [comm.get_local_rank()],
         broadcast_buffers=False,
         find_unused_parameters=cfg.MODEL.DDP_FIND_UNUSED_PARAMETERS,
@@ -88,6 +88,7 @@ def run_with_cmdline_args(args):
         args=(cfg, output_dir, runner, args.eval_only, args.resume),
     )
 
+
 def cli():
     parser = basic_argument_parser(requires_output_dir=False)
     parser.add_argument(
@@ -100,5 +101,6 @@ def cli():
     )
     run_with_cmdline_args(parser.parse_args())
 
+
 if __name__ == "__main__":
     cli()