Pass the zero_grad_before_forward flag to the trainer

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/525

In d2go, pass the argument ZERO_GRAD_BEFORE_FORWARD to the detectron runtime.

Reviewed By: tglik

Differential Revision: D44267319

fbshipit-source-id: 3bd5874bea96ac381fb49972a2dfe9bb52005a7d

d2go/runner/config_defaults.py

@@ -118,6 +118,9 @@ def _add_detectron2go_runner_default_cfg(_C: CN) -> None:
     # Specify whether to perform NUMA binding
     _C.NUMA_BINDING = False
 
+    # Specify whether to zero the gradients before forward
+    _C.ZERO_GRAD_BEFORE_FORWARD = False
+
 
 def _add_rcnn_default_config(_C: CN) -> None:
     _C.EXPORT_CAFFE2 = CN()

d2go/runner/default_runner.py

@@ -548,6 +548,7 @@ class Detectron2GoRunner(D2GoDataAPIMixIn, BaseRunner):
                 data_loader,
                 optimizer,
                 gather_metric_period=cfg.GATHER_METRIC_PERIOD,
+                zero_grad_before_forward=cfg.ZERO_GRAD_BEFORE_FORWARD,
                 grad_scaler=get_grad_scaler(cfg),
                 precision=parse_precision_from_string(
                     cfg.SOLVER.AMP.PRECISION, lightning=False
@@ -560,6 +561,7 @@ class Detectron2GoRunner(D2GoDataAPIMixIn, BaseRunner):
                 data_loader,
                 optimizer,
                 gather_metric_period=cfg.GATHER_METRIC_PERIOD,
+                zero_grad_before_forward=cfg.ZERO_GRAD_BEFORE_FORWARD,
             )
 
         if cfg.SOLVER.AMP.ENABLED and torch.cuda.is_available():