print grad scaler as part of the metric.

Summary:
Pull Request resolved: #501

X-link: facebookresearch/detectron2#4851

print grad scaler as part of the metric.
* Controlled by a flag "SOLVER.AMP.LOG_GRAD_SCALER"

Reviewed By: tax313

Differential Revision: D43585363

fbshipit-source-id: 495b37ff524c47e515cea0b3c677ee81b34ad4ca

Author: Peizhao Zhang

d2go/runner/config_defaults.py

@@ -81,6 +81,8 @@ def _add_detectron2go_runner_default_cfg(_C: CN) -> None:
     assert not _C.SOLVER.AMP.ENABLED
     # AMP precision is used by both D2 and lightning backend. Can be "float16" or "bfloat16".
     _C.SOLVER.AMP.PRECISION = "float16"
+    # log the grad scalar to the output
+    _C.SOLVER.AMP.LOG_GRAD_SCALER = False
 
     # Betas are used in the AdamW optimizer
     _C.SOLVER.BETAS = (0.9, 0.999)

d2go/runner/default_runner.py

@@ -551,6 +551,7 @@ def _get_model_with_abnormal_checker(model):
                 precision=parse_precision_from_string(
                     cfg.SOLVER.AMP.PRECISION, lightning=False
                 ),
+                log_grad_scaler=cfg.SOLVER.AMP.LOG_GRAD_SCALER,
             )
         else:
             trainer = SimpleTrainer(