Add support for keeping running bn stats the same across distributed training nodes before eval/save

timm/utils.py

@@ -21,11 +21,15 @@ except ImportError:
 from torch import distributed as dist
 
 
-def get_state_dict(model):
+def unwrap_model(model):
     if isinstance(model, ModelEma):
-        return get_state_dict(model.ema)
+        return unwrap_model(model.ema)
     else:
-        return model.module.state_dict() if hasattr(model, 'module') else model.state_dict()
+        return model.module if hasattr(model, 'module') else model
+
+
+def get_state_dict(model):
+    return unwrap_model(model).state_dict()
 
 
 class CheckpointSaver:
@@ -206,6 +210,14 @@ def reduce_tensor(tensor, n):
     return rt
 
 
+def reduce_bn(model, world_size):
+    # ensure every node has the same running bn stats
+    for bn_name, bn_buf in unwrap_model(model).named_buffers(recurse=True):
+        if ('running_mean' in bn_name) or ('running_var' in bn_name):
+            torch.distributed.all_reduce(bn_buf, op=dist.ReduceOp.SUM)
+            bn_buf /= float(world_size)
+
+
 class ModelEma:
     """ Model Exponential Moving Average
     Keep a moving average of everything in the model state_dict (parameters and buffers).

train.py

@@ -145,6 +145,8 @@ parser.add_argument('--amp', action='store_true', default=False,
                     help='use NVIDIA amp for mixed precision training')
 parser.add_argument('--sync-bn', action='store_true',
                     help='enabling apex sync BN.')
+parser.add_argument('--reduce-bn', action='store_true',
+                    help='average BN running stats across all distributed nodes between train and validation.')
 parser.add_argument('--no-prefetcher', action='store_true', default=False,
                     help='disable fast prefetcher')
 parser.add_argument('--output', default='', type=str, metavar='PATH',
@@ -256,7 +258,7 @@ def main():
             if args.local_rank == 0:
                 logging.info('Restoring NVIDIA AMP state from checkpoint')
             amp.load_state_dict(resume_state['amp'])
-    resume_state = None  # clear it
+    del resume_state
 
     model_ema = None
     if args.model_ema:
@@ -388,9 +390,17 @@ def main():
                 lr_scheduler=lr_scheduler, saver=saver, output_dir=output_dir,
                 use_amp=use_amp, model_ema=model_ema)
 
+            if args.distributed and args.reduce_bn:
+                if args.local_rank == 0:
+                    logging.info("Averaging bn running means and vars")
+                reduce_bn(model, args.world_size)
+
             eval_metrics = validate(model, loader_eval, validate_loss_fn, args)
 
             if model_ema is not None and not args.model_ema_force_cpu:
+                if args.distributed and args.reduce_bn:
+                    reduce_bn(model_ema, args.world_size)
+
                 ema_eval_metrics = validate(
                     model_ema.ema, loader_eval, validate_loss_fn, args, log_suffix=' (EMA)')
                 eval_metrics = ema_eval_metrics