Add --fp16-scale-tolerance (facebookresearch#397)

Summary:
Let's only decrease the loss scale if a large enough percentage of batches overflow.
Pull Request resolved: facebookresearch#397

Differential Revision: D13355159

Pulled By: myleott

fbshipit-source-id: e17dde73d34a639519b4348c013fdd19d2b314e6

fairseq/optim/fp16_optimizer.py

@@ -12,19 +12,29 @@
 
 class DynamicLossScaler:
 
-    def __init__(self, init_scale=2.**15, scale_factor=2., scale_window=2000):
+    def __init__(self, init_scale=2.**15, scale_factor=2., scale_window=2000, tolerance=0.05):
         self.loss_scale = init_scale
         self.scale_factor = scale_factor
         self.scale_window = scale_window
+        self.tolerance = tolerance
         self._iter = 0
         self._last_overflow_iter = -1
+        self._last_rescale_iter = -1
+        self._overflows_since_rescale = 0
 
     def update_scale(self, overflow):
+        iter_since_rescale = self._iter - self._last_rescale_iter
         if overflow:
-            self.loss_scale /= self.scale_factor
             self._last_overflow_iter = self._iter
+            self._overflows_since_rescale += 1
+            pct_overflow = self._overflows_since_rescale / float(iter_since_rescale)
+            if pct_overflow >= self.tolerance:
+                self.loss_scale /= self.scale_factor
+                self._last_rescale_iter = self._iter
+                self._overflows_since_rescale = 0
         elif (self._iter - self._last_overflow_iter) % self.scale_window == 0:
             self.loss_scale *= self.scale_factor
+            self._last_rescale_iter = self._iter
         self._iter += 1
 
     @staticmethod
@@ -55,6 +65,7 @@ def __init__(self, args, params, fp32_optimizer, fp32_params):
         self.scaler = DynamicLossScaler(
             init_scale=args.fp16_init_scale,
             scale_window=scale_window,
+            tolerance=args.fp16_scale_tolerance,
         )
 
     @staticmethod

fairseq/options.py

@@ -133,6 +133,8 @@ def get_parser(desc, default_task='translation'):
                         help='default FP16 loss scale')
     parser.add_argument('--fp16-scale-window', type=int,
                         help='number of updates before increasing loss scale')
+    parser.add_argument('--fp16-scale-tolerance', default=0.0, type=float,
+                        help='pct of updates that can overflow before decreasing the loss scale')
 
     # Task definitions can be found under fairseq/tasks/
     parser.add_argument('--task', metavar='TASK', default=default_task,