fix documentation of center argument in BN layers (tensorflow#6792)

fixes tensorflow#6736

tensorflow/contrib/layers/python/layers/layers.py

@@ -176,7 +176,8 @@ def _fused_batch_norm(
       to 1.0, typically in the multiple-nines range: 0.999, 0.99, 0.9, etc. Lower
       `decay` value (recommend trying `decay`=0.9) if model experiences reasonably
       good training performance but poor validation and/or test performance.
-    center: If True, subtract `beta`. If False, `beta` is ignored.
+    center: If True, add offset of `beta` to normalized tensor. If False, `beta`
+      is ignored.
     scale: If True, multiply by `gamma`. If False, `gamma` is
       not used. When the next layer is linear (also e.g. `nn.relu`), this can be
       disabled since the scaling can be done by the next layer.
@@ -407,7 +408,8 @@ def batch_norm(
       Lower `decay` value (recommend trying `decay`=0.9) if model experiences
       reasonably good training performance but poor validation and/or test
       performance. Try zero_debias_moving_mean=True for improved stability.
-    center: If True, subtract `beta`. If False, `beta` is ignored.
+    center: If True, add offset of `beta` to normalized tensor. If False, `beta`
+      is ignored.
     scale: If True, multiply by `gamma`. If False, `gamma` is
       not used. When the next layer is linear (also e.g. `nn.relu`), this can be
       disabled since the scaling can be done by the next layer.
@@ -1447,7 +1449,8 @@ def layer_norm(inputs,
   Args:
     inputs: a tensor with 2 or more dimensions. The normalization
             occurs over all but the first dimension.
-    center: If True, subtract `beta`. If False, `beta` is ignored.
+    center: If True, add offset of `beta` to normalized tensor. If False, `beta`
+      is ignored.
     scale: If True, multiply by `gamma`. If False, `gamma` is
       not used. When the next layer is linear (also e.g. `nn.relu`), this can be
       disabled since the scaling can be done by the next layer.

tensorflow/python/layers/normalization.py

@@ -55,7 +55,8 @@ class BatchNormalization(base._Layer):  # pylint: disable=protected-access
       `data_format="channels_first"`, set `axis=1` in `BatchNormalization`.
     momentum: Momentum for the moving average.
     epsilon: Small float added to variance to avoid dividing by zero.
-    center: If True, subtract `beta`. If False, `beta` is ignored.
+    center: If True, add offset of `beta` to normalized tensor. If False, `beta`
+      is ignored.
     scale: If True, multiply by `gamma`. If False, `gamma` is
       not used. When the next layer is linear (also e.g. `nn.relu`), this can be
       disabled since the scaling can be done by the next layer.
@@ -276,7 +277,8 @@ def batch_normalization(inputs,
       `data_format="channels_first"`, set `axis=1` in `BatchNormalization`.
     momentum: Momentum for the moving average.
     epsilon: Small float added to variance to avoid dividing by zero.
-    center: If True, subtract `beta`. If False, `beta` is ignored.
+    center: If True, add offset of `beta` to normalized tensor. If False, `beta`
+      is ignored.
     scale: If True, multiply by `gamma`. If False, `gamma` is
       not used. When the next layer is linear (also e.g. `nn.relu`), this can be
       disabled since the scaling can be done by the next layer.