diff --git a/keras_vggface/models.py b/keras_vggface/models.py
index 62804ad..7e384c1 100644
--- a/keras_vggface/models.py
+++ b/keras_vggface/models.py
@@ -339,6 +339,7 @@ def senet_conv_block(input_tensor, kernel_size, filters,
         bn_axis = 3
     else:
         bn_axis = 1
+    bn_eps = 0.0001
 
     conv1_reduce_name = 'conv' + str(stage) + "_" + str(block) + "_1x1_reduce"
     conv1_increase_name = 'conv' + str(stage) + "_" + str(
@@ -348,22 +349,22 @@ def senet_conv_block(input_tensor, kernel_size, filters,
 
     x = Conv2D(filters1, (1, 1), use_bias=bias, strides=strides,
                name=conv1_reduce_name)(input_tensor)
-    x = BatchNormalization(axis=bn_axis, name=conv1_reduce_name + "/bn")(x)
+    x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name=conv1_reduce_name + "/bn")(x)
     x = Activation('relu')(x)
 
     x = Conv2D(filters2, kernel_size, padding='same', use_bias=bias,
                name=conv3_name)(x)
-    x = BatchNormalization(axis=bn_axis, name=conv3_name + "/bn")(x)
+    x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name=conv3_name + "/bn")(x)
     x = Activation('relu')(x)
 
     x = Conv2D(filters3, (1, 1), name=conv1_increase_name, use_bias=bias)(x)
-    x = BatchNormalization(axis=bn_axis, name=conv1_increase_name + "/bn")(x)
+    x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name=conv1_increase_name + "/bn")(x)
 
     se = senet_se_block(x, stage=stage, block=block, bias=True)
 
     shortcut = Conv2D(filters3, (1, 1), use_bias=bias, strides=strides,
                       name=conv1_proj_name)(input_tensor)
-    shortcut = BatchNormalization(axis=bn_axis,
+    shortcut = BatchNormalization(axis=bn_axis, epsilon=bn_eps,
                                   name=conv1_proj_name + "/bn")(shortcut)
 
     m = layers.add([se, shortcut])
@@ -378,6 +379,7 @@ def senet_identity_block(input_tensor, kernel_size,
         bn_axis = 3
     else:
         bn_axis = 1
+    bn_eps = 0.0001
 
     conv1_reduce_name = 'conv' + str(stage) + "_" + str(block) + "_1x1_reduce"
     conv1_increase_name = 'conv' + str(stage) + "_" + str(
@@ -386,20 +388,20 @@ def senet_identity_block(input_tensor, kernel_size,
 
     x = Conv2D(filters1, (1, 1), use_bias=bias,
                name=conv1_reduce_name)(input_tensor)
-    x = BatchNormalization(axis=bn_axis, name=conv1_reduce_name + "/bn")(x)
+    x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name=conv1_reduce_name + "/bn")(x)
     x = Activation('relu')(x)
 
     x = Conv2D(filters2, kernel_size, padding='same', use_bias=bias,
                name=conv3_name)(x)
-    x = BatchNormalization(axis=bn_axis, name=conv3_name + "/bn")(x)
+    x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name=conv3_name + "/bn")(x)
     x = Activation('relu')(x)
 
     x = Conv2D(filters3, (1, 1), name=conv1_increase_name, use_bias=bias)(x)
-    x = BatchNormalization(axis=bn_axis, name=conv1_increase_name + "/bn")(x)
+    x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name=conv1_increase_name + "/bn")(x)
 
     se = senet_se_block(x, stage=stage, block=block, bias=True)
 
-    m = layers.add([x, se])
+    m = layers.add([se, input_tensor])
     m = Activation('relu')(m)
 
     return m
@@ -427,11 +429,12 @@ def SENET50(include_top=True, weights='vggface',
         bn_axis = 3
     else:
         bn_axis = 1
+    bn_eps = 0.0001
 
     x = Conv2D(
         64, (7, 7), use_bias=False, strides=(2, 2), padding='same',
         name='conv1/7x7_s2')(img_input)
-    x = BatchNormalization(axis=bn_axis, name='conv1/7x7_s2/bn')(x)
+    x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name='conv1/7x7_s2/bn')(x)
     x = Activation('relu')(x)
     x = MaxPooling2D((3, 3), strides=(2, 2))(x)