Add python wrapper for label smoothing

doc/fluid/api/layers.rst

@@ -473,6 +473,12 @@ multiplex
 ..  autofunction:: paddle.fluid.layers.multiplex
     :noindex:
 
+label_smooth
+------------
+
+..  autofunction:: paddle.fluid.layers.label_smooth
+    :noindex:
+
 ops
 ===
 

python/paddle/fluid/layers/nn.py

@@ -77,6 +77,7 @@
     'lod_reset',
     'lrn',
     'pad',
+    'label_smooth',
 ]
 
 
@@ -3678,3 +3679,68 @@ def pad(x, paddings, pad_value=0., name=None):
         attrs={'paddings': paddings,
                'pad_value': float(pad_value)})
     return out
+
+
+def label_smooth(label,
+                 prior_dist=None,
+                 epsilon=0.1,
+                 dtype="float32",
+                 name=None):
+    """
+    Label smoothing is a mechanism to regularize the classifier layer and is
+    called label-smoothing regularization (LSR). 
+
+    Label smoothing is proposed to encourage the model to be less confident,
+    since optimizing the log-likelihood of the correct label directly may
+    cause overfitting and reduce the ability of the model to adapt. Label
+    smoothing replaces the ground-truth label :math:`y` with the weighted sum
+    of itself and some fixed distribution :math:`\mu`. For class :math:`k`,
+    i.e.
+
+    .. math::
+
+        \\tilde{y_k} = (1 - \epsilon) * y_k + \epsilon * \mu_k,
+
+    where :math:`1 - \epsilon` and :math:`\epsilon` are the weights
+    respectively, and :math:`\\tilde{y}_k` is the smoothed label. Usually
+    uniform distribution is used for :math:`\mu`.
+
+    See more details about label smoothing in https://arxiv.org/abs/1512.00567.
+
+    Args:
+        label(Variable): The input variable containing the label data. The
+                          label data should use one-hot representation.
+        prior_dist(Variable): The prior distribution to be used to smooth
+                              labels. If not provided, an uniform distribution
+                              is used. The shape of :attr:`prior_dist` should
+                              be :math:`(1, class\_num)`. 
+        epsilon(float): The weight used to mix up the original ground-truth
+                        distribution and the fixed distribution.
+        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, 
+                                                  float_64, int etc.
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+
+    Returns:
+        Variable: The tensor variable containing the smoothed labels.
+
+    Examples:
+        .. code-block:: python
+
+            label = layers.data(name="label", shape=[1], dtype="float32")
+            one_hot_label = layers.one_hot(input=label, depth=10)
+            smooth_label = layers.label_smooth(
+                label=one_hot_label, epsilon=0.1, dtype="float32")
+    """
+    if epsilon > 1. or epsilon < 0.:
+        raise ValueError("The value of epsilon must be between 0 and 1.")
+    helper = LayerHelper("label_smooth", **locals())
+    label.stop_gradient = True
+    smooth_label = helper.create_tmp_variable(dtype)
+    helper.append_op(
+        type="label_smooth",
+        inputs={"X": label,
+                "PriorDist": prior_dist} if prior_dist else {"X": label},
+        outputs={"Out": smooth_label},
+        attrs={"epsilon": float(epsilon)})
+    return smooth_label

python/paddle/fluid/tests/unittests/test_layers.py

@@ -340,6 +340,16 @@ def test_lod_reset(self):
             print(layers.lod_reset(x=x, y=y))
         print(str(program))
 
+    def test_label_smooth(self):
+        program = Program()
+        with program_guard(program):
+            label = layers.data(name="label", shape=[1], dtype="float32")
+            one_hot_label = layers.one_hot(input=label, depth=10)
+            smooth_label = layers.label_smooth(
+                label=one_hot_label, epsilon=0.1, dtype="float32")
+            self.assertIsNotNone(smooth_label)
+        print(str(program))
+
 
 if __name__ == '__main__':
     unittest.main()