Merge pull request #14894 from velconia/add_huber_regression_loss_op

Add python interface for huber loss

paddle/fluid/API.spec

@@ -201,6 +201,7 @@ paddle.fluid.layers.merge_selected_rows ArgSpec(args=['x', 'name'], varargs=None
 paddle.fluid.layers.get_tensor_from_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1))
 paddle.fluid.layers.psroi_pool ArgSpec(args=['input', 'rois', 'output_channels', 'spatial_scale', 'pooled_height', 'pooled_width', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.huber_loss ArgSpec(args=['input', 'label', 'delta'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
 paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)

paddle/fluid/operators/huber_loss_op.cc

@@ -124,8 +124,9 @@ REGISTER_OPERATOR(huber_loss, ops::HuberLossOp, ops::HuberLossOpMaker<float>,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 REGISTER_OPERATOR(huber_loss_grad, ops::HuberLossGradOp);
 REGISTER_OP_CPU_KERNEL(
-    huber_loss,
-    ops::HuberLossKernel<paddle::platform::CPUDeviceContext, float>);
+    huber_loss, ops::HuberLossKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::HuberLossKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     huber_loss_grad,
-    ops::HuberLossGradKernel<paddle::platform::CPUDeviceContext, float>);
+    ops::HuberLossGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::HuberLossGradKernel<paddle::platform::CPUDeviceContext, double>);

python/paddle/fluid/layers/nn.py

@@ -176,6 +176,7 @@ __all__ = [
     'get_tensor_from_selected_rows',
     'lstm',
     'psroi_pool',
+    'huber_loss',
 ]
 
 kIgnoreIndex = -100
@@ -9377,3 +9378,51 @@ def psroi_pool(input,
             'pooled_width': pooled_width
         })
     return out
+
+
+def huber_loss(input, label, delta):
+    """
+    Huber loss is a loss function used in robust.
+    Huber loss can evaluate the fitness of input to label.
+    Different from MSE loss, Huber loss is more robust for outliers.
+
+    When the difference between input and label is large than delta
+    .. math::
+
+        huber\_loss = delta * (label - input) - 0.5 * delta * delta
+
+    When the difference between input and label is less than delta
+    .. math::
+
+        huber\_loss = 0.5 * (label - input) * (label - input)
+
+
+    Args:
+        input (Variable): This input is a probability computed by the previous operator.
+                          The first dimension is batch size, and the last dimension is 1.
+        label (Variable): The groud truth whose first dimension is batch size
+                          and last dimension is 1.
+        delta (float): The parameter of huber loss, which controls
+                       the range of outliers
+
+    Returns:
+        huber\_loss (Variable): The huber loss with shape [batch_size, 1].
+
+    Examples:
+        .. code-block:: python
+
+            predictions = fluid.layers.softmax(x)
+            loss = fluid.layers.huber_loss(input=predictions, label=label, 1.0)
+    """
+    helper = LayerHelper('huber_loss', **locals())
+    residual = helper.create_variable_for_type_inference(
+        dtype=helper.input_dtype())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
+    helper.append_op(
+        type='huber_loss',
+        inputs={'X': input,
+                'Y': label},
+        outputs={'Out': out,
+                 'Residual': residual},
+        attrs={'delta': delta})
+    return out