Add python API for backward regularization ops (#5135)

* Add regularizer code
* Fix code

python/paddle/v2/framework/framework.py

@@ -505,6 +505,8 @@ class Parameter(Variable):
 
         self.optimize_attr = kwargs.get('optimize_attr', {'learning_rate': 1.0})
 
+        self.regularizer = kwargs.get('regularizer', None)
+
 
 # program is a global instance.
 g_program = Program()

python/paddle/v2/framework/optimizer.py

@@ -2,6 +2,7 @@ from collections import defaultdict
 
 import paddle.v2.framework.framework as framework
 from paddle.v2.framework.backward import append_backward_ops
+from paddle.v2.framework.regularizer import append_regularization_ops
 
 __all__ = [
     'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer',
@@ -161,6 +162,8 @@ class Optimizer(object):
         """
         params_grads = append_backward_ops(loss, parameter_list, no_grad_set or
                                            set())
+        # Add regularization if any 
+        params_grads = append_regularization_ops(params_grads)
         optimize_ops = self.create_optimization_pass(params_grads, loss)
         return optimize_ops
 

python/paddle/v2/framework/regularizer.py

+import paddle.v2.framework.framework as framework
+
+__all__ = ['append_regularization_ops', 'L2DecayRegularizer']
+
+
+def append_regularization_ops(parameters_and_grads):
+    """Create and add backward regularization Operators
+
+    Creates and adds backward regularization operators in the BlockDesc.
+    This will add gradients of the regularizer function to the gradients
+    of the parameters and return these modified gradients. This is the
+    same as implementing weight decay in optimizers for regularization.
+
+    Args:
+        parameters_and_grads: A list of (parameters, gradients) pairs
+                              that need to be regularized.
+
+    Returns:
+        list of (parameters, gradients) pair with the regularized gradient
+
+    Raises:
+        Exception: Unknown regularization type
+    """
+    params_and_grads = []
+    for param, grad in parameters_and_grads:
+        # If no gradient or no regularization specified,
+        # then we don't need to do anything
+        if grad is None or param.regularizer is None:
+            params_and_grads.append((param, grad))
+            continue
+
+        # Add variable for regularization term in grad block
+        regularization_term = param.regularizer(param, grad.block)
+        assert grad.shape == regularization_term.shape
+
+        grad.block.append_op(
+            type='elementwise_add',
+            inputs={"X": grad,
+                    "Y": regularization_term},
+            outputs={"Out": grad})
+        params_and_grads.append((param, grad))
+
+    return params_and_grads
+
+
+class WeightDecayRegularizer(object):
+    """Base class for weight decay regularizers
+
+    Defines the common interface of weight-decay regularizers.
+    Weight-decay regularizers are added only during the backward
+    pass for faster regularization. They add operations to the network
+    that correspond to gradient of the regularization function.
+    Users should not use this class directly, but need to use one
+    of its implementations
+    """
+
+    def __init__(self):
+        pass
+
+    def __call__(self, param, block):
+        """Add corresponding weight decay operations to the network
+        """
+        raise NotImplementedError()
+
+
+class L2DecayRegularizer(WeightDecayRegularizer):
+    """Implements the L2 Weight Decay Regularization
+    """
+
+    def __init__(self, regularization_coeff=0.0):
+        assert regularization_coeff is not None
+        super(L2DecayRegularizer, self).__init__()
+        self._regularization_coeff = regularization_coeff
+
+    def __call__(self, param, block):
+        """Add L2 weight decay ops to network
+
+        Adds L2 weight decay ops.
+        L2WeightDecay = reg_coeff * parameter
+
+        Args:
+            param: parameter variable for which regularization is applied
+            block: block in which variable is to be created
+
+        Returns:
+            new variable for weight decay
+        """
+        assert isinstance(param, framework.Parameter)
+        assert isinstance(block, framework.Block)
+        decay = block.create_var(
+            dtype="float32", shape=param.shape, lod_level=param.lod_level)
+        # Append Op to calculate decay
+        block.append_op(
+            type='scale',
+            inputs={"X": param},
+            outputs={"Out": decay},
+            attrs={"scale": self._regularization_coeff})
+
+        return decay

python/paddle/v2/framework/tests/test_regularizer.py

+import unittest
+
+import paddle.v2.framework.framework as framework
+import paddle.v2.framework.optimizer as optimizer
+import paddle.v2.framework.regularizer as regularizer
+from paddle.v2.framework.backward import append_backward_ops
+
+
+class TestL2DecayRegularizer(unittest.TestCase):
+    def test_l2decay_regularizer(self):
+        program = framework.Program()
+        block = program.global_block()
+        mul_x = block.create_parameter(
+            dtype="float32",
+            shape=[5, 10],
+            lod_level=0,
+            name="mul.x",
+            regularizer=regularizer.L2DecayRegularizer(0.5))
+        self.assertTrue(mul_x.regularizer is not None)
+        self.assertTrue(
+            isinstance(mul_x.regularizer, regularizer.L2DecayRegularizer))
+        mul_y = block.create_var(
+            dtype="float32", shape=[10, 8], lod_level=0, name="mul.y")
+        mul_out = block.create_var(
+            dtype="float32", shape=[5, 8], lod_level=0, name="mul.out")
+        block.append_op(
+            type="mul",
+            inputs={"X": mul_x,
+                    "Y": mul_y},
+            outputs={"Out": mul_out},
+            attrs={"x_num_col_dims": 1})
+        params_grads = append_backward_ops(mul_out)
+        self.assertEqual(len(params_grads), 1)
+        count_ops = len(block.ops)
+        params_grads = optimizer.append_regularization_ops(params_grads)
+        self.assertEqual(len(params_grads), 1)
+        self.assertEqual(len(block.ops), count_ops + 2)
+        self.assertEqual(block.ops[-1].type, 'elementwise_add')
+        self.assertEqual(block.ops[-2].type, 'scale')
+
+
+if __name__ == '__main__':
+    unittest.main()