Adding interface for decayed adagrad optimizer (#5644)

* add decayed adagrad python code

* fix typo and order

* small fix

python/paddle/v2/fluid/optimizer.py

@@ -9,7 +9,7 @@ from paddle.v2.fluid.layer_helper import LayerHelper
 
 __all__ = [
     'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer',
-    'AdamaxOptimizer'
+    'AdamaxOptimizer', 'DecayedAdagradOptimizer'
 ]
 
 
@@ -85,7 +85,7 @@ class Optimizer(object):
         """
         if (name in self._accumulators and
                 param.name in self._accumulators[name]):
-            raise Exception("Accumulator {} already exists for parmeter {}".
+            raise Exception("Accumulator {} already exists for parameter {}".
                             format(name, param.name))
 
         assert isinstance(self.helper, LayerHelper)
@@ -307,7 +307,7 @@ class AdagradOptimizer(Optimizer):
         moment_acc = self._get_accumulator(self._moment_acc_str,
                                            param_and_grad[0])
 
-        # create the adagrad optimizer op
+        # Create the adagrad optimizer op
         adagrad_op = block.append_op(
             type=self.type,
             inputs={
@@ -510,3 +510,51 @@ class AdamaxOptimizer(Optimizer):
             attrs={"scale": self._beta1})
 
         return [scale_beta1]
+
+
+class DecayedAdagradOptimizer(Optimizer):
+    """Simple Decayed Adagrad optimizer with moment state
+    """
+    _moment_acc_str = "moment"
+
+    def __init__(self,
+                 learning_rate,
+                 decay=0.95,
+                 epsilon=1.0e-6,
+                 global_step=None):
+        assert learning_rate is not None
+        assert decay is not None
+        assert epsilon is not None
+
+        super(DecayedAdagradOptimizer, self).__init__(global_step)
+        self.type = "decayed_adagrad"
+        self._learning_rate = learning_rate
+        self._decay = decay
+        self._epsilon = epsilon
+
+    def _create_accumulators(self, block, parameters):
+        assert isinstance(block, framework.Block)
+
+        for p in parameters:
+            self._add_accumulator(self._moment_acc_str, p)
+
+    def _append_optimize_op(self, block, param_and_grad):
+        assert isinstance(block, framework.Block)
+
+        moment_acc = self._get_accumulator(self._moment_acc_str,
+                                           param_and_grad[0])
+
+        # Create the decayed adagrad optimizer op
+        decayed_adagrad_op = block.append_op(
+            type=self.type,
+            inputs={
+                "Param": param_and_grad[0],
+                "Grad": param_and_grad[1],
+                "Moment": moment_acc,
+                "LearningRate": self._create_param_lr(param_and_grad)
+            },
+            outputs={"ParamOut": param_and_grad[0],
+                     "MomentOut": moment_acc},
+            attrs={"epsilon": self._epsilon})
+
+        return decayed_adagrad_op

python/paddle/v2/fluid/tests/test_optimizer.py

@@ -198,7 +198,7 @@ class TestAdagradOptimizer(unittest.TestCase):
         adagrad_op = opts[0]
         self.assertEqual(adagrad_op.type, "adagrad")
 
-        # check accumulators
+        # Check accumulators
         accumulators = adagrad_optimizer.get_accumulators()
         self.assertEqual(len(accumulators), 1)
         self.assertTrue(adagrad_optimizer.get_moment_str() in accumulators)
@@ -331,5 +331,59 @@ class TestAdamaxOptimizer(unittest.TestCase):
         self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate)
 
 
+class TestDecayedAdagradOptimizer(unittest.TestCase):
+    class MockDecayedAdagrad(optimizer.DecayedAdagradOptimizer):
+        def get_accumulators(self):
+            return self._accumulators
+
+        def get_moment_str(self):
+            return self._moment_acc_str
+
+    def test_decayed_adagrad_optimizer(self):
+        init_program = framework.Program()
+        program = framework.Program()
+        block = program.global_block()
+        mul_x = block.create_parameter(
+            dtype="float32", shape=[5, 10], lod_level=0, name="mul.x")
+        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})
+        learning_rate = 0.01
+        decayed_adagrad_optimizer = self.MockDecayedAdagrad(
+            learning_rate=learning_rate, decay=0.95, epsilon=1.0e-6)
+        params_grads = append_backward_ops(mul_out)
+        self.assertEqual(len(params_grads), 1)
+        self.assertEqual(len(decayed_adagrad_optimizer.get_accumulators()), 0)
+        opts = decayed_adagrad_optimizer.create_optimization_pass(
+            params_grads, mul_out, init_program)
+        self.assertEqual(len(opts), 1)
+        decayed_adagrad_op = opts[0]
+        self.assertEqual(decayed_adagrad_op.type, "decayed_adagrad")
+
+        # Check accumulators
+        accumulators = decayed_adagrad_optimizer.get_accumulators()
+        self.assertEqual(len(accumulators), 1)
+        self.assertTrue(
+            decayed_adagrad_optimizer.get_moment_str() in accumulators)
+        moment_acc = accumulators[decayed_adagrad_optimizer.get_moment_str()]
+        self.assertEqual(len(moment_acc), 1)
+        self.assertTrue(mul_x.name in moment_acc)
+
+        # Check init_program
+        init_ops = init_program.global_block().ops
+        self.assertEqual(len(init_ops), 2)
+        self.assertEqual(init_ops[0].type, "fill_constant")
+        self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate)
+        self.assertEqual(init_ops[1].type, "fill_constant")
+        self.assertAlmostEqual(init_ops[1].attr('value'), 0.0)
+
+
 if __name__ == '__main__':
     unittest.main()