Merge pull request #15630 from xuezhong/fix_bug_adagrad2

add initial_accumulator_value for adagrad

paddle/fluid/API.spec

@@ -427,7 +427,7 @@ paddle.fluid.optimizer.MomentumOptimizer.__init__ ArgSpec(args=['self', 'learnin
 paddle.fluid.optimizer.MomentumOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.optimizer.MomentumOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.MomentumOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
-paddle.fluid.optimizer.AdagradOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1e-06, None, None))
+paddle.fluid.optimizer.AdagradOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'regularization', 'name', 'initial_accumulator_value'], varargs=None, keywords=None, defaults=(1e-06, None, None, 0.0))
 paddle.fluid.optimizer.AdagradOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.optimizer.AdagradOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.AdagradOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))

python/paddle/fluid/optimizer.py

@@ -649,6 +649,7 @@ class AdagradOptimizer(Optimizer):
         regularization: A Regularizer, such as
                         fluid.regularizer.L2DecayRegularizer.
         name: A optional name prefix.
+        initial_accumulator_value (float): Initial value for moment accumulator.
 
     Examples:
         .. code-block:: python
@@ -662,7 +663,8 @@ class AdagradOptimizer(Optimizer):
                  learning_rate,
                  epsilon=1.0e-6,
                  regularization=None,
-                 name=None):
+                 name=None,
+                 initial_accumulator_value=0.0):
         assert learning_rate is not None
         assert epsilon is not None
         super(AdagradOptimizer, self).__init__(
@@ -671,6 +673,7 @@ class AdagradOptimizer(Optimizer):
             name=name)
         self.type = "adagrad"
         self._epsilon = epsilon
+        self.initial_accumulator_value = initial_accumulator_value
 
     def _create_accumulators(self, block, parameters):
         assert isinstance(block, framework.Block)
@@ -683,6 +686,16 @@ class AdagradOptimizer(Optimizer):
 
         moment_acc = self._get_accumulator(self._moment_acc_str,
                                            param_and_grad[0])
+        startup_block = framework.default_startup_program().global_block()
+        startup_block.append_op(
+            type='fill_constant',
+            inputs={},
+            outputs={'Out': [moment_acc]},
+            attrs={
+                'dtype': moment_acc.dtype,
+                'value': self.initial_accumulator_value,
+                'shape': moment_acc.shape,
+            })
 
         # Create the adagrad optimizer op
         adagrad_op = block.append_op(

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

@@ -274,7 +274,7 @@ class TestAdagradOptimizer(unittest.TestCase):
 
         # Check init_program
         init_ops = init_program.global_block().ops
-        self.assertEqual(len(init_ops), 2)
+        self.assertEqual(len(init_ops), 3)
         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")