Extract apply_backward_pass to backward.py (#5026)

* Extract apply_backward_pass to backward.py

Rename apply_backward_pass to append_backward_ops

* Fix CI

* Update design doc

doc/design/optimizer.md

@@ -65,20 +65,6 @@ class Optimizer(object):
     def __init__(self):
         pass
 
-    def create_backward_pass(self, loss, parameter_list=None):
-        """
-        create and add gradient Operators in BlockDesc to Compute gradients of `loss`
-        for parameters in parameter_list
-
-        Args:
-          loss: an variable generated by cost function.
-          parameter_list: parameters that need to compute gradient and update to optimize the lost.
-
-        Returns:
-          list of (parameters, gradients) pair.
-        """
-        return None
-
     def create_optimization_pass(self, parameters_and_grads):
         """Add optimization operators to update gradients to variables.
 
@@ -93,7 +79,7 @@ class Optimizer(object):
     def minimize(self, loss, parameter_list):
         """Add operations to minimize `loss` by updating `parameter_list`.
 
-        This method combines interface `create_backward_pass()` and
+        This method combines interface `append_backward_ops()` and
         `create_optimization_pass()` into one.
         """
         params_grads = self.create_backward_pass(loss, parameter_list)

python/paddle/v2/framework/backward.py

@@ -0,0 +1,45 @@
+from paddle.v2.framework import framework as framework
+
+__all__ = ['append_backward_ops']
+
+
+def append_backward_ops(loss, parameter_list=None, no_grad_set=None):
+    """
+    Create and add gradient Operators in BlockDesc to compute
+    gradients of `loss` for parameters in parameter_list
+
+    :param loss: an variable generated by cost function.
+    :type loss: Variable
+    :param no_grad_set: variable that should not create gradient
+    :type no_grad_set: set
+    :param parameter_list: parameters that need to compute gradient and 
+    update to optimize the lost.
+    :type: list
+    :return: list of (parameters, gradients) pair.
+    :rtype: list[Variable]
+    """
+    assert isinstance(loss, framework.Variable)
+    param_grad_map = loss.block.program.append_backward(loss, no_grad_set or
+                                                        set())
+    if parameter_list is not None:
+        parameters = parameter_list
+    else:
+        params = loss.block.program.global_block().all_parameters()
+        parameters = [param.name for param in params]
+    params_and_grads = []
+    for param in parameters:
+        if param not in param_grad_map:
+            raise ValueError("param %s is not in map" % param)
+        grad_info = param_grad_map[param]
+        grad_block = loss.block.program.block(grad_info[1])
+        if not grad_block.has_var(grad_info[0]):
+            raise ValueError("grad block[{0}] did not have grad var {1}".format(
+                grad_info[1], grad_info[0]))
+        # Get the param var from the global block
+        param_var = loss.block.program.global_block().var(param)
+        grad_var = grad_block.var(grad_info[0])
+        if loss.block.has_var(grad_info[0]):
+            params_and_grads.append((param_var, grad_var))
+        else:
+            params_and_grads.append((param_var, None))
+    return params_and_grads

python/paddle/v2/framework/optimizer.py

@@ -1,6 +1,8 @@
-import paddle.v2.framework.framework as framework
 from collections import defaultdict
 
+import paddle.v2.framework.framework as framework
+from paddle.v2.framework.backward import append_backward_ops
+
 __all__ = [
     'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer'
 ]
@@ -105,45 +107,6 @@ class Optimizer(object):
                             format(name, param.name))
         return self._accumulators[name][param.name]
 
-    def create_backward_pass(self, loss, parameter_list=None, no_grad_set=None):
-        """Create and add gradient Operators in BlockDesc to compute
-        gradients of `loss` for parameters in parameter_list
-
-        Args:
-          loss: an variable generated by cost function.
-          no_grad_set: variable that should not create gradient
-          parameter_list: parameters that need to compute gradient and
-          update to optimize the lost.
-
-        Returns:
-          list of (parameters, gradients) pair.
-        """
-        assert isinstance(loss, framework.Variable)
-        param_grad_map = loss.block.program.append_backward(loss, no_grad_set or
-                                                            set())
-        if parameter_list is not None:
-            parameters = parameter_list
-        else:
-            params = loss.block.program.global_block().all_parameters()
-            parameters = [param.name for param in params]
-        params_and_grads = []
-        for param in parameters:
-            if param not in param_grad_map:
-                raise Exception("param %s is not in map" % param)
-            grad_info = param_grad_map[param]
-            grad_block = loss.block.program.block(grad_info[1])
-            if not grad_block.has_var(grad_info[0]):
-                raise Exception("grad block[%d] did not have grad var %s" %
-                                grad_info[1], grad_info[0])
-            # Get the param var from the global block
-            param_var = loss.block.program.global_block().var(param)
-            grad_var = grad_block.var(grad_info[0])
-            if loss.block.has_var(grad_info[0]):
-                params_and_grads.append((param_var, grad_var))
-            else:
-                params_and_grads.append((param_var, None))
-        return params_and_grads
-
     def create_optimization_pass(self, parameters_and_grads, loss):
         """Add optimization operators to update gradients to variables.
 
@@ -192,11 +155,11 @@ class Optimizer(object):
     def minimize(self, loss, parameter_list=None, no_grad_set=None):
         """Add operations to minimize `loss` by updating `parameter_list`.
 
-        This method combines interface `create_backward_pass()` and
+        This method combines interface `append_backward_ops()` and
         `create_optimization_pass()` into one.
         """
-        params_grads = self.create_backward_pass(loss, parameter_list,
+        params_grads = append_backward_ops(loss, parameter_list, no_grad_set or
-                                                 no_grad_set or set())
+                                           set())
         optimize_ops = self.create_optimization_pass(params_grads, loss)
         return optimize_ops
 

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

@@ -2,6 +2,7 @@ import unittest
 
 import paddle.v2.framework.framework as framework
 import paddle.v2.framework.optimizer as optimizer
+from paddle.v2.framework.backward import append_backward_ops
 
 
 class TestOptimizer(unittest.TestCase):
@@ -51,7 +52,7 @@ class TestMomentumOptimizer(unittest.TestCase):
             outputs={"Out": mul_out},
             attrs={"x_num_col_dims": 1})
         momentum_optimizer = self.MockMomentum(learning_rate=0.01, momentum=0.2)
-        params_grads = momentum_optimizer.create_backward_pass(mul_out)
+        params_grads = append_backward_ops(mul_out)
         self.assertEqual(len(params_grads), 1)
         self.assertEqual(len(momentum_optimizer.get_accumulators()), 0)
         opts = momentum_optimizer.create_optimization_pass(params_grads,
@@ -93,7 +94,7 @@ class TestAdagradOptimizer(unittest.TestCase):
             outputs={"Out": mul_out},
             attrs={"x_num_col_dims": 1})
         adagrad_optimizer = self.MockAdagrad(learning_rate=0.01, epsilon=1.0e-6)
-        params_grads = adagrad_optimizer.create_backward_pass(mul_out)
+        params_grads = append_backward_ops(mul_out)
         self.assertEqual(len(params_grads), 1)
         self.assertEqual(len(adagrad_optimizer.get_accumulators()), 0)
         opts = adagrad_optimizer.create_optimization_pass(params_grads, mul_out)
@@ -138,7 +139,7 @@ class TestAdamOptimizer(unittest.TestCase):
             attrs={"x_num_col_dims": 1})
         adam_optimizer = self.MockAdam(
             learning_rate=0.01, beta1=0.9, beta2=0.999)
-        params_grads = adam_optimizer.create_backward_pass(mul_out)
+        params_grads = append_backward_ops(mul_out)
         self.assertEqual(len(params_grads), 1)
         self.assertEqual(len(adam_optimizer.get_accumulators()), 0)
         opts = adam_optimizer.create_optimization_pass(params_grads, mul_out)