Restore the param infos in Program.clone() (#5873)

* Restore the param infos in Program.clone()

The Program.clone only clone the variables and ops
in the program into a new program. However, the
information of Parameter is not clone.

So we need restore the information of Parameters.

Fix #5871

* Follow comments

* Fix CI

* Fix CI

* Fix CI

Author: reyoung

python/paddle/v2/fluid/framework.py

@@ -395,7 +395,11 @@ def var(self, name):
         return v
 
     def all_parameters(self):
-        return {v for k, v in self.vars.iteritems() if isinstance(v, Parameter)}
+        return list(self.iter_parameters())
+
+    def iter_parameters(self):
+        return (item[1] for item in self.vars.iteritems()
+                if isinstance(item[1], Parameter))
 
     def create_var(self, *args, **kwargs):
         var = Variable(self, *args, **kwargs)
@@ -469,6 +473,37 @@ def sync_with_cpp(self):
         for index in range(len(self.ops)):
             assert self.ops[index].desc == ops_in_cpp[index]
 
+    def copy_param_info_from(self, other):
+        """
+        Copy the information of parameters from other block
+        Args:
+            other(Block): other block 
+
+        Returns:
+            None
+        """
+        if not isinstance(other, Block):
+            raise TypeError("copy_param_info_from should be invoked with Block")
+        for p in other.iter_parameters():
+            assert isinstance(p, Parameter)
+            v = self.vars.get(p.name, None)
+            if v is None:
+                raise ValueError("copy_param_info_from should be invoked with "
+                                 "same topology")
+            assert isinstance(v, Variable)
+            new_p = Parameter(
+                block=self,
+                shape=v.shape,
+                dtype=v.dtype,
+                type=v.type,
+                lod_level=v.lod_level,
+                stop_gradient=p.stop_gradient,
+                trainable=p.trainable,
+                optimize_attr=p.optimize_attr,
+                regularizer=p.regularizer,
+                name=v.name)
+            self.vars[new_p.name] = new_p
+
 
 class Program(object):
     def __init__(self):
@@ -489,6 +524,7 @@ def clone(self):
         p.desc = core.ProgramDesc(self.desc)
         p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())]
         p.sync_with_cpp()
+        p.copy_param_info_from(self)
         return p
 
     def prune(self, targets):
@@ -572,6 +608,24 @@ def sync_with_cpp(self):
         for block in self.blocks:
             block.sync_with_cpp()
 
+    def copy_param_info_from(self, other):
+        """
+        Copy the information of parameters from other program. 
+        Args:
+            other(Program): Other program
+
+        Returns:
+            None
+        """
+        if not isinstance(other, Program):
+            raise TypeError("copy_param_info_from should be invoked with "
+                            "Program")
+
+        if len(self.blocks) != len(other.blocks):
+            raise ValueError("copy_param_info_from should be invoked with two "
+                             "program, with represent the same topology")
+        self.global_block().copy_param_info_from(other.global_block())
+
     def list_vars(self):
         for each_block in self.blocks:
             for each_var in each_block.vars.itervalues():

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

@@ -1,7 +1,9 @@
+from __future__ import print_function
 import unittest
 
 from paddle.v2.fluid.framework import Program
 from paddle.v2.fluid.framework import g_main_program
+import paddle.v2.fluid.layers as layers
 
 
 class TestProgram(unittest.TestCase):
@@ -48,8 +50,8 @@ def test_program_clone(self):
 
         # FIXME(yuyang18): We manual compare the output string, since the order
         # of variable could be changed.
-        print prog
-        print prog.clone()
+        print(prog)
+        print(prog.clone())
 
     def test_parse_program_from_string(self):
         prog = Program()
@@ -67,8 +69,8 @@ def test_parse_program_from_string(self):
         binary_str = prog.desc.serialize_to_string()
         prog_restored = Program.parse_from_string(binary_str)
 
-        print prog
-        print prog_restored
+        print(prog)
+        print(prog_restored)
 
     def test_append_backward(self):
         prog = Program()
@@ -123,6 +125,20 @@ def grad_name(name):
             actual_ops.append(op.type)
         self.assertEqual(actual_ops, expect_ops)
 
+    def test_program_clone_with_parameter(self):
+        main_program = Program()
+        startup_program = Program()
+        kwargs = {
+            'main_program': main_program,
+            'startup_program': startup_program
+        }
+        d = layers.data(name='x', shape=[784], dtype='float32', **kwargs)
+        hidden = layers.fc(input=d, size=100, **kwargs)
+        layers.fc(input=hidden, size=100, **kwargs)
+
+        new_program = main_program.clone()
+        self.assertNotEqual(0, len(new_program.blocks[0].all_parameters()))
+
 
 if __name__ == '__main__':
     unittest.main()