Workaround the optimizer learning rate map issue

optimizer use `default_main_program()` as key for learning rate map, `Program`
seem to rely on object hash for equality check, and cloned program is not equal
to the original

model.py

@@ -78,7 +78,7 @@ class StaticGraphAdapter(object):
         # with `_build_once` gone, parameters are now created in `__init__`
         # so we need to keep track of the parameters already created
         self._startup_prog = fluid.default_startup_program()
-        self._main_prog = fluid.default_main_program()
+        self._orig_prog = fluid.default_main_program()
 
         self._label_vars = None  # label variables
         self._endpoints = {}
@@ -241,7 +241,12 @@ class StaticGraphAdapter(object):
             return out[:num_output], out[num_output:]
 
     def _make_program(self, inputs):
-        prog = self._main_prog.clone(self.mode != 'train')
+        prog = self._orig_prog.clone(for_test=self.mode != 'train')
+        if self.mode == 'train':
+            # HACK workaround learning rate map issue
+            lr_var = self.model._optimizer._learning_rate_map[self._orig_prog]
+            self.model._optimizer._learning_rate_map[prog] = lr_var
+        losses = []
         with fluid.program_guard(prog, self._startup_prog):
             outputs = to_list(self.model.forward(*inputs))
             losses = []