support fetch the feed var when use_prune=True, test=develop (#24110)

python/paddle/fluid/framework.py

@@ -4237,6 +4237,14 @@ class Program(object):
                     raise ValueError(
                         "All targets of Program._prune_with_input() can only be "
                         "Variable or Operator, but received %s." % type(t))
+
+                # NOTEZ(zhiqiu): For variable to be fed in fetch_list, there two cases:
+                # (1) the variable is leaf, it has no op that generates it;
+                # (2) the variable is not leaf, and we need to prune the op that generates it.
+                # In both cases, wo can just skip target_op of that it.
+                if name in feeded_var_names:
+                    continue
+
                 # After transpiler processing, the op that output this
                 # variable maybe has been changed, so t.op is not reliable
                 # and we need to find the current op that generate this
@@ -4253,10 +4261,13 @@ class Program(object):
                         else:
                             target_op = op
                             break
-                t = target_op
-                if t is None:
-                    raise ValueError("The target variable must have an "
+                if target_op is None:
+                    raise ValueError(
+                        "The target variable used for pruning should have an "
                         "associated operator that generates it.")
+                else:
+                    targets_idx.append([target_op.block.idx, target_op.idx])
+            else:
                 targets_idx.append([t.block.idx, t.idx])
 
         res = Program()

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

@@ -725,6 +725,59 @@ class TestExecutorRunAutoPrune(unittest.TestCase):
         self.assertTrue(np.array_equal(weight_with_prune, weight_expected))
         self.assertFalse(np.array_equal(weight_without_prune, weight_expected))
 
+    def test_prune_feed_var_in_fetchlist_1(self):
+        # the variable to be fed is not leaf
+        program = framework.Program()
+        startup_program = framework.Program()
+        scope = fluid.Scope()
+        with fluid.scope_guard(scope):
+            with fluid.program_guard(program, startup_program):
+                (x, y, label, loss1, loss2, w_param_attrs) = self.net1()
+                exe = fluid.Executor(fluid.CPUPlace())
+                exe.run(startup_program)
+                weight_init = np.array(
+                    scope.find_var(w_param_attrs.name).get_tensor())
+                x_np = np.random.random(size=(10, 2)).astype('float32')
+                label_np = np.random.randint(1, size=(10, 1)).astype('int64')
+                res = exe.run(program,
+                              feed={y.name: x_np,
+                                    'label': label_np},
+                              fetch_list=[y.name, loss1.name],
+                              use_prune=True)
+                self.assertIsNotNone(scope.find_var(loss1.name))
+                self.assertIsNone(scope.find_var(loss2.name))
+                self.assertIsNone(scope.find_var(x.name))
+                weight = np.array(
+                    scope.find_var(w_param_attrs.name).get_tensor())
+                self.assertTrue(np.array_equal(weight_init,
+                                               weight))  # weight unchanged
+
+    def test_prune_feed_var_in_fetchlist_2(self):
+        # the variable to be fed is leaf
+        program = framework.Program()
+        startup_program = framework.Program()
+        scope = fluid.Scope()
+        with fluid.scope_guard(scope):
+            with fluid.program_guard(program, startup_program):
+                (x, y, label, loss1, loss2, w_param_attrs) = self.net1()
+                exe = fluid.Executor(fluid.CPUPlace())
+                exe.run(startup_program)
+                weight_init = np.array(
+                    scope.find_var(w_param_attrs.name).get_tensor())
+                x_np = np.random.random(size=(10, 2)).astype('float32')
+                label_np = np.random.randint(1, size=(10, 1)).astype('int64')
+                res = exe.run(program,
+                              feed={x.name: x_np,
+                                    'label': label_np},
+                              fetch_list=[x.name, loss1.name],
+                              use_prune=True)
+                self.assertIsNotNone(scope.find_var(loss1.name))
+                self.assertIsNone(scope.find_var(loss2.name))
+                weight = np.array(
+                    scope.find_var(w_param_attrs.name).get_tensor())
+                self.assertTrue(np.array_equal(weight_init,
+                                               weight))  # weight unchanged
+
 
 if __name__ == '__main__':
     unittest.main()