Add return_numpy for PE (#11792)

Author: chengduo

python/paddle/fluid/executor.py

@@ -78,6 +78,8 @@ def as_numpy(tensor):
     Returns:
         numpy.ndarray
     """
+    if isinstance(tensor, core.LoDTensorArray):
+        return [as_numpy(t) for t in tensor]
     if isinstance(tensor, list):
         return [as_numpy(t) for t in tensor]
     assert isinstance(tensor, core.LoDTensor)

python/paddle/fluid/parallel_executor.py

@@ -160,7 +160,7 @@ def __init__(self,
             build_strategy, num_trainers, trainer_id)
         self.scope = scope
 
-    def run(self, fetch_list, feed=None, feed_dict=None):
+    def run(self, fetch_list, feed=None, feed_dict=None, return_numpy=True):
         """
         Run a parallel executor with fetch_list.
 
@@ -196,6 +196,8 @@ def run(self, fetch_list, feed=None, feed_dict=None):
                 to each device. Default None.
             feed_dict: Alias for feed parameter, for backward compatibility.
                 This parameter has been deprecated. Default None.
+            return_numpy(bool): Whether converts the fetched tensor to numpy.
+                Default: True.
 
         Returns:
             List: The fetched result list.
@@ -270,6 +272,9 @@ def run(self, fetch_list, feed=None, feed_dict=None):
         if self.is_dist:
             self.bcast_params()
 
+        if return_numpy:
+            return executor.as_numpy(arr)
+
         return [arr[i] for i in range(len(arr))]
 
     def bcast_params(self):

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

@@ -81,7 +81,6 @@ def run_executor(exe, feed, fetch_list, program=None):
             begin = time.time()
             first_loss, = run_executor(
                 exe=exe, feed=feed_dict, fetch_list=[loss.name])
-            first_loss = np.array(first_loss)
 
             for i in xrange(iter):
                 run_executor(exe=exe, feed=feed_dict, fetch_list=[])
@@ -94,8 +93,6 @@ def run_executor(exe, feed, fetch_list, program=None):
                 print "%.4f Instance per second" % (
                     (batch_size * iter + 2) / (end - begin))
 
-            last_loss = np.array(last_loss)
-
             print first_loss, last_loss
             # self.assertGreater(first_loss[0], last_loss[0])
             return first_loss, last_loss

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

@@ -169,9 +169,8 @@ def check_network_convergence(self,
             data = train_data()
             for i in xrange(10):
                 cur_batch = next(data)
-                print map(np.array,
-                          pe.run(feed=feeder.feed(cur_batch),
-                                 fetch_list=[avg_cost.name]))[0]
+                print pe.run(feed=feeder.feed(cur_batch),
+                             fetch_list=[avg_cost.name])[0]
 
     @unittest.skip(reason="CI hangs")
     def test_update_sparse_parameter_all_reduce(self):

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

@@ -75,7 +75,9 @@ def parallel_exe(self, train_inputs, seed, use_cuda):
                     fetch_list.append(k)
 
             for data in train_inputs:
-                ret = pe.run(fetch_list, feed=feeder.feed(data))
+                ret = pe.run(fetch_list,
+                             feed=feeder.feed(data),
+                             return_numpy=True)
                 for i in range(len(fetch_list)):
                     assert not math.isnan(np.sum(ret[i])) and \
                            not math.isinf(np.sum(ret[i]))
@@ -128,7 +130,7 @@ def parallel_exe(self, use_cuda, seed):
             use_cuda=use_cuda, loss_name=loss.name, main_program=main)
 
         for batch_id, data in enumerate(reader()):
-            loss_np = np.array(pe.run(feed=data, fetch_list=[loss.name])[0])
+            loss_np = pe.run(feed=data, fetch_list=[loss.name])[0]
             print batch_id, loss_np
             if batch_id == 2:
                 break

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

@@ -70,10 +70,9 @@ def check_network_convergence(self, use_cuda, build_strategy=None):
 
             for i in xrange(5):
                 test_loss, = test_exe.run([loss.name], feed=feed_dict)
-                test_loss = np.array(test_loss)
 
                 train_loss, = train_exe.run([loss.name], feed=feed_dict)
-                train_loss = np.array(train_loss)
+
                 self.assertTrue(
                     np.allclose(
                         train_loss, test_loss, atol=1e-8),