[cherry pick] dataloader fix (#31028)

* fix dataloader collate return list mix tensor and numpy array. test=develop

* remove numpy array check in single-process dataloader. test=develop

python/paddle/fluid/dataloader/dataloader_iter.py

@@ -320,7 +320,6 @@ class _DataLoaderIterSingleProcess(_DataLoaderIterBase):
                 array = core.LoDTensorArray()
                 for slot in batch:
                     if not isinstance(slot, core.LoDTensor):
-                        self._check_input_array(slot)
                         # FIXME(dkp): blocking_queue only support
                         #             core.LoDTensorArray as input now, read
                         #             numpy data into a LoDTensorArray here,
@@ -346,19 +345,6 @@ class _DataLoaderIterSingleProcess(_DataLoaderIterBase):
             logging.warning("DataLoader reader thread raised an exception.")
             six.reraise(*sys.exc_info())
 
-    @classmethod
-    def _check_input_array(cls, item):
-        if isinstance(item, paddle.Tensor):
-            return
-        arr = np.array(item)
-        if arr.dtype == np.object:
-            raise TypeError((
-                "\n\tFaild to convert input data to a regular ndarray :\n\t* Usually "
-                "this means the input data contains nested lists with different lengths. "
-                "\n\t* Check the reader function passed to 'decorate_batch_generator'"
-                " to locate the data causes this issue.\n\t* Please consider using "
-                "'fluid.create_lod_tensor' to convert it to a LoD-Tensor."))
-
     def __next__(self):
         try:
             if in_dygraph_mode():
@@ -454,11 +440,16 @@ def _worker_loop(dataset, dataset_kind, indices_queue, out_queue, done_event,
                 if use_shared_memory:
                     # FIXME(dkp): _convert_to_tensor_list only support np.array
                     #             list now, should support paddle.Tensor list
-                    if isinstance(batch[0][0], paddle.Tensor):
-                        np_batch = []
-                        for sample in batch:
-                            np_batch.append([s.numpy() for s in sample])
-                        batch = np_batch
+                    new_batch = []
+                    for sample in batch:
+                        new_sample = []
+                        for s in sample:
+                            if isinstance(s, paddle.Tensor):
+                                new_sample.append(s.numpy())
+                            else:
+                                new_sample.append(s)
+                        new_batch.append(new_sample)
+                    batch = new_batch
 
                     tensor_list = core._convert_to_tensor_list(batch)
                     out_queue.put((idx, tensor_list))

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

@@ -142,5 +142,43 @@ class TestChainDataset(unittest.TestCase):
             self.run_main(num_workers=0, places=p)
 
 
+class NumpyMixTensorDataset(Dataset):
+    def __init__(self, sample_num):
+        self.sample_num = sample_num
+
+    def __len__(self):
+        return self.sample_num
+
+    def __getitem__(self, idx):
+        np.random.seed(idx)
+        image = np.random.random([IMAGE_SIZE]).astype('float32')
+        label = np.random.randint(0, 9, (1, )).astype('int64')
+        return paddle.to_tensor(image, place=paddle.CPUPlace()), label
+
+
+class TestNumpyMixTensorDataset(TestTensorDataset):
+    def run_main(self, num_workers, places):
+        paddle.static.default_startup_program().random_seed = 1
+        paddle.static.default_main_program().random_seed = 1
+        place = paddle.CPUPlace()
+        with fluid.dygraph.guard(place):
+            dataset = NumpyMixTensorDataset(16)
+            assert len(dataset) == 16
+            dataloader = DataLoader(
+                dataset,
+                places=place,
+                num_workers=num_workers,
+                batch_size=1,
+                drop_last=True)
+
+            for i, (input, label) in enumerate(dataloader()):
+                assert len(input) == 1
+                assert len(label) == 1
+                assert input.shape == [1, IMAGE_SIZE]
+                assert label.shape == [1, 1]
+                assert isinstance(input, paddle.Tensor)
+                assert isinstance(label, paddle.Tensor)
+
+
 if __name__ == '__main__':
     unittest.main()