Replace legacy tensor.Tensor with torch.tensor/torch.empty (#12027)

* Replace legacy torch.Tensor constructor with torch.{tensor, empty}

* Remove torch.Tensor in examples

examples/research_projects/lxmert/modeling_frcnn.py

@@ -1426,7 +1426,7 @@ class AnchorGenerator(nn.Module):
                 h = aspect_ratio * w
                 x0, y0, x1, y1 = -w / 2.0, -h / 2.0, w / 2.0, h / 2.0
                 anchors.append([x0, y0, x1, y1])
-        return nn.Parameter(torch.Tensor(anchors))
+        return nn.Parameter(torch.tensor(anchors))
 
     def forward(self, features):
         """

examples/research_projects/lxmert/utils.py

@@ -532,7 +532,7 @@ def load_frcnn_pkl_from_url(url):
     for k, v in model.items():
         new[k] = torch.from_numpy(v)
         if "running_var" in k:
-            zero = torch.Tensor([0])
+            zero = torch.tensor([0])
             k2 = k.replace("running_var", "num_batches_tracked")
             new[k2] = zero
     return new

examples/research_projects/movement-pruning/emmental/modules/masked_nn.py

@@ -72,7 +72,7 @@ class MaskedLinear(nn.Linear):
         if self.pruning_method in ["topK", "threshold", "sigmoied_threshold", "l0"]:
             self.mask_scale = mask_scale
             self.mask_init = mask_init
-            self.mask_scores = nn.Parameter(torch.Tensor(self.weight.size()))
+            self.mask_scores = nn.Parameter(torch.empty(self.weight.size()))
             self.init_mask()
 
     def init_mask(self):

examples/research_projects/rag-end2end-retriever/finetune_rag.py

@@ -223,7 +223,7 @@ class GenerativeQAModule(BaseTransformer):
                 decoder_start_token_id = generator.config.decoder_start_token_id
                 decoder_input_ids = (
                     torch.cat(
-                        [torch.Tensor([[decoder_start_token_id]] * target_ids.shape[0]).to(target_ids), target_ids],
+                        [torch.tensor([[decoder_start_token_id]] * target_ids.shape[0]).to(target_ids), target_ids],
                         dim=1,
                     )
                     if target_ids.shape[0] < self.target_lens["train"]

examples/research_projects/rag/finetune_rag.py

@@ -222,7 +222,7 @@ class GenerativeQAModule(BaseTransformer):
                 decoder_start_token_id = generator.config.decoder_start_token_id
                 decoder_input_ids = (
                     torch.cat(
-                        [torch.Tensor([[decoder_start_token_id]] * target_ids.shape[0]).to(target_ids), target_ids],
+                        [torch.tensor([[decoder_start_token_id]] * target_ids.shape[0]).to(target_ids), target_ids],
                         dim=1,
                     )
                     if target_ids.shape[0] < self.target_lens["train"]

src/transformers/models/convbert/modeling_convbert.py

@@ -473,8 +473,8 @@ class GroupedLinearLayer(nn.Module):
         self.num_groups = num_groups
         self.group_in_dim = self.input_size // self.num_groups
         self.group_out_dim = self.output_size // self.num_groups
-        self.weight = nn.Parameter(torch.Tensor(self.num_groups, self.group_in_dim, self.group_out_dim))
+        self.weight = nn.Parameter(torch.empty(self.num_groups, self.group_in_dim, self.group_out_dim))
-        self.bias = nn.Parameter(torch.Tensor(output_size))
+        self.bias = nn.Parameter(torch.empty(output_size))
 
     def forward(self, hidden_states):
         batch_size = list(hidden_states.size())[0]

tests/test_activations.py

@@ -27,7 +27,7 @@ if is_torch_available():
 @require_torch
 class TestActivations(unittest.TestCase):
     def test_gelu_versions(self):
-        x = torch.Tensor([-100, -1, -0.1, 0, 0.1, 1.0, 100])
+        x = torch.tensor([-100, -1, -0.1, 0, 0.1, 1.0, 100])
         torch_builtin = get_activation("gelu")
         self.assertTrue(torch.eq(_gelu_python(x), torch_builtin(x)).all().item())
         self.assertFalse(torch.eq(_gelu_python(x), gelu_new(x)).all().item())

tests/test_modeling_bart.py

@@ -308,14 +308,16 @@ class BartHeadTests(unittest.TestCase):
             max_position_embeddings=48,
         )
         lm_model = BartForConditionalGeneration(config).to(torch_device)
-        context = torch.Tensor([[71, 82, 18, 33, 46, 91, 2], [68, 34, 26, 58, 30, 2, 1]]).long().to(torch_device)
+        context = torch.tensor(
-        summary = torch.Tensor([[82, 71, 82, 18, 2], [58, 68, 2, 1, 1]]).long().to(torch_device)
+            [[71, 82, 18, 33, 46, 91, 2], [68, 34, 26, 58, 30, 2, 1]], device=torch_device, dtype=torch.long
+        )
+        summary = torch.tensor([[82, 71, 82, 18, 2], [58, 68, 2, 1, 1]], device=torch_device, dtype=torch.long)
         outputs = lm_model(input_ids=context, decoder_input_ids=summary, labels=summary)
         expected_shape = (*summary.shape, config.vocab_size)
         self.assertEqual(outputs["logits"].shape, expected_shape)
 
     def test_generate_beam_search(self):
-        input_ids = torch.Tensor([[71, 82, 2], [68, 34, 2]]).long().to(torch_device)
+        input_ids = torch.tensor([[71, 82, 2], [68, 34, 2]], device=torch_device, dtype=torch.long)
         config = BartConfig(
             vocab_size=self.vocab_size,
             d_model=24,
@@ -345,7 +347,7 @@ class BartHeadTests(unittest.TestCase):
         self.assertEqual(generated_ids.shape, (input_ids.shape[0], max_length))
 
     def test_shift_tokens_right(self):
-        input_ids = torch.Tensor([[71, 82, 18, 33, 2, 1, 1], [68, 34, 26, 58, 30, 82, 2]]).long()
+        input_ids = torch.tensor([[71, 82, 18, 33, 2, 1, 1], [68, 34, 26, 58, 30, 82, 2]], dtype=torch.long)
         shifted = shift_tokens_right(input_ids, 1, 2)
         n_pad_before = input_ids.eq(1).float().sum()
         n_pad_after = shifted.eq(1).float().sum()
@@ -358,8 +360,8 @@ class BartHeadTests(unittest.TestCase):
         tokenizer = BartTokenizer.from_pretrained("facebook/bart-large")
         examples = [" Hello world", " DomDramg"]  # need leading spaces for equality
         fairseq_results = [
-            torch.Tensor([0, 20920, 232, 2]),
+            torch.tensor([0, 20920, 232, 2]),
-            torch.Tensor([0, 11349, 495, 4040, 571, 2]),
+            torch.tensor([0, 11349, 495, 4040, 571, 2]),
         ]
         for ex, desired_result in zip(examples, fairseq_results):
             bart_toks = tokenizer.encode(ex, return_tensors="pt").squeeze()
@@ -614,7 +616,7 @@ class BartModelIntegrationTests(unittest.TestCase):
         batched_logits = outputs.logits
         expected_shape = torch.Size((2, 3))
         self.assertEqual(batched_logits.shape, expected_shape)
-        expected_slice = torch.Tensor([[0.1907, 1.4342, -1.0289]]).to(torch_device)
+        expected_slice = torch.tensor([[0.1907, 1.4342, -1.0289]], device=torch_device)
         logits_arr = batched_logits[0].detach()
 
         # Test that padding does not change results

tests/test_modeling_clip.py

@@ -556,6 +556,6 @@ class CLIPModelIntegrationTest(unittest.TestCase):
             torch.Size((inputs.input_ids.shape[0], inputs.pixel_values.shape[0])),
         )
 
-        expected_logits = torch.Tensor([[24.5056, 18.8076]]).to(torch_device)
+        expected_logits = torch.tensor([[24.5056, 18.8076]], device=torch_device)
 
         self.assertTrue(torch.allclose(outputs.logits_per_image, expected_logits, atol=1e-3))

tests/test_modeling_fsmt.py

@@ -305,7 +305,7 @@ class FSMTHeadTests(unittest.TestCase):
         return config, input_ids, batch_size
 
     def test_generate_beam_search(self):
-        input_ids = torch.Tensor([[71, 82, 2], [68, 34, 2]]).long().to(torch_device)
+        input_ids = torch.tensor([[71, 82, 2], [68, 34, 2]], dtype=torch.long, device=torch_device)
         config = self._get_config()
         lm_model = FSMTForConditionalGeneration(config).to(torch_device)
         lm_model.eval()
@@ -322,7 +322,7 @@ class FSMTHeadTests(unittest.TestCase):
         self.assertEqual(new_input_ids.shape, (input_ids.shape[0], max_length))
 
     def test_shift_tokens_right(self):
-        input_ids = torch.Tensor([[71, 82, 18, 33, 2, 1, 1], [68, 34, 26, 58, 30, 82, 2]]).long()
+        input_ids = torch.tensor([[71, 82, 18, 33, 2, 1, 1], [68, 34, 26, 58, 30, 82, 2]], dtype=torch.long)
         shifted = shift_tokens_right(input_ids, 1)
         n_pad_before = input_ids.eq(1).float().sum()
         n_pad_after = shifted.eq(1).float().sum()

tests/test_modeling_ibert.py

@@ -285,7 +285,7 @@ class IBertModelTest(ModelTesterMixin, unittest.TestCase):
         config = self.model_tester.prepare_config_and_inputs()[0]
         embeddings = IBertEmbeddings(config=config)
 
-        inputs_embeds = torch.Tensor(2, 4, 30)
+        inputs_embeds = torch.empty(2, 4, 30)
         expected_single_positions = [
             0 + embeddings.padding_idx + 1,
             1 + embeddings.padding_idx + 1,

tests/test_modeling_mbart.py

@@ -396,8 +396,10 @@ class MBartEnroIntegrationTest(AbstractSeq2SeqIntegrationTest):
             add_final_layer_norm=True,
         )
         lm_model = MBartForConditionalGeneration(config).to(torch_device)
-        context = torch.Tensor([[71, 82, 18, 33, 46, 91, 2], [68, 34, 26, 58, 30, 2, 1]]).long().to(torch_device)
+        context = torch.tensor(
-        summary = torch.Tensor([[82, 71, 82, 18, 2], [58, 68, 2, 1, 1]]).long().to(torch_device)
+            [[71, 82, 18, 33, 46, 91, 2], [68, 34, 26, 58, 30, 2, 1]], device=torch_device, dtype=torch.long
+        )
+        summary = torch.tensor([[82, 71, 82, 18, 2], [58, 68, 2, 1, 1]], device=torch_device, dtype=torch.long)
         result = lm_model(input_ids=context, decoder_input_ids=summary, labels=summary)
         expected_shape = (*summary.shape, config.vocab_size)
         self.assertEqual(result.logits.shape, expected_shape)

tests/test_modeling_roberta.py

@@ -461,7 +461,7 @@ class RobertaModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCas
         config = self.model_tester.prepare_config_and_inputs()[0]
         embeddings = RobertaEmbeddings(config=config)
 
-        inputs_embeds = torch.Tensor(2, 4, 30)
+        inputs_embeds = torch.empty(2, 4, 30)
         expected_single_positions = [
             0 + embeddings.padding_idx + 1,
             1 + embeddings.padding_idx + 1,