[EncoderDecoder] Make tests more aggressive (#9256)

* add tests

* make style and fix bart bug

* fix bart past key value edge case

* correct tf bart test

* fix gpt2 tf

* fix t5 test

src/transformers/models/bart/modeling_bart.py

@@ -95,9 +95,7 @@ def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_
     return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
 
 
-def _expand_mask(
-    mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None, past_key_values_length: int = 0
-):
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
     """
     Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
     """
@@ -106,16 +104,6 @@ def _expand_mask(
 
     expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
 
-    if past_key_values_length > 0:
-        # concat fully attendend attention_mask to the beginning if `past_key_values` are used
-        expanded_mask = torch.cat(
-            [
-                torch.ones(bsz, 1, tgt_len, past_key_values_length, device=expanded_mask.device, dtype=dtype),
-                expanded_mask,
-            ],
-            dim=-1,
-        )
-
     inverted_mask = 1.0 - expanded_mask
 
     return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(dtype).min)
@@ -941,11 +929,21 @@ class BartDecoder(BartPretrainedModel):
             attention_mask = input_ids.ne(self.config.pad_token_id).to(torch.long)
             # never mask leading token, even if it is pad
             attention_mask[:, 0] = attention_mask[:, 1]
+            if past_key_values_length > 0:
+                attention_mask = torch.cat(
+                    [
+                        torch.ones(
+                            (input_shape[0], past_key_values_length), dtype=torch.long, device=input_ids.device
+                        ),
+                        attention_mask,
+                    ],
+                    dim=-1,
+                )
 
         if attention_mask is not None and combined_attention_mask is not None:
             # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
             combined_attention_mask = combined_attention_mask + _expand_mask(
-                attention_mask, inputs_embeds.dtype, past_key_values_length=past_key_values_length
+                attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
             )
 
         # expand encoder attention mask

src/transformers/models/bart/modeling_tf_bart.py

@@ -94,7 +94,7 @@ def _make_causal_mask(input_ids_shape: tf.TensorShape, past_key_values_length: i
     return tf.broadcast_to(mask[None, None, :, :], (bsz, 1, tgt_len, tgt_len + past_key_values_length))
 
 
-def _expand_mask(mask: tf.Tensor, tgt_len: Optional[int] = None, past_key_values_length: int = 0):
+def _expand_mask(mask: tf.Tensor, tgt_len: Optional[int] = None):
     """
     Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
     """
@@ -103,16 +103,6 @@ def _expand_mask(mask: tf.Tensor, tgt_len: Optional[int] = None, past_key_values
 
     expanded_mask = tf.cast(tf.broadcast_to(mask[:, None, None, :], (bsz, 1, tgt_len, src_len)), tf.float32)
 
-    if past_key_values_length > 0:
-        # concat fully attendend attention_mask to the beginning if `past_key_values` are used
-        expanded_mask = tf.concat(
-            [
-                tf.ones((bsz, 1, tgt_len, past_key_values_length), dtype=tf.float32),
-                expanded_mask,
-            ],
-            axis=-1,
-        )
-
     return (1.0 - expanded_mask) * LARGE_NEGATIVE
 
 
@@ -902,14 +892,16 @@ class TFBartDecoder(tf.keras.layers.Layer):
             attention_mask = tf.cast(
                 tf.math.not_equal(inputs["input_ids"], self.config.pad_token_id), inputs["input_ids"].dtype
             )
+            attention_mask = tf.concat(
+                [tf.ones((input_shape[0], past_key_values_length), dtype=attention_mask.dtype), attention_mask],
+                axis=-1,
+            )
         else:
-            attention_mask = tf.ones(input_shape, dtype=tf.int32)
+            attention_mask = tf.ones((input_shape[0], input_shape[1] + past_key_values_length), dtype=tf.int32)
 
         if attention_mask is not None and combined_attention_mask is not None:
             # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
-            combined_attention_mask = combined_attention_mask + _expand_mask(
-                attention_mask, past_key_values_length=past_key_values_length
-            )
+            combined_attention_mask = combined_attention_mask + _expand_mask(attention_mask, tgt_len=input_shape[-1])
 
         encoder_hidden_states = inputs["encoder_hidden_states"]
         if encoder_hidden_states is not None and inputs["encoder_attention_mask"] is not None:

tests/test_modeling_bart.py

@@ -156,12 +156,16 @@ class BartModelTester:
 
         # create hypothetical multiple next token and extent to next_input_ids
         next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
 
         # append to next input_ids and
         next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+        next_attention_mask = torch.cat([inputs_dict["attention_mask"], next_mask], dim=-1)
 
-        output_from_no_past = model(next_input_ids)["last_hidden_state"]
-        output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
+        output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
+        output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[
+            "last_hidden_state"
+        ]
 
         # select random slice
         random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()

tests/test_modeling_encoder_decoder.py

@@ -568,15 +568,27 @@ class BertEncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCase):
         model.to(torch_device)
         tokenizer = BertTokenizer.from_pretrained("patrickvonplaten/bert2bert-cnn_dailymail-fp16")
 
-        ARTICLE = """(CNN)Sigma Alpha Epsilon is under fire for a video showing party-bound fraternity members singing a racist chant. SAE's national chapter suspended the students, but University of Oklahoma President David Boren took it a step further, saying the university's affiliation with the fraternity is permanently done. The news is shocking, but it's not the first time SAE has faced controversy. SAE was founded March 9, 1856, at the University of Alabama, five years before the American Civil War, according to the fraternity website. When the war began, the group had fewer than 400 members, of which "369 went to war for the Confederate States and seven for the Union Army," the website says. The fraternity now boasts more than 200,000 living alumni, along with about 15,000 undergraduates populating 219 chapters and 20 "colonies" seeking full membership at universities. SAE has had to work hard to change recently after a string of member deaths, many blamed on the hazing of new recruits, SAE national President Bradley Cohen wrote in a message on the fraternity's website. The fraternity's website lists more than 130 chapters cited or suspended for "health and safety incidents" since 2010. At least 30 of the incidents involved hazing, and dozens more involved alcohol. However, the list is missing numerous incidents from recent months. Among them, according to various media outlets: Yale University banned the SAEs from campus activities last month after members allegedly tried to interfere with a sexual misconduct investigation connected to an initiation rite. Stanford University in December suspended SAE housing privileges after finding sorority members attending a fraternity function were subjected to graphic sexual content. And Johns Hopkins University in November suspended the fraternity for underage drinking. "The media has labeled us as the 'nation's deadliest fraternity,' " Cohen said. In 2011, for example, a student died while being coerced into excessive alcohol consumption, according to a lawsuit. SAE's previous insurer dumped the fraternity. "As a result, we are paying Lloyd's of London the highest insurance rates in the Greek-letter world," Cohen said. Universities have turned down SAE's attempts to open new chapters, and the fraternity had to close 12 in 18 months over hazing incidents."""
+        ARTICLE_SIGMA = """(CNN)Sigma Alpha Epsilon is under fire for a video showing party-bound fraternity members singing a racist chant. SAE's national chapter suspended the students, but University of Oklahoma President David Boren took it a step further, saying the university's affiliation with the fraternity is permanently done. The news is shocking, but it's not the first time SAE has faced controversy. SAE was founded March 9, 1856, at the University of Alabama, five years before the American Civil War, according to the fraternity website. When the war began, the group had fewer than 400 members, of which "369 went to war for the Confederate States and seven for the Union Army," the website says. The fraternity now boasts more than 200,000 living alumni, along with about 15,000 undergraduates populating 219 chapters and 20 "colonies" seeking full membership at universities. SAE has had to work hard to change recently after a string of member deaths, many blamed on the hazing of new recruits, SAE national President Bradley Cohen wrote in a message on the fraternity's website. The fraternity's website lists more than 130 chapters cited or suspended for "health and safety incidents" since 2010. At least 30 of the incidents involved hazing, and dozens more involved alcohol. However, the list is missing numerous incidents from recent months. Among them, according to various media outlets: Yale University banned the SAEs from campus activities last month after members allegedly tried to interfere with a sexual misconduct investigation connected to an initiation rite. Stanford University in December suspended SAE housing privileges after finding sorority members attending a fraternity function were subjected to graphic sexual content. And Johns Hopkins University in November suspended the fraternity for underage drinking. "The media has labeled us as the 'nation's deadliest fraternity,' " Cohen said. In 2011, for example, a student died while being coerced into excessive alcohol consumption, according to a lawsuit. SAE's previous insurer dumped the fraternity. "As a result, we are paying Lloyd's of London the highest insurance rates in the Greek-letter world," Cohen said. Universities have turned down SAE's attempts to open new chapters, and the fraternity had to close 12 in 18 months over hazing incidents."""
 
-        EXPECTED_SUMMARY = """sae was founded in 1856, five years before the civil war. the fraternity has had to work hard to change recently. the university of oklahoma president says the university's affiliation with the fraternity is permanently done. the sae has had a string of members in recent months."""
+        ARTICLE_AMERICA = """(CNN) -- The 2013 America's Cup will be faster than ever after organizers announced that wingsail catamarans will be the vessels of choice. The race has historically been between yachts with a single hull, however the 34th edition of the contest will be between multi-hull vessels with wings rather than traditional sails. This means the boats will travel faster through the water, with top speeds in excess of 30 knots, almost three times as fast as in the past. The Golden Gate Yacht Club, hosts of the 2013 race and holders of the cup, have also announced a new, shorter race format for the competition. In an attempt to boost interest in one of sailing's showpiece events an annual World Series will also take place, starting in 2011, resulting a world champion team being crowned. In addition, a youth America's Cup will also be introduced, set to begin in 2012. In a statement on the International Sailing Federation (ISAF) website, the CEO of 2010's winning syndicate BMW ORACLE Racing Russell Coutts explained the reasons behind the changes. "We believe this new format and new boat will put the America's Cup back at the pinnacle of our sport," said Coutts. "These changes will give equal opportunity to competitors and long-term economic stability to all teams and all commercial partners. We promised fairness and innovation and this is what we've delivered." The statement also explained how, in addition to generating interest in the contest, the new annual America's Cup World Series will provide increased commercial revenue for the teams and their sponsors. The venue for the 2013 contest is not due to be announced until the end of the year, with San Francisco, Valencia and a location near Rome believed to be under consideration. Vincenzo Onorato, President of the 2013 challengers Mascalzone Latino, supported the changes: "I think that we need to acknowledge that the Defender has kept its word. The America's Cup is going to have fair rules and a truly independent management of the racing."""
 
-        input_ids = tokenizer(ARTICLE, return_tensors="pt").input_ids.to(torch_device)
-        output_ids = model.generate(input_ids)
-        summary = tokenizer.decode(output_ids[0], skip_special_tokens=True)
+        EXPECTED_SUMMARY_SIGMA = """sae was founded in 1856, five years before the civil war. the fraternity has had to work hard to change recently. the university of oklahoma president says the university's affiliation with the fraternity is permanently done. the sae has had a string of members in recent months."""
 
-        self.assertEqual(summary, EXPECTED_SUMMARY)
+        EXPECTED_SUMMARY_AMERICA = """the 2013 america's cup will be faster than ever. the 34th edition of the competition will be held in 2011. the 2013 race will be between multi - hull vessels with wings rather than traditional sails. the new america'' cup will provide increased commercial revenue. the event will also be expanded to a youth america'cup."""
+
+        input_dict = tokenizer(
+            [ARTICLE_SIGMA, ARTICLE_AMERICA],
+            padding="max_length",
+            pad_to_max_length=True,
+            max_length=512,
+            return_tensors="pt",
+        )
+        output_ids = model.generate(
+            input_dict["input_ids"].to(torch_device), attention_mask=input_dict["attention_mask"].to(torch_device)
+        )
+        summary = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
+
+        self.assertEqual(summary, [EXPECTED_SUMMARY_SIGMA, EXPECTED_SUMMARY_AMERICA])
 
 
 @require_torch
@@ -630,15 +642,23 @@ class BertGenerationEncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCa
         model.to(torch_device)
         tokenizer = AutoTokenizer.from_pretrained("google/roberta2roberta_L-24_bbc")
 
-        ARTICLE = """The problem is affecting people using the older versions of the PlayStation 3, called the "Fat" model.The problem isn't affecting the newer PS3 Slim systems that have been on sale since September last year.Sony have also said they are aiming to have the problem fixed shortly but is advising some users to avoid using their console for the time being."We hope to resolve this problem within the next 24 hours," a statement reads. "In the meantime, if you have a model other than the new slim PS3, we advise that you do not use your PS3 system, as doing so may result in errors in some functionality, such as recording obtained trophies, and not being able to restore certain data."We believe we have identified that this problem is being caused by a bug in the clock functionality incorporated in the system."The PlayStation Network is used by millions of people around the world.It allows users to play their friends at games like Fifa over the internet and also do things like download software or visit online stores."""
+        ARTICLE_PS3 = """The problem is affecting people using the older versions of the PlayStation 3, called the "Fat" model.The problem isn't affecting the newer PS3 Slim systems that have been on sale since September last year.Sony have also said they are aiming to have the problem fixed shortly but is advising some users to avoid using their console for the time being."We hope to resolve this problem within the next 24 hours," a statement reads. "In the meantime, if you have a model other than the new slim PS3, we advise that you do not use your PS3 system, as doing so may result in errors in some functionality, such as recording obtained trophies, and not being able to restore certain data."We believe we have identified that this problem is being caused by a bug in the clock functionality incorporated in the system."The PlayStation Network is used by millions of people around the world.It allows users to play their friends at games like Fifa over the internet and also do things like download software or visit online stores."""
 
-        EXPECTED_SUMMARY = """Sony has said that a bug in its PlayStation 3 console is preventing them from using the machine as a computer."""
+        ARTICLE_TOSHIBA = """An independent panel appointed by Toshiba found institutional accounting irregularities, the firm said in a statement to investors. Toshiba said it "takes the situation it has caused very seriously" and that it "deeply apologised" to shareholders. The overstatement was roughly triple an initial Toshiba estimate. The probe could lead to a restatement of earnings, a board overhaul and potential action by regulators. "Within Toshiba, there was a corporate culture in which one could not go against the wishes of superiors," the report said. "Therefore, when top management presented 'challenges', division presidents, line managers and employees below them continually carried out inappropriate accounting practices to meet targets in line with the wishes of their superiors." The improper accounting practices stretched back to 2008."""
 
-        input_ids = tokenizer(ARTICLE, return_tensors="pt").input_ids.to(torch_device)
-        output_ids = model.generate(input_ids)
-        summary = tokenizer.decode(output_ids[0], skip_special_tokens=True)
+        EXPECTED_SUMMARY_PS3 = """Sony has said that a bug in its PlayStation 3 console is preventing them from using the machine as a computer."""
+
+        EXPECTED_SUMMARY_TOSHIBA = """Japanese electronics giant Toshiba overstated its annual earnings by more than a third last year, according to a report."""
+
+        input_dict = tokenizer(
+            [ARTICLE_PS3, ARTICLE_TOSHIBA], max_length=512, padding="max_length", return_tensors="pt"
+        )
+        output_ids = model.generate(
+            input_dict["input_ids"].to(torch_device), attention_mask=input_dict["attention_mask"].to(torch_device)
+        )
+        summary = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
 
-        self.assertEqual(summary, EXPECTED_SUMMARY)
+        self.assertEqual(summary, [EXPECTED_SUMMARY_PS3, EXPECTED_SUMMARY_TOSHIBA])
 
 
 @require_torch

tests/test_modeling_gpt2.py

@@ -278,22 +278,26 @@ class GPT2ModelTester:
         model.eval()
 
         # first forward pass
-        outputs = model(input_ids, token_type_ids=token_type_ids, use_cache=True)
+        outputs = model(input_ids, token_type_ids=token_type_ids, attention_mask=input_mask, use_cache=True)
 
         output, past = outputs.to_tuple()
 
         # create hypothetical next token and extent to next_input_ids
         next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
         next_token_types = ids_tensor([self.batch_size, 3], self.type_vocab_size)
+        next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
 
         # append to next input_ids and token_type_ids
         next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
         next_token_type_ids = torch.cat([token_type_ids, next_token_types], dim=-1)
-
-        output_from_no_past = model(next_input_ids, token_type_ids=next_token_type_ids)["last_hidden_state"]
-        output_from_past = model(next_tokens, token_type_ids=next_token_types, past_key_values=past)[
-            "last_hidden_state"
-        ]
+        next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
+
+        output_from_no_past = model(
+            next_input_ids, token_type_ids=next_token_type_ids, attention_mask=next_attention_mask
+        )["last_hidden_state"]
+        output_from_past = model(
+            next_tokens, token_type_ids=next_token_types, attention_mask=next_attention_mask, past_key_values=past
+        )["last_hidden_state"]
         self.parent.assertTrue(output_from_past.shape[1] == next_tokens.shape[1])
 
         # select random slice

tests/test_modeling_t5.py

@@ -314,18 +314,22 @@ class T5ModelTester:
     ):
         model = T5Model(config=config).get_decoder().to(torch_device).eval()
         # first forward pass
-        outputs = model(input_ids, use_cache=True)
+        outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
 
         output, past_key_values = outputs.to_tuple()
 
         # create hypothetical multiple next token and extent to next_input_ids
         next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
 
         # append to next input_ids and
         next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+        next_attention_mask = torch.cat([attention_mask, next_mask], dim=-1)
 
-        output_from_no_past = model(next_input_ids)["last_hidden_state"]
-        output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
+        output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
+        output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[
+            "last_hidden_state"
+        ]
 
         # select random slice
         random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()

tests/test_modeling_tf_bart.py

@@ -90,6 +90,7 @@ class TFBartModelTester:
         input_ids = inputs_dict["input_ids"]
 
         input_ids = input_ids[:1, :]
+        attention_mask = inputs_dict["attention_mask"][:1, :]
         self.batch_size = 1
 
         # first forward pass
@@ -100,12 +101,14 @@ class TFBartModelTester:
 
         # create hypothetical next token and extent to next_input_ids
         next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_attn_mask = tf.cast(ids_tensor((self.batch_size, 3), 2), tf.int8)
 
         # append to next input_ids and
         next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
+        next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)
 
-        output_from_no_past = model(next_input_ids)[0]
-        output_from_past = model(next_tokens, past_key_values=past_key_values)[0]
+        output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)[0]
+        output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[0]
 
         self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])
 

tests/test_modeling_tf_gpt2.py

@@ -219,20 +219,26 @@ class TFGPT2ModelTester:
         model = TFGPT2Model(config=config)
 
         # first forward pass
-        outputs = model(input_ids, token_type_ids=token_type_ids, use_cache=True)
+        outputs = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, use_cache=True)
 
         output, past = outputs.to_tuple()
 
         # create hypothetical next token and extent to next_input_ids
         next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
         next_token_types = ids_tensor((self.batch_size, 3), self.type_vocab_size)
+        next_attn_mask = ids_tensor((self.batch_size, 3), 2)
 
         # append to next input_ids and token_type_ids
         next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
         next_token_type_ids = tf.concat([token_type_ids, next_token_types], axis=-1)
-
-        output_from_no_past = model(next_input_ids, token_type_ids=next_token_type_ids)["last_hidden_state"]
-        output_from_past = model(next_tokens, token_type_ids=next_token_types, past=past)["last_hidden_state"]
+        next_attention_mask = tf.concat([input_mask, next_attn_mask], axis=-1)
+
+        output_from_no_past = model(
+            next_input_ids, token_type_ids=next_token_type_ids, attention_mask=next_attention_mask
+        )["last_hidden_state"]
+        output_from_past = model(
+            next_tokens, token_type_ids=next_token_types, attention_mask=next_attention_mask, past=past
+        )["last_hidden_state"]
         self.parent.assertTrue(output_from_past.shape[1] == next_tokens.shape[1])
 
         # select random slice

tests/test_modeling_tf_t5.py

@@ -201,19 +201,24 @@ class TFT5ModelTester:
         model = TFT5Model(config=config).get_decoder()
 
         input_ids = input_ids[:1, :]
+        attention_mask = attention_mask[:1, :]
         self.batch_size = 1
 
         # first forward pass
-        outputs = model(input_ids, use_cache=True)
+        outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
 
         # create hypothetical next token and extent to next_input_ids
         next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_attn_mask = ids_tensor((self.batch_size, 3), 2)
 
         # append to next input_ids and
         next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
+        next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)
 
-        output_from_no_past = model(next_input_ids)[0]
-        output_from_past = model(next_tokens, past_key_values=outputs.past_key_values)[0]
+        output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)[0]
+        output_from_past = model(
+            next_tokens, attention_mask=next_attention_mask, past_key_values=outputs.past_key_values
+        )[0]
 
         self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])