[Wav2Vec2] Correctly pad mask indices for PreTraining (#12748)

* fix_torch_device_generate_test

* remove @

* start adding tests

* correct wav2vec2 pretraining

* up

* up

Co-authored-by: Patrick von Platen <patrick@huggingface.co>

examples/research_projects/jax-projects/wav2vec2/run_wav2vec2_pretrain_flax.py

@@ -174,11 +174,23 @@ class FlaxDataCollatorForWav2Vec2Pretraining:
         )
         mask_indices_seq_length = self.model._get_feat_extract_output_lengths(batch["input_values"].shape[-1])
 
+        batch_size = batch["input_values"].shape[0]
+
+        if batch["attention_mask"] is not None:
+            output_lengths = self.model._get_feat_extract_output_lengths(batch["attention_mask"].sum(-1))
+            attention_mask = np.zeros((batch_size, mask_indices_seq_length), dtype=np.int8)
+
+            # these two operations makes sure that all values
+            # before the output lengths indices are attended to
+            attention_mask[(np.arange(attention_mask.shape[0]), output_lengths - 1)] = 1
+            attention_mask = jnp.flip(jnp.flip(attention_mask, -1).cumsum(-1), -1).astype("bool")
+
         # sample randomly masked indices
         batch["mask_time_indices"] = _compute_mask_indices(
-            (batch["input_values"].shape[0], mask_indices_seq_length),
+            (batch_size, mask_indices_seq_length),
             self.model.config.mask_time_prob,
             self.model.config.mask_time_length,
+            attention_mask=attention_mask,
             min_masks=2,
         )
 

examples/research_projects/wav2vec2/run_pretrain.py

@@ -172,12 +172,33 @@ class DataCollatorForWav2Vec2Pretraining:
         )
         mask_indices_seq_length = self.model._get_feat_extract_output_lengths(batch["input_values"].shape[-1])
 
+        batch_size = batch["input_values"].shape[0]
+
+        # make sure that no loss is computed on padded inputs
+        if batch["attention_mask"] is not None:
+            # compute real output lengths according to convolution formula
+            output_lengths = self.model._get_feat_extract_output_lengths(batch["attention_mask"].sum(-1)).to(
+                torch.long
+            )
+
+            attention_mask = torch.zeros(
+                (batch_size, mask_indices_seq_length), dtype=torch.long, device=batch["input_values"].device
+            )
+
+            # these two operations makes sure that all values
+            # before the output lengths indices are attended to
+            attention_mask[
+                (torch.arange(attention_mask.shape[0], device=batch["input_values"].device), output_lengths - 1)
+            ] = 1
+            attention_mask = attention_mask.flip([-1]).cumsum(-1).flip([-1]).bool()
+
         # sample randomly masked indices
         batch["mask_time_indices"] = _compute_mask_indices(
-            (batch["input_values"].shape[0], mask_indices_seq_length),
+            (batch_size, mask_indices_seq_length),
             self.model.config.mask_time_prob,
             self.model.config.mask_time_length,
             device=batch["input_values"].device,
+            attention_mask=attention_mask,
             min_masks=2,
         )
 

src/transformers/models/hubert/modeling_hubert.py

@@ -47,6 +47,7 @@ def _compute_mask_indices(
     mask_prob: float,
     mask_length: int,
     device: torch.device,
+    attention_mask: Optional[torch.tensor] = None,
     min_masks: int = 0,
 ) -> torch.tensor:
     """
@@ -813,7 +814,10 @@ class HubertModel(HubertPreTrainedModel):
 
     # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2Model._mask_hidden_states
     def _mask_hidden_states(
-        self, hidden_states: torch.FloatTensor, mask_time_indices: Optional[torch.FloatTensor] = None
+        self,
+        hidden_states: torch.FloatTensor,
+        mask_time_indices: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
     ):
         """
         Masks extracted features along time axis and/or along feature axis according to `SpecAugment
@@ -836,6 +840,7 @@ class HubertModel(HubertPreTrainedModel):
                 mask_prob=self.config.mask_time_prob,
                 mask_length=self.config.mask_time_length,
                 device=hidden_states.device,
+                attention_mask=attention_mask,
                 min_masks=2,
             )
             hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
@@ -847,6 +852,7 @@ class HubertModel(HubertPreTrainedModel):
                 mask_prob=self.config.mask_feature_prob,
                 mask_length=self.config.mask_feature_length,
                 device=hidden_states.device,
+                attention_mask=attention_mask,
             )
             hidden_states[mask_feature_indices[:, None].expand(-1, sequence_length, -1)] = 0
 

src/transformers/models/wav2vec2/modeling_flax_wav2vec2.py

@@ -107,6 +107,7 @@ def _compute_mask_indices(
     shape: Tuple[int, int],
     mask_prob: float,
     mask_length: int,
+    attention_mask: Optional[np.ndarray] = None,
     min_masks: int = 0,
 ) -> np.ndarray:
     """
@@ -166,6 +167,10 @@ def _compute_mask_indices(
     # scatter indices to mask
     np.put_along_axis(spec_aug_mask, spec_aug_mask_idxs, 1, -1)
 
+    if attention_mask is not None:
+        # make sure padded input ids cannot be masked
+        spec_aug_mask = np.where(attention_mask, spec_aug_mask, False)
+
     return spec_aug_mask
 
 
@@ -873,6 +878,7 @@ class FlaxWav2Vec2Module(nn.Module):
         """
         extract_features = self.feature_extractor(input_values)
 
+        # make sure that no loss is computed on padded inputs
         if attention_mask is not None:
             # compute real output lengths according to convolution formula
             output_lengths = self._get_feat_extract_output_lengths(attention_mask.sum(-1).astype("i4"))

src/transformers/models/wav2vec2/modeling_wav2vec2.py

@@ -120,6 +120,7 @@ def _compute_mask_indices(
     mask_prob: float,
     mask_length: int,
     device: torch.device,
+    attention_mask: Optional[torch.tensor] = None,
     min_masks: int = 0,
 ) -> torch.tensor:
     """
@@ -179,6 +180,10 @@ def _compute_mask_indices(
     # scatter indices to mask
     spec_aug_mask = spec_aug_mask.scatter(1, spec_aug_mask_idxs, True)
 
+    if attention_mask is not None:
+        # make sure padded input ids cannot be masked
+        spec_aug_mask = torch.where(attention_mask.bool(), spec_aug_mask, False)
+
     return spec_aug_mask
 
 
@@ -950,7 +955,10 @@ class Wav2Vec2Model(Wav2Vec2PreTrainedModel):
         self.init_weights()
 
     def _mask_hidden_states(
-        self, hidden_states: torch.FloatTensor, mask_time_indices: Optional[torch.FloatTensor] = None
+        self,
+        hidden_states: torch.FloatTensor,
+        mask_time_indices: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
     ):
         """
         Masks extracted features along time axis and/or along feature axis according to `SpecAugment
@@ -973,6 +981,7 @@ class Wav2Vec2Model(Wav2Vec2PreTrainedModel):
                 mask_prob=self.config.mask_time_prob,
                 mask_length=self.config.mask_time_length,
                 device=hidden_states.device,
+                attention_mask=attention_mask,
                 min_masks=2,
             )
             hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
@@ -984,6 +993,7 @@ class Wav2Vec2Model(Wav2Vec2PreTrainedModel):
                 mask_prob=self.config.mask_feature_prob,
                 mask_length=self.config.mask_feature_length,
                 device=hidden_states.device,
+                attention_mask=attention_mask,
             )
             hidden_states[mask_feature_indices[:, None].expand(-1, sequence_length, -1)] = 0
 
@@ -1049,7 +1059,9 @@ class Wav2Vec2Model(Wav2Vec2PreTrainedModel):
             attention_mask = attention_mask.flip([-1]).cumsum(-1).flip([-1]).bool()
 
         hidden_states, extract_features = self.feature_projection(extract_features)
-        hidden_states = self._mask_hidden_states(hidden_states, mask_time_indices=mask_time_indices)
+        hidden_states = self._mask_hidden_states(
+            hidden_states, mask_time_indices=mask_time_indices, attention_mask=attention_mask
+        )
 
         encoder_outputs = self.encoder(
             hidden_states,

tests/test_modeling_flax_wav2vec2.py

@@ -245,6 +245,24 @@ class FlaxWav2Vec2UtilsTest(unittest.TestCase):
         for batch_sum in mask.sum(axis=-1):
             self.assertTrue(int(batch_sum) <= mask_prob * sequence_length)
 
+    def test_compute_mask_indices_attn_mask_overlap(self):
+        batch_size = 4
+        sequence_length = 80
+        mask_prob = 0.5
+        mask_length = 4
+
+        attention_mask = np.ones((batch_size, sequence_length), dtype=np.int32)
+        attention_mask[:2, sequence_length // 2 :] = 0
+
+        mask = _compute_mask_indices(
+            (batch_size, sequence_length), mask_prob, mask_length, attention_mask=attention_mask
+        )
+
+        for batch_sum in mask.sum(axis=-1):
+            self.assertTrue(int(batch_sum) <= mask_prob * sequence_length)
+
+        self.assertTrue(mask[:2, sequence_length // 2 :].sum() == 0)
+
     def test_compute_perplexity(self):
         probs = np.arange(100).reshape(2, 5, 10) / 100
 

tests/test_modeling_wav2vec2.py

@@ -580,6 +580,24 @@ class Wav2Vec2UtilsTest(unittest.TestCase):
         for batch_sum in mask.sum(axis=-1):
             self.assertTrue(int(batch_sum) <= mask_prob * sequence_length)
 
+    def test_compute_mask_indices_attn_mask_overlap(self):
+        batch_size = 4
+        sequence_length = 80
+        mask_prob = 0.5
+        mask_length = 4
+
+        attention_mask = torch.ones((batch_size, sequence_length), dtype=torch.long, device=torch_device)
+        attention_mask[:2, sequence_length // 2 :] = 0
+
+        mask = _compute_mask_indices(
+            (batch_size, sequence_length), mask_prob, mask_length, device=torch_device, attention_mask=attention_mask
+        )
+
+        for batch_sum in mask.sum(axis=-1):
+            self.assertTrue(int(batch_sum) <= mask_prob * sequence_length)
+
+        self.assertTrue(mask[:2, sequence_length // 2 :].sum() == 0)
+
     def test_compute_perplexity(self):
         probs = torch.arange(100, device=torch_device).reshape(2, 5, 10) / 100