Early labels validation (#31240)

* Move label validation checks - fail early

* Remove some formatting changes - add back labels change wav2vec2

src/transformers/models/bark/modeling_bark.py

@@ -763,6 +763,12 @@ class BarkCausalModel(BarkPreTrainedModel):
         use_cache = use_cache if use_cache is not None else self.config.use_cache
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        loss = None
+        if labels is not None:
+            raise NotImplementedError(
+                "Training is not implemented yet for Bark - ensure you do not pass `labels` to the model."
+            )
+
         # Verify if input_embeds already exists
         # then compute embeddings.
         if input_ids is not None and input_embeds is not None:
@@ -870,12 +876,6 @@ class BarkCausalModel(BarkPreTrainedModel):
 
         logits = self.lm_head(hidden_states)
 
-        loss = None
-        if labels is not None:
-            raise NotImplementedError(
-                "Training is not implemented yet for Bark - ensure you do not pass `labels` to the model."
-            )
-
         if not return_dict:
             return tuple(
                 v for v in [None, logits, present_key_values, all_hidden_states, all_self_attentions] if v is not None
@@ -1393,6 +1393,10 @@ class BarkFineModel(BarkPreTrainedModel):
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        loss = None
+        if labels is not None:
+            raise NotImplementedError("Training is not implemented yet")
+
         if codebook_idx == 0:
             raise ValueError("Cannot predict 0th codebook - 0th codebook should be predicted by the coarse model")
 
@@ -1470,10 +1474,6 @@ class BarkFineModel(BarkPreTrainedModel):
 
         logits = self.lm_heads[codebook_idx - self.config.n_codes_given](hidden_states)
 
-        loss = None
-        if labels is not None:
-            raise NotImplementedError("Training is not implemented yet")
-
         if not return_dict:
             return tuple(v for v in [None, logits, all_hidden_states, all_self_attentions] if v is not None)
 

src/transformers/models/beit/modeling_beit.py

@@ -1247,6 +1247,9 @@ class BeitForSemanticSegmentation(BeitPreTrainedModel):
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
 
+        if labels is not None and self.config.num_labels == 1:
+            raise ValueError("The number of labels should be greater than one")
+
         outputs = self.beit(
             pixel_values,
             head_mask=head_mask,
@@ -1279,9 +1282,6 @@ class BeitForSemanticSegmentation(BeitPreTrainedModel):
 
         loss = None
         if labels is not None:
-            if self.config.num_labels == 1:
-                raise ValueError("The number of labels should be greater than one")
-            else:
             loss = self.compute_loss(logits, auxiliary_logits, labels)
 
         if not return_dict:

src/transformers/models/data2vec/modeling_data2vec_audio.py

@@ -1372,9 +1372,11 @@ class Data2VecAudioForCTC(Data2VecAudioPreTrainedModel):
             All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ...,
             config.vocab_size - 1]`.
         """
-
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        if labels is not None and labels.max() >= self.config.vocab_size:
+            raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}")
+
         outputs = self.data2vec_audio(
             input_values,
             attention_mask=attention_mask,
@@ -1390,9 +1392,6 @@ class Data2VecAudioForCTC(Data2VecAudioPreTrainedModel):
 
         loss = None
         if labels is not None:
-            if labels.max() >= self.config.vocab_size:
-                raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}")
-
             # retrieve loss input_lengths from attention_mask
             attention_mask = (
                 attention_mask if attention_mask is not None else torch.ones_like(input_values, dtype=torch.long)

src/transformers/models/data2vec/modeling_data2vec_vision.py

@@ -1173,6 +1173,9 @@ class Data2VecVisionForSemanticSegmentation(Data2VecVisionPreTrainedModel):
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
 
+        if labels is not None and self.config.num_labels == 1:
+            raise ValueError("The number of labels should be greater than one")
+
         outputs = self.data2vec_vision(
             pixel_values,
             head_mask=head_mask,
@@ -1205,9 +1208,6 @@ class Data2VecVisionForSemanticSegmentation(Data2VecVisionPreTrainedModel):
 
         loss = None
         if labels is not None:
-            if self.config.num_labels == 1:
-                raise ValueError("The number of labels should be greater than one")
-            else:
             loss = self.compute_loss(logits, auxiliary_logits, labels)
 
         if not return_dict:

src/transformers/models/data2vec/modeling_tf_data2vec_vision.py

@@ -1633,6 +1633,9 @@ class TFData2VecVisionForSemanticSegmentation(TFData2VecVisionPreTrainedModel):
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
 
+        if labels is not None and self.config.num_labels == 1:
+            raise ValueError("The number of labels should be greater than one")
+
         outputs = self.data2vec_vision(
             pixel_values,
             head_mask=head_mask,
@@ -1672,9 +1675,6 @@ class TFData2VecVisionForSemanticSegmentation(TFData2VecVisionPreTrainedModel):
 
         loss = None
         if labels is not None:
-            if self.config.num_labels == 1:
-                raise ValueError("The number of labels should be greater than one")
-            else:
             loss = self.compute_loss(logits, auxiliary_logits, labels)
 
         if not return_dict:

src/transformers/models/deprecated/mctct/modeling_mctct.py

@@ -732,6 +732,8 @@ class MCTCTForCTC(MCTCTPreTrainedModel):
             All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ...,
             config.vocab_size - 1]`.
         """
+        if labels is not None and labels.max() >= self.config.vocab_size:
+            raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}")
 
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
         outputs = self.mctct(
@@ -749,9 +751,6 @@ class MCTCTForCTC(MCTCTPreTrainedModel):
 
         loss = None
         if labels is not None:
-            if labels.max() >= self.config.vocab_size:
-                raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}")
-
             # retrieve loss input_lengths from attention_mask
             attention_mask = (
                 attention_mask

src/transformers/models/deprecated/realm/modeling_realm.py

@@ -1440,9 +1440,13 @@ class RealmKnowledgeAugEncoder(RealmPreTrainedModel):
         >>> outputs = model(**inputs)
         >>> logits = outputs.logits
         ```"""
-
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        if labels is not None and relevance_score is None:
+            raise ValueError(
+                "You have to specify `relevance_score` when `labels` is specified in order to compute loss."
+            )
+
         (flattened_input_ids, flattened_attention_mask, flattened_token_type_ids) = self._flatten_inputs(
             input_ids, attention_mask, token_type_ids
         )
@@ -1468,11 +1472,6 @@ class RealmKnowledgeAugEncoder(RealmPreTrainedModel):
 
         masked_lm_loss = None
         if labels is not None:
-            if candidate_score is None:
-                raise ValueError(
-                    "You have to specify `relevance_score` when `labels` is specified in order to compute loss."
-                )
-
             batch_size, seq_length = labels.size()
 
             if mlm_mask is None:

src/transformers/models/depth_anything/modeling_depth_anything.py

@@ -424,6 +424,10 @@ class DepthAnythingForDepthEstimation(DepthAnythingPreTrainedModel):
         >>> formatted = (output * 255 / np.max(output)).astype("uint8")
         >>> depth = Image.fromarray(formatted)
         ```"""
+        loss = None
+        if labels is not None:
+            raise NotImplementedError("Training is not implemented yet")
+
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
@@ -444,10 +448,6 @@ class DepthAnythingForDepthEstimation(DepthAnythingPreTrainedModel):
 
         predicted_depth = self.head(hidden_states, patch_height, patch_width)
 
-        loss = None
-        if labels is not None:
-            raise NotImplementedError("Training is not implemented yet")
-
         if not return_dict:
             if output_hidden_states:
                 output = (predicted_depth,) + outputs[1:]

src/transformers/models/dpt/modeling_dpt.py

@@ -1136,6 +1136,10 @@ class DPTForDepthEstimation(DPTPreTrainedModel):
         >>> formatted = (output * 255 / np.max(output)).astype("uint8")
         >>> depth = Image.fromarray(formatted)
         ```"""
+        loss = None
+        if labels is not None:
+            raise NotImplementedError("Training is not implemented yet")
+
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
@@ -1183,10 +1187,6 @@ class DPTForDepthEstimation(DPTPreTrainedModel):
 
         predicted_depth = self.head(hidden_states)
 
-        loss = None
-        if labels is not None:
-            raise NotImplementedError("Training is not implemented yet")
-
         if not return_dict:
             if output_hidden_states:
                 output = (predicted_depth,) + outputs[1:]
@@ -1308,6 +1308,9 @@ class DPTForSemanticSegmentation(DPTPreTrainedModel):
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
 
+        if labels is not None and self.config.num_labels == 1:
+            raise ValueError("The number of labels should be greater than one")
+
         outputs = self.dpt(
             pixel_values,
             head_mask=head_mask,
@@ -1342,9 +1345,6 @@ class DPTForSemanticSegmentation(DPTPreTrainedModel):
 
         loss = None
         if labels is not None:
-            if self.config.num_labels == 1:
-                raise ValueError("The number of labels should be greater than one")
-            else:
             # upsample logits to the images' original size
             upsampled_logits = nn.functional.interpolate(
                 logits, size=labels.shape[-2:], mode="bilinear", align_corners=False

src/transformers/models/gptj/modeling_tf_gptj.py

@@ -921,6 +921,8 @@ class TFGPTJForSequenceClassification(TFGPTJPreTrainedModel, TFSequenceClassific
             config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
             `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
         """
+        if labels is not None and self.config.pad_token_id is None and input_ids.shape[0] != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
 
         transformer_outputs = self.transformer(
             input_ids=input_ids,
@@ -963,9 +965,6 @@ class TFGPTJForSequenceClassification(TFGPTJPreTrainedModel, TFSequenceClassific
         loss = None
 
         if labels is not None:
-            if self.config.pad_token_id is None and logits_shape[0] != 1:
-                raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
-
             if not tf.is_tensor(sequence_lengths):
                 in_logits = logits[0 : logits_shape[0], sequence_lengths]
 

src/transformers/models/hubert/modeling_hubert.py

@@ -1574,9 +1574,11 @@ class HubertForCTC(HubertPreTrainedModel):
             All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ...,
             config.vocab_size - 1]`.
         """
-
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        if labels is not None and labels.max() >= self.config.vocab_size:
+            raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}")
+
         outputs = self.hubert(
             input_values,
             attention_mask=attention_mask,
@@ -1592,9 +1594,6 @@ class HubertForCTC(HubertPreTrainedModel):
 
         loss = None
         if labels is not None:
-            if labels.max() >= self.config.vocab_size:
-                raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}")
-
             # retrieve loss input_lengths from attention_mask
             attention_mask = (
                 attention_mask if attention_mask is not None else torch.ones_like(input_values, dtype=torch.long)

src/transformers/models/hubert/modeling_tf_hubert.py

@@ -1600,6 +1600,8 @@ class TFHubertForCTC(TFHubertPreTrainedModel):
 
         >>> loss = model(input_values, labels=labels).loss
         ```"""
+        if labels is not None and tf.reduce_max(labels) >= self.config.vocab_size:
+            raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}")
 
         outputs = self.hubert(
             input_values=input_values,
@@ -1619,9 +1621,6 @@ class TFHubertForCTC(TFHubertPreTrainedModel):
         logits = self.lm_head(hidden_states)
 
         if labels is not None:
-            if tf.reduce_max(labels) >= self.config.vocab_size:
-                raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}")
-
             attention_mask = (
                 attention_mask if attention_mask is not None else tf.ones_like(input_values, dtype=tf.float32)
             )

src/transformers/models/mobilenet_v2/modeling_mobilenet_v2.py

@@ -822,6 +822,9 @@ class MobileNetV2ForSemanticSegmentation(MobileNetV2PreTrainedModel):
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        if labels is not None and self.config.num_labels == 1:
+            raise ValueError("The number of labels should be greater than one")
+
         outputs = self.mobilenet_v2(
             pixel_values,
             output_hidden_states=True,  # we need the intermediate hidden states
@@ -834,9 +837,6 @@ class MobileNetV2ForSemanticSegmentation(MobileNetV2PreTrainedModel):
 
         loss = None
         if labels is not None:
-            if self.config.num_labels == 1:
-                raise ValueError("The number of labels should be greater than one")
-            else:
             # upsample logits to the images' original size
             upsampled_logits = nn.functional.interpolate(
                 logits, size=labels.shape[-2:], mode="bilinear", align_corners=False

src/transformers/models/mobilevit/modeling_mobilevit.py

@@ -1026,6 +1026,9 @@ class MobileViTForSemanticSegmentation(MobileViTPreTrainedModel):
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        if labels is not None and self.config.num_labels == 1:
+            raise ValueError("The number of labels should be greater than one")
+
         outputs = self.mobilevit(
             pixel_values,
             output_hidden_states=True,  # we need the intermediate hidden states
@@ -1038,9 +1041,6 @@ class MobileViTForSemanticSegmentation(MobileViTPreTrainedModel):
 
         loss = None
         if labels is not None:
-            if self.config.num_labels == 1:
-                raise ValueError("The number of labels should be greater than one")
-            else:
             # upsample logits to the images' original size
             upsampled_logits = nn.functional.interpolate(
                 logits, size=labels.shape[-2:], mode="bilinear", align_corners=False

src/transformers/models/mobilevit/modeling_tf_mobilevit.py

@@ -1323,6 +1323,9 @@ class TFMobileViTForSemanticSegmentation(TFMobileViTPreTrainedModel):
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        if labels is not None and not self.config.num_labels > 1:
+            raise ValueError("The number of labels should be greater than one")
+
         outputs = self.mobilevit(
             pixel_values,
             output_hidden_states=True,  # we need the intermediate hidden states
@@ -1336,9 +1339,6 @@ class TFMobileViTForSemanticSegmentation(TFMobileViTPreTrainedModel):
 
         loss = None
         if labels is not None:
-            if not self.config.num_labels > 1:
-                raise ValueError("The number of labels should be greater than one")
-            else:
             loss = self.hf_compute_loss(logits=logits, labels=labels)
 
         # make logits of shape (batch_size, num_labels, height, width) to

src/transformers/models/mobilevitv2/modeling_mobilevitv2.py

@@ -990,6 +990,9 @@ class MobileViTV2ForSemanticSegmentation(MobileViTV2PreTrainedModel):
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        if labels is not None and self.config.num_labels == 1:
+            raise ValueError("The number of labels should be greater than one")
+
         outputs = self.mobilevitv2(
             pixel_values,
             output_hidden_states=True,  # we need the intermediate hidden states
@@ -1002,9 +1005,6 @@ class MobileViTV2ForSemanticSegmentation(MobileViTV2PreTrainedModel):
 
         loss = None
         if labels is not None:
-            if self.config.num_labels == 1:
-                raise ValueError("The number of labels should be greater than one")
-            else:
             # upsample logits to the images' original size
             upsampled_logits = nn.functional.interpolate(
                 logits, size=labels.shape[-2:], mode="bilinear", align_corners=False

src/transformers/models/perceiver/modeling_perceiver.py

@@ -1740,6 +1740,10 @@ class PerceiverForOpticalFlow(PerceiverPreTrainedModel):
         ```"""
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        loss = None
+        if labels is not None:
+            raise NotImplementedError("Optical flow training is not yet supported")
+
         outputs = self.perceiver(
             inputs=inputs,
             attention_mask=attention_mask,
@@ -1750,10 +1754,6 @@ class PerceiverForOpticalFlow(PerceiverPreTrainedModel):
         )
         logits = outputs.logits if return_dict else outputs[0]
 
-        loss = None
-        if labels is not None:
-            raise NotImplementedError("Optical flow training is not yet supported")
-
         if not return_dict:
             output = (logits,) + outputs[2:]
             return ((loss,) + output) if loss is not None else output
@@ -1974,6 +1974,10 @@ class PerceiverForMultimodalAutoencoding(PerceiverPreTrainedModel):
         ```"""
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        loss = None
+        if labels is not None:
+            raise NotImplementedError("Multimodal autoencoding training is not yet supported")
+
         outputs = self.perceiver(
             inputs=inputs,
             attention_mask=attention_mask,
@@ -1985,10 +1989,6 @@ class PerceiverForMultimodalAutoencoding(PerceiverPreTrainedModel):
         )
         logits = outputs.logits if return_dict else outputs[0]
 
-        loss = None
-        if labels is not None:
-            raise NotImplementedError("Multimodal autoencoding training is not yet supported")
-
         if not return_dict:
             output = (logits,) + outputs[2:]
             return ((loss,) + output) if loss is not None else output

src/transformers/models/segformer/modeling_segformer.py

@@ -784,6 +784,9 @@ class SegformerForSemanticSegmentation(SegformerPreTrainedModel):
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
 
+        if labels is not None and self.config.num_labels < 1:
+            raise ValueError(f"Number of labels should be >=0: {self.config.num_labels}")
+
         outputs = self.segformer(
             pixel_values,
             output_attentions=output_attentions,
@@ -809,8 +812,6 @@ class SegformerForSemanticSegmentation(SegformerPreTrainedModel):
                 loss_fct = BCEWithLogitsLoss(reduction="none")
                 loss = loss_fct(upsampled_logits.squeeze(1), labels.float())
                 loss = (loss * valid_mask).mean()
-            else:
-                raise ValueError(f"Number of labels should be >=0: {self.config.num_labels}")
 
         if not return_dict:
             if output_hidden_states:

src/transformers/models/segformer/modeling_tf_segformer.py

@@ -988,6 +988,9 @@ class TFSegformerForSemanticSegmentation(TFSegformerPreTrainedModel):
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
 
+        if labels is not None and not self.config.num_labels > 1:
+            raise ValueError("The number of labels should be greater than one")
+
         outputs = self.segformer(
             pixel_values,
             output_attentions=output_attentions,
@@ -1001,9 +1004,6 @@ class TFSegformerForSemanticSegmentation(TFSegformerPreTrainedModel):
 
         loss = None
         if labels is not None:
-            if not self.config.num_labels > 1:
-                raise ValueError("The number of labels should be greater than one")
-            else:
             loss = self.hf_compute_loss(logits=logits, labels=labels)
 
         # make logits of shape (batch_size, num_labels, height, width) to

src/transformers/models/sew/modeling_sew.py

@@ -1418,9 +1418,11 @@ class SEWForCTC(SEWPreTrainedModel):
             All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ...,
             config.vocab_size - 1]`.
         """
-
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        if labels is not None and labels.max() >= self.config.vocab_size:
+            raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}")
+
         outputs = self.sew(
             input_values,
             attention_mask=attention_mask,
@@ -1436,9 +1438,6 @@ class SEWForCTC(SEWPreTrainedModel):
 
         loss = None
         if labels is not None:
-            if labels.max() >= self.config.vocab_size:
-                raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}")
-
             # retrieve loss input_lengths from attention_mask
             attention_mask = (
                 attention_mask if attention_mask is not None else torch.ones_like(input_values, dtype=torch.long)