add mobilebert onnx configs (huggingface#17029)

* update docs of length_penalty

* Revert "update docs of length_penalty"

This reverts commit 466bf48.

* add mobilebert onnx config

* address suggestions

* Update auto.mdx

* Update __init__.py

* Update features.py

docs/source/en/model_doc/auto.mdx

@@ -194,6 +194,10 @@ Likewise, if your `NewModel` is a subclass of [`PreTrainedModel`], make sure its
 
 [[autodoc]] TFAutoModelForMultipleChoice
 
+## TFAutoModelForNextSentencePrediction
+
+[[autodoc]] TFAutoModelForNextSentencePrediction
+
 ## TFAutoModelForTableQuestionAnswering
 
 [[autodoc]] TFAutoModelForTableQuestionAnswering

docs/source/en/serialization.mdx

@@ -68,6 +68,7 @@ Ready-made configurations include the following architectures:
 - M2M100
 - Marian
 - mBART
+- MobileBert
 - OpenAI GPT-2
 - PLBart
 - RoBERTa

src/transformers/__init__.py

@@ -1798,6 +1798,7 @@
             "TFAutoModelForSeq2SeqLM",
             "TFAutoModelForSequenceClassification",
             "TFAutoModelForSpeechSeq2Seq",
+            "TFAutoModelForNextSentencePrediction",
             "TFAutoModelForTableQuestionAnswering",
             "TFAutoModelForTokenClassification",
             "TFAutoModelForVision2Seq",
@@ -3964,6 +3965,7 @@
             TFAutoModelForImageClassification,
             TFAutoModelForMaskedLM,
             TFAutoModelForMultipleChoice,
+            TFAutoModelForNextSentencePrediction,
             TFAutoModelForPreTraining,
             TFAutoModelForQuestionAnswering,
             TFAutoModelForSeq2SeqLM,

src/transformers/models/auto/__init__.py

@@ -108,6 +108,7 @@
         "TFAutoModelForSeq2SeqLM",
         "TFAutoModelForSequenceClassification",
         "TFAutoModelForSpeechSeq2Seq",
+        "TFAutoModelForNextSentencePrediction",
         "TFAutoModelForTableQuestionAnswering",
         "TFAutoModelForTokenClassification",
         "TFAutoModelForVision2Seq",
@@ -224,6 +225,7 @@
             TFAutoModelForImageClassification,
             TFAutoModelForMaskedLM,
             TFAutoModelForMultipleChoice,
+            TFAutoModelForNextSentencePrediction,
             TFAutoModelForPreTraining,
             TFAutoModelForQuestionAnswering,
             TFAutoModelForSeq2SeqLM,

src/transformers/models/mobilebert/__init__.py

@@ -22,7 +22,11 @@
 
 
 _import_structure = {
-    "configuration_mobilebert": ["MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "MobileBertConfig"],
+    "configuration_mobilebert": [
+        "MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "MobileBertConfig",
+        "MobileBertOnnxConfig",
+    ],
     "tokenization_mobilebert": ["MobileBertTokenizer"],
 }
 
@@ -62,7 +66,11 @@
 
 
 if TYPE_CHECKING:
-    from .configuration_mobilebert import MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, MobileBertConfig
+    from .configuration_mobilebert import (
+        MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        MobileBertConfig,
+        MobileBertOnnxConfig,
+    )
     from .tokenization_mobilebert import MobileBertTokenizer
 
     if is_tokenizers_available():

src/transformers/models/mobilebert/configuration_mobilebert.py

@@ -13,8 +13,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """ MobileBERT model configuration"""
+from collections import OrderedDict
+from typing import Mapping
 
 from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfig
 from ...utils import logging
 
 
@@ -165,3 +168,20 @@ def __init__(
             self.true_hidden_size = hidden_size
 
         self.classifier_dropout = classifier_dropout
+
+
+# Copied from transformers.models.bert.configuration_bert.BertOnnxConfig with Bert->MobileBert
+class MobileBertOnnxConfig(OnnxConfig):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        if self.task == "multiple-choice":
+            dynamic_axis = {0: "batch", 1: "choice", 2: "sequence"}
+        else:
+            dynamic_axis = {0: "batch", 1: "sequence"}
+        return OrderedDict(
+            [
+                ("input_ids", dynamic_axis),
+                ("attention_mask", dynamic_axis),
+                ("token_type_ids", dynamic_axis),
+            ]
+        )

src/transformers/onnx/features.py

@@ -25,6 +25,7 @@
 from ..models.m2m_100 import M2M100OnnxConfig
 from ..models.marian import MarianOnnxConfig
 from ..models.mbart import MBartOnnxConfig
+from ..models.mobilebert import MobileBertOnnxConfig
 from ..models.roberta import RobertaOnnxConfig
 from ..models.roformer import RoFormerOnnxConfig
 from ..models.t5 import T5OnnxConfig
@@ -44,6 +45,7 @@
         AutoModelForMaskedImageModeling,
         AutoModelForMaskedLM,
         AutoModelForMultipleChoice,
+        AutoModelForNextSentencePrediction,
         AutoModelForQuestionAnswering,
         AutoModelForSeq2SeqLM,
         AutoModelForSequenceClassification,
@@ -55,6 +57,7 @@
         TFAutoModelForCausalLM,
         TFAutoModelForMaskedLM,
         TFAutoModelForMultipleChoice,
+        TFAutoModelForNextSentencePrediction,
         TFAutoModelForQuestionAnswering,
         TFAutoModelForSeq2SeqLM,
         TFAutoModelForSequenceClassification,
@@ -108,6 +111,7 @@ class FeaturesManager:
             "question-answering": AutoModelForQuestionAnswering,
             "image-classification": AutoModelForImageClassification,
             "masked-im": AutoModelForMaskedImageModeling,
+            "next-sentence-prediction": AutoModelForNextSentencePrediction,
         }
     if is_tf_available():
         _TASKS_TO_TF_AUTOMODELS = {
@@ -119,6 +123,7 @@ class FeaturesManager:
             "token-classification": TFAutoModelForTokenClassification,
             "multiple-choice": TFAutoModelForMultipleChoice,
             "question-answering": TFAutoModelForQuestionAnswering,
+            "next-sentence-prediction": TFAutoModelForNextSentencePrediction,
         }
 
     # Set of model topologies we support associated to the features supported by each topology and the factory
@@ -153,6 +158,7 @@ class FeaturesManager:
             "multiple-choice",
             "token-classification",
             "question-answering",
+            "next-sentence-prediction",
             onnx_config_cls=BertOnnxConfig,
         ),
         "big-bird": supported_features_mapping(
@@ -316,6 +322,16 @@ class FeaturesManager:
             "question-answering",
             onnx_config_cls=MBartOnnxConfig,
         ),
+        "mobilebert": supported_features_mapping(
+            "default",
+            "masked-lm",
+            "next-sentence-prediction",
+            "sequence-classification",
+            "multiple-choice",
+            "token-classification",
+            "question-answering",
+            onnx_config_cls=MobileBertOnnxConfig,
+        ),
         "m2m-100": supported_features_mapping(
             "default", "default-with-past", "seq2seq-lm", "seq2seq-lm-with-past", onnx_config_cls=M2M100OnnxConfig
         ),

tests/onnx/test_onnx_v2.py

@@ -180,6 +180,7 @@ def test_values_override(self):
     ("electra", "google/electra-base-generator"),
     ("roberta", "roberta-base"),
     ("roformer", "junnyu/roformer_chinese_base"),
+    ("mobilebert", "google/mobilebert-uncased"),
     ("xlm-roberta", "xlm-roberta-base"),
     ("layoutlm", "microsoft/layoutlm-base-uncased"),
     ("vit", "google/vit-base-patch16-224"),