Add OnnxConfig for SqueezeBert iss17314 (#17315)

* add onnx config for SqueezeBert

* add test for onnx config for SqueezeBert

* add automatically updated doc for onnx config for SqueezeBert

* Update src/transformers/onnx/features.py

Co-authored-by: lewtun <lewis.c.tunstall@gmail.com>

* Update src/transformers/models/squeezebert/configuration_squeezebert.py

Co-authored-by: lewtun <lewis.c.tunstall@gmail.com>

Co-authored-by: lewtun <lewis.c.tunstall@gmail.com>

docs/source/en/serialization.mdx

@@ -73,6 +73,7 @@ Ready-made configurations include the following architectures:
 - PLBart
 - RoBERTa
 - RoFormer
+- SqueezeBERT
 - T5
 - ViT
 - XLM

src/transformers/models/squeezebert/__init__.py

@@ -22,7 +22,11 @@ from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_
 
 
 _import_structure = {
-    "configuration_squeezebert": ["SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "SqueezeBertConfig"],
+    "configuration_squeezebert": [
+        "SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "SqueezeBertConfig",
+        "SqueezeBertOnnxConfig",
+    ],
     "tokenization_squeezebert": ["SqueezeBertTokenizer"],
 }
 
@@ -54,7 +58,11 @@ else:
 
 
 if TYPE_CHECKING:
-    from .configuration_squeezebert import SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, SqueezeBertConfig
+    from .configuration_squeezebert import (
+        SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        SqueezeBertConfig,
+        SqueezeBertOnnxConfig,
+    )
     from .tokenization_squeezebert import SqueezeBertTokenizer
 
     try:

src/transformers/models/squeezebert/configuration_squeezebert.py

@@ -13,8 +13,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """ SqueezeBERT model configuration"""
+from collections import OrderedDict
+from typing import Mapping
 
 from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfig
 from ...utils import logging
 
 
@@ -154,3 +157,20 @@ class SqueezeBertConfig(PretrainedConfig):
         self.post_attention_groups = post_attention_groups
         self.intermediate_groups = intermediate_groups
         self.output_groups = output_groups
+
+
+# # Copied from transformers.models.bert.configuration_bert.BertOnxxConfig with Bert->SqueezeBert
+class SqueezeBertOnnxConfig(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

@@ -28,6 +28,7 @@ from ..models.mbart import MBartOnnxConfig
 from ..models.mobilebert import MobileBertOnnxConfig
 from ..models.roberta import RobertaOnnxConfig
 from ..models.roformer import RoFormerOnnxConfig
+from ..models.squeezebert import SqueezeBertOnnxConfig
 from ..models.t5 import T5OnnxConfig
 from ..models.vit import ViTOnnxConfig
 from ..models.xlm import XLMOnnxConfig
@@ -352,6 +353,15 @@ class FeaturesManager:
             "token-classification",
             onnx_config_cls=RoFormerOnnxConfig,
         ),
+        "squeezebert": supported_features_mapping(
+            "default",
+            "masked-lm",
+            "sequence-classification",
+            "multiple-choice",
+            "token-classification",
+            "question-answering",
+            onnx_config_cls=SqueezeBertOnnxConfig,
+        ),
         "t5": supported_features_mapping(
             "default", "default-with-past", "seq2seq-lm", "seq2seq-lm-with-past", onnx_config_cls=T5OnnxConfig
         ),

tests/onnx/test_onnx_v2.py

@@ -180,6 +180,7 @@ PYTORCH_EXPORT_MODELS = {
     ("electra", "google/electra-base-generator"),
     ("roberta", "roberta-base"),
     ("roformer", "junnyu/roformer_chinese_base"),
+    ("squeezebert", "squeezebert/squeezebert-uncased"),
     ("mobilebert", "google/mobilebert-uncased"),
     ("xlm", "xlm-clm-ende-1024"),
     ("xlm-roberta", "xlm-roberta-base"),