Export to ONNX

If you need to deploy 🤗 Transformers models in production environments, we recommend exporting them to a serialized format that can be loaded and executed on specialized runtimes and hardware. In this guide, we’ll show you how to export 🤗 Transformers models to ONNX (Open Neural Network eXchange).

ONNX is an open standard that defines a common set of operators and a common file format to represent deep learning models in a wide variety of frameworks, including PyTorch and TensorFlow. When a model is exported to the ONNX format, these operators are used to construct a computational graph (often called an intermediate representation) which represents the flow of data through the neural network.

By exposing a graph with standardized operators and data types, ONNX makes it easy to switch between frameworks. For example, a model trained in PyTorch can be exported to ONNX format and then imported in TensorFlow (and vice versa).

🤗 Transformers provides a transformers.onnx package that enables you to convert model checkpoints to an ONNX graph by leveraging configuration objects. These configuration objects come ready made for a number of model architectures, and are designed to be easily extendable to other architectures.

Ready-made configurations include the following architectures:

• ALBERT
• BART
• BEiT
• BERT
• BigBird
• BigBird-Pegasus
• Blenderbot
• BlenderbotSmall
• BLOOM
• CamemBERT
• Chinese-CLIP
• CLIP
• CodeGen
• Conditional DETR
• ConvBERT
• ConvNeXT
• Data2VecText
• Data2VecVision
• DeBERTa
• DeBERTa-v2
• DeiT
• DETR
• DistilBERT
• EfficientNet
• ELECTRA
• ERNIE
• FlauBERT
• GPT Neo
• GPT-J
• GPT-Sw3
• GroupViT
• I-BERT
• ImageGPT
• LayoutLM
• LayoutLMv3
• LeViT
• Longformer
• LongT5
• M2M100
• Marian
• mBART
• MobileBERT
• MobileNetV1
• MobileNetV2
• MobileViT
• MT5
• OpenAI GPT-2
• OWL-ViT
• Perceiver
• PLBart
• PoolFormer
• RemBERT
• ResNet
• RoBERTa
• RoBERTa-PreLayerNorm
• RoFormer
• SegFormer
• SqueezeBERT
• Swin Transformer
• T5
• Table Transformer
• Vision Encoder decoder
• ViT
• Whisper
• X-MOD
• XLM
• XLM-RoBERTa
• XLM-RoBERTa-XL
• YOLOS

In the next two sections, we’ll show you how to:

• Export a supported model using the transformers.onnx package.
• Export a custom model for an unsupported architecture.

Exporting a model to ONNX

To export a 🤗 Transformers model to ONNX, you’ll first need to install some extra dependencies:

pip install transformers[onnx]

The transformers.onnx package can then be used as a Python module:

python -m transformers.onnx --help

usage: Hugging Face Transformers ONNX exporter [-h] -m MODEL [--feature {causal-lm, ...}] [--opset OPSET] [--atol ATOL] output

positional arguments:
  output                Path indicating where to store generated ONNX model.

optional arguments:
  -h, --help            show this help message and exit
  -m MODEL, --model MODEL
                        Model ID on huggingface.co or path on disk to load model from.
  --feature {causal-lm, ...}
                        The type of features to export the model with.
  --opset OPSET         ONNX opset version to export the model with.
  --atol ATOL           Absolute difference tolerance when validating the model.

Exporting a checkpoint using a ready-made configuration can be done as follows:

python -m transformers.onnx --model=distilbert-base-uncased onnx/

You should see the following logs:

Validating ONNX model...
        -[✓] ONNX model output names match reference model ({'last_hidden_state'})
        - Validating ONNX Model output "last_hidden_state":
                -[✓] (2, 8, 768) matches (2, 8, 768)
                -[✓] all values close (atol: 1e-05)
All good, model saved at: onnx/model.onnx

This exports an ONNX graph of the checkpoint defined by the --model argument. In this example, it is distilbert-base-uncased, but it can be any checkpoint on the Hugging Face Hub or one that’s stored locally.

The resulting model.onnx file can then be run on one of the many accelerators that support the ONNX standard. For example, we can load and run the model with ONNX Runtime as follows:

>>> from transformers import AutoTokenizer
>>> from onnxruntime import InferenceSession

>>> tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
>>> session = InferenceSession("onnx/model.onnx")
>>> # ONNX Runtime expects NumPy arrays as input
>>> inputs = tokenizer("Using DistilBERT with ONNX Runtime!", return_tensors="np")
>>> outputs = session.run(output_names=["last_hidden_state"], input_feed=dict(inputs))

The required output names (like ["last_hidden_state"]) can be obtained by taking a look at the ONNX configuration of each model. For example, for DistilBERT we have:

>>> from transformers.models.distilbert import DistilBertConfig, DistilBertOnnxConfig

>>> config = DistilBertConfig()
>>> onnx_config = DistilBertOnnxConfig(config)
>>> print(list(onnx_config.outputs.keys()))
["last_hidden_state"]

The process is identical for TensorFlow checkpoints on the Hub. For example, we can export a pure TensorFlow checkpoint from the Keras organization as follows:

python -m transformers.onnx --model=keras-io/transformers-qa onnx/

To export a model that’s stored locally, you’ll need to have the model’s weights and tokenizer files stored in a directory. For example, we can load and save a checkpoint as follows:

>>> from transformers import AutoTokenizer, AutoModelForSequenceClassification

>>> # Load tokenizer and PyTorch weights form the Hub
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
>>> pt_model = AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased")
>>> # Save to disk
>>> tokenizer.save_pretrained("local-pt-checkpoint")
>>> pt_model.save_pretrained("local-pt-checkpoint")

python -m transformers.onnx --model=local-pt-checkpoint onnx/

>>> from transformers import AutoTokenizer, TFAutoModelForSequenceClassification

>>> # Load tokenizer and TensorFlow weights from the Hub
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
>>> tf_model = TFAutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased")
>>> # Save to disk
>>> tokenizer.save_pretrained("local-tf-checkpoint")
>>> tf_model.save_pretrained("local-tf-checkpoint")

python -m transformers.onnx --model=local-tf-checkpoint onnx/

Selecting features for different model tasks

Each ready-made configuration comes with a set of features that enable you to export models for different types of tasks. As shown in the table below, each feature is associated with a different AutoClass:

For each configuration, you can find the list of supported features via the FeaturesManager. For example, for DistilBERT we have:

>>> from transformers.onnx.features import FeaturesManager

>>> distilbert_features = list(FeaturesManager.get_supported_features_for_model_type("distilbert").keys())
>>> print(distilbert_features)
["default", "masked-lm", "causal-lm", "sequence-classification", "token-classification", "question-answering"]

You can then pass one of these features to the --feature argument in the transformers.onnx package. For example, to export a text-classification model we can pick a fine-tuned model from the Hub and run:

python -m transformers.onnx --model=distilbert-base-uncased-finetuned-sst-2-english \
                            --feature=sequence-classification onnx/

This displays the following logs:

Validating ONNX model...
        -[✓] ONNX model output names match reference model ({'logits'})
        - Validating ONNX Model output "logits":
                -[✓] (2, 2) matches (2, 2)
                -[✓] all values close (atol: 1e-05)
All good, model saved at: onnx/model.onnx

Notice that in this case, the output names from the fine-tuned model are logits instead of the last_hidden_state we saw with the distilbert-base-uncased checkpoint earlier. This is expected since the fine-tuned model has a sequence classification head.

Exporting a model for an unsupported architecture

If you wish to export a model whose architecture is not natively supported by the library, there are three main steps to follow:

1. Implement a custom ONNX configuration.
2. Export the model to ONNX.
3. Validate the outputs of the PyTorch and exported models.

In this section, we’ll look at how DistilBERT was implemented to show what’s involved with each step.

Implementing a custom ONNX configuration

Let’s start with the ONNX configuration object. We provide three abstract classes that you should inherit from, depending on the type of model architecture you wish to export:

• Encoder-based models inherit from OnnxConfig
• Decoder-based models inherit from OnnxConfigWithPast
• Encoder-decoder models inherit from OnnxSeq2SeqConfigWithPast

Since DistilBERT is an encoder-based model, its configuration inherits from OnnxConfig:

>>> from typing import Mapping, OrderedDict
>>> from transformers.onnx import OnnxConfig


>>> class DistilBertOnnxConfig(OnnxConfig):
...     @property
...     def inputs(self) -> Mapping[str, Mapping[int, str]]:
...         return OrderedDict(
...             [
...                 ("input_ids", {0: "batch", 1: "sequence"}),
...                 ("attention_mask", {0: "batch", 1: "sequence"}),
...             ]
...         )

Every configuration object must implement the inputs property and return a mapping, where each key corresponds to an expected input, and each value indicates the axis of that input. For DistilBERT, we can see that two inputs are required: input_ids and attention_mask. These inputs have the same shape of (batch_size, sequence_length) which is why we see the same axes used in the configuration.

Once you have implemented an ONNX configuration, you can instantiate it by providing the base model’s configuration as follows:

>>> from transformers import AutoConfig

>>> config = AutoConfig.from_pretrained("distilbert-base-uncased")
>>> onnx_config = DistilBertOnnxConfig(config)

The resulting object has several useful properties. For example, you can view the ONNX operator set that will be used during the export:

>>> print(onnx_config.default_onnx_opset)
11

You can also view the outputs associated with the model as follows:

>>> print(onnx_config.outputs)
OrderedDict([("last_hidden_state", {0: "batch", 1: "sequence"})])

Notice that the outputs property follows the same structure as the inputs; it returns an OrderedDict of named outputs and their shapes. The output structure is linked to the choice of feature that the configuration is initialised with. By default, the ONNX configuration is initialized with the default feature that corresponds to exporting a model loaded with the AutoModel class. If you want to export a model for another task, just provide a different feature to the task argument when you initialize the ONNX configuration. For example, if we wished to export DistilBERT with a sequence classification head, we could use:

>>> from transformers import AutoConfig

>>> config = AutoConfig.from_pretrained("distilbert-base-uncased")
>>> onnx_config_for_seq_clf = DistilBertOnnxConfig(config, task="sequence-classification")
>>> print(onnx_config_for_seq_clf.outputs)
OrderedDict([('logits', {0: 'batch'})])

Exporting the model

Once you have implemented the ONNX configuration, the next step is to export the model. Here we can use the export() function provided by the transformers.onnx package. This function expects the ONNX configuration, along with the base model and tokenizer, and the path to save the exported file:

>>> from pathlib import Path
>>> from transformers.onnx import export
>>> from transformers import AutoTokenizer, AutoModel

>>> onnx_path = Path("model.onnx")
>>> model_ckpt = "distilbert-base-uncased"
>>> base_model = AutoModel.from_pretrained(model_ckpt)
>>> tokenizer = AutoTokenizer.from_pretrained(model_ckpt)

>>> onnx_inputs, onnx_outputs = export(tokenizer, base_model, onnx_config, onnx_config.default_onnx_opset, onnx_path)

The onnx_inputs and onnx_outputs returned by the export() function are lists of the keys defined in the inputs and outputs properties of the configuration. Once the model is exported, you can test that the model is well formed as follows:

>>> import onnx

>>> onnx_model = onnx.load("model.onnx")
>>> onnx.checker.check_model(onnx_model)

Validating the model outputs

The final step is to validate that the outputs from the base and exported model agree within some absolute tolerance. Here we can use the validate_model_outputs() function provided by the transformers.onnx package as follows:

>>> from transformers.onnx import validate_model_outputs

>>> validate_model_outputs(
...     onnx_config, tokenizer, base_model, onnx_path, onnx_outputs, onnx_config.atol_for_validation
... )

This function uses the generate_dummy_inputs() method to generate inputs for the base and exported model, and the absolute tolerance can be defined in the configuration. We generally find numerical agreement in the 1e-6 to 1e-4 range, although anything smaller than 1e-3 is likely to be OK.

Contributing a new configuration to 🤗 Transformers

We are looking to expand the set of ready-made configurations and welcome contributions from the community! If you would like to contribute your addition to the library, you will need to:

• Implement the ONNX configuration in the corresponding configuration_<model_name>.py file
• Include the model architecture and corresponding features in ~onnx.features.FeatureManager
• Add your model architecture to the tests in test_onnx_v2.py

Check out how the configuration for IBERT was contributed to get an idea of what’s involved.