Added ONNX docs

CHANGELOG.md

@@ -4,7 +4,7 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [0.49.0] - 2020-09-17
+## [0.48.5] - 2020-09-17
 
 ### Added
 
@@ -1133,7 +1133,9 @@ Model checkpoint is now saved for all epochs again.
 
 - This CHANGELOG file to hopefully serve as an evolving example of a standardized open source project CHANGELOG.
 
-[0.48.4]: https://github.com/ThilinaRajapakse/simpletransformers/compare/7ef56b0...HEAD
+[0.48.5]: https://github.com/ThilinaRajapakse/simpletransformers/compare/39d25d0...HEAD
+
+[0.48.4]: https://github.com/ThilinaRajapakse/simpletransformers/compare/7ef56b0...39d25d0
 
 [0.48.1]: https://github.com/ThilinaRajapakse/simpletransformers/compare/8c1ae68...7ef56b0
 

docs/_docs/03.5-tips-and-tricks.md

@@ -2,7 +2,7 @@
 title: "Tips and Tricks"
 permalink: /docs/tips-and-tricks/
 excerpt: "Various tips and tricks applicable to most tasks."
-last_modified_at: 2020/08/10 00:42:16
+last_modified_at: 2020/09/24 15:06:27
 toc: true
 ---
 
@@ -483,3 +483,134 @@ model = ClassficationModel(
     proxies={"http": "foo.bar:3128", "http://hostname": "foo.bar:4012"}
 )
 ```
+
+## ONNX Support (Beta)
+
+Simple Transformers has ONNX support for Classification and NER tasks. These models can be converted to an ONNX model and run through the ONNX-runtime.
+
+**Heads up:** ONNX support should be considered experimental at this time. If you encounter any problems, please open an issue in the repo. Please provide a detailed explanation and the **minimal** code necessary to replicate the issue.
+{: .notice--warning}
+
+### ONNX setup
+
+Please refer to the following pages for instructions on installing ONNX.
+
+- [ONNX](https://github.com/onnx/onnx#installation)
+- [ONNX Runtime](https://github.com/microsoft/onnxruntime#get-started)
+
+### Converting a Simple Transformers model to the ONNX format.
+
+The following models are currently compatible:
+
+- ClassificationModel
+- NERModel
+
+These models can be converted by calling the `convert_to_onnx()` method. You can change the output directory by specifying `output_dir` when calling this method.
+
+```python
+from simpletransformers.classification import (
+    ClassificationModel,
+    ClassificationArgs,
+)
+
+
+model = ClassificationModel(
+    "roberta",
+    "roberta-base",
+)
+
+model.convert_to_onnx("onnx_outputs")
+
+```
+
+### Loading a converted ONNX model
+
+You can load the ONNX model just as you would load any other model in Simple Transformers.
+
+```python
+from simpletransformers.classification import (
+    ClassificationModel,
+    ClassificationArgs,
+)
+
+
+model = ClassificationModel(
+    "roberta",
+    "onnx_outputs",
+)
+
+model.convert_to_onnx("onnx_outputs")
+
+```
+
+After the model is loaded, you can use the `predict()` method to make predictions.
+
+### Code example
+
+```python
+from time import time
+
+from simpletransformers.classification import (
+    ClassificationModel,
+    ClassificationArgs,
+)
+
+
+model_args = ClassificationArgs()
+model_args.overwrite_output_dir = True
+
+
+# Create a TransformerModel
+model = ClassificationModel(
+    "roberta",
+    "roberta-base",
+    use_cuda=False,
+    args=model_args,
+)
+
+start = time()
+print(model.predict(["test " * 450]))
+end = time()
+print(f"Pytorch CPU: {end - start}")
+
+model.convert_to_onnx("onnx_outputs")
+
+model_args.dynamic_quantize = True
+
+model = ClassificationModel(
+    "roberta",
+    "onnx_outputs",
+    args=model_args,
+)
+
+start = time()
+print(model.predict(["test " * 450]))
+end = time()
+print(f"ONNX CPU (Cold): {end - start}")
+
+start = time()
+print(model.predict(["test " * 450]))
+end = time()
+print(f"ONNX CPU (Warm): {end - start}")
+
+```
+
+### Execution Providers
+
+ONNX-Runtime supports many different [Execution Providers](https://github.com/microsoft/onnxruntime/tree/master/docs/execution_providers).
+
+If `use_cuda` is `True`, `CUDAExecutionProvider` will be used. If it is `False`, the `CPUExecutionProvider` will be used.
+
+You can manually specify the provider using the `onnx_execution_provider` argument when loading a model.
+
+```python
+model = ClassificationModel(
+    "roberta",
+    "onnx_outputs",
+    args=model_args,
+    onnx_execution_provider="CPUExecutionProvider",
+)
+
+```
+
+*Note that the library is only tested with CPU and CUDA Execution Providers*

setup.py

@@ -5,7 +5,7 @@
 
 setup(
     name="simpletransformers",
-    version="0.48.4",
+    version="0.48.5",
     author="Thilina Rajapakse",
     author_email="chaturangarajapakshe@gmail.com",
     description="An easy-to-use wrapper library for the Transformers library.",