Table of Contents
- πΉοΈ User-friendly menu based on the dialoguer crate.
- π₯· Under the hood .onnx and .pb parser to load ONNX models and protobuf data.
- π¦ Delivered with a set of validated operators, models and datasets.
- π¦ Network inference can be performed in both verbose and non-verbose mode.
- β Easy to add new operators to extend model compatibility.
- πΌοΈ Images & labels serialization to extend the set of available datasets.
- π Rayon-powered image-based and intra-network parallelization.
- π Python bindings for a complete Rusty experience.
We recommend connecting to GitHub with SSH.
After the setup, it will be possible to clone the repo by doing:
git clone git@github.com:claudiotancredi/RustONNX.git
Depending on the Rust toolchain you decide to use, please be aware that you may experience build-time errors caused by spaces in the project path.
To avoid any build issues, it's recommended to clone or place the project in a directory without spaces in its path.
To run the project, you need to have rustup installed on your PC.
After cloning the repo, you can run our application by doing:
cd RustONNX/rust-onnx
cargo run
If everything goes smoothly, an interactive menu will appear to let you decide how to perform inference.
Let's see the flow when executing our RustONNX.
At first, you will see this interactive menu. Let's choose MNIST for simplicity.
We'll be asked for the dataset to perform the inference on. Let's go with a custom dataset.
We'll be asked for the folder name of the custom dataset. Let's use the provided custom dataset, so our input should be "dataset".
We'll be asked for the mode: verbose or non-verbose. We'll go with non-verbose this time.
RustONNX will keep us updated about the runtime status for the different operations (model loading, serialization if any, dataset loading, model optimization for intra-network parallel execution). At the end, some general info will be printed in a tabular way and the execution in non-verbose mode will start.
When finished, the network will display some info: execution time and results for each input image. In this case, the limit of 50 images has been exceeded, so we can only see a sneak peek. The complete table has been stored in the results.txt file.
After the table, error rate and accuracy will be printed too, also in the results.txt file.
Now, just to give you a hint, this is what the execution in verbose mode looks like. As you can see, the difference with the non-verbose mode is that here also the names and the shapes of inputs, initializers and outputs are displayed in a pretty and colorful table.
Here is an overview of the project structure with the most important folders & files, along with some useful information.
RustONNX
ββββimages //Assets for README.md
ββββrust-onnx
ββββmodels //Folder where supported models are located
β ββββmnist-12
β ββββmobilenet-v2-7
β β ββββdataset //Sample custom ImageNet-like dataset
β β β ββββ566 //Each folder contains images associated with the label given by the folder name
β β β ββββ569
β β β ββββ571
β β β ββββ574
β β β ββββ701
β β ββββdataset_serialized //Generated at run-time via serialization on the sample custom dataset
β β ββββtest_data_set_0 //Test dataset provided by ONNX
| | ββββmodel.onnx //.onnx file for the model
β ββββresnet18-v1-7
β ββββresnet34-v2-7
ββββprotos //.proto3 files used at build-time
ββββpy_rust_onnx
β ββββpython-binding //Python virtual environment folder generated while doing the setup for the binding
β ββββtest.py //Python demo for binding
ββββsrc
β ββββdatasets //Dataset related functions and data structures
β ββββmodels //Models related functions and data structures
β ββββonnx_parser
β β ββββprotoc_generated //Generated at build-time using protoc. Parser modules.
β β ββββparser.rs //ONNX parser functions
β ββββoperators //Operators implementation
β ββββutils //Generic utilities module (metrics, errors, menu, execution serialization, etc.)
β β ββββauxiliary_functions.rs
β β ββββerrors.rs
β β ββββmenu.rs
β β ββββrun.rs
β β ββββserialization_utils.rs
β β ββββshared_constants.rs
β ββββlib.rs //Used for Python bindings
β ββββmain.rs
ββββbuild.rs
ββββCargo.lock
ββββCargo.toml
ββββresults.txt //Generated at run-time. Execution results when the dataset has # images > 50
The folder RustONNX/rust-onnx/src/operators/
should be your single-point-of-change-in-code (SPOCIC) when implementing new operators.
The folders RustONNX/rust-onnx/src/models/
and RustONNX/rust-onnx/src/datasets/
should be your SPOCICs when adding new models that are compatible with new datasets.
The folder RustONNX/rust-onnx/models/
is where ONNX models and datasets must reside.
Reference: ONNX Model Zoo
Model name | Description | Version | Opset version |
---|---|---|---|
mnist-12 | Handwritten digits recognition model specifically designed for the MNIST dataset. | N/A | 12 |
resnet18-v1-7 | Image classification model with 18 layers for ImageNet-like datasets. | 1 | 7 |
resnet34-v2-7 | Image classification model with 34 layers for ImageNet-like datasets. | 2 | 7 |
mobilenet-v2-7 | Efficient and lightweight image classification model for ImageNet-like datasets. | 2 | 7 |
During the development of this project, the ONNX Model Zoo was being expanded by incorporating additional models. Since new models had not been validated yet, the models supported by our RustONNX were taken from the set of validated networks.
Please, be aware that new models may be more performant and that old models, included the ones supported by our RustONNX, may be deleted in future updates of the ONNX Model Zoo.
Reference: ONNX Operators
Operator | Version | Description |
---|---|---|
Add | 14 | Performs element-wise binary addition (with Numpy-style broadcasting support). |
BatchNorm | 15 | Carries out batch normalization as described in the paper https://arxiv.org/abs/1502.03167. |
Convolution | 11 | The convolution operator performs the convolution given an input tensor and a filter, and computes the output. RustONNX provides the optimized version whose Python implementation can be found here. |
Flatten | 21 | Flattens the input tensor into a 2D matrix. If input tensor has shape (d_0, d_1, β¦ d_n) then the output will have shape (d_0 X d_1 β¦ d_(axis-1), d_axis X d_(axis+1) β¦ X dn). |
Gemm | 13 | General Matrix multiplication: https://en.wikipedia.org/wiki/Basic_Linear_Algebra_Subprograms#Level_3 Aβ = transpose(A) if transA else A Bβ = transpose(B) if transB else B Compute Y = alpha * Aβ * Bβ + beta * C |
GlobalAveragePooling | 1 | Given an input tensor X, applies average pooling across the values in the same channel. |
MatMul | 13 | Matrix product that behaves like numpy.matmul: https://numpy.org/doc/stable/reference/generated/numpy.matmul.html |
MaxPool | 12 | Given an input tensor X, applies max pooling across the tensor according to kernel sizes, stride sizes, and pad lengths. Max pooling consists of computing the max on all values of a subset of the input tensor according to the kernel size and downsampling the data into the output tensor Y for further processing. |
ReLU | 14 | Relu takes one input data (Tensor) and produces one output data (Tensor) where the rectified linear function, y = max(0, x), is applied to the tensor elementwise. |
Reshape | 21 | Reshape the input tensor similar to numpy.reshape. First input is the data tensor, second input is a shape tensor which specifies the output shape. It outputs the reshaped tensor. |
For all models supported by RustONNX, two sample datasets are provided:
- a test dataset provided by ONNX, placed in
RustONNX/rust-onnx/models/<model_to_run>/test_data_set_0/
, with an image and a label already serialized into .pb files. The datasets validated by ONNX are especially useful when implementing new operators to extend the model compatibility of RustONNX. - a custom dataset, placed in
RustONNX/rust-onnx/models/<model_to_run>/dataset/
, with .jpg, .jpeg or .png images to be serialized into .pb files, on which the network will be executed. Thanks to the serialization feature of RustONNX, you can smoothly perform inference on your own datasets. But if you're lazy and just want to test the feature, at least we've got you covered! π¦₯
The foundation of our project is encapsulated in the Operator
trait, which outlines the essential capabilities each operator must provide.
Here is the list of functions to implement.
fn execute(&self, inputs: &HashMap<String, ArrayD<f32>>) -> Result<Vec<ArrayD<f32>>, OnnxError>;
The execute contains the logic of the operator. It accepts a &HashMap<String, ArrayD<f32>>
where the key is the name of the input of the network or the name of a previously executed node, and the value is the associated data/output. It returns a Result<Vec<ArrayD<f32>>, OnnxError>
, which either yields a vector of multi-dimensional arrays
upon success or an OnnxError in case of failure.
fn to_string(&self, inputs: &HashMap<String, ArrayD<f32>>, outputs: &Vec<ArrayD<f32>>, image_index: String) -> String;
This is used during the run in verbose mode. This method provides descriptive information about the inputs, initializers, outputs and the image being processed, for a specific node. It takes a &HashMap<String, ArrayD<f32>>
and a Vec<ArrayD<f32>>
as inputs, along with an
image index of type String
that represents the index of the image that the model is currently performing inference on. It returns a String
that can be printed on the console.
fn get_inputs(&self) -> Vec<String>;
fn get_output_names(&self) -> Vec<String>;
fn get_node_name(&self) -> String;
fn get_op_type(&self) -> String;
fn get_initializers_arr(&self) -> Vec<Initializer>;
In this order:
- Returns a
Vec<String>
listing the names of the inputs. - Provides a
Vec<String>
with the names of the operator's outputs. - Returns a
String
containing the operator's node name. - Returns a
String
containing the type of the operator (e.g. "Conv"). - Retrieves a
Vec
, which contains the initializers associated with the node, if any.
fn clone_box(&self) -> Box<dyn Operator>;
This function is essential for Python bindings. It returns a Box<dyn Operator>
, enabling the cloning of operator instances.
The rayon crate has been used to implement both an image-based and an intra-network parallelization.
The image-based parallelization is trivial: when loading different .pb files, where each represents an image, the inference on these images will be performed in parallel.
As for the intra-network parallelization, different strategies could be investigated. Our RustONNX employs an initial processing of the network to optimize it for the subsequent execution. This process can be divided into two steps:
- creation of a graph structure using the petgraph crate. This is done to ease the analysis of the network architecture;
- graph traversal to collect the nodes belonging to the same layers. A layer is a subset of nodes that don't show any dependencies and can, therefore, run in parallel.
Let's do an example. The MNIST (opset-version=12)
architecture is:
The result of the initial optimization on MNIST (opset-version=12)
is:
[["Convolution28", "Times212_reshape1"], ["Plus30"], ["ReLU32"], ["Pooling66"], ["Convolution110"], ["Plus112"], ["ReLU114"], ["Pooling160"], ["Times212_reshape0"], ["Times212"], ["Plus214"]]
Indeed, the nodes with names Convolution28
and Times212_reshape1
can run in parallel.
RustONNX provides Python bindings using the maturin crate and the PyO3 library. This way, if you want to work in the Python environment you can still benefit from Rust's performance.
To ensure a smooth testing experience of our Python bindings, we advise using a virtual environment. Below are the instructions to set up and activate a Python virtual environment. After setting up the environment, you will be able to run the example file we have provided to demonstrate the usage of our binding.
- Create the virtual environment:
cd rust-onnx python3 -m venv ./py_rust_onnx/python-binding/
- Activate the virtual environment:
- For Linux and macOS:
source py_rust_onnx/python-binding/bin/activate
- For Windows:
.\py_rust_onnx\python-binding\Scripts\activate
- For Linux and macOS:
- Install the necessary packages:
pip install maturin numpy
- Build the Rust code and install it:
maturin develop --release
- Run our Python script to test the binding:
python3 ./py_rust_onnx/test.py
The exported Rust functions are encapsulated in a module called rust_onnx_lib
, that you can import as:
import rust_onnx_lib
Here are the 5 exported functions:
pub fn py_load_model(_py: Python, file_path_str: String) -> PyResult<(PyModel, String, String)>
This function loads the ONNX model from the provided path and returns the model, the expected name for the input of the network and the name of the final node of the network.
pub fn py_serialize_custom_dataset(_py: Python, model_id: usize, dir_path: String) -> PyResult<()>
If a custom dataset is selected to infer on, this function performs the serialization of images and labels into .pb files that can then be loaded by the framework.
pub fn py_load_images_and_labels(py: Python, model_id: usize, folder_name: String, test_dataset: bool)
->PyResult<(PyObject, PyObject, Vec<String>)>
This function retrieves and loads the images and the labels of the selected dataset, returning the images data, the labels data and a vec containing the file paths of the loaded .pb files.
pub fn py_run_model(py:Python, inputs: Vec<PyObject>, model: &PyModel, input_name: String,
verbose: bool, final_layer_name: String, id_model: usize)
-> PyResult<PyObject>
This is the core function of the framework. It runs the model returning the final output of the network.
pub fn py_print_results (py: Python, model_id: usize, files: Vec<String>, model_output: PyObject,
labels: PyObject) -> PyResult<()>
This last function allows displaying the result of the network for each input image, along with global metrics, such as accuracy and error rate.
To integrate a new operator into our framework, you can follow these steps into the SPOCIC directory operators/
:
-
Create a new file: generate a new Rust source file (
.rs
) within theoperators/
directory. This file will contain your operator's implementation. -
Export the operator: update the
mod.rs
file in theoperators/
folder to export your new operator, making it accessible to other parts of the project. -
Edit create_operator() function: the function is located within the
operators.rs
file and you should add a new arm, for your new operator, to the existing match, as done for the others.
Each operator must adhere to the Operator
trait, which outlines essential functionalities. Your operator should implement the following methods:
execute()
: defines the core logic for the operator;get_inputs()
: returns aVec<String>
listing the names of the inputs;get_output_names()
: provides aVec<String>
with the names of the outputs;get_node_name()
: gives back aString
with the operator's node name;get_op_type()
: delivers aString
specifying the operator's type;get_initializers_arr()
: retrieves aVec
, which contains any initializers associated with the node;clone_box()
: this method is necessary for creating Python bindings.
By following these specifications, you ensure that the new operator is compatible with our framework.
To add a new model to our project, follow these organized steps carefully:
- Navigate to
rust-onnx/models/
. - Create a new folder named after your model.
Inside your model's folder, you should include:
- The ONNX model file renamed to
model.onnx
. - A subfolder named
test_data_set_0
which will contain the following files provided by ONNX:input_0.pb
: the input protobuf file.output_0.pb
: the output protobuf file.
In a following section you'll know more about the possibility of adding a new dataset.
Here your SPOCIC is the src/models/
folder.
- Open the file
src/models/models.rs
. - Update the constant
MODELS_NAMES
with the name of your new model. - Add a new variant to the
Model
enum representing your model. - Extend the methods of the Model enum implementation with the new alternative.
By following these steps, you'll successfully integrate a new model into our framework.
It is possible to run a network on a new dataset, provided that the latter conforms to some rules:
- the dataset folder must be placed under the folder of the model you want to run (e.g.
models/mnist-12/my-dataset/
); - the dataset folder must include subfolders whose names match the label, in numeric format, of the images they contain;
- at least one subfolder that follows the naming convention mentioned above must reside within the dataset folder;
- accepted image formats are .jpg, .jpeg or .png;
- all subfolders following the expected naming convention (i.e. label in numeric format) must include at least a .jpg, .jpeg or .png file.
For example, you may have a "my-dataset/"
folder under "resnet18-v1-7/"
with a "207/"
subfolder that includes a .jpg, .jpeg or .png image of a golden retriever, since 207 is the label for a golden retriever in the ImageNet dataset.
There are a few other things you must check before going wild with your images. Here your SPOCIC is the src/datasets/
folder:
- if the dataset classes are different from the ones in MNIST and ImageNet, you should provide the mapping of the labels in the file
label_mappings.rs
; - if you're also going to run a new model on these images, you may want to check the ONNX Model Zoo documentation and, if required by the model, add a new serialization function for the new dataset. You can do it in the
images_serialization.rs
file.
- ONNX Model Zoo
- ONNX Operators
- Netron
- protobuf_codegen crate
- ndarray crate
- rayon crate
- maturin crate
- PyO3 user guide
Contributions are what make the open source community such an amazing place to learn, inspire, and create. Any contributions you make will benefit everybody else and are greatly appreciated.
If you're interested in opening an issue, please be sure to make it:
- Scoped to a single topic. One topic per issue.
- Specific. Include as many details as possible.
- Unique. Do not duplicate existing opened issues: search for the topic you're interested in before creating any issues.
- Reproducible. If you're reporting a problem, include the necessary steps to reproduce it.
If you have a suggestion that would make this better, please fork the repo and create a pull request.
Don't forget to give the project a star! Thanks again!
Distributed under the MIT License. See LICENSE
for more information.
RustONNX is delivered with sample custom datasets:
- for MNIST, images are sourced from MNIST-JPG;
- for models that work on ImageNet-like datasets, images are sourced from Imagenette and are licensed under the Apache License 2.0.