Skip to content

Latest commit

 

History

History
460 lines (377 loc) · 62.4 KB

README.md

File metadata and controls

460 lines (377 loc) · 62.4 KB

WONNX

GitHub Workflow Status docs.rs Crates.io (latest) Crates.io

Wonnx is a GPU-accelerated ONNX inference run-time written 100% in Rust, ready for the web.

Supported Platforms (enabled by wgpu)

API Windows Linux & Android macOS & iOS
Vulkan
Metal
DX12 ✅ (W10 only)
DX11 🚧
GLES3 🆗

✅ = First Class Support — 🆗 = Best Effort Support — 🚧 = Unsupported, but support in progress

Getting started

From the command line

Ensure your system supports either Vulkan, Metal or DX12 for access to the GPU. Then either download a binary release, or install Rust and run cargo install --git https://github.com/webonnx/wonnx.git wonnx-cli to install the CLI.

The CLI tool (nnx) provides a convenient interface for tinkering with models (see the README for more information):

nnx info ./data/models/opt-squeeze.onnx
nnx infer ./data/models/opt-squeeze.onnx -i data=./data/images/pelican.jpeg --labels ./data/models/squeeze-labels.txt --top 3

From Rust

Add the wonnx crate as dependency (cargo add wonnx if you have cargo-add). Then, see the examples for usage examples, or browse the API docs.

From Python

pip install wonnx

And then, to use:

from wonnx import Session
session = Session.from_path(
    "../data/models/single_relu.onnx"
)
inputs = {"x": [-1.0, 2.0]}
assert session.run(inputs) == {"y": [0.0, 2.0]}

Then run python3 with the above Python code!

For more details on the Python package including build instructions, see wonnx-py.

In the browser, using WebGPU + WebAssembly

npm install @webonnx/wonnx-wasm

And then, on the client side:

import init, { Session, Input } from "@webonnx/wonnx-wasm";

// Check for WebGPU availability first: if(navigator.gpu) { .. }
await init();
const session = await Session.fromBytes(modelBytes /* Uint8Array containing the ONNX file */);
const input = new Input();
input.insert("x", [13.0, -37.0]);
const result = await session.run(input); // This will be an object where the keys are the names of the model outputs and the values are arrays of numbers.
session.free();
input.free();

The package @webonnx/wonnx-wasm provides an interface to WONNX, which is included as WebAssembly module and will use the browser's WebGPU implementation. See wonnx-wasm-example for a more complete usage example involving a bundler.

For more details on the JS/WASM package including build instructions, see wonnx-wasm.

For development

To work on wonnx itself, follow the following steps:

  • Install Rust
  • Install Vulkan, Metal, or DX12 for the GPU API.
  • git clone this repo.
git clone https://github.com/webonnx/wonnx.git

Then, you're all set! You can run one of the included examples through cargo:

cargo run --example squeeze --release

Running other models

  • To run an onnx model, first simplify it with nnx prepare (substitute with cargo run -- prepare when inside this repo):
nnx prepare -i ./some-model.onnx ./some-model-prepared.onnx

To specify dynamic dimension parameters, add e.g. --set batch_size=1.

You can also use an external tool, such as onnx-simplifier, with the command:

# pip install -U pip && pip install onnx-simplifier
python -m onnxsim mnist-8.onnx opt-mnist.onnx
cargo run --example mnist --release

Tested models

  • Squeezenet
  • MNIST
  • BERT

GPU selection

Except when running in WebAssembly, you may set the following environment variables to influence GPU selection by WGPU:

  • WGPU_ADAPTER_NAME with a substring of the name of the adapter you want to use (e.g. 1080 will match NVIDIA GeForce 1080ti).
  • WGPU_BACKEND with a comma separated list of the backends you want to use (vulkan, metal, dx12, dx11, or gl).
  • WGPU_POWER_PREFERENCE with the power preference to choose when a specific adapter name isn't specified (high or low)

Contribution: On implementing a new Operator

Contributions are very much welcomed even without large experience in DL, WGSL, or Rust. I hope that this project can be a sandbox for all of us to learn more about those technologies beyond this project's initial scope.

To implement an operator all you have to do is:

  1. Add a new matching pattern in compiler.rs
  2. Retrieve its attributes values using the get_attribute function:
    let alpha = get_attribute("alpha", Some(1.0), node);
    // or without default value
    let alpha = get_attribute::<f32>("alpha", None, node);
  1. Add any variable you want to use in the WGSL shader using context.
  2. Write a new WGSL template in the templates folder.

Available types are in structs.wgsl but you can also generate new ones within your templates.

  1. Respect the binding layout that each entry is incremented by 1 starting from 0, with input first and output last. If the number of binding is above 4. Increment the binding group. You can change the input within sequencer.rs
  2. Write the logic.

There is default variables in the context:

  • {{ i_lens[0] }}: the length of the input 0. This also work for output: {{ o_lens[0] }} and other input {{ i_lens[1] }}
  • {{ i_shape[0] }}: the array of dimensions of input 0. To get the first dimension of the array, just use: {{ i_shape[0][0] }}
  • {{ i_chunks[0] }}: the size of the chunks of each dimensions of input 0. By default, each variable is represented as a long array of values where to get to specific values you have to move by chunks. Those chunks are represented within this variable. To get the size of the chunks of the first dimensions use: {{ i_chunks[0][0] }}.
  • {{ op_type }} the op type as some op_type like activation are using the same template.
  1. Test it using the utils function and place it in the tests folder. The test can look as follows:
#[test]
fn test_matmul_square_matrix() {
    // USER INPUT

    let n = 16;
    let mut input_data = HashMap::new();

    let data_a = ndarray::Array2::eye(n);
    let mut data_b = ndarray::Array2::<f32>::zeros((n, n));
    data_b[[0, 0]] = 0.2;
    data_b[[0, 1]] = 0.5;

    let sum = data_a.dot(&data_b);

    input_data.insert("A".to_string(), data_a.as_slice().unwrap());
    input_data.insert("B".to_string(), data_b.as_slice().unwrap());

    let n = n as i64;
    let model = model(graph(
        vec![tensor("A", &[n, n]), tensor("B", &[n, n])],
        vec![tensor("C", &[n, n])],
        vec![],
        vec![],
        vec![node(vec!["A", "B"], vec!["C"], "MatMul", "MatMul", vec![])],
    ));

    let session =
        pollster::block_on(wonnx::Session::from_model(model)).expect("Session did not create");

    let result = pollster::block_on(session.run(input_data)).unwrap();

    // Note: it is better to use a method that compares floats with a tolerance to account for differences
    // between implementations; see `wonnx/tests/common/mod.rs` for an example.
    assert_eq!((&result["C"]).try_into().unwrap(),sum.as_slice().unwrap());
}

Check out tera documentation for other templating operation: https://tera.netlify.app/docs/

  1. If at any point you want to do optimisation of several nodes you can do it within sequencer.rs.

Supported Operators (ref ONNX IR)

Operator Since version Implemented Shape inference supported
Abs 13, 6, 1
Acos 7
Acosh 9
Add 14, 13, 7, 6, 1
And 7, 1
ArgMax 13, 12, 11, 1
ArgMin 13, 12, 11, 1
Asin 7
Asinh 9
Atan 7
Atanh 9
AveragePool 11, 10, 7, 1
BatchNormalization 15, 14, 9, 7, 6, 1
BitShift 11
Cast 13, 9, 6, 1
Ceil 13, 6, 1
Clip 13, 12, 11, 6, 1
Compress 11, 9
Concat 13, 11, 4, 1
ConcatFromSequence 11
Constant 13, 12, 11, 9, 1
ConstantOfShape 9
Conv 11, 1
ConvInteger 10
ConvTranspose 11, 1
Cos 7
Cosh 9
CumSum 14, 11
DepthToSpace 13, 11, 1
DequantizeLinear 13, 10
Det 11
Div 14, 13, 7, 6, 1
Dropout 13, 12, 10, 7, 6, 1
Einsum 12
Elu 6, 1
Equal 13, 11, 7, 1
Erf 13, 9
Exp 13, 6, 1
Expand 13, 8
EyeLike 9
Flatten 13, 11, 9, 1
Floor 13, 6, 1
GRU 14, 7, 3, 1
Gather 13, 11, 1 ✅ (axis=0)
GatherElements 13, 11
GatherND 13, 12, 11
Gemm 13, 11, 9, 7, 6, 1 ✅*
GlobalAveragePool 1
GlobalLpPool 2, 1
GlobalMaxPool 1
Greater 13, 9, 7, 1
GridSample 16
HardSigmoid 6, 1
Hardmax 13, 11, 1
Identity 16, 14, 13, 1
If 16, 13, 11, 1
InstanceNormalization 6, 1
IsInf 10
IsNaN 13, 9
LRN 13, 1
LSTM 14, 7, 1
LeakyRelu 6, 1
Less 13, 9, 7, 1
Log 13, 6, 1
Loop 16, 13, 11, 1
LpNormalization 1
LpPool 11, 2, 1
MatMul 13, 9, 1
MatMulInteger 10
Max 13, 12, 8, 6, 1
MaxPool 12, 11, 10, 8, 1
MaxRoiPool 1
MaxUnpool 11, 9
Mean 13, 8, 6, 1
Min 13, 12, 8, 6, 1
Mod 13, 10
Mul 14, 13, 7, 6, 1
Multinomial 7
Neg 13, 6, 1
NonMaxSuppression 11, 10
NonZero 13, 9
Not 1
OneHot 11, 9 ✅ (axis=-1)
Optional 15
OptionalGetElement 15
OptionalHasElement 15
Or 7, 1
PRelu 9, 7, 6, 1
Pad 13, 11, 2, 1 ✅ (mode=constant, pads>=0)
Pow 15, 13, 12, 7, 1 ✅ (broadcast=0 and data type is f32)
QLinearConv 10
QLinearMatMul 10
QuantizeLinear 13, 10
RNN 14, 7, 1
RandomNormal 1
RandomNormalLike 1
RandomUniform 1
RandomUniformLike 1
Reciprocal 13, 6, 1
ReduceL1 13, 11, 1
ReduceL2 13, 11, 1
ReduceLogSum 13, 11, 1
ReduceLogSumExp 13, 11, 1
ReduceMax 13, 12, 11, 1
ReduceMean 13, 11, 1
ReduceMin 13, 12, 11, 1
ReduceProd 13, 11, 1
ReduceSum 13, 11, 1
ReduceSumSquare 13, 11, 1
Relu 14, 13, 6, 1
Reshape 14, 13, 5, 1
Resize 13, 11, 10
ReverseSequence 10
RoiAlign 16, 10
Round 11
Scan 11, 9, 8
Scatter (deprecated) 11, 9
ScatterElements 16, 13, 11
ScatterND 16, 13, 11
Selu 6, 1
SequenceAt 11
SequenceConstruct 11
SequenceEmpty 11
SequenceErase 11
SequenceInsert 11
SequenceLength 11
Shape 15, 13, 1
Shrink 9
Sigmoid 13, 6, 1
Sign 13, 9
Sin 7
Sinh 9
Size 13, 1
Slice 13, 11, 10, 1
Softplus 1
Softsign 1
SpaceToDepth 13, 1
Split 13, 11, 2, 1
SplitToSequence 11
Sqrt 13, 6, 1
Squeeze 13, 11, 1
StringNormalizer 10
Sub 14, 13, 7, 6, 1
Sum 13, 8, 6, 1
Tan 7
Tanh 13, 6, 1
TfIdfVectorizer 9
ThresholdedRelu 10
Tile 13, 6, 1
TopK 11, 10, 1
Transpose 13, 1
Trilu 14
Unique 11
Unsqueeze 13, 11, 1
Upsample (deprecated) 10, 9, 7
Where 16, 9
Xor 7, 1
Function Since version
Bernoulli 15
CastLike 15
Celu 12
DynamicQuantizeLinear 11
GreaterOrEqual 12
HardSwish 14
LessOrEqual 12
LogSoftmax 13, 11, 1
MeanVarianceNormalization 13, 9
NegativeLogLikelihoodLoss 13, 12
Range 11
Softmax 13, 11, 1
SoftmaxCrossEntropyLoss 13, 12

Known limitations

  • The Clip, Resize, Reshape, Split, Pad and ReduceSum ops accept (typically optional) secondary inputs to set various parameters (i.e. axis). These inputs are only supported if they are supplied as initializer tensors (i.e. do not depend on inputs and are not outputs of other ops), because wonnx pre-compiles all operations to shaders in advance (and must know these parameters up front).

  • Internally 64-bit integers are not supported (the reason is they are not supported in the current version of WGSL); inputs and initializers with 64-bit scalars are converted to 32-bit values (possibly overflowing).

  • For MatMul and Gemm, the matrix dimensions must be divisible by 2, or the output matrix must be of size (1, N). Matrix multiplication only supports floats, not integers (this is a WebGPU/WGSL limitation).

Shape inference

WONNX needs to know the shape of input and output tensors for each operation in order to generate shader code for executing it. ONNX models however do not always contain this information for intermediate values. Shape inference is the process of deducing the shape of intermediate values from the shape of inputs and outputs and the characteristics of each operation.

WONNX supports a limited form of shape inference (the process of determining what the shapes are of the various nodes in a model's graph). Shape inference is available programmatically as well as through the CLI. Before shape inference can be performed, all dynamic dimension parameters need to be replaced with static values. Shape inference only infers output shapes from input shapes for specific supported ops (see the table above). Inference cannot succeed if the shape for any input of a node is not known. Nodes that already have fully defined shapes for their outputs are left unchanged (and the outputs are used for shape inference on nodes that use these outputs as inputs).

To perform shape inference using the CLI, run a command similar to this (here batch_size and sequence_length are dynamic dimension parameters; the -i flag enables shape inference):

nnx prepare model.onnx model-prepared.onnx --set batch_size=1 --set sequence_length=255 -i

To perform shape inference programmatically, use apply_dynamic_dimensions and infer_shapes from the wonnx_preprocessing::shape_inference module.

Constant folding

Some models contain subgraphs whose output can be determined statically, as they do not depend on the specific inputs provided during inference. WONNX can replace such constant intermediate values with static values ('constant folding'). This is supported in the following cases:

  • Output of nodes of the Constant op type (these are replaced with initializers)
  • Output of nodes of the Shape op type where the shape of the input is known (up front or during inference)
  • Output of nodes of which all inputs are constant (possibly after folding), and for which the operator is supported by WONNX.

Constant folding is performed as part of shape inference, unless disabled (from the CLI pass --no-fold-constants to disable). This is done in order to support models that dynamically calculate shapes using operators such as Shape/Squeeze/Unsqueeze depending on dynamically set dimension parameters (e.g. batch size).

License

Licensed under either of

Except for the following files:

  • data/models:

    • mobilenetv2-7.onnx: source, Apache-2.0 license only.
    • squeezenet-labels.txt: source, Apache-2.0 license only.
  • data/images:

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you shall be dual licensed as above, without any additional terms or conditions.