FFSVM stands for "Really Fast Support Vector Machine"


Latest Version Travis-CI Status deps.svg docs MIT

In One Sentence

You trained a SVM using libSVM, now you want the highest possible performance during (real-time) classification, like games or VR.


  • loads almost all libSVM types (C-SVC, ν-SVC, ε-SVR, ν-SVR) and kernels (linear, poly, RBF and sigmoid)
  • produces practically same classification results as libSVM
  • optimized for SIMD and can be mixed seamlessly with Rayon
  • written in 100% Rust
  • allocation-free during classification for dense SVMs
  • 2.5x - 14x faster than libSVM for dense SVMs
  • extremely low classification times for small models (e.g., 128 SV, 16 dense attributes, linear ~ 500ns)
  • successfully used in Unity and VR projects (Windows & Android)

Note: Currently requires Rust nightly (March 2019 and later), because we depend on RFC 2366 (portable SIMD). Once that stabilizes we'll also go stable.


Train with libSVM (e.g., using the tool svm-train), then classify with ffsvm-rust.

From Rust:

// Replace `SAMPLE_MODEL` with a `&str` to your model.
let svm = DenseSVM::try_from(SAMPLE_MODEL)?;

let mut problem = Problem::from(&svm);
let features = problem.features();

features[0] = 0.55838;
features[1] = -0.157895;
features[2] = 0.581292;
features[3] = -0.221184;

svm.predict_value(&mut problem)?;

assert_eq!(problem.solution(), Solution::Label(42));


  • June 7, 2019: Gave up on 'no unsafe', but gained runtime SIMD selection.
  • March 10, 2019: As soon as we can move away from nightly we'll go beta.
  • Aug 5, 2018: Still in alpha, but finally on crates.io.
  • May 27, 2018: We're in alpha. Successfully used internally on Windows, Mac, Android and Linux on various machines and devices. Once SIMD stabilizes and we can cross-compile to WASM we'll move to beta.
  • December 16, 2017: We're in pre-alpha. It will probably not even work on your machine.



Classification time vs. libSVM for dense models.


Performance milestones during development.

All performance numbers reported for the DenseSVM. We also have support for SparseSVMs, which are slower for "mostly dense" models, and faster for "mostly sparse" models (and generally on the performance level of libSVM).

See here for details.


  • For an x-fold performance increase, create a number of Problem structures, and process them with Rayon's par_iter.


See here for details.

  • Suggestions for better results

    Suggestions for better results

    Hi I am training regression svm using this library, and I get horrible results. I verified with libsvm and it provides the same results. What is a good approach to svm regression? Here is what I should get: Screen Shot 2019-03-20 at 1.17.59 AM.pdf Here is the predicted: Screen Shot 2019-03-20 at 1.18.35 AM.pdf I tried all possible kernels but I cannot get it to work. Screen Shot 2019-03-20 at 1.18.56 AM.pdf Screen Shot 2019-03-20 at 1.21.41 AM.pdf

    I have 219k data points with 100 dimensions.

    opened by karasjoh000 2
  • Lack of empty line in a model causes kernel panic

    Lack of empty line in a model causes kernel panic

    The model below causes a panic. Please note the lack of empty line at the end of the file.

    kernel_type rbf
    gamma 1
    nr_class 2
    total_sv 2
    rho -0.109665
    label 1 0
    probA -4.83344
    probB 0.590724
    nr_sv 1 1
    0 0:0 1:0 2:0 3:0 4:0 5:0 6:0 7:0 8:0 9:0 10:0 11:0 12:0 13:0 14:0 15:0 16:0 17:0 18:0 19:0 20:0 21:0 22:0 23:0
    1 0:0 1:0 2:0 3:0 4:0 5:0 6:0 7:0 8:0 9:0 10:0 11:0 12:0 13:0 14:0 15:0 16:0 17:0 18:0 19:0 20:0 21:0 22:0 23:0
    opened by Hoodad 2
  • breaks on label with labels 1 -1

    breaks on label with labels 1 -1

    I have a model file with label 1 -1 however the parser only parses label 1 and therefore it breaks at this part:

                // Construct vector of classes
                let classes = match svm_type {
                    // TODO: CLEAN THIS UP ... We can probably unify the logic
                    SVMType::CSvc | SVMType::NuSvc => (0..num_classes)
                        .map(|c| {
                            let label = header.label[c];
                            let num_sv = nr_sv[c] as usize;
                            Class::<$m32>::with_parameters(num_classes, num_sv, num_attributes, label)
                    SVMType::ESvr | SVMType::NuSvr => vec![Class::<$m32>::with_parameters(

    here is my file header:

    svm_type c_svc
    kernel_type rbf
    gamma 0.076923076923076927
    nr_class 2
    total_sv 132
    rho 0.42446205176771579
    label 1 -1
    nr_sv 64 68

    however the header gets parsed like this:

    [src/svm/core/sparse.rs:41] header = Header {
        svm_type: "c_svc",
        kernel_type: "rbf",
        gamma: Some(
        coef0: None,
        degree: None,
        nr_class: 2,
        total_sv: 132,
        rho: [
        label: [
        prob_a: None,
        prob_b: None,
        nr_sv: [

    probably does not like the negative sign

    opened by karasjoh000 1
  • Status Update

    Status Update

    Various notes to self:

    • the crate is not really dead, the cross-platform SIMD situation has just been a bit ... frustrating
    • once something like packed SIMD ever stabilizes I want to move to that
    • I consider classification / regression mostly feature complete
    • IIRC there are one or two rarer kernels not implemented yet, PRs welcome
    • Training is missing too, probably won't implement it myself as I don't really need it, PRs also welcome #2
    • runtime SIMD selection should probably be removed again (came to generally prefer unsafe free over a few percent more performance)
    • once safe transmutes are a thing simd_aligned dependency should be removed
    • if meanwhile anyone feels repelled by fact last commit is 2+ years old and / or project would not compile in specific constellation I'm also happy to accept PRs and / or add co-maintainers
    opened by ralfbiedert 0
  • Should support training

    Should support training

    Blocked by

    • [ ] Need high-performance matrix library that abstracts away sparse and dense (SIMD) matrices.
    • [ ] In turn blocked by num-traits abstraction to handle scalar and packed types / math transparently.
    opened by ralfbiedert 2
Ralf Biedert
Ralf Biedert
Yarte stands for Yet Another Rust Template Engine

Should we start to worry? bytes-buf feature can produce SIGILL. avx and sse flags are in almost all cpus of x86 and x86_64 architectures. More details

Juan Aguilar 249 Dec 19, 2022