Merged documentation and fixes for integration into internal projects.

Cargo.toml

@@ -21,7 +21,8 @@ maintenance = { status = "experimental" }
 
 
 [dependencies]
-simd_aligned = "0.1.1"
+# simd_aligned = "0.1.1"
+simd_aligned = { path = "../simd_aligned" }
 packed_simd = "0.1"
 rand = "0.5"
 pest = "1.0"

README.md

@@ -22,7 +22,7 @@ You trained a SVM using [libSVM](https://github.com/cjlin1/libsvm), now you want
 * free of `unsafe` code ;)
 
 
-# Principal Usage
+# Usage
 
 Train with [libSVM](https://github.com/cjlin1/libsvm) (e.g., using the tool `svm-train`), then classify with `ffsvm-rust`.
 
@@ -42,7 +42,7 @@ features[3] = -0.221184;
 
 svm.predict_value(&mut problem)?;
 
-assert_eq!(problem.result(), Outcome::Label(42));
+assert_eq!(problem.solution(), Solution::Label(42));
 ```
 
 From C / FFI:

benches/svm_dense.rs

@@ -4,23 +4,23 @@ extern crate ffsvm;
 extern crate test;
 
 mod svm_dense {
-    use ffsvm::{DenseSVM, Linear, ModelFile, Poly, Predict, Problem, Rbf, SVMType, Sigmoid};
+    use ffsvm::{DenseSVM, ModelFile, Predict, Problem};
     use std::convert::TryFrom;
     use test::Bencher;
 
     /// Produces a test case run for benchmarking
     #[allow(dead_code)]
     fn produce_testcase(svm_type: &str, kernel_type: &str, total_sv: u32, num_attributes: u32) -> impl FnMut() {
         let raw_model = ModelFile::random_dense(svm_type, kernel_type, total_sv, num_attributes);
-        let mut svm = DenseSVM::try_from(&raw_model).unwrap();
+        let svm = DenseSVM::try_from(&raw_model).unwrap();
         let mut problem = Problem::from(&svm);
         let problem_mut = problem.features().as_slice_mut();
 
         for i in 0 .. num_attributes {
             problem_mut[i as usize] = i as f32;
         }
 
-        move || (&mut svm).predict_value(&mut problem).expect("This should work")
+        move || (&svm).predict_value(&mut problem).expect("This should work")
     }
 
     // RBF

benches/svm_sparse.rs

@@ -4,7 +4,7 @@ extern crate ffsvm;
 extern crate test;
 
 mod svm_sparse {
-    use ffsvm::{Linear, ModelFile, Poly, Predict, Problem, Rbf, SVMType, Sigmoid, SparseSVM};
+    use ffsvm::{ModelFile, Predict, Problem, SparseSVM};
     use std::convert::TryFrom;
     use test::Bencher;
 
@@ -20,7 +20,7 @@ mod svm_sparse {
             problem_mut[i as usize] = i as f32;
         }
 
-        move || (&mut svm).predict_value(&mut problem).expect("This should work")
+        move || (&svm).predict_value(&mut problem).expect("This should work")
     }
 
     // RBF

docs/performance.md

@@ -2,7 +2,7 @@
 
 # Performance vs. LibSVM
 
-Benchmarks are a tricky thing, but for classifying dense RBF-C-SVMs `ffsvm` should be between 2.5x and 14x faster than `libSVM` on reasonably modern x86 CPUs (supporting AVX2).
+Benchmarks are a tricky thing, but for classifying dense SVMs `ffsvm` should be between 2.5x and 14x faster than `libSVM` on reasonably modern x86 CPUs (supporting AVX2).
 
 ![performance](performance_absolute.v3.png)
 
@@ -14,7 +14,7 @@ There are 3 major factors contributing to this:
 
 * no allocation during classification
 * cache-friendly memory layout
-* usage of SIMD / AVX
+* usage of SIMD
 
 In addition, ffsvm mixes seamlessly with Rayon for _batch classification_, providing even higher speed ups if you classify more than one problem at a time.
 

examples/basic.rs

@@ -7,7 +7,7 @@ fn main() -> Result<(), Error> {
     let svm = DenseSVM::try_from(SAMPLE_MODEL)?;
 
     let mut problem = Problem::from(&svm);
-    let mut features = problem.features();
+    let features = problem.features();
 
     features[0] = 0.55838;
     features[1] = -0.157895;
@@ -16,7 +16,7 @@ fn main() -> Result<(), Error> {
 
     svm.predict_value(&mut problem)?;
 
-    assert_eq!(problem.result(), Outcome::Label(12));
+    assert_eq!(problem.solution(), Solution::Label(12));
 
     Ok(())
 }

src/errors.rs

@@ -40,7 +40,7 @@ pub enum Error {
     /// If the model does not have a `degree` set this error may be raised.
     NoDegree,
 
-    /// Wrapper for [ModelError] when unifiying error handling.
+    /// Wrapper for internal parsing error when unifiying error handling.
     ParsingError(String),
 }
 

src/lib.rs

@@ -60,7 +60,7 @@
 //!
 //!     svm.predict_value(&mut problem)?;
 //!
-//!     assert_eq!(problem.result(), Outcome::Label(42));
+//!     assert_eq!(problem.solution(), Solution::Label(42));
 //!
 //!     Ok(())
 //! }
@@ -86,9 +86,9 @@ pub use crate::{
     parser::ModelFile,
     svm::{
         core::SVMCore,
-        kernel::{KernelDense, Linear, Poly, Rbf, Sigmoid},
+        kernel::{KernelDense, KernelSparse, Linear, Poly, Rbf, Sigmoid},
         predict::Predict,
-        problem::{Outcome, Problem},
+        problem::{DenseProblem, Problem, Solution, SparseProblem},
         DenseSVM, SVMType, SparseSVM,
     },
 };

src/sparse.rs

@@ -1,28 +1,25 @@
-use std::{
-    collections::{btree_map::Iter, BTreeMap},
-    ops::{Index, IndexMut},
-};
+use std::ops::{Index, IndexMut};
 
 #[derive(Clone, Debug)]
 struct Entry<T>
 where
-    T: Copy,
+    T: Copy + Clone + Default,
 {
     index: u32,
     value: T,
 }
 
-#[derive(Clone, Debug)]
+#[derive(Clone, Debug, Default)]
 pub struct SparseVector<T>
 where
-    T: Clone + Copy,
+    T: Clone + Copy + Default,
 {
     entries: Vec<Entry<T>>,
 }
 
 impl<T> SparseVector<T>
 where
-    T: Clone + Copy,
+    T: Clone + Copy + Default,
 {
     pub fn new() -> Self { SparseVector { entries: Vec::new() } }
 
@@ -35,7 +32,7 @@ where
 #[derive(Clone, Debug)]
 pub struct SparseVectorIter<'a, T: 'a>
 where
-    T: Clone + Copy,
+    T: Clone + Copy + Default,
 {
     /// Reference to the matrix we iterate over.
     crate vector: &'a SparseVector<T>,
@@ -46,7 +43,7 @@ where
 
 impl<'a, T> Iterator for SparseVectorIter<'a, T>
 where
-    T: Clone + Copy,
+    T: Clone + Copy + Default,
 {
     type Item = (u32, T);
 
@@ -64,14 +61,14 @@ where
 
 impl<T> Index<usize> for SparseVector<T>
 where
-    T: Copy + Sized,
+    T: Copy + Sized + Default,
 {
     type Output = T;
 
     fn index(&self, index: usize) -> &T {
         // TODO: Beautify me
 
-        for (i, e) in self.entries.iter().enumerate() {
+        for e in self.entries.iter() {
             if e.index == index as u32 {
                 return &e.value;
             }
@@ -111,14 +108,14 @@ where
 #[derive(Clone, Debug)]
 pub struct SparseMatrix<T>
 where
-    T: Clone + Copy,
+    T: Clone + Copy + Default,
 {
     vectors: Vec<SparseVector<T>>,
 }
 
 impl<T> SparseMatrix<T>
 where
-    T: Clone + Copy,
+    T: Clone + Copy + Default,
 {
     pub fn with(rows: usize) -> Self {
         SparseMatrix {
@@ -134,7 +131,7 @@ where
 
 impl<T> Index<(usize, usize)> for SparseMatrix<T>
 where
-    T: Copy + Sized,
+    T: Copy + Sized + Default,
 {
     type Output = T;
 
@@ -152,7 +149,7 @@ where
 #[derive(Clone, Debug)]
 pub struct SparseMatrixIter<'a, T: 'a>
 where
-    T: Clone + Copy,
+    T: Clone + Copy + Default,
 {
     /// Reference to the matrix we iterate over.
     crate matrix: &'a SparseMatrix<T>,
@@ -163,7 +160,7 @@ where
 
 impl<'a, T> Iterator for SparseMatrixIter<'a, T>
 where
-    T: Clone + Copy,
+    T: Clone + Copy + Default,
 {
     type Item = &'a SparseVector<T>;
 

src/svm/class.rs

@@ -32,7 +32,7 @@ impl Class<SimdMatrix<f32s, RowOptimized>> {
 
 impl Class<SparseMatrix<f32>> {
     /// Creates a new class with the given parameters.
-    pub fn with_parameters(classes: usize, support_vectors: usize, attributes: usize, label: u32) -> Class<SparseMatrix<f32>> {
+    pub fn with_parameters(classes: usize, support_vectors: usize, _attributes: usize, label: u32) -> Class<SparseMatrix<f32>> {
         Class {
             label,
             num_support_vectors: support_vectors,