Merge pull request #16 from Polochon-street/add-cosine-distance

Add cosine distance and formatter

CHANGELOG.md

@@ -1,5 +1,11 @@
 # Changelog
 
+## bliss 0.3.1
+* Show error message when song storage fails in the Library trait.
+* Added a `distance` module containing euclidean and cosine distance.
+* Added various custom_distance functions to avoid being limited to the
+  euclidean distance only.
+
 ## bliss 0.3.0
 * Changed `Song.path` from `String` to `PathBuf`.
 * Made `Song` metadata (artist, album, etc) `Option`s.

Cargo.lock

@@ -75,7 +75,7 @@ checksum = "cf1de2fe8c75bc145a2f577add951f8134889b4795d47466a54a5c846d691693"
 
 [[package]]
 name = "bliss-audio"
-version = "0.3.0"
+version = "0.3.1"
 dependencies = [
  "bliss-audio-aubio-rs",
  "crossbeam",

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "bliss-audio"
-version = "0.3.0"
+version = "0.3.1"
 authors = ["Polochon-street <polochonstreet@gmx.fr>"]
 edition = "2018"
 license = "GPL-3.0-only"

src/chroma.rs

@@ -556,8 +556,8 @@ mod bench {
     use ndarray::{arr2, Array1, Array2};
     use ndarray_npy::ReadNpyExt;
     use std::fs::File;
-    use test::Bencher;
     use std::path::Path;
+    use test::Bencher;
 
     #[bench]
     fn bench_estimate_tuning(b: &mut Bencher) {

src/distance.rs

@@ -0,0 +1,75 @@
+//! Module containing various distance metric functions.
+//!
+//! All of these functions are intended to be used with the
+//! [custom_distance](Song::custom_distance) method, or with
+//! [playlist_from_songs_custom_distance](Library::playlist_from_song_custom_distance).
+//!
+//! They will yield different styles of playlists, so don't hesitate to
+//! experiment with them if the default (euclidean distance for now) doesn't
+//! suit you.
+use crate::NUMBER_FEATURES;
+#[cfg(doc)]
+use crate::{Library, Song};
+use ndarray::{Array, Array1};
+
+/// Convenience trait for user-defined distance metrics.
+pub trait DistanceMetric: Fn(&Array1<f32>, &Array1<f32>) -> f32 {}
+impl<F> DistanceMetric for F where F: Fn(&Array1<f32>, &Array1<f32>) -> f32 {}
+
+/// Return the [euclidean
+/// distance](https://en.wikipedia.org/wiki/Euclidean_distance#Higher_dimensions)
+/// between two vectors.
+pub fn euclidean_distance(a: &Array1<f32>, b: &Array1<f32>) -> f32 {
+    // Could be any square symmetric positive semi-definite matrix;
+    // just no metric learning has been done yet.
+    // See https://lelele.io/thesis.pdf chapter 4.
+    let m = Array::eye(NUMBER_FEATURES);
+
+    (a - b).dot(&m).dot(&(a - b)).sqrt()
+}
+
+/// Return the [cosine
+/// distance](https://en.wikipedia.org/wiki/Cosine_similarity#Angular_distance_and_similarity)
+/// between two vectors.
+pub fn cosine_distance(a: &Array1<f32>, b: &Array1<f32>) -> f32 {
+    let similarity = a.dot(b) / (a.dot(a).sqrt() * b.dot(b).sqrt());
+    1. - similarity
+}
+
+#[cfg(test)]
+mod test {
+    use super::*;
+    use ndarray::arr1;
+
+    #[test]
+    fn test_euclidean_distance() {
+        let a = arr1(&[
+            1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 0.,
+        ]);
+        let b = arr1(&[
+            0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0.,
+        ]);
+        assert_eq!(euclidean_distance(&a, &b), 4.242640687119285);
+
+        let a = arr1(&[0.5; 20]);
+        let b = arr1(&[0.5; 20]);
+        assert_eq!(euclidean_distance(&a, &b), 0.);
+        assert_eq!(euclidean_distance(&a, &b), 0.);
+    }
+
+    #[test]
+    fn test_cosine_distance() {
+        let a = arr1(&[
+            1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 0.,
+        ]);
+        let b = arr1(&[
+            0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0.,
+        ]);
+        assert_eq!(cosine_distance(&a, &b), 0.7705842661294382);
+
+        let a = arr1(&[0.5; 20]);
+        let b = arr1(&[0.5; 20]);
+        assert_eq!(cosine_distance(&a, &b), 0.);
+        assert_eq!(cosine_distance(&a, &b), 0.);
+    }
+}

src/lib.rs

@@ -65,6 +65,7 @@
 #![warn(missing_docs)]
 #![warn(missing_doc_code_examples)]
 mod chroma;
+pub mod distance;
 mod library;
 mod misc;
 mod song;
@@ -80,7 +81,7 @@ extern crate serde;
 use thiserror::Error;
 
 pub use library::Library;
-pub use song::{Analysis, AnalysisIndex, NUMBER_FEATURES, Song};
+pub use song::{Analysis, AnalysisIndex, Song, NUMBER_FEATURES};
 
 const CHANNELS: u16 = 1;
 const SAMPLE_RATE: u32 = 22050;
@@ -178,7 +179,11 @@ mod tests {
 
         let mut analysed_songs: Vec<String> = results
             .iter()
-            .filter_map(|x| x.as_ref().ok().map(|x| x.path.to_str().unwrap().to_string()))
+            .filter_map(|x| {
+                x.as_ref()
+                    .ok()
+                    .map(|x| x.path.to_str().unwrap().to_string())
+            })
             .collect();
         analysed_songs.sort_by(|a, b| a.cmp(b));
 

src/library.rs

@@ -1,5 +1,8 @@
 //! Module containing the Library trait, useful to get started to implement
 //! a plug-in for an audio player.
+#[cfg(doc)]
+use crate::distance;
+use crate::distance::DistanceMetric;
 use crate::{BlissError, BlissResult, Song};
 use log::{debug, error, info};
 use noisy_float::prelude::*;
@@ -42,15 +45,71 @@ pub trait Library {
         first_song: Song,
         playlist_length: usize,
     ) -> BlissResult<Vec<Song>> {
-        let analysis_current_song = first_song.analysis;
         let mut songs = self.get_stored_songs()?;
-        songs.sort_by_cached_key(|song| n32(analysis_current_song.distance(&song.analysis)));
+        songs.sort_by_cached_key(|song| n32(first_song.distance(&song)));
 
         let playlist = songs
             .into_iter()
             .take(playlist_length)
             .collect::<Vec<Song>>();
-        debug!("Playlist created: {:?}", playlist);
+        debug!(
+            "Playlist created: {}",
+            playlist
+                .iter()
+                .map(|s| format!("{:?}", &s))
+                .collect::<Vec<String>>()
+                .join("\n"),
+        );
+        Ok(playlist)
+    }
+
+    /// Return a list of songs that are similar to ``first_song``, using a
+    /// custom distance metric.
+    ///
+    /// # Arguments
+    ///
+    /// * `first_song` - The song the playlist will be built from.
+    /// * `playlist_length` - The playlist length. If there are not enough
+    /// songs in the library, it will be truncated to the size of the library.
+    /// * `distance` - a user-supplied valid distance metric, either taken
+    /// from the [distance](distance) module, or made from scratch.
+    ///
+    /// # Returns
+    ///
+    /// A vector of `playlist_length` Songs, including `first_song`, that you
+    /// most likely want to plug in your audio player by using something like
+    /// `ret.map(|song| song.path.to_owned()).collect::<Vec<String>>()`.
+    ///
+    /// # Custom distance example
+    ///
+    /// ```
+    /// use ndarray::Array1;
+    ///
+    /// fn manhattan_distance(a: &Array1<f32>, b: &Array1<f32>) -> f32 {
+    ///     (a - b).mapv(|x| x.abs()).sum()
+    /// }
+    /// ```
+    fn playlist_from_song_custom_distance(
+        &self,
+        first_song: Song,
+        playlist_length: usize,
+        distance: impl DistanceMetric,
+    ) -> BlissResult<Vec<Song>> {
+        let mut songs = self.get_stored_songs()?;
+        songs.sort_by_cached_key(|song| n32(first_song.custom_distance(&song, &distance)));
+
+        let playlist = songs
+            .into_iter()
+            .take(playlist_length)
+            .collect::<Vec<Song>>();
+        debug!(
+            "Playlist created: {}",
+            playlist
+                .iter()
+                .map(|s| format!("{:?}", &s))
+                .collect::<Vec<String>>()
+                .join("\n"),
+        );
         Ok(playlist)
     }
 
@@ -95,9 +154,13 @@ pub trait Library {
             // A storage fail should just warn the user, but not abort the whole process
             match song {
                 Ok(song) => {
-                    self.store_song(&song)
+                    self.store_song(&song).unwrap_or_else(|e| {
-                        .unwrap_or_else(|_| error!("Error while storing song '{}'", song.path.display()));
+                        error!("Error while storing song '{}': {}", song.path.display(), e)
-                    info!("Analyzed and stored song '{}' successfully.", song.path.display())
+                    });
+                    info!(
+                        "Analyzed and stored song '{}' successfully.",
+                        song.path.display()
+                    )
                 }
                 Err(e) => {
                     self.store_error_song(path.to_string(), e.to_owned())
@@ -135,6 +198,7 @@ pub trait Library {
 mod test {
     use super::*;
     use crate::song::Analysis;
+    use ndarray::Array1;
     use std::path::Path;
 
     #[derive(Default)]
@@ -158,11 +222,7 @@ mod test {
             Ok(())
         }
 
-        fn store_error_song(
+        fn store_error_song(&mut self, song_path: String, error: BlissError) -> BlissResult<()> {
-            &mut self,
-            song_path: String,
-            error: BlissError,
-        ) -> BlissResult<()> {
             self.failed_files.push((song_path, error.to_string()));
             Ok(())
         }
@@ -221,11 +281,7 @@ mod test {
             Ok(vec![])
         }
 
-        fn store_error_song(
+        fn store_error_song(&mut self, song_path: String, error: BlissError) -> BlissResult<()> {
-            &mut self,
-            song_path: String,
-            error: BlissError,
-        ) -> BlissResult<()> {
             Err(BlissError::ProviderError(format!(
                 "Could not store errored song: {}, with error: {}",
                 song_path, error
@@ -384,4 +440,52 @@ mod test {
         let mut test_library = TestLibrary::default();
         assert!(test_library.analyze_paths(vec![]).is_ok());
     }
+
+    fn custom_distance(a: &Array1<f32>, b: &Array1<f32>) -> f32 {
+        if a == b {
+            return 0.;
+        }
+        1. / (a.first().unwrap() - b.first().unwrap()).abs()
+    }
+
+    #[test]
+    fn test_playlist_from_song_custom_distance() {
+        let mut test_library = TestLibrary::default();
+        let first_song = Song {
+            path: Path::new("path-to-first").to_path_buf(),
+            analysis: Analysis::new([0.; 20]),
+            ..Default::default()
+        };
+
+        let second_song = Song {
+            path: Path::new("path-to-second").to_path_buf(),
+            analysis: Analysis::new([0.1; 20]),
+            ..Default::default()
+        };
+
+        let third_song = Song {
+            path: Path::new("path-to-third").to_path_buf(),
+            analysis: Analysis::new([10.; 20]),
+            ..Default::default()
+        };
+
+        let fourth_song = Song {
+            path: Path::new("path-to-fourth").to_path_buf(),
+            analysis: Analysis::new([20.; 20]),
+            ..Default::default()
+        };
+
+        test_library.internal_storage = vec![
+            first_song.to_owned(),
+            fourth_song.to_owned(),
+            third_song.to_owned(),
+            second_song.to_owned(),
+        ];
+        assert_eq!(
+            vec![first_song.to_owned(), fourth_song, third_song],
+            test_library
+                .playlist_from_song_custom_distance(first_song, 3, custom_distance)
+                .unwrap()
+        );
+    }
 }

src/song.rs

@@ -14,6 +14,7 @@ extern crate ndarray_npy;
 
 use super::CHANNELS;
 use crate::chroma::ChromaDesc;
+use crate::distance::{euclidean_distance, DistanceMetric};
 use crate::misc::LoudnessDesc;
 use crate::temporal::BPMDesc;
 use crate::timbral::{SpectralDesc, ZeroCrossingRateDesc};
@@ -31,13 +32,13 @@ use ffmpeg_next::util::format::sample::{Sample, Type};
 use ffmpeg_next::util::frame::audio::Audio;
 use ffmpeg_next::util::log;
 use ffmpeg_next::util::log::level::Level;
-use ndarray::{arr1, Array, Array1};
+use ndarray::{arr1, Array1};
 use std::convert::TryInto;
 use std::fmt;
-use std::sync::mpsc;
-use std::sync::mpsc::Receiver;
 use std::path::Path;
 use std::path::PathBuf;
+use std::sync::mpsc;
+use std::sync::mpsc::Receiver;
 use std::thread as std_thread;
 use strum::{EnumCount, IntoEnumIterator};
 use strum_macros::{EnumCount, EnumIter};
@@ -173,34 +174,55 @@ impl Analysis {
         self.internal_analysis.to_vec()
     }
 
-    /// Return the [euclidean
+    /// Compute distance between two analysis using a user-provided distance
-    /// distance](https://en.wikipedia.org/wiki/Euclidean_distance#Higher_dimensions)
+    /// metric. You most likely want to use `song.custom_distance` directly
-    /// between two analysis.
+    /// rather than this function.
     ///
-    /// Note that it is usually easier to just use [`song.distance(song2)`](Song::distance)
+    /// For this function to be integrated properly with the rest
-    /// (which calls this function in turn).
+    /// of bliss' parts, it should be a valid distance metric, i.e.:
-    pub fn distance(&self, other: &Self) -> f32 {
+    /// 1. For X, Y real vectors, d(X, Y) = 0 ⇔ X = Y
-        let a1 = self.to_arr1();
+    /// 2. For X, Y real vectors, d(X, Y) >= 0
-        let a2 = other.to_arr1();
+    /// 3. For X, Y real vectors, d(X, Y) = d(Y, X)
-        // Could be any square symmetric positive semi-definite matrix;
+    /// 4. For X, Y, Z real vectors d(X, Y) ≤ d(X + Z) + d(Z, Y)
-        // just no metric learning has been done yet.
+    ///
-        // See https://lelele.io/thesis.pdf chapter 4.
+    /// Note that almost all distance metrics you will find obey these
-        let m = Array::eye(NUMBER_FEATURES);
+    /// properties, so don't sweat it too much.
-
+    pub fn custom_distance(&self, other: &Self, distance: impl DistanceMetric) -> f32 {
-        (self.to_arr1() - &a2).dot(&m).dot(&(&a1 - &a2)).sqrt()
+        distance(&self.to_arr1(), &other.to_arr1())
     }
 }
 
 impl Song {
     #[allow(dead_code)]
-    /// Compute the distance between the current song and any given Song.
+    /// Compute the distance between the current song and any given
+    /// Song.
     ///
     /// The smaller the number, the closer the songs; usually more useful
     /// if compared between several songs
     /// (e.g. if song1.distance(song2) < song1.distance(song3), then song1 is
     /// closer to song2 than it is to song3.
+    ///
+    /// Currently uses the euclidean distance, but this can change in an
+    /// upcoming release if another metric performs better.
     pub fn distance(&self, other: &Self) -> f32 {
-        self.analysis.distance(&other.analysis)
+        self.analysis
+            .custom_distance(&other.analysis, euclidean_distance)
+    }
+
+    /// Compute distance between two songs using a user-provided distance
+    /// metric.
+    ///
+    /// For this function to be integrated properly with the rest
+    /// of bliss' parts, it should be a valid distance metric, i.e.:
+    /// 1. For X, Y real vectors, d(X, Y) = 0 ⇔ X = Y
+    /// 2. For X, Y real vectors, d(X, Y) >= 0
+    /// 3. For X, Y real vectors, d(X, Y) = d(Y, X)
+    /// 4. For X, Y, Z real vectors d(X, Y) ≤ d(X + Z) + d(Z, Y)
+    ///
+    /// Note that almost all distance metrics you will find obey these
+    /// properties, so don't sweat it too much.
+    pub fn custom_distance(&self, other: &Self, distance: impl DistanceMetric) -> f32 {
+        self.analysis.custom_distance(&other.analysis, distance)
     }
 
     /// Returns a decoded Song given a file path, or an error if the song
@@ -261,25 +283,23 @@ impl Song {
         }
 
         thread::scope(|s| {
-            let child_tempo: thread::ScopedJoinHandle<'_, BlissResult<f32>> =
+            let child_tempo: thread::ScopedJoinHandle<'_, BlissResult<f32>> = s.spawn(|_| {
-                s.spawn(|_| {
+                let mut tempo_desc = BPMDesc::new(SAMPLE_RATE)?;
-                    let mut tempo_desc = BPMDesc::new(SAMPLE_RATE)?;
+                let windows = sample_array
-                    let windows = sample_array
+                    .windows(BPMDesc::WINDOW_SIZE)
-                        .windows(BPMDesc::WINDOW_SIZE)
+                    .step_by(BPMDesc::HOP_SIZE);
-                        .step_by(BPMDesc::HOP_SIZE);
 
-                    for window in windows {
+                for window in windows {
-                        tempo_desc.do_(&window)?;
+                    tempo_desc.do_(&window)?;
-                    }
+                }
-                    Ok(tempo_desc.get_value())
+                Ok(tempo_desc.get_value())
-                });
+            });
 
-            let child_chroma: thread::ScopedJoinHandle<'_, BlissResult<Vec<f32>>> =
+            let child_chroma: thread::ScopedJoinHandle<'_, BlissResult<Vec<f32>>> = s.spawn(|_| {
-                s.spawn(|_| {
+                let mut chroma_desc = ChromaDesc::new(SAMPLE_RATE, 12);
-                    let mut chroma_desc = ChromaDesc::new(SAMPLE_RATE, 12);
+                chroma_desc.do_(&sample_array)?;
-                    chroma_desc.do_(&sample_array)?;
+                Ok(chroma_desc.get_values())
-                    Ok(chroma_desc.get_values())
+            });
-                });
 
             #[allow(clippy::type_complexity)]
             let child_timbral: thread::ScopedJoinHandle<
@@ -305,8 +325,8 @@ impl Song {
                 Ok(zcr_desc.get_value())
             });
 
-            let child_loudness: thread::ScopedJoinHandle<'_, BlissResult<Vec<f32>>> = s
+            let child_loudness: thread::ScopedJoinHandle<'_, BlissResult<Vec<f32>>> =
-                .spawn(|_| {
+                s.spawn(|_| {
                     let mut loudness_desc = LoudnessDesc::default();
                     let windows = sample_array.chunks(LoudnessDesc::WINDOW_SIZE);
 
@@ -390,7 +410,6 @@ impl Song {
                 "" => None,
                 a => Some(a.to_string()),
             };
-
         };
         if let Some(album) = format.metadata().get("album") {
             song.album = match album {
@@ -804,14 +823,35 @@ mod tests {
             format!("{:?}", song.analysis),
         );
     }
+
+    fn dummy_distance(_: &Array1<f32>, _: &Array1<f32>) -> f32 {
+        0.
+    }
+    #[test]
+    fn test_custom_distance() {
+        let mut a = Song::default();
+        a.analysis = Analysis::new([
+            0.16391512, 0.11326739, 0.96868552, 0.8353934, 0.49867523, 0.76532606, 0.63448005,
+            0.82506196, 0.71457147, 0.62395476, 0.69680329, 0.9855766, 0.41369333, 0.13900452,
+            0.68001012, 0.11029723, 0.97192943, 0.57727861, 0.07994821, 0.88993185,
+        ]);
+
+        let mut b = Song::default();
+        b.analysis = Analysis::new([
+            0.5075758, 0.36440256, 0.28888011, 0.43032829, 0.62387977, 0.61894916, 0.99676086,
+            0.11913155, 0.00640396, 0.15943407, 0.33829514, 0.34947174, 0.82927523, 0.18987604,
+            0.54437275, 0.22076826, 0.91232151, 0.29233168, 0.32846024, 0.04522147,
+        ]);
+        assert_eq!(a.custom_distance(&b, dummy_distance), 0.);
+    }
 }
 
 #[cfg(all(feature = "bench", test))]
 mod bench {
     extern crate test;
     use crate::Song;
-    use test::Bencher;
     use std::path::Path;
+    use test::Bencher;
 
     #[bench]
     fn bench_resample_multi(b: &mut Bencher) {

src/utils.rs

@@ -525,8 +525,8 @@ mod bench {
     use super::*;
     use crate::Song;
     use ndarray::Array;
-    use test::Bencher;
     use std::path::Path;
+    use test::Bencher;
 
     #[bench]
     fn bench_convolve(b: &mut Bencher) {
@@ -540,7 +540,9 @@ mod bench {
 
     #[bench]
     fn bench_compute_stft(b: &mut Bencher) {
-        let signal = Song::decode(Path::new("data/piano.flac")).unwrap().sample_array;
+        let signal = Song::decode(Path::new("data/piano.flac"))
+            .unwrap()
+            .sample_array;
 
         b.iter(|| {
             stft(&signal, 2048, 512);