Merge pull request #66 from Polochon-street/add-chroma

Add chroma filter generator stuff

bliss-rs/Cargo.toml

@@ -10,14 +10,15 @@ edition = "2018"
 build-ffmpeg = ["ffmpeg4/build"]
 
 [dependencies]
-ripemd160 = "*"
+ripemd160 = "0.9.0"
 
 [dependencies.aubio-rs]
 git = "https://github.com/katyo/aubio-rs"
 
 [dependencies.aubio-lib]
-git = "https://github.com/katyo/aubio-rs"
+git = "https://github.com/Polochon-street/aubio-rs"
 optional = true
+branch = "fix-python-version"
 
 [dependencies.ffmpeg4]
 # Like this until https://github.com/bacek/rust-ffmpeg4/pull/5 is merged

bliss-rs/src/chroma.rs

@@ -0,0 +1,155 @@
+#[cfg(feature = "aubio-lib")]
+extern crate aubio_lib;
+
+use crate::utils::hz_to_octs;
+
+// chroma(22050, n_fft=5, n_chroma=12)
+pub fn chroma_filter(sample_rate: u32, n_fft: u32, n_chroma: u32, tuning: f64) -> Vec<Vec<f32>> {
+    let step = sample_rate as f64 / n_fft as f64;
+    let ctroct = 5.0;
+    let octwidth = 2;
+
+    // [4410.0, 8820.0, 13230.0, 17640.0]
+    let frequencies = (1..n_fft).map(|i| i as f64 * step).collect::<Vec<f64>>();
+
+    // [87.90243872,  99.90243872, 106.92198873, 111.90243872]
+    let temp_freq_bins = hz_to_octs(&frequencies, tuning, n_chroma)
+        .iter()
+        .map(|i| i * n_chroma as f64)
+        .collect::<Vec<f64>>();
+    // [69.90243872,  87.90243872,  99.90243872, 106.92198873, 111.90243872]
+    let mut freq_bins = vec![temp_freq_bins[0] - 1.5 * n_chroma as f64];
+    freq_bins.extend_from_slice(temp_freq_bins.as_slice());
+
+    // [18., 12., 7.01955001, 4.98044999, 1.]
+    let mut binwidth_bin = (&freq_bins[1..freq_bins.len()])
+        .iter()
+        .zip(&freq_bins[0..freq_bins.len() - 1])
+        .map(|(x, y)| (x - y).max(1.0))
+        .collect::<Vec<f64>>();
+    binwidth_bin.extend_from_slice(&[1.0]);
+
+    // [[69.90243872, 87.90243872, 99.90243872, 106.92198873, 111.90243872]
+    // ...
+    // [58.90243872, 76.90243872, 88.90243872, 95.92198873, 100.90243872]]
+    let d = (0..n_chroma)
+        .map(|i| freq_bins.iter().map(|f| f - i as f64).collect::<Vec<f64>>())
+        .collect::<Vec<Vec<f64>>>();
+
+    let n_chroma2 = (n_chroma as f64 / 2.0).round() as u32;
+
+    // [[-2.09756128,  3.90243872,  3.90243872, -1.07801127,  3.90243872]
+    // ...
+    // [-1.09756128,  4.90243872,  4.90243872, -0.07801127,  4.90243872]]
+    let d = d
+        .iter()
+        .map(|s| {
+            s.iter()
+                .map(|v| {
+                    (v + n_chroma2 as f64 + 10. * n_chroma as f64) % n_chroma as f64
+                        - n_chroma2 as f64
+                })
+                .collect::<Vec<f64>>()
+        })
+        .collect::<Vec<Vec<f64>>>();
+
+    // [[9.73206457e-01, 8.09357725e-01, 5.38948409e-01, 9.10555920e-01, 5.91880948e-14],
+    // ...
+    // [9.92591525e-01, 7.16193969e-01, 3.76996585e-01, 9.99509430e-01, 1.33172632e-21]]
+    let wts = d
+        .iter()
+        .map(|s| {
+            s.iter()
+                .zip(&binwidth_bin)
+                .map(|(v, b)| (-0.5 * ((2.0 * v / b) as f64).powf(2.0)).exp())
+                .collect::<Vec<f64>>()
+        })
+        .collect::<Vec<Vec<f64>>>();
+
+    // Normalize by computing the l2-norm over the columns
+    let mut length = vec![0.; n_fft as usize];
+    for i in 0..(n_fft as usize) {
+        let mut vec = vec![0.; n_fft as usize];
+        for x in &wts {
+            vec.push(x[i].powf(2.0));
+        }
+        length[i] = vec.iter().map(|x| x).sum::<f64>().sqrt();
+    }
+
+    // [3.01362739e-01, 2.70504696e-01, 2.17863058e-01, 4.33578438e-01, 5.89176039e-14]
+    // ...
+    // [3.07365511e-01, 2.39367373e-01, 1.52396088e-01, 4.75935336e-01, 1.32564030e-21]
+    let wts = wts
+        .iter()
+        .map(|s| {
+            s.iter()
+                .enumerate()
+                .map(|(i, v)| v / (&length)[i])
+                .collect::<Vec<f64>>()
+        })
+        .collect::<Vec<Vec<f64>>>();
+
+    // [0.91840203, 0.50873844, 0.25104525, 0.14790657, 0.09647968]
+    let scaling = &freq_bins
+        .iter()
+        .map(|f| (-0.5 * (((f / n_chroma as f64 - ctroct) / octwidth as f64).powf(2.0))).exp())
+        .collect::<Vec<f64>>();
+
+    // [[2.76772150e-01, 1.37616138e-01, 5.46934868e-02, 6.41290987e-02, 5.68435153e-15]
+    // ...
+    // [2.82285109e-01, 1.21775385e-01, 3.82583145e-02, 7.03939621e-02, 1.27897351e-22]]
+    let mut wts = wts
+        .iter()
+        .map(|s| {
+            s.iter()
+                .zip(scaling)
+                .map(|(v, f)| v * f)
+                .collect::<Vec<f64>>()
+        })
+        .collect::<Vec<Vec<f64>>>();
+    // "if base_c"
+    // [[2.42245481e-01, 1.68118894e-01, 9.81821393e-02, 1.84251933e-02, 1.88395920e-02]
+    // ...
+    // [2.56393048e-01, 1.61695478e-01, 8.76171131e-02, 3.28090351e-02, 6.89899611e-05]]
+    wts.rotate_left(3);
+
+    let non_aliased = 1 + n_fft / 2;
+
+    wts.iter()
+        .map(|s| {
+            s.iter()
+                .take(non_aliased as usize)
+                .map(|f| *f as f32)
+                .collect::<Vec<f32>>()
+        })
+        .collect::<Vec<Vec<f32>>>()
+}
+
+mod test {
+    use super::*;
+
+    #[test]
+    fn test_generate_chroma() {
+        let expected_chroma: Vec<Vec<f32>> = vec![
+            vec![0.24224548, 0.16811889, 0.09818214],
+            vec![0.22936151, 0.17000881, 0.10144266],
+            vec![0.24518087, 0.16721014, 0.09663921],
+            vec![0.25887551, 0.15995214, 0.08488495],
+            vec![0.2699813, 0.14881742, 0.06874682],
+            vec![0.27810881, 0.13466469, 0.05133566],
+            vec![0.28296593, 0.11851955, 0.03534518],
+            vec![0.28437531, 0.10480594, 0.0246752],
+            vec![0.28228511, 0.12177538, 0.03825831],
+            vec![0.27677215, 0.13761614, 0.05469349],
+            vec![0.26803725, 0.151257, 0.07209248],
+            vec![0.25639305, 0.16169548, 0.08761711],
+        ];
+        let filter = chroma_filter(22050, 5, 12, 0.0);
+
+        for (column1, column2) in expected_chroma.iter().zip(filter.iter()) {
+            for (val1, val2) in column1.iter().zip(column2) {
+                assert!(0.0001 > (val1 - val2).abs());
+            }
+        }
+    }
+}

bliss-rs/src/decode.rs

@@ -147,10 +147,10 @@ mod tests {
         let song = decode_song(path).unwrap();
         let mut hasher = Ripemd160::new();
         for sample in song.sample_array.iter() {
-            hasher.input(sample.to_le_bytes().to_vec());
+            hasher.update(sample.to_le_bytes().to_vec());
         }
 
-        assert_eq!(expected_hash, hasher.result().as_slice());
+        assert_eq!(expected_hash, hasher.finalize().as_slice());
     }
 
     #[test]

bliss-rs/src/lib.rs

@@ -1,5 +1,6 @@
 // temporarily pub
 pub mod analyze;
+pub mod chroma;
 pub mod decode;
 pub mod misc;
 pub mod temporal;

bliss-rs/src/utils.rs

@@ -44,6 +44,12 @@ pub fn geometric_mean(input: &[f32]) -> f32 {
     mean.exp()
 }
 
+pub fn hz_to_octs(frequencies: &[f64], tuning: f64, bins_per_octave: u32) -> Vec<f64> {
+    let a440 = 440.0 * (2_f64.powf(tuning / bins_per_octave as f64) as f64);
+
+    frequencies.iter().map(|freq| (freq / (a440 / 16.)).log2()).collect()
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -62,4 +68,13 @@ mod tests {
         let numbers = vec![4.0, 1.0, 0.03125];
         assert_eq!(0.5, geometric_mean(&numbers));
     }
+
+    #[test]
+    fn test_hz_to_octs() {
+        let frequencies = vec![32., 64., 128., 256.];
+        let expected = vec![0.16864029, 1.16864029, 2.16864029, 3.16864029];
+
+        let octs = hz_to_octs(&frequencies, 0.5, 10);
+        octs.iter().zip(expected.iter()).for_each(|(x, y) | assert!(0.0001 > (x - y).abs()));
+    }
 }