Merge pull request #117 from napulen/rntxt_bug

Evaluation bugs

AugmentedNet/annotation_parser.py

@@ -40,7 +40,7 @@
 
 
 def _m21Parse(f):
-    return music21.converter.parse(f, format="romantext")
+    return music21.converter.parse(f, format="romantext", forceSource=True)
 
 
 def from_tsv(tsv, sep="\t"):

AugmentedNet/inference.py

@@ -11,7 +11,7 @@
 
 from . import __version__
 from . import cli
-from .chord_vocabulary import frompcset
+from .chord_vocabulary import frompcset, cosineSimilarity
 from .cache import forceTonicization, getTonicizationScaleDegree
 from .score_parser import parseScore
 from .input_representations import available_representations as availableInputs
@@ -56,15 +56,27 @@ def solveChordSegmentation(df):
     return df.dropna()[df.HarmonicRhythm7 == 0]
 
 
-def resolveRomanNumeral(b, t, a, s, pcs, key, tonicizedKey):
+def resolveRomanNumeralCosine(b, t, a, s, pcs, key, numerator, tonicizedKey):
+    pcsetVector = np.zeros(12)
     chord = music21.chord.Chord(f"{b}2 {t}3 {a}4 {s}5")
-    pcset = tuple(sorted(set(chord.pitchClasses)))
-    # if the SATB notes don't make sense, use the pcset classifier
-    if pcset not in frompcset:
-        # which is guaranteed to exist in the chord vocabulary
-        pcset = pcs
-    # if the chord is nondiatonic to the tonicizedKey
-    # force a tonicization where the chord does exist
+    for n in chord.pitches:
+        pcsetVector[n.pitchClass] += 1
+    for pc in pcs:
+        pcsetVector[pc] += 1
+    chordNumerator = music21.roman.RomanNumeral(
+        numerator.replace("Cad", "Cad64"), tonicizedKey
+    ).pitchClasses
+    for pc in chordNumerator:
+        pcsetVector[pc] += 1
+    smallestDistance = -2
+    for pcs in frompcset:
+        v2 = np.zeros(12)
+        for p in pcs:
+            v2[p] = 1
+        similarity = cosineSimilarity(pcsetVector, v2)
+        if similarity > smallestDistance:
+            pcset = pcs
+            smallestDistance = similarity
     if tonicizedKey not in frompcset[pcset]:
         # print("Forcing a tonicization")
         candidateKeys = list(frompcset[pcset].keys())
@@ -83,6 +95,8 @@ def resolveRomanNumeral(b, t, a, s, pcs, key, tonicizedKey):
     elif invfigure in ["6", "64"]:
         rnfigure += invfigure
     rn = rnfigure
+    if numerator == "Cad" and inv == 2:
+        rn = "Cad64"
     if tonicizedKey != key:
         denominator = getTonicizationScaleDegree(key, tonicizedKey)
         rn = f"{rn}/{denominator}"
@@ -92,22 +106,17 @@ def resolveRomanNumeral(b, t, a, s, pcs, key, tonicizedKey):
     return rn, chordLabel
 
 
-def generateRomanText(h):
+def generateRomanText(h, ts):
     metadata = h.metadata
     metadata.composer = metadata.composer or "Unknown"
     metadata.title = metadata.title or "Unknown"
     composer = metadata.composer.split("\n")[0]
     title = metadata.title.split("\n")[0]
-    ts = {
-        (ts.measureNumber, float(ts.beat)): ts.ratioString
-        for ts in h.flat.getElementsByClass("TimeSignature")
-    }
     rntxt = f"""\
 Composer: {composer}
 Title: {title}
 Analyst: AugmentedNet v{__version__} - https://github.com/napulen/AugmentedNet
 """
-    setKey = False
     currentMeasure = -1
     for n in h.flat.notes:
         if not n.lyric:
@@ -160,14 +169,20 @@ def predict(model, inputPath):
     dfout["offset"] = paddedIndex
     dfout["measure"] = paddedMeasure
     chords = solveChordSegmentation(dfout)
-    s = music21.converter.parse(inputPath)
+    s = music21.converter.parse(inputPath, forceSource=True)
+    ts = {
+        (ts.measureNumber, float(ts.beat)): ts.ratioString
+        for ts in s.flat.getElementsByClass("TimeSignature")
+    }
+    schord = s.chordify().flat.notesAndRests
+    schord.metadata = s.metadata
     # remove all lyrics from score
-    for note in s.recurse().notes:
-        note.lyrics = []
+    # for note in s.recurse().notes:
+    #     note.lyrics = []
     prevkey = ""
     for analysis in chords.itertuples():
         notes = []
-        for n in s.flat.notes.getElementsByOffset(analysis.offset):
+        for n in schord.getElementsByOffset(analysis.offset):
             if isinstance(n, music21.note.Note):
                 notes.append((n, n.pitch.midi))
             elif isinstance(n, music21.chord.Chord) and not isinstance(
@@ -180,13 +195,15 @@ def predict(model, inputPath):
         thiskey = analysis.LocalKey38
         tonicizedKey = analysis.TonicizedKey38
         pcset = analysis.PitchClassSet121
-        rn2, chordLabel = resolveRomanNumeral(
+        numerator = analysis.RomanNumeral31
+        rn2, chordLabel = resolveRomanNumeralCosine(
             analysis.Bass35,
             analysis.Tenor35,
             analysis.Alto35,
             analysis.Soprano35,
             pcset,
             thiskey,
+            numerator,
             tonicizedKey,
         )
         if thiskey != prevkey:
@@ -196,12 +213,12 @@ def predict(model, inputPath):
             rn2fig = rn2
         bass.addLyric(formatRomanNumeral(rn2fig, thiskey))
         bass.addLyric(formatChordLabel(chordLabel))
-    rntxt = generateRomanText(s)
+    rntxt = generateRomanText(schord, ts)
     filename, _ = inputPath.rsplit(".", 1)
     annotatedScore = f"{filename}_annotated.musicxml"
     annotationCSV = f"{filename}_annotated.csv"
     annotatedRomanText = f"{filename}_annotated.rntxt"
-    s.write(fp=annotatedScore)
+    schord.write(fp=annotatedScore)
     dfout.to_csv(annotationCSV)
     with open(annotatedRomanText, "w") as fd:
         fd.write(rntxt)
@@ -224,6 +241,7 @@ def batch(inputPath, dir, modelPath, useGpu):
                 # do not recursively annotate an annotated_file
                 continue
             filepath = os.path.join(root, f)
+            print(filepath)
             predict(model, inputPath=filepath)
 
 

AugmentedNet/score_parser.py

@@ -38,7 +38,7 @@
 
 
 def _m21Parse(f, fmt=None):
-    s = music21.converter.parse(f, format=fmt)
+    s = music21.converter.parse(f, format=fmt, forceSource=True)
     perc = [
         p
         for p in s.parts

misc/compare_rntxt.py

@@ -2,15 +2,12 @@
 import pandas as pd
 import numpy as np
 import matplotlib.pyplot as plt
-from requests import PreparedRequest
 
 from AugmentedNet.annotation_parser import parseAnnotation
 from AugmentedNet.feature_representation import COMMON_ROMAN_NUMERALS
 
 
 def compare_annotation_df(a, b):
-    b = b.reindex_like(a, method="ffill")
-    b = b.fillna(method="bfill")
     pcset_accuracy = sum(a.a_pcset == b.a_pcset) / len(a.index)
     key_accuracy = sum(a.a_tonicizedKey == b.a_tonicizedKey) / len(a.index)
     inversion_accuracy = sum(a.a_inversion == b.a_inversion) / len(a.index)
@@ -19,8 +16,6 @@ def compare_annotation_df(a, b):
 
 
 def compute_confusion_matrix(a, b):
-    b = b.reindex_like(a, method="ffill")
-    b = b.fillna(method="bfill")
     roman_numeral_classes = len(COMMON_ROMAN_NUMERALS)
     matrix = np.zeros((roman_numeral_classes, roman_numeral_classes))
     for gt, pred in zip(a.a_romanNumeral, b.a_romanNumeral):
@@ -35,18 +30,6 @@ def compute_confusion_matrix(a, b):
     groundtruth = "predictions_groundtruth"
     models = [m for m in os.listdir(root) if m != groundtruth and "rntxt" in m]
     gtdir = os.path.join(root, groundtruth)
-    crashing_july_7_2022 = [
-        "abc-op127-2.rntxt",
-        "abc-op18-no1-1.rntxt",
-        "bps-07-op010-no3-1.rntxt",
-        "bps-10-op014-no2-1.rntxt",
-        "bps-23-op057-appassionata-1.rntxt",
-        "tavern-beethoven-woo-75-a.rntxt",
-        "tavern-beethoven-woo-75-b.rntxt",
-        "wir-openscore-liedercorpus-hensel-6-lieder-op-9-1-die-ersehnte.rntxt",
-        "wir-openscore-liedercorpus-schumann-dichterliebe-op-48-16-die-alten-bosen-lieder.rntxt",
-        "wirwtc-bach-wtc-i-8.rntxt",
-    ]
     dfdict = {
         "file": [],
         "model": [],
@@ -58,8 +41,8 @@ def compute_confusion_matrix(a, b):
     }
     i = 0
     for f in sorted(os.listdir(gtdir)):
-        if f in crashing_july_7_2022 or "wir-bach-chorales-19" not in f:
-            continue
+        # if "bps-07" not in f:
+        #     continue
         print(f)
         path = os.path.join(root, groundtruth, f)
         a = parseAnnotation(path)
@@ -68,7 +51,13 @@ def compute_confusion_matrix(a, b):
             if not os.path.exists(mpath):
                 continue
             print(f"\t{mpath}", end=" ")
-            b = parseAnnotation(mpath)
+            try:
+                b = parseAnnotation(mpath)
+            except:
+                print("\tFAILED")
+                continue
+            b = b.reindex_like(a, method="ffill")
+            b = b.fillna(method="bfill")
             diff = compare_annotation_df(a, b)
             confm = compute_confusion_matrix(a, b)
             plt.imshow(confm, cmap="Blues")