def benchmark(net, benchmark_dir):
r"""
Run a benchmark to gauge the quality of the ANN approach.
This function takes a fully trained ANN as well as a directory
containing only benchmark files as parameters.
"""
tracks = read_standard_tracks
hits = 0
misses = 0
# **note that there is one *chord context* per chord**
contexts = [context for piece in os.listdir(benchmark_dir)
for track in tracks(os.path.join(benchmark_dir, piece))
for context in chord_contexts(track)]
for context in contexts:
preceding = context[0]
theoretical_chord = physical2theoretical_chord(context[1])
following = context[2]
recent_notes = context[3]
chord_index = context[4]
limiting = context[5]
best_config = _best_sane_chord(preceding, theoretical_chord, following,
recent_notes, chord_index, limiting,
net)
LOG.info('Best config: {best_config}.'.format(**locals()))
LOG.info('Truth: {0}'.format(context[1]))
if best_config == context[1]:
hits += 1
else:
misses += 1
print('Hits: {hits}.\nMisses: {misses}'.format(**locals()))
print('Accuracy: {ratio}'.format(ratio = float(hits) / (hits + misses)))
def _lowest_note_chord(chord):
r"""
Return the highest note in a chord, in its theoretical notation.
Chords are given in their physical form.
>>> _lowest_note_chord({5: 5, 4: 5})
('A', 2)
"""
notes = physical2theoretical_chord(chord)
if notes:
return sorted(notes.keys(), key=cmp_to_key(compare_notes))[0]
def test_two_note_chord(self):
physical = {3: 7, 4: 7}
result = notemappings.physical2theoretical_chord(physical)
expected = Counter([('E', 3), ('A', 3)])
self.assertEqual(result, expected)
