import pickle

import librosa

import matplotlib.pyplot as plt

import numpy as np

from matplotlib.ticker import MultipleLocator

from sklearn.linear_model import LogisticRegression

from sklearn.metrics import confusion_matrix

from sklearn.model_selection import cross_val_predict

from sklearn.preprocessing import StandardScaler

from umap import UMAP

from xgboost import XGBClassifier

from representations import represent_w_lstsq_order, shift_root_to_first_column

import ml

import plotting

from pitches_problem import beats_chroma, analyse_audio, PitchesProblem, SongAnalysis

#%%

life_on_mars_fname = "/Users/bgeelen/Music/iTunes/iTunes Media/Music/Compilations/Life on Mars/02 Life on Mars_.mp3"

# life_on_mars_fname = '/Users/bgeelen/Downloads/Tones and I - Dance Monkey (Lyrics).mp3'

life_on_mars_wav, sr = librosa.load(life_on_mars_fname)

my_funny_valentine_fname = "/Users/bgeelen/Music/iTunes/iTunes Media/Music/Chet Baker/Chet Baker Sings/10 My Funny Valentine.mp3"

#%%

chromae = ["C", "C#", "D", "D#", "E", "F", "F#", "G", "G#", "A", "A#", "B"]

n_octaves = 5

n_bins = 12 * n_octaves + 1

hop_length = 128

sparsity = 0.9

fmin = librosa.note_to_hz("C2")

C = librosa.cqt(

)

abs_C = np.abs(C)

log_C = np.log(abs_C + 0.01)

#%%

h, w = C.shape

plt.figure(figsize=(7.5, 3), dpi=300)

plt.imshow(log_C[:, :10 * sr // hop_length], interpolation="nearest", aspect="auto")

plt.gca().invert_yaxis()

plt.ylabel("Harmonic Frequency")

# plt.title(f"sparsity: {sparsity}")

ytick_locations = list(range(0, h, 12))

ytick_texts = [f"C{2+i}" for i in range(len(ytick_locations))]

plt.yticks(ytick_locations, ytick_texts)

plt.xlabel("Time (s)")

xtick_locations = np.floor(np.arange(0, 11) * sr / hop_length)

xtick_texts = range(11)

plt.xticks(xtick_locations, xtick_texts)

plt.gca().yaxis.set_minor_locator(MultipleLocator(1))

plt.show()

#%%

tempo, beats = librosa.beat.beat_track(

onset_envelope=librosa.onset.onset_strength(S=abs_C), sr=sr, hop_length=hop_length,

trim=False

)

print(beats[:2])

#%%

plt.figure(figsize=(7.5, 3), dpi=300)

plt.imshow(log_C[:, :10*sr//hop_length], interpolation="nearest", aspect="auto")

plt.gca().invert_yaxis()

plt.ylabel("Harmonic Frequency")

plt.yticks(ytick_locations, ytick_texts)

plt.xlabel("Time (s)")

plt.xticks(xtick_locations, xtick_texts)

plt.gca().yaxis.set_minor_locator(MultipleLocator(1))

plt.vlines(beats[beats < 10 * sr // hop_length], 0, h - 1, "white")

plt.show()

#%%

plt.figure(figsize=(7.5, 3), dpi=300)

plt.imshow(log_C[:, :10*sr//hop_length], interpolation="nearest", aspect="auto", origin='lower')

plt.ylabel("Harmonic Frequency")

plt.yticks(ytick_locations, ytick_texts)

plt.xlabel("Time (s)")

plt.xticks(xtick_locations, xtick_texts)

plt.gca().yaxis.set_minor_locator(MultipleLocator(1))

plt.hlines(np.arange(0, h, 12), 0, 10*sr//hop_length, "white")

plt.show()

#%%

beat_aligned_chromagram = beats_chroma(C, beats, hop_length)

#%%

n_beats_to_show = 20

plt.figure(figsize=(7.5, 3), dpi=300)

plt.imshow(beat_aligned_chromagram[:n_beats_to_show].T, interpolation="nearest", aspect="auto")

plt.gca().invert_yaxis()

plt.ylabel("Chroma")

plt.yticks(range(12), chromae)

plt.xlabel("Time (Beats)")

plt.xticks(

range(0, n_beats_to_show, 4),

np.arange(0, n_beats_to_show, 4) + 1,

)

plt.gca().xaxis.set_minor_locator(MultipleLocator(1))

plt.show()

#%%

sa = analyse_audio(life_on_mars_fname, include_timbres=False)

sa_valentine = analyse_audio(my_funny_valentine_fname, include_timbres=False)

#%%

plt.figure(figsize=(10, 3), dpi=300)

plt.imshow(beat_aligned_chromagram.T, interpolation="nearest", aspect="auto")

plt.gca().invert_yaxis()

plt.ylabel("Chroma")

plt.yticks(range(12), chromae)

plt.xlabel("Measures")

plt.xticks(np.arange(0, len(beat_aligned_chromagram), 16), np.arange(0, len(beat_aligned_chromagram), 16) // 4 + 1)

plt.gca().xaxis.set_minor_locator(MultipleLocator(1))

plt.show()

#%%

def first_order_matrix(beats: np.ndarray):

x, _, _, _ = np.linalg.lstsq(beats[:-1, :], beats[1:, :])

return x

x = first_order_matrix(sa.beats)

x_valentine = first_order_matrix(sa_valentine.beats)

#%%

plt.figure(figsize=(4, 4), dpi=300)

plt.imshow(x, origin="upper")

plt.title("Life on Mars - David Bowie")

plt.xlim(-0.5, 11.5)

plt.ylim(-0.5, 11.5)

plt.xticks(range(12), chromae)

plt.yticks(range(12), chromae)

plt.ylabel('"Input" chromae')

plt.xlabel('"Predicted" chromae')

plt.gca().invert_yaxis()

plt.show()

#%%

plt.figure(figsize=(4, 4), dpi=300)

plt.imshow(x_valentine, origin="upper")

plt.title("My Funny Valentine - Chet Baker")

plt.xlim(-0.5, 11.5)

plt.ylim(-0.5, 11.5)

plt.ylabel('"Input" chromae')

plt.xlabel('"Predicted" chromae')

plt.xticks(range(12), chromae)

plt.yticks(range(12), chromae)

plt.gca().invert_yaxis()

plt.show()

#%%

# with open('ismir_genre_dataset.p', 'rb') as f:

# ismir_genre = pickle.load(f)

with open("gtzan.p", "rb") as f:

gtzan: PitchesProblem = pickle.load(f)

#%%

xs = np.vstack(

[first_order_matrix(sa.beats[:, :12]).flatten() for sa in gtzan.song_analyses]

)

#%%

linearmodel = LogisticRegression(fit_intercept=False)

linearmodel.fit(xs, gtzan.y)

#%%

plt.figure(figsize=(4, 4), dpi=300)

plt.imshow(

linearmodel.coef_[gtzan.labelnames.index("jazz")].reshape(12, 12), origin="upper"

)

plt.title('"Jazz" transitions')

plt.xlim(-0.5, 11.5)

plt.ylim(-0.5, 11.5)

plt.ylabel('"Input" chromae')

plt.xlabel('"Predicted" chromae')

plt.xticks(range(12), chromae)

plt.yticks(range(12), chromae)

plt.gca().invert_yaxis()

plt.show()

#%%

plt.figure(figsize=(4, 4), dpi=300)

plt.imshow(

linearmodel.coef_[gtzan.labelnames.index("rock")].reshape(12, 12), origin="upper"

)

plt.title('"Rock" transitions')

plt.xlim(-0.5, 11.5)

plt.ylim(-0.5, 11.5)

plt.ylabel('"Input" chromae')

plt.xlabel('"Predicted" chromae')

plt.xticks(range(12), chromae)

plt.yticks(range(12), chromae)

plt.gca().invert_yaxis()

plt.show()

#%%

model = XGBClassifier(n_estimators=1000)

predictions = cross_val_predict(model, xs, gtzan.y, cv=5)

#%%

mat = confusion_matrix(gtzan.y, predictions)

accuracy = ml.accuracy_from_confusion_matrix(mat)

plotting.imshow_confusion_matrix(

mat, gtzan.labelnames, f"GTZAN cross validation accuracy: {accuracy * 100:2.1f}%"

)

#%%

def fourth_order_lstsq(song_analysis):

return representation.flatten()

xs_lstsq_2 = np.vstack([fourth_order_lstsq(sa) for sa in gtzan.song_analyses])

#%%=

model = XGBClassifier(n_estimators=1000)

predictions = cross_val_predict(model, xs_lstsq_2, gtzan.y, cv=5)

#%%

mat = confusion_matrix(gtzan.y, predictions)

accuracy = ml.accuracy_from_confusion_matrix(mat)

plotting.imshow_confusion_matrix(

mat, gtzan.labelnames, f"GTZAN cross validation accuracy: {accuracy * 100:2.1f}%"

)

#%%

def represent_w_correlation(sa: SongAnalysis, order: int = 1):

return representation.flatten()

corr_bowie = represent_w_correlation(sa)

corr_valentine = represent_w_correlation(sa_valentine)

#%%

plt.figure(figsize=(4, 4), dpi=300)

plt.imshow(

corr_bowie.reshape(12, 12), origin="upper"

)

plt.title('Chroma covariances: Life on Mars ')

plt.xlim(-0.5, 11.5)

plt.ylim(-0.5, 11.5)

plt.ylabel('"Input" chromae')

plt.xlabel('"Predicted" chromae')

plt.xticks(range(12), chromae)

plt.yticks(range(12), chromae)

plt.gca().invert_yaxis()

plt.show()

#%%

plt.figure(figsize=(4, 4), dpi=300)

plt.imshow(

corr_valentine.reshape(12, 12), origin="upper"

)

plt.title('Chroma covariances: My Funny Valentine')

plt.xlim(-0.5, 11.5)

plt.ylim(-0.5, 11.5)

plt.ylabel('"Input" chromae')

plt.xlabel('"Predicted" chromae')

plt.xticks(range(12), chromae)

plt.yticks(range(12), chromae)

plt.gca().invert_yaxis()

plt.show()

#%%

shifter = np.roll(np.arange(0, 7*12, 7) % 12, 6)

#%%

n_beats_to_show = 30

plt.figure(figsize=(4, 5), dpi=300)

plt.imshow(beat_aligned_chromagram[:n_beats_to_show], interpolation="nearest", aspect="auto")

plt.xlabel("Chroma")

plt.xticks(range(12), chromae)

plt.ylabel("Time (Beats)")

plt.yticks(

range(0, len(beat_aligned_chromagram[:n_beats_to_show]), 4),

np.arange(0, len(beat_aligned_chromagram[:n_beats_to_show]), 4) + 1,

)

plt.gca().yaxis.set_minor_locator(MultipleLocator(1))

# plt.xlim(-.5, len())

plt.show()

#%%

chromae_roman = ['I', '', 'II', '', 'III', 'IV', '', 'V', '', 'VI', '', 'VII']

shift = -(beat_aligned_chromagram.sum(0).argmax())

n_beats_to_show = 30

plt.figure(figsize=(4, 5), dpi=300)

plt.imshow(np.roll(beat_aligned_chromagram[:n_beats_to_show], shift, axis=1), interpolation="nearest", aspect="auto")

plt.xlabel("Chroma")

plt.xticks(range(12), chromae_roman)

plt.ylabel("Time (Beats)")

plt.yticks(

range(0, len(beat_aligned_chromagram[:n_beats_to_show]), 4),

np.arange(0, len(beat_aligned_chromagram[:n_beats_to_show]), 4) + 1,

)

plt.gca().yaxis.set_minor_locator(MultipleLocator(1))

# plt.xlim(-.5, len())

plt.show()

#%%

plt.figure(figsize=(4, 4), dpi=300)

plt.imshow(

corr_valentine.reshape(12, 12), origin="upper"

)

plt.title('Chroma covariances: My Funny Valentine')

plt.xlim(-0.5, 11.5)

plt.ylim(-0.5, 11.5)

plt.ylabel('"Input" chromae')

plt.xlabel('"Predicted" chromae')

plt.xticks(range(12), chromae)

plt.yticks(range(12), chromae)

plt.gca().invert_yaxis()

plt.show()

#%%

plt.figure(figsize=(4, 4), dpi=300)

plt.imshow(

corr_valentine.reshape(12, 12)[shifter][:, shifter], origin="upper"

)

plt.title('Chroma covariances: My Funny Valentine')

plt.xlim(-0.5, 11.5)

plt.ylim(-0.5, 11.5)

plt.ylabel('"Input" chromae')

plt.xlabel('"Predicted" chromae')

plt.xticks(range(12), np.array(chromae)[shifter][:, shifter])

plt.yticks(range(12), np.array(chromae)[shifter][:, shifter])

plt.gca().invert_yaxis()

plt.show()

#%%

xs = np.vstack(

represent_w_correlation(sa, order=1) for sa in gtzan.song_analyses

)

model = XGBClassifier(n_estimators=1000)

predictions = cross_val_predict(model, xs, gtzan.y, cv=5)

mat = confusion_matrix(gtzan.y, predictions)

accuracy = ml.accuracy_from_confusion_matrix(mat)

plotting.imshow_confusion_matrix(

mat, gtzan.labelnames, f"GTZAN cross validation accuracy: {accuracy * 100:2.1f}%"

)

#%%

xs = np.vstack(

represent_w_correlation(sa, order=4) for sa in gtzan.song_analyses

)

model = XGBClassifier(n_estimators=1000)

predictions = cross_val_predict(model, xs, gtzan.y, cv=5, n_jobs=-1)

mat = confusion_matrix(gtzan.y, predictions)

accuracy = ml.accuracy_from_confusion_matrix(mat)

plotting.imshow_confusion_matrix(

mat, gtzan.labelnames, f"GTZAN cross validation accuracy: {accuracy * 100:2.1f}%"

)

#%%

def represent_w_correlation_shifted(sa: SongAnalysis, order: int = 1):

return representation.flatten()

xs = np.vstack(

represent_w_correlation_shifted(sa, order=4) for sa in gtzan.song_analyses

)

model = XGBClassifier(n_estimators=1000)

predictions = cross_val_predict(model, xs, gtzan.y, cv=5, n_jobs=-1)

mat = confusion_matrix(gtzan.y, predictions)

accuracy = ml.accuracy_from_confusion_matrix(mat)

plotting.imshow_confusion_matrix(

mat, gtzan.labelnames, f"GTZAN cross validation accuracy: {accuracy * 100:2.1f}%"

)

#%%

def represent_w_correlation_shifted_pitches(sa: SongAnalysis, order: int = 1):

return representation.flatten()

xs = np.vstack(

represent_w_correlation_shifted_pitches(sa, order=4) for sa in gtzan.song_analyses

)

model = XGBClassifier(n_estimators=1000)

predictions = cross_val_predict(model, xs, gtzan.y, cv=5, n_jobs=-1)

mat = confusion_matrix(gtzan.y, predictions)

accuracy = ml.accuracy_from_confusion_matrix(mat)

plotting.imshow_confusion_matrix(

mat, gtzan.labelnames, f"GTZAN cross validation accuracy: {accuracy * 100:2.1f}%"

)

#%%

def plot_umap(X, y, labelnames):

plt.show()

plot_umap(xs, gtzan.y, gtzan.labelnames)