Info predict file #2
Comments
|
Hello Thomas,
I guess I've the file that you mentioned. Is it "predict.py" or
"prediction.py"?
I can send it to you. But I am not sure about the content. It was 3+ years
ago))
Kind regards,
Artem
…On Sat, 27 May 2023, 11:02 TomasSand, ***@***.***> wrote:
Dear Artyom Ivasik, your project is very interesting. I try to run your
code and it's ok. But i don't find the predict file that you mentions in
your Thesis. It's possible view this file?
Thanks for your support.
—
Reply to this email directly, view it on GitHub
<#2>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AQGSOQJFRAJIYHMCOXAVBFDXIGYJHANCNFSM6AAAAAAYRA47R4>
.
You are receiving this because you are subscribed to this thread.Message
ID: ***@***.***>
|
|
Dear Artem, Thanks for your answer.
I would like to try using the network after the training phase to
separate the sources. I guess it's prediction :)
Thanks again for the kind reply and congrats again for your project.
King regards
Tommaso
… Il 30/05/2023 09:07 CEST Artyom Ivasik ***@***.***> ha scritto:
Hello Thomas,
I guess I've the file that you mentioned. Is it "predict.py" or
"prediction.py"?
I can send it to you. But I am not sure about the content. It was 3+ years
ago))
Kind regards,
Artem
On Sat, 27 May 2023, 11:02 TomasSand, ***@***.***> wrote:
> Dear Artyom Ivasik, your project is very interesting. I try to run your
> code and it's ok. But i don't find the predict file that you mentions in
> your Thesis. It's possible view this file?
> Thanks for your support.
>
> —
> Reply to this email directly, view it on GitHub
> <#2>,
> or unsubscribe
> <https://github.com/notifications/unsubscribe-auth/AQGSOQJFRAJIYHMCOXAVBFDXIGYJHANCNFSM6AAAAAAYRA47R4>
> .
> You are receiving this because you are subscribed to this thread.Message
> ID: ***@***.***>
>
—
Reply to this email directly, view it on GitHub #2 (comment), or unsubscribe https://github.com/notifications/unsubscribe-auth/AJCBRJOS3JBC7YKPZYRSNRLXIWME7ANCNFSM6AAAAAAYRA47R4.
You are receiving this because you authored the thread.Message ID: ***@***.***>
|
|
Hello Tommaso,
Sorry for the delay. I attached two files. I hope one of them is what you
are looking for.
Kind regards,
Artem
вт, 30 мая 2023 г. в 17:34, TomasSand ***@***.***>:
… Dear Artem, Thanks for your answer.
I would like to try using the network after the training phase to
separate the sources. I guess it's prediction :)
Thanks again for the kind reply and congrats again for your project.
King regards
Tommaso
> Il 30/05/2023 09:07 CEST Artyom Ivasik ***@***.***> ha scritto:
>
>
>
>
>
> Hello Thomas,
>
> I guess I've the file that you mentioned. Is it "predict.py" or
> "prediction.py"?
> I can send it to you. But I am not sure about the content. It was 3+
years
> ago))
>
> Kind regards,
> Artem
>
> On Sat, 27 May 2023, 11:02 TomasSand, ***@***.***> wrote:
>
> > Dear Artyom Ivasik, your project is very interesting. I try to run your
> > code and it's ok. But i don't find the predict file that you mentions
in
> > your Thesis. It's possible view this file?
> > Thanks for your support.
> >
> > —
> > Reply to this email directly, view it on GitHub
> > <
#2>,
> > or unsubscribe
> > <
https://github.com/notifications/unsubscribe-auth/AQGSOQJFRAJIYHMCOXAVBFDXIGYJHANCNFSM6AAAAAAYRA47R4
>
> > .
> > You are receiving this because you are subscribed to this
thread.Message
> > ID: ***@***.***>
> >
>
> —
> Reply to this email directly, view it on GitHub
#2 (comment),
or unsubscribe
https://github.com/notifications/unsubscribe-auth/AJCBRJOS3JBC7YKPZYRSNRLXIWME7ANCNFSM6AAAAAAYRA47R4
.
> You are receiving this because you authored the thread.Message ID:
***@***.***>
>
—
Reply to this email directly, view it on GitHub
<#2 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AQGSOQLMD5GQIAHW6O5IMUTXIYAO7ANCNFSM6AAAAAAYRA47R4>
.
You are receiving this because you commented.Message ID:
***@***.***
com>
|
|
Dear Artem, thanks for your answer :)
but i don't see the file.
I apologize if i do something wrong :)
I look forward to your reply.
Thanks again
Tommaso
Il 06/06/2023 21:47 CEST Artyom Ivasik ***@***.***> ha scritto:
…
Hello Tommaso,
Sorry for the delay. I attached two files. I hope one of them is what you
are looking for.
Kind regards,
Artem
вт, 30 мая 2023 г. в 17:34, TomasSand ***@***.***>:
> Dear Artem, Thanks for your answer.
>
> I would like to try using the network after the training phase to
> separate the sources. I guess it's prediction :)
> Thanks again for the kind reply and congrats again for your project.
>
> King regards
> Tommaso
>
>
>
> > Il 30/05/2023 09:07 CEST Artyom Ivasik ***@***.***> ha scritto:
> >
> >
> >
> >
> >
> > Hello Thomas,
> >
> > I guess I've the file that you mentioned. Is it "predict.py" or
> > "prediction.py"?
> > I can send it to you. But I am not sure about the content. It was 3+
> years
> > ago))
> >
> > Kind regards,
> > Artem
> >
> > On Sat, 27 May 2023, 11:02 TomasSand, ***@***.***> wrote:
> >
> > > Dear Artyom Ivasik, your project is very interesting. I try to run your
> > > code and it's ok. But i don't find the predict file that you mentions
> in
> > > your Thesis. It's possible view this file?
> > > Thanks for your support.
> > >
> > > —
> > > Reply to this email directly, view it on GitHub
> > > <
> #2>,
> > > or unsubscribe
> > > <
> https://github.com/notifications/unsubscribe-auth/AQGSOQJFRAJIYHMCOXAVBFDXIGYJHANCNFSM6AAAAAAYRA47R4
> >
> > > .
> > > You are receiving this because you are subscribed to this
> thread.Message
> > > ID: ***@***.***>
> > >
> >
> > —
> > Reply to this email directly, view it on GitHub
> #2 (comment),
> or unsubscribe
> https://github.com/notifications/unsubscribe-auth/AJCBRJOS3JBC7YKPZYRSNRLXIWME7ANCNFSM6AAAAAAYRA47R4
> .
> > You are receiving this because you authored the thread.Message ID:
> ***@***.***>
> >
>
> —
> Reply to this email directly, view it on GitHub
> <#2 (comment)>,
> or unsubscribe
> <https://github.com/notifications/unsubscribe-auth/AQGSOQLMD5GQIAHW6O5IMUTXIYAO7ANCNFSM6AAAAAAYRA47R4>
> .
> You are receiving this because you commented.Message ID:
> ***@***.***
> com>
>
—
Reply to this email directly, view it on GitHub #2 (comment), or unsubscribe https://github.com/notifications/unsubscribe-auth/AJCBRJLPGQT2IFHC5HRIOETXJ6COZANCNFSM6AAAAAAYRA47R4.
You are receiving this because you authored the thread.Message ID: ***@***.***>
|
|
prediction.py
from keras.models import load_model
import numpy as np
import musdb
import librosa
import librosa.display
import math
from time import time
def P2R(radii, angles):
return radii * (math.cos(angles) + 1j*math.sin(angles))
def get_power_and_angle(audio, orig_sr, new_sr, win_len, hop_len):
if orig_sr!= new_sr:
audio = librosa.core.resample(audio, orig_sr, new_sr)
stft_ar = librosa.stft(audio, win_len, hop_len, center = False)
return np.abs(stft_ar), np.angle(stft_ar)
def normalize_pow_spec(pow_spec):
# Масштабирование спектрограммы к [0;1]
pow_spec_norm = np.zeros((pow_spec.shape[0], pow_spec.shape[1]))
max = np.max(pow_spec)
if max != 0:
pow_spec_norm = pow_spec / max
return pow_spec_norm
def save_audio(power_spec, phases, hop_length, win_length, sampling_rate,
path, mask = None):
if mask is None:
mask = np.ones(power_spec.shape)
stft_array = np.empty(power_spec.shape, dtype=complex)
for i in range(stft_array.shape[0]):
for j in range(stft_array.shape[1]):
stft_array[i, j] = P2R(power_spec[i][j]*mask[i][j],
phases[i][j])
y = librosa.core.istft(stft_array, hop_length, win_length,)
librosa.output.write_wav(path, y, sampling_rate)
ORIG_SR = 44100 # Оригинальная частота дискретизации
SR = 22050 # Частота дискретизации, с которой работает сеть
WIN = 1024 # Размер окна дискретного преобразования Фурье
HOP = 256 # Размер шага дискретного преобразования Фурье
sample_len = 27 # Размер отрезка спектрограммы, на котором обучена сеть
TRGT = 'vocals' # Выделяемый источник
# Загрузка модели
t0 = time()
print("<----------[INFO] model loading...")
model = load_model('output/model final 4.hdf5')
t1 = time()
print("<----------[INFO] model was loaded in " + str(round((t1-t0), 1)) + "
seconds...")
#Загрузка тестового набора
mus = musdb.DB(root='./musdb18', subsets='test', split=None)
for track in mus:
print(track.name)
mix = librosa.core.to_mono(track.audio.T)
mix_power_spec, mix_phases = get_power_and_angle(mix, ORIG_SR, SR, WIN,
HOP)
target = librosa.core.to_mono(track.targets[TRGT].audio.T)
target_power_spec, target_phases = get_power_and_angle(target, ORIG_SR,
SR, WIN, HOP)
# Нормализация и расширение спектрограммы
mix_power_spec_norm = normalize_pow_spec(mix_power_spec)
pad = np.zeros((mix_power_spec_norm.shape[0], sample_len//2))
mix_power_spec_norm = np.concatenate((pad, mix_power_spec_norm), axis=1)
mix_power_spec_norm = np.concatenate((mix_power_spec_norm, pad), axis=1)
mask = np.zeros((mix_power_spec_norm.shape[0], 0))
bin_mask = np.zeros((mix_power_spec_norm.shape[0], 0))
set_to_zero_mask = np.zeros((mix_power_spec_norm.shape[0], 0))
for i in range(mix_power_spec.shape[1]):
mix = mix_power_spec_norm[:,i:i+sample_len]
mix = np.expand_dims(mix, 0)
mix = np.expand_dims(mix, len(mix.shape))
pred = model.predict(mix)[0]
mask = np.concatenate((mask, np.expand_dims(pred,
len(pred.shape))), axis=1)
for i in range(mask.shape[0]):
for j in range(mask.shape[1]):
if mask[i][j] >= 0.5:
bin_mask[i][j] = 1
set_to_zero_mask[i][j] = mask[i][j]
save_audio(mix_power_spec, mix_phases, HOP, WIN, SR,
'output/estimation/soft/'+ track.name +'.wav', mask)
save_audio(mix_power_spec, mix_phases, HOP, WIN, SR,
'output/estimation/bin/'+ track.name +'.wav', mask)
save_audio(mix_power_spec, mix_phases, HOP, WIN, SR,
'output/estimation/zero/'+ track.name +'.wav', mask)
break;
…_______________________________________________________________________________________________________-
predict.py
from keras.models import load_model
import numpy as np
#import musdb
import librosa
import librosa.display
import math
from time import time
import matplotlib.pyplot as plt
def P2R(radii, angles):
return radii * (math.cos(angles) + 1j*math.sin(angles))
def get_power_and_angle(audio, orig_sr, new_sr, win_len, hop_len):
if orig_sr!= new_sr:
audio = librosa.core.resample(audio, orig_sr, new_sr)
stft_ar = librosa.stft(audio, win_len, hop_len, center = False)
return np.abs(stft_ar), np.angle(stft_ar)
def normalize_pow_spec(pow_spec):
# Масштабирование спектрограммы к [0;1]
pow_spec_norm = np.zeros((pow_spec.shape[0], pow_spec.shape[1]))
max = np.max(pow_spec)
if max != 0:
pow_spec_norm = pow_spec / max
return pow_spec_norm
def save_audio(power_spec, phases, hop_length, win_length, sampling_rate,
path, mask = None):
if mask is None:
mask = np.ones(power_spec.shape)
stft_array = np.empty(power_spec.shape, dtype=complex)
for i in range(stft_array.shape[0]):
for j in range(stft_array.shape[1]):
stft_array[i, j] = P2R(power_spec[i][j]*mask[i][j],
phases[i][j])
y = librosa.core.istft(stft_array, hop_length, win_length,)
librosa.output.write_wav(path, y, sampling_rate)
def plot_spec(power):
plt.subplots(figsize=(power.shape[1]/120,power.shape[0]/120), dpi = 150)
librosa.display.specshow(data = librosa.core.amplitude_to_db(power,
ref=np.max),
y_axis='linear', #log or linear
x_axis='time',
sr = SR,
hop_length = HOP)
ORIG_SR = 44100
SR = 22050 # Частота дискретизации, с которой работает сеть
WIN = 1024 # Размер окна дискретного преобразования Фурье
HOP = 256 # Размер шага дискретного преобразования Фурье
sample_len = 27
TRGT = 'vocals'
##Загрузка тестового набора
#mus = musdb.DB(root='./musdb18', subsets='test', split=None)
#track = mus[1]
#print(track.name)
#ORIG_SR = track.rate
#track.chunk_duration = 15
#track.chunk_start = 0
#mix = librosa.core.to_mono(track.audio.T)
#target = librosa.core.to_mono(track.targets[TRGT].audio.T)
#
#mix_power_spec, mix_phases = get_power_and_angle(mix, ORIG_SR, SR, WIN,
HOP)
#save_audio(mix_power_spec, mix_phases, HOP, WIN, SR, 'D:/mix.wav')
#
#target_power_spec, target_phases = get_power_and_angle(target, ORIG_SR,
SR, WIN, HOP)
#save_audio(target_power_spec, target_phases, HOP, WIN, SR,
'D:/orig_source.wav')
#
#other_power_spec = np.zeros(mix_power_spec.shape)
#for i in range(mix_power_spec.shape[0]):
# for j in range(mix_power_spec.shape[1]):
# if mix_power_spec[i][j] >= target_power_spec[i][j]:
# other_power_spec[i][j] = mix_power_spec[i][j] -
target_power_spec[i][j]
# else:
# other_power_spec[i][j] = 0
#save_audio(other_power_spec, mix_phases, HOP, WIN, SR, 'D:/orig_other.wav')
name = 'Кино'
mix, ORIG_SR = librosa.core.load(path = 'D:\\tracks\\' + name + '.mp3',
sr = 44100,
mono = True,
offset=21,
duration=34-21)
mix_power_spec, mix_phases = get_power_and_angle(mix, ORIG_SR, SR, WIN, HOP)
save_audio(mix_power_spec, mix_phases, HOP, WIN, SR, 'D:\\tracks\\result\\'
+ name + ' mix.wav')
plot_spec(mix_power_spec)
# Нормализация и расширение спектрограммы
mix_power_spec_norm = normalize_pow_spec(mix_power_spec)
pad = np.zeros((mix_power_spec_norm.shape[0], sample_len//2))
mix_power_spec_norm = np.concatenate((pad, mix_power_spec_norm), axis=1)
mix_power_spec_norm = np.concatenate((mix_power_spec_norm, pad), axis=1)
# Загрузка модели
t0 = time()
print("<----------[INFO] model loading...")
model = load_model('output/model final 10.hdf5')
t1 = time()
print("<----------[INFO] model was loaded in " + str(round((t1-t0), 1)) + "
seconds...")
t0 = time()
print("<----------[INFO] mask forming...")
mask = np.zeros((mix_power_spec_norm.shape[0], 0))
for i in range(mix_power_spec.shape[1]):
mix = mix_power_spec_norm[:,i:i+sample_len]
mix = np.expand_dims(mix, 0)
mix = np.expand_dims(mix, len(mix.shape))
pred = model.predict(mix)[0]
mask = np.concatenate((mask, np.expand_dims(pred, len(pred.shape))),
axis=1)
# Обработка маски
for i in range(mask.shape[0]):
for j in range(mask.shape[1]):
if mask[i][j] >= 0.5:
mask[i][j] = 1
else:
mask[i][j] = 0
# if mask[i][j] <= 0.5:
# mask[i][j] = 0
plot_spec(mask*1000)
t1 = time()
print("<----------[INFO] mask was formed in " + str(round((t1-t0), 1)) + "
seconds...")
save_audio(mix_power_spec, mix_phases, HOP, WIN, SR, 'D:\\tracks\\result\\'
+ name + ' source.wav', mask)
plot_spec(mix_power_spec*mask)
reverse_mask = np.zeros(mask.shape)
for i in range(mask.shape[0]):
for j in range(mask.shape[1]):
reverse_mask[i,j] = 1 - mask[i,j]
save_audio(mix_power_spec, mix_phases, HOP, WIN, SR, 'D:\\tracks\\result\\'
+ name + ' other.wav', reverse_mask)
сб, 27 мая 2023 г. в 11:02, TomasSand ***@***.***>:
Dear Artyom Ivasik, your project is very interesting. I try to run your
code and it's ok. But i don't find the predict file that you mentions in
your Thesis. It's possible view this file?
Thanks for your support.
—
Reply to this email directly, view it on GitHub
<#2>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AQGSOQJFRAJIYHMCOXAVBFDXIGYJHANCNFSM6AAAAAAYRA47R4>
.
You are receiving this because you are subscribed to this thread.Message
ID: ***@***.***>
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Dear Artyom Ivasik, your project is very interesting. I try to run your code and it's ok. But i don't find the predict file that you mentions in your Thesis. It's possible view this file?
Thanks for your support.
The text was updated successfully, but these errors were encountered: