from plottools import plotall
import pylab as P
sys.path.append('../PythonSrc')
import evaluation
import masking
import imputation
if __name__ == '__main__':
btchroma145 = sio.loadmat(
'/home/thierry/Columbia/covers80/coversongs/covers32k/Caroline_No/beach_boys+Pet_Sounds+13-Caroline_No.mp3.mat'
)['btchroma']
mask, masked_cols = masking.random_col_mask(btchroma145, ncols=1, win=30)
recon, lt = imputation.lintransform_cols(btchroma145,
mask,
masked_cols,
win=1)
pos1 = masked_cols[0] - 7
pos2 = masked_cols[0] + 7
im1 = btchroma145[:, pos1:pos2].copy()
im2 = (btchroma145 * mask)[:, pos1:pos2].copy()
im3 = recon[:, pos1:pos2].copy()
# plot all this
fig = P.figure()
fig.subplots_adjust(hspace=0.4)
blackbarsfun = lambda: P.gca().axvline(linewidth=2, color='0.', x=6.5
) and P.gca().axvline(
linewidth=2, color='0.', x=7.5)
plotall([im1, im2, im3],
subplot=(3, 1),
cmap='gray_r',
def compare_all(btchroma, mask, masked_cols, codebook=None):
"""
Compare all the algorithms we have so far.
A lot of parameters hard-coded, but...
It will get improved.
We display:
1) original beat chromagram
2) masked beat chromagram
3) imputation by random
4) imputation by averaging nearby columns
5) imputation by knn on the rest of the song
6) imputation by linear prediction
7) imputation by NMF
8) imputation by codebook, if provided
To init mask and masked_cols, something like:
mask,masked_cols = masking.random_col_mask(btchroma,ncols=1,win=30)
"""
# for displaying purposes, display window size 50
pos1 = masked_cols[0] - 5
pos2 = masked_cols[0] + 5
allimages = []
titles = []
xlabels = []
# original beat chroma and masked one
im1 = btchroma[:, pos1:pos2].copy()
im2 = (btchroma * mask)[:, pos1:pos2].copy()
allimages.append(im1)
titles.append('original beat chroma')
xlabels.append('')
allimages.append(im2)
titles.append('masked beat chroma')
xlabels.append('')
# 3) random
recon = IMPUTATION.random_col(btchroma, mask, masked_cols)
im3_err = EVAL.recon_error(btchroma, mask, recon, measure='eucl')
im3 = recon[:, pos1:pos2].copy()
allimages.append(im3)
titles.append('random')
xlabels.append('err=' + str(im3_err))
# 4) average nearby columns
recon = IMPUTATION.average_col(btchroma, mask, masked_cols, win=3)
im4_err = EVAL.recon_error(btchroma, mask, recon, measure='eucl')
im4 = recon[:, pos1:pos2].copy()
allimages.append(im4)
titles.append('average 2 nearby cols')
xlabels.append('err=' + str(im4_err))
# 5) knn
recon, used_cols = IMPUTATION.eucldist_cols(btchroma,
mask,
masked_cols,
win=7)
im5_err = EVAL.recon_error(btchroma, mask, recon, measure='eucl')
im5 = recon[:, pos1:pos2].copy()
allimages.append(im5)
titles.append('knn for the whole song')
xlabels.append('err=' + str(im5_err))
# 6) linear prediction
recon, proj = IMPUTATION.lintransform_cols(btchroma,
mask,
masked_cols,
win=1)
im6_err = EVAL.recon_error(btchroma, mask, recon, measure='eucl')
im6 = recon[:, pos1:pos2].copy()
allimages.append(im6)
titles.append('linear prediction (1 col)')
xlabels.append('err=' + str(im6_err))
# 7) SIPLCA
res = IMPUTATION_PLCA.SIPLCA_mask.analyze((btchroma * mask).copy(),
4,
mask,
win=5)
W, Z, H, norm, recon, logprob = res
im7_err = EVAL.recon_error(btchroma, mask, recon, measure='eucl')
im7 = recon[:, pos1:pos2].copy()
allimages.append(im7)
titles.append('SIPLCA, rank=4, win=5')
xlabels.append('err=' + str(im7_err))
# 7) SIPLCA 2
res = IMPUTATION_PLCA.SIPLCA_mask.analyze((btchroma * mask).copy(),
25,
mask,
win=10)
W, Z, H, norm, recon, logprob = res
im7_err_bis = EVAL.recon_error(btchroma, mask, recon, measure='eucl')
im7_bis = recon[:, pos1:pos2].copy()
allimages.append(im7_bis)
titles.append('SIPLCA, rank=25, win=10')
xlabels.append('err=' + str(im7_err_bis))
# 8) codebook
if codebook != None:
cb = [p.reshape(12, p.size / 12) for p in codebook]
recon, used_codes = IMPUTATION.codebook_cols(btchroma, mask,
masked_cols, cb)
im8_err = EVAL.recon_error(btchroma, mask, recon, measure='eucl')
im8 = recon[:, pos1:pos2].copy()
allimages.append(im8)
titles.append('codebook, ' + str(codebook.shape[0]) +
' codes of length ' + str(codebook.shape[1] / 12))
xlabels.append('err=' + str(im8_err))
# ALL IMAGES CREATED
fig = plt.figure()
fig.subplots_adjust(hspace=0.4)
blackbarsfun = lambda: plt.gca().axvline(
linewidth=2, color='0.', x=4.5) and plt.gca().axvline(
linewidth=2, color='0.', x=5.5)
# plotall
plotall(allimages,
subplot=(3, 3),
cmap='gray_r',
title=titles,
xlabel=xlabels,
axvlines=blackbarsfun,
colorbar=False)
def test_maskedcol_on_dataset(datasetdir,
method='random',
ncols=1,
win=3,
rank=4,
codebook=None,
**kwargs):
"""
General method to test a method on a whole dataset for one masked column
Methods are:
- random
- randomfromsong
- average
- codebook
- knn_eucl
- knn_kl
- lintrans
- siplca
- siplca2
Used arguments vary based on the method. For SIPLCA, we can use **kwargs
to set priors.
"""
# get all matfiles
matfiles = get_all_matfiles(datasetdir)
# init
total_cnt = 0
errs_eucl = []
errs_kl = []
# some specific inits
if codebook != None and not type(codebook) == type([]):
codebook = [p.reshape(12, codebook.shape[1] / 12) for p in codebook]
print 'codebook in ndarray format transformed to list'
# iterate
for matfile in matfiles:
btchroma = sio.loadmat(matfile)['btchroma']
if len(btchroma.shape) < 2:
continue
if btchroma.shape[1] < MINLENGTH or np.isnan(btchroma).any():
continue
mask, masked_cols = MASKING.random_col_mask(btchroma,
ncols=ncols,
win=25)
########## ALGORITHM DEPENDENT
if method == 'random':
recon = IMPUTATION.random_col(btchroma, mask, masked_cols)
elif method == 'randomfromsong':
recon = IMPUTATION.random_col_from_song(btchroma, mask,
masked_cols)
elif method == 'average':
recon = IMPUTATION.average_col(btchroma,
mask,
masked_cols,
win=win)
elif method == 'codebook':
recon, used_codes = IMPUTATION.codebook_cols(
btchroma, mask, masked_cols, codebook)
elif method == 'knn_eucl':
recon, used_cols = IMPUTATION.knn_cols(btchroma,
mask,
masked_cols,
win=win,
measure='eucl')
elif method == 'knn_kl':
recon, used_cols = IMPUTATION.knn_cols(btchroma,
mask,
masked_cols,
win=win,
measure='kl')
elif method == 'lintrans':
recon, proj = IMPUTATION.lintransform_cols(btchroma,
mask,
masked_cols,
win=win)
elif method == 'siplca':
res = IMPUTATION_PLCA.SIPLCA_mask.analyze((btchroma * mask).copy(),
rank,
mask,
win=win,
convergence_thresh=1e-15,
**kwargs)
W, Z, H, norm, recon, logprob = res
elif method == 'siplca2':
res = IMPUTATION_PLCA.SIPLCA2_mask.analyze(
(btchroma * mask).copy(),
rank,
mask,
win=win,
convergence_thresh=1e-15,
**kwargs)
W, Z, H, norm, recon, logprob = res
else:
print 'unknown method:', method
return
########## ALGORITHM DEPENDENT END
# measure recon
err = recon_error(btchroma, mask, recon, measure='eucl')
if err > 100:
print 'huge EUCL error:', err, ', method =', method, ',file =', matfile
errs_eucl.append(err)
err = recon_error(btchroma, mask, recon, measure='kl')
if err > 100:
print 'huge KL error:', err, ', method =', method, ',file =', matfile
errs_kl.append(err)
total_cnt += 1
# done
print 'number of songs tested:', total_cnt
print 'average sq euclidean dist:', np.mean(errs_eucl), '(', np.std(
errs_eucl), ')'
print 'average kl divergence:', np.mean(errs_kl), '(', np.std(errs_kl), ')'
def test_maskedcol_on_dataset(datasetdir,method='random',ncols=1,win=3,rank=4,codebook=None,**kwargs):
"""
General method to test a method on a whole dataset for one masked column
Methods are:
- random
- randomfromsong
- average
- codebook
- knn_eucl
- knn_kl
- lintrans
- siplca
- siplca2
Used arguments vary based on the method. For SIPLCA, we can use **kwargs
to set priors.
"""
# get all matfiles
matfiles = get_all_matfiles(datasetdir)
# init
total_cnt = 0
errs_eucl = []
errs_kl = []
# some specific inits
if codebook != None and not type(codebook) == type([]):
codebook = [p.reshape(12,codebook.shape[1]/12) for p in codebook]
print 'codebook in ndarray format transformed to list'
# iterate
for matfile in matfiles:
btchroma = sio.loadmat(matfile)['btchroma']
if len(btchroma.shape) < 2:
continue
if btchroma.shape[1] < MINLENGTH or np.isnan(btchroma).any():
continue
mask,masked_cols = MASKING.random_col_mask(btchroma,ncols=ncols,win=25)
########## ALGORITHM DEPENDENT
if method == 'random':
recon = IMPUTATION.random_col(btchroma,mask,masked_cols)
elif method == 'randomfromsong':
recon = IMPUTATION.random_col_from_song(btchroma,mask,masked_cols)
elif method == 'average':
recon = IMPUTATION.average_col(btchroma,mask,masked_cols,win=win)
elif method == 'codebook':
recon,used_codes = IMPUTATION.codebook_cols(btchroma,mask,masked_cols,codebook)
elif method == 'knn_eucl':
recon,used_cols = IMPUTATION.knn_cols(btchroma,mask,masked_cols,win=win,measure='eucl')
elif method == 'knn_kl':
recon,used_cols = IMPUTATION.knn_cols(btchroma,mask,masked_cols,win=win,measure='kl')
elif method == 'lintrans':
recon,proj = IMPUTATION.lintransform_cols(btchroma,mask,masked_cols,win=win)
elif method == 'siplca':
res = IMPUTATION_PLCA.SIPLCA_mask.analyze((btchroma*mask).copy(),
rank,mask,win=win,
convergence_thresh=1e-15,
**kwargs)
W, Z, H, norm, recon, logprob = res
elif method == 'siplca2':
res = IMPUTATION_PLCA.SIPLCA2_mask.analyze((btchroma*mask).copy(),
rank,mask,win=win,
convergence_thresh=1e-15,
**kwargs)
W, Z, H, norm, recon, logprob = res
else:
print 'unknown method:',method
return
########## ALGORITHM DEPENDENT END
# measure recon
err = recon_error(btchroma,mask,recon,measure='eucl')
if err > 100:
print 'huge EUCL error:',err,', method =',method,',file =',matfile
errs_eucl.append( err )
err = recon_error(btchroma,mask,recon,measure='kl')
if err > 100:
print 'huge KL error:',err,', method =',method,',file =',matfile
errs_kl.append( err )
total_cnt += 1
# done
print 'number of songs tested:',total_cnt
print 'average sq euclidean dist:',np.mean(errs_eucl),'(',np.std(errs_eucl),')'
print 'average kl divergence:',np.mean(errs_kl),'(',np.std(errs_kl),')'
import numpy as np
import scipy.io as sio
from plottools import plotall
import pylab as P
sys.path.append('../PythonSrc')
import evaluation
import masking
import imputation
if __name__ == '__main__':
btchroma145 = sio.loadmat('/home/thierry/Columbia/covers80/coversongs/covers32k/Caroline_No/beach_boys+Pet_Sounds+13-Caroline_No.mp3.mat')['btchroma']
mask,masked_cols = masking.random_col_mask(btchroma145,ncols=1,win=30)
recon,lt = imputation.lintransform_cols(btchroma145,mask,masked_cols,win=1)
pos1 = masked_cols[0] - 7
pos2 = masked_cols[0] + 7
im1 = btchroma145[:,pos1:pos2].copy()
im2 = (btchroma145 * mask)[:,pos1:pos2].copy()
im3 = recon[:,pos1:pos2].copy()
# plot all this
fig = P.figure()
fig.subplots_adjust(hspace=0.4)
blackbarsfun = lambda: P.gca().axvline(linewidth=2,color='0.',x=6.5) and P.gca().axvline(linewidth=2,color='0.',x=7.5)
plotall([im1,im2,im3],subplot=(3,1),cmap='gray_r',
title=['original','original masked','reconstruction'],
axvlines=blackbarsfun,colorbar=False,xticks=[()]*3)
P.show()
def compare_all(btchroma,mask,masked_cols,codebook=None):
"""
Compare all the algorithms we have so far.
A lot of parameters hard-coded, but...
It will get improved.
We display:
1) original beat chromagram
2) masked beat chromagram
3) imputation by random
4) imputation by averaging nearby columns
5) imputation by knn on the rest of the song
6) imputation by linear prediction
7) imputation by NMF
8) imputation by codebook, if provided
To init mask and masked_cols, something like:
mask,masked_cols = masking.random_col_mask(btchroma,ncols=1,win=30)
"""
# for displaying purposes, display window size 50
pos1 = masked_cols[0] - 5
pos2 = masked_cols[0] + 5
allimages = []
titles = []
xlabels = []
# original beat chroma and masked one
im1 = btchroma[:,pos1:pos2].copy()
im2 = (btchroma * mask)[:,pos1:pos2].copy()
allimages.append(im1)
titles.append('original beat chroma')
xlabels.append('')
allimages.append(im2)
titles.append('masked beat chroma')
xlabels.append('')
# 3) random
recon = IMPUTATION.random_col(btchroma,mask,masked_cols)
im3_err = EVAL.recon_error(btchroma,mask,recon,measure='eucl')
im3 = recon[:,pos1:pos2].copy()
allimages.append(im3)
titles.append('random')
xlabels.append('err='+str(im3_err))
# 4) average nearby columns
recon = IMPUTATION.average_col(btchroma,mask,masked_cols,win=3)
im4_err = EVAL.recon_error(btchroma,mask,recon,measure='eucl')
im4 = recon[:,pos1:pos2].copy()
allimages.append(im4)
titles.append('average 2 nearby cols')
xlabels.append('err='+str(im4_err))
# 5) knn
recon,used_cols = IMPUTATION.eucldist_cols(btchroma,mask,masked_cols,win=7)
im5_err = EVAL.recon_error(btchroma,mask,recon,measure='eucl')
im5 = recon[:,pos1:pos2].copy()
allimages.append(im5)
titles.append('knn for the whole song')
xlabels.append('err='+str(im5_err))
# 6) linear prediction
recon,proj = IMPUTATION.lintransform_cols(btchroma,mask,masked_cols,win=1)
im6_err = EVAL.recon_error(btchroma,mask,recon,measure='eucl')
im6 = recon[:,pos1:pos2].copy()
allimages.append(im6)
titles.append('linear prediction (1 col)')
xlabels.append('err='+str(im6_err))
# 7) SIPLCA
res = IMPUTATION_PLCA.SIPLCA_mask.analyze((btchroma*mask).copy(),
4,mask,win=5)
W, Z, H, norm, recon, logprob = res
im7_err = EVAL.recon_error(btchroma,mask,recon,measure='eucl')
im7 = recon[:,pos1:pos2].copy()
allimages.append(im7)
titles.append('SIPLCA, rank=4, win=5')
xlabels.append('err='+str(im7_err))
# 7) SIPLCA 2
res = IMPUTATION_PLCA.SIPLCA_mask.analyze((btchroma*mask).copy(),
25,mask,win=10)
W, Z, H, norm, recon, logprob = res
im7_err_bis = EVAL.recon_error(btchroma,mask,recon,measure='eucl')
im7_bis = recon[:,pos1:pos2].copy()
allimages.append(im7_bis)
titles.append('SIPLCA, rank=25, win=10')
xlabels.append('err='+str(im7_err_bis))
# 8) codebook
if codebook != None:
cb = [p.reshape(12,p.size/12) for p in codebook]
recon,used_codes = IMPUTATION.codebook_cols(btchroma,mask,masked_cols,cb)
im8_err = EVAL.recon_error(btchroma,mask,recon,measure='eucl')
im8 = recon[:,pos1:pos2].copy()
allimages.append(im8)
titles.append('codebook, '+str(codebook.shape[0])+' codes of length '+str(codebook.shape[1]/12))
xlabels.append('err='+str(im8_err))
# ALL IMAGES CREATED
fig = plt.figure()
fig.subplots_adjust(hspace=0.4)
blackbarsfun = lambda: plt.gca().axvline(linewidth=2,color='0.',x=4.5) and plt.gca().axvline(linewidth=2,color='0.',x=5.5)
# plotall
plotall(allimages,subplot=(3,3),cmap='gray_r',title=titles,xlabel=xlabels,axvlines=blackbarsfun,colorbar=False)