Goal: assess the usefulness of the auto-encoder with and without the graph when faced with noiseless data. The euclidean distance metric is used to construct the graph.
Conclusion: In a noiseless setting, our structuring auto-encoder improves the accuracy by 7.3% while a sparse auto-encoder imporves the performance by 6.2% compared to the baseline (raw spectrograms).
Observations:
13e_dm
, 13c_novoting
and 13b_noise_lg
(voting).13c_novoting
and 13b_noise_lg
(voting).13e_dm
. The only difference I see is a different kNN from FLANN (see the numbers in dist
or w
).Pname = 'lg'
Pvalues = [None, 100]
# Regenerate the graph or the features at each iteration.
regen_graph = False
regen_features = True
p = {}
# Preprocessing.
# Graph.
p['data_scaling_graph'] = 'features'
p['K'] = 10 + 1 # 5 to 10 + 1 for self-reference
p['dm'] = 'euclidean'
p['Csigma'] = 1
p['diag'] = True
p['laplacian'] = 'normalized'
# Feature extraction.
p['m'] = 128 # 64, 128, 512
p['ls'] = 1
p['ld'] = 10
p['le'] = None
p['lg'] = 100
# Classification.
p['scale'] = None
p['Nvectors'] = 6
p['svm_type'] = 'C'
p['kernel'] = 'linear'
p['C'] = 1
p['nu'] = 0.5
p['majority_voting'] = False
# HDF5 data stores.
p['folder'] = 'data'
p['filename_gtzan'] = 'gtzan.hdf5'
p['filename_audio'] = 'audio.hdf5'
p['filename_graph'] = 'graph.hdf5'
p['filename_features'] = 'features.hdf5'
# Dataset (10,100,644 | 5,100,149 | 2,10,644).
p['Ngenres'] = 5
p['Nclips'] = 100
p['Nframes'] = 149
# Added white noise.
p['noise_std'] = 0
# Graph.
p['tol'] = 1e-5
# Feature extraction.
p['rtol'] = 1e-5 # 1e-3, 1e-5, 1e-7
p['N_inner'] = 500
p['N_outer'] = 50
# Classification.
p['Nfolds'] = 10
p['Ncv'] = 20
p['dataset_classification'] = 'Z'
import numpy as np
import time
texperiment = time.time()
# Result dictionary.
res = ['accuracy', 'accuracy_std']
res += ['sparsity', 'atoms']
res += ['objective_g', 'objective_h', 'objective_i', 'objective_j']
res += ['time_features', 'iterations_inner', 'iterations_outer']
res = dict.fromkeys(res)
for key in res.keys():
res[key] = []
def separator(name, parameter=False):
if parameter:
name += ', {} = {}'.format(Pname, p[Pname])
dashes = 20 * '-'
print('\n {} {} {} \n'.format(dashes, name, dashes))
# Fair comparison when tuning parameters.
# Randomnesses: dictionary initialization, training and testing sets.
np.random.seed(1)
#%run gtzan.ipynb
#%run audio_preprocessing.ipynb
if not regen_graph:
separator('Graph')
%run audio_graph.ipynb
if not regen_features:
separator('Features')
%run audio_features.ipynb
# Hyper-parameter under test.
for p[Pname] in Pvalues:
if regen_graph:
separator('Graph', True)
%run audio_graph.ipynb
if regen_features:
separator('Features', True)
p['filename_features'] = 'features_{}_{}.hdf5'.format(Pname, p[Pname])
%run audio_features.ipynb
separator('Classification', True)
%run audio_classification.ipynb
# Collect results.
for key in res:
res[key].append(globals()[key])
# Baseline, i.e. classification with spectrograms.
p['dataset_classification'] = 'X'
p['scale'] = 'minmax' # Todo: should be done in pre-processing.
if not regen_graph and not regen_features:
# Classifier parameters are being tested.
for p[Pname] in Pvalues:
separator('Baseline', True)
%run audio_classification.ipynb
else:
separator('Baseline')
%run audio_classification.ipynb
res['baseline'] = len(Pvalues) * [accuracy]
res['baseline_std'] = accuracy_std
-------------------- Graph -------------------- Data: (149000, 96), float32 Elapsed time: 188.94 seconds All self-referenced in the first column: True dist in [0.0, 1.47963690758] w in [0.00555037613958, 1.0] Ones on the diagonal: 149000 (over 149000) assert: True W in [0.0, 1.0] Datasets: L_data : (2397786,), float32 L_indices : (2397786,), int32 L_indptr : (149001,) , int32 L_shape : (2,) , int64 W_data : (2397786,), float32 W_indices : (2397786,), int32 W_indptr : (149001,) , int32 W_shape : (2,) , int64 Attributes: K = 11 dm = euclidean Csigma = 1 diag = True laplacian = normalized Overall time: 198.11 seconds -------------------- Features, lg = None -------------------- Attributes: sr = 22050 labels = ['blues' 'classical' 'country' 'disco' 'hiphop' 'jazz' 'metal' 'pop' 'reggae' 'rock'] Datasets: Xa: (10, 100, 644, 2, 1024) , float32 Xs: (10, 100, 644, 2, 96) , float32 Full dataset: size: N=1,288,000 x n=96 -> 123,648,000 floats dim: 123,648 features per clip shape: (10, 100, 644, 2, 96) <class 'h5py._hl.dataset.Dataset'> Reduced dataset: size: N=149,000 x n=96 -> 14,304,000 floats dim: 28,608 features per clip shape: (5, 100, 149, 2, 96) <type 'numpy.ndarray'> Data: (149000, 96), float32 Attributes: K = 11 dm = euclidean Csigma = 1 diag = True laplacian = normalized Datasets: L_data : (2397786,), float32 L_indices : (2397786,), int32 L_indptr : (149001,) , int32 L_shape : (2,) , int64 W_data : (2397786,), float32 W_indices : (2397786,), int32 W_indptr : (149001,) , int32 W_shape : (2,) , int64 Size X: 13.6 M --> 54.6 MiB Size Z: 18.2 M --> 72.8 MiB Size D: 12.0 k --> 48.0 kiB Size E: 12.0 k --> 48.0 kiB Elapsed time: 1143 seconds
Inner loop: 898 iterations g(Z) = ||X-DZ||_2^2 = 4.231717e+04 rdiff: 0.00242506942781 i(Z) = ||Z||_1 = 6.887042e+04
Global objective: 1.111876e+05
Outer loop: 14 iterations Z in [-1.15960109234, 1.81123566628] Sparsity of Z: 905,112 non-zero entries out of 19,072,000 entries, i.e. 4.7%.
D in [-0.760848701, 0.938695788383] d in [0.999999701977, 1.00000035763] Constraints on D: True
Datasets: D : (128, 96) , float32 X : (5, 100, 149, 2, 96) , float32 Z : (5, 100, 149, 2, 128) , float32 Attributes: sr = 22050 labels = ['blues' 'classical' 'country' 'disco' 'hiphop' 'jazz' 'metal' 'pop' 'reggae' 'rock'] Overall time: 1150 seconds -------------------- Classification, lg = None -------------------- Software versions: numpy: 1.8.2 sklearn: 0.14.1 Attributes: sr = 22050 labels = ['blues' 'classical' 'country' 'disco' 'hiphop' 'jazz' 'metal' 'pop' 'reggae' 'rock'] Datasets: D : (128, 96) , float32 X : (5, 100, 149, 2, 96) , float32 Z : (5, 100, 149, 2, 128) , float32 Full dataset: size: N=149,000 x n=128 -> 19,072,000 floats dim: 38,144 features per clip shape: (5, 100, 149, 2, 128) <class 'h5py._hl.dataset.Dataset'> Reduced dataset: size: N=149,000 x n=128 -> 19,072,000 floats dim: 38,144 features per clip shape: (5, 100, 149, 2, 128) <type 'numpy.ndarray'> Flattened frames: size: N=149,000 x n=128 -> 19,072,000 floats dim: 38,144 features per clip shape: (5, 100, 298, 128) Truncated and grouped: size: N=135,000 x n=128 -> 17,280,000 floats dim: 34,560 features per clip shape: (5, 100, 6, 45, 128) Truncated and grouped: size: N=135,000 x n=128 -> 17,280,000 floats dim: 34,560 features per clip shape: (5, 100, 6, 45, 128) Feature vectors: size: N=6,000 x n=128 -> 768,000 floats dim: 1,536 features per clip shape: (5, 100, 6, 2, 128)
5 genres: blues, classical, country, disco, hiphop Training data: (3600, 128), float64 Testing data: (2400, 128), float64 Training labels: (3600,), uint8 Testing labels: (2400,), uint8 Accuracy: 74.8 % 5 genres: blues, classical, country, disco, hiphop Training data: (300, 1536), float64 Testing data: (200, 1536), float64 Training labels: (300,), uint8 Testing labels: (200,), uint8 Feature vectors accuracy: 61.4 % Clips accuracy: 71.0 % 5 genres: blues, classical, country, disco, hiphop Data: (6000, 128), float64 Labels: (6000,), uint8 76 (+/- 1.3) <- [75 77 75 77 77 76 76 73 77 76] 76 (+/- 1.2) <- [74 77 75 76 74 76 77 73 75 75] 76 (+/- 1.6) <- [77 74 77 78 75 75 72 74 76 76] 75 (+/- 1.4) <- [76 73 75 75 72 73 77 76 76 75] 76 (+/- 1.8) <- [74 73 76 77 75 74 78 78 77 75] 76 (+/- 1.6) <- [77 77 73 74 78 75 77 78 76 76] 76 (+/- 1.9) <- [78 74 75 76 76 73 75 73 77 79] 76 (+/- 1.3) <- [76 75 73 76 77 75 76 75 78 74] 76 (+/- 1.8) <- [77 72 76 76 73 74 78 77 76 77] 76 (+/- 1.6) <- [77 73 76 75 74 77 78 74 76 73] 76 (+/- 2.2) <- [74 78 78 75 76 78 74 76 71 74] 76 (+/- 2.0) <- [74 78 75 77 72 75 74 78 75 78] 76 (+/- 1.8) <- [77 76 73 76 76 77 79 75 73 74] 76 (+/- 2.0) <- [76 75 74 78 76 78 76 73 75 80] 76 (+/- 2.4) <- [74 75 75 74 79 72 76 78 80 75] 76 (+/- 1.7) <- [74 78 74 78 74 74 75 77 76 76] 76 (+/- 1.7) <- [76 75 77 77 77 74 76 77 72 74] 76 (+/- 2.4) <- [75 71 76 73 79 79 74 77 76 76] 77 (+/- 1.5) <- [75 78 77 73 77 76 77 76 76 78] 76 (+/- 0.9) <- [77 75 76 75 77 76 76 75 74 76] Accuracy: 76.0 (+/- 1.77) Mean time (20 cv): 30.97 seconds Overall time: 624.45 seconds -------------------- Features, lg = 100 -------------------- Attributes: sr = 22050 labels = ['blues' 'classical' 'country' 'disco' 'hiphop' 'jazz' 'metal' 'pop' 'reggae' 'rock'] Datasets: Xa: (10, 100, 644, 2, 1024) , float32 Xs: (10, 100, 644, 2, 96) , float32 Full dataset: size: N=1,288,000 x n=96 -> 123,648,000 floats dim: 123,648 features per clip shape: (10, 100, 644, 2, 96) <class 'h5py._hl.dataset.Dataset'> Reduced dataset: size: N=149,000 x n=96 -> 14,304,000 floats dim: 28,608 features per clip shape: (5, 100, 149, 2, 96) <type 'numpy.ndarray'> Data: (149000, 96), float32 Attributes: K = 11 dm = euclidean Csigma = 1 diag = True laplacian = normalized Datasets: L_data : (2397786,), float32 L_indices : (2397786,), int32 L_indptr : (149001,) , int32 L_shape : (2,) , int64 W_data : (2397786,), float32 W_indices : (2397786,), int32 W_indptr : (149001,) , int32 W_shape : (2,) , int64 Size X: 13.6 M --> 54.6 MiB Size Z: 18.2 M --> 72.8 MiB Size D: 12.0 k --> 48.0 kiB Size E: 12.0 k --> 48.0 kiB Elapsed time: 2573 seconds
Inner loop: 1243 iterations g(Z) = ||X-DZ||_2^2 = 6.872115e+04 rdiff: 0.00151142965386 i(Z) = ||Z||_1 = 5.812207e+04 j(Z) = tr(Z^TLZ) = 9.091110e+03
Global objective: 1.359343e+05
Outer loop: 6 iterations Z in [-0.233149766922, 1.01959300041] Sparsity of Z: 3,418,431 non-zero entries out of 19,072,000 entries, i.e. 17.9%.
D in [-0.137461602688, 0.915598213673] d in [0.999999642372, 1.00000035763] Constraints on D: True
Datasets: D : (128, 96) , float32 X : (5, 100, 149, 2, 96) , float32 Z : (5, 100, 149, 2, 128) , float32 Attributes: sr = 22050 labels = ['blues' 'classical' 'country' 'disco' 'hiphop' 'jazz' 'metal' 'pop' 'reggae' 'rock'] Overall time: 2581 seconds -------------------- Classification, lg = 100 -------------------- Software versions: numpy: 1.8.2 sklearn: 0.14.1 Attributes: sr = 22050 labels = ['blues' 'classical' 'country' 'disco' 'hiphop' 'jazz' 'metal' 'pop' 'reggae' 'rock'] Datasets: D : (128, 96) , float32 X : (5, 100, 149, 2, 96) , float32 Z : (5, 100, 149, 2, 128) , float32 Full dataset: size: N=149,000 x n=128 -> 19,072,000 floats dim: 38,144 features per clip shape: (5, 100, 149, 2, 128) <class 'h5py._hl.dataset.Dataset'> Reduced dataset: size: N=149,000 x n=128 -> 19,072,000 floats dim: 38,144 features per clip shape: (5, 100, 149, 2, 128) <type 'numpy.ndarray'> Flattened frames: size: N=149,000 x n=128 -> 19,072,000 floats dim: 38,144 features per clip shape: (5, 100, 298, 128) Truncated and grouped: size: N=135,000 x n=128 -> 17,280,000 floats dim: 34,560 features per clip shape: (5, 100, 6, 45, 128) Truncated and grouped: size: N=135,000 x n=128 -> 17,280,000 floats dim: 34,560 features per clip shape: (5, 100, 6, 45, 128) Feature vectors: size: N=6,000 x n=128 -> 768,000 floats dim: 1,536 features per clip shape: (5, 100, 6, 2, 128)
5 genres: blues, classical, country, disco, hiphop Training data: (3600, 128), float64 Testing data: (2400, 128), float64 Training labels: (3600,), uint8 Testing labels: (2400,), uint8 Accuracy: 75.6 % 5 genres: blues, classical, country, disco, hiphop Training data: (300, 1536), float64 Testing data: (200, 1536), float64 Training labels: (300,), uint8 Testing labels: (200,), uint8 Feature vectors accuracy: 63.5 % Clips accuracy: 69.0 % 5 genres: blues, classical, country, disco, hiphop Data: (6000, 128), float64 Labels: (6000,), uint8 77 (+/- 1.7) <- [76 77 75 80 76 79 74 77 76 78] 77 (+/- 1.7) <- [72 77 78 77 76 77 79 78 76 76] 77 (+/- 1.2) <- [77 74 77 77 78 78 78 76 76 77] 77 (+/- 1.8) <- [75 74 78 76 77 75 81 76 78 76] 77 (+/- 2.2) <- [75 72 79 75 78 77 79 77 78 80] 77 (+/- 1.5) <- [75 76 75 76 79 78 78 79 76 75] 77 (+/- 1.5) <- [78 74 76 77 78 75 77 76 75 80] 77 (+/- 1.5) <- [80 78 76 76 78 77 76 75 78 75] 77 (+/- 1.1) <- [76 77 77 76 75 75 78 78 76 79] 77 (+/- 1.2) <- [78 77 77 78 78 77 76 76 77 74] 77 (+/- 2.5) <- [76 76 81 75 76 81 74 76 74 75] 77 (+/- 1.7) <- [75 79 78 77 73 78 75 77 78 77] 77 (+/- 1.7) <- [75 80 77 75 78 76 79 75 75 77] 77 (+/- 2.2) <- [77 75 74 78 75 77 80 73 77 80] 76 (+/- 2.4) <- [73 80 76 74 79 73 76 77 79 73] 77 (+/- 1.2) <- [78 77 76 77 77 77 75 80 76 76] 77 (+/- 1.4) <- [76 75 76 77 78 76 78 77 77 74] 77 (+/- 1.8) <- [77 73 78 75 77 80 76 78 78 77] 77 (+/- 1.0) <- [75 76 77 75 76 77 77 78 78 77] 77 (+/- 1.5) <- [77 78 75 77 80 76 77 75 76 77] Accuracy: 77.1 (+/- 1.70) Mean time (20 cv): 19.31 seconds Overall time: 390.47 seconds -------------------- Baseline -------------------- Software versions: numpy: 1.8.2 sklearn: 0.14.1 Attributes: sr = 22050 labels = ['blues' 'classical' 'country' 'disco' 'hiphop' 'jazz' 'metal' 'pop' 'reggae' 'rock'] Datasets: D : (128, 96) , float32 X : (5, 100, 149, 2, 96) , float32 Z : (5, 100, 149, 2, 128) , float32 Full dataset: size: N=149,000 x n=96 -> 14,304,000 floats dim: 28,608 features per clip shape: (5, 100, 149, 2, 96) <class 'h5py._hl.dataset.Dataset'> Reduced dataset: size: N=149,000 x n=96 -> 14,304,000 floats dim: 28,608 features per clip shape: (5, 100, 149, 2, 96) <type 'numpy.ndarray'> Flattened frames: size: N=149,000 x n=96 -> 14,304,000 floats dim: 28,608 features per clip shape: (5, 100, 298, 96) Truncated and grouped: size: N=135,000 x n=96 -> 12,960,000 floats dim: 25,920 features per clip shape: (5, 100, 6, 45, 96) Truncated and grouped: size: N=135,000 x n=96 -> 12,960,000 floats dim: 25,920 features per clip shape: (5, 100, 6, 45, 96) Feature vectors: size: N=6,000 x n=96 -> 576,000 floats dim: 1,152 features per clip shape: (5, 100, 6, 2, 96)
5 genres: blues, classical, country, disco, hiphop Training data: (3600, 96), float64 Testing data: (2400, 96), float64 Training labels: (3600,), uint8 Testing labels: (2400,), uint8 Accuracy: 68.8 % 5 genres: blues, classical, country, disco, hiphop Training data: (300, 1152), float64 Testing data: (200, 1152), float64 Training labels: (300,), uint8 Testing labels: (200,), uint8 Feature vectors accuracy: 62.1 % Clips accuracy: 68.5 % 5 genres: blues, classical, country, disco, hiphop Data: (6000, 96), float64 Labels: (6000,), uint8 70 (+/- 1.9) <- [70 71 66 71 68 71 67 68 69 72] 69 (+/- 1.9) <- [66 70 69 71 67 68 72 68 70 70] 70 (+/- 1.5) <- [69 70 68 72 69 71 68 68 69 72] 70 (+/- 1.6) <- [68 70 69 69 67 72 72 69 70 67] 70 (+/- 2.2) <- [67 68 67 68 69 66 73 71 71 72] 70 (+/- 2.9) <- [67 70 62 70 72 69 73 71 71 70] 70 (+/- 2.1) <- [71 69 71 72 69 68 67 66 70 73] 70 (+/- 2.3) <- [69 72 66 71 69 68 70 68 73 66] 70 (+/- 1.9) <- [71 67 68 69 67 69 70 73 68 73] 70 (+/- 1.7) <- [73 68 71 69 69 71 70 68 70 67] 70 (+/- 1.8) <- [68 71 72 67 70 72 67 70 69 67] 70 (+/- 1.8) <- [68 69 70 73 70 70 68 69 66 70] 70 (+/- 1.6) <- [67 71 66 70 69 71 71 69 67 70] 70 (+/- 1.5) <- [71 69 70 70 69 67 71 69 67 72] 70 (+/- 1.8) <- [70 70 67 67 69 66 70 71 72 69] 70 (+/- 1.9) <- [68 72 68 72 69 71 69 73 68 68] 70 (+/- 1.5) <- [66 69 69 69 71 69 69 69 72 71] 70 (+/- 1.2) <- [69 67 70 68 69 70 69 71 69 70] 70 (+/- 2.3) <- [65 71 68 67 70 72 72 67 69 71] 70 (+/- 2.7) <- [69 67 66 71 73 72 70 66 68 74] Accuracy: 69.8 (+/- 1.96) Mean time (20 cv): 15.09 seconds Overall time: 305.83 seconds
print('{} = {}'.format(Pname, Pvalues))
for key, value in res.items():
if key is not 'atoms':
print('res[\'{}\'] = {}'.format(key, value))
def plot(*args, **kwargs):
plt.figure(figsize=(8,5))
x = range(len(Pvalues))
log = 'log' in kwargs and kwargs['log'] is True
pltfunc = plt.semilogy if log else plt.plot
params = {}
params['linestyle'] = '-'
params['marker'] = '.'
params['markersize'] = 10
for i, var in enumerate(args):
if 'err' in kwargs:
pltfunc = plt.errorbar
params['yerr'] = res[kwargs['err'][i]]
params['capsize'] = 5
pltfunc(x, res[var], label=var, **params)
for i,j in zip(x, res[var]):
plt.annotate('{:.2f}'.format(j), xy=(i,j), xytext=(5,5), textcoords='offset points')
margin = 0.25
params['markersize'] = 10
plt.xlim(-margin, len(Pvalues)-1+margin)
if 'ylim' in kwargs:
plt.ylim(kwargs['ylim'])
plt.title('{} vs {}'.format(', '.join(args), Pname))
plt.xlabel(Pname)
plt.ylabel(' ,'.join(args))
plt.xticks(x, Pvalues)
plt.grid(True); plt.legend(loc='best'); plt.show()
def div(l):
div = Pvalues if Pname is l else [p[l]]
return np.array([1 if v is None else v for v in div])
# Classification results.
res['chance'] = len(Pvalues) * [100./p['Ngenres']]
res['chance_std'] = 0
err=['accuracy_std', 'baseline_std', 'chance_std']
plot('accuracy', 'baseline', 'chance', err=err, ylim=[0,100])
# Features extraction results.
if regen_features:
plot('objective_g', 'objective_i', 'objective_j', log=True)
# Unweighted objectives.
print('g(Z) = ||X-DZ||_2^2, h(Z) = ||Z-EX||_2^2, i(Z) = ||Z||_1, j(Z) = tr(Z^TLZ)')
res['objective_g_un'] = res['objective_g'] / div('ld')
res['objective_i_un'] = res['objective_i'] / div('ls')
res['objective_j_un'] = res['objective_j'] / div('lg')
plot('objective_g_un', 'objective_i_un', 'objective_j_un', log=True)
plot('sparsity', ylim=[0,100])
plot('time_features')
plot('iterations_inner')
plot('iterations_outer')
for i, fig in enumerate(res['atoms']):
print('Dictionary atoms for {} = {}'.format(Pname, Pvalues[i]))
fig.show()
print('Experiment time: {:.0f} seconds'.format(time.time() - texperiment))
lg = [None, 100] res['accuracy_std'] = [1.7664541931979625, 1.696892584828056] res['objective_j'] = [0, 9091.1102294921875] res['objective_i'] = [68870.421875, 58122.0703125] res['objective_h'] = [0, 0] res['objective_g'] = [42317.16796875, 68721.15234375] res['baseline'] = [69.785833333333315, 69.785833333333315] res['time_features'] = [1142.8163068294525, 2572.735995054245] res['baseline_std'] = 1.95713633631 res['sparsity'] = [4.745763422818792, 17.923820260067114] res['iterations_inner'] = [898, 1243] res['iterations_outer'] = [14, 6] res['accuracy'] = [76.034166666666607, 77.083333333333357]
/usr/lib/python2.7/dist-packages/numpy/ma/core.py:3847: UserWarning: Warning: converting a masked element to nan. warnings.warn("Warning: converting a masked element to nan.")
g(Z) = ||X-DZ||_2^2, h(Z) = ||Z-EX||_2^2, i(Z) = ||Z||_1, j(Z) = tr(Z^TLZ)
Dictionary atoms for lg = None Dictionary atoms for lg = 100 Experiment time: 5257 seconds
/usr/lib/pymodules/python2.7/matplotlib/figure.py:371: UserWarning: matplotlib is currently using a non-GUI backend, so cannot show the figure "matplotlib is currently using a non-GUI backend, "