def parse_file(file, targetfile):
print "print parsing %s" %file
with open(file, 'r') as f:
global tracks, labels, features
tracks = list()
labels = list()
for line in f:
track, label = line.strip().split(' ')
tracks.append(track)
labels.append(int(label))
labels = np.array(labels)
features = np.empty((len(tracks),NUM_FEATURES),dtype='float')#hardcode number of features otherwise problem with scaling
track_info = np.empty((len(tracks)),dtype='object')
for i,track in enumerate(tracks):
#print track +' - ' +ut.get_track_info(track)
if(i%100 ==0):
print "processing track:%s\t%s" %(str(i),str(track))
if(file == SOURCE_DATA_FILE): #fetch h5 file from small dataset
h5 = GETTERS.open_h5_file_read("../../msd_dense_subset/dense/"+track[2]+"/"+track[3]+"/"+track[4]+"/"+track+".h5") #fetch h5 file to allow faster preprocessing
else:
h5 = GETTERS.open_h5_file_read("../../msd_dense_subset/mood/"+track[2]+"/"+track[3]+"/"+track[4]+"/"+track+".h5") #fetch h5 file to allow faster preprocessing
track_info[i] = ut.get_track_info(track,h5)
timbre = ut.get_timbre(track,h5) #returns a tuple with 5 elements (12*5 = 60)
tempo = ut.get_tempo_feature(track,h5) #(1)
loudness = ut.get_loudness(track,h5)#returns a tuple with 3 elements (3)
energy = ut.get_energy_feature(track) #(1)
pitches = ut.get_pitches(track, h5) #(12)
features[i] = np.concatenate((timbre[0], timbre[1], timbre[2], timbre[3], timbre[4],pitches[0], pitches[1],pitches[2], pitches[3],pitches[4],np.array([tempo]),np.array(loudness),np.array([energy])))
h5.close()
print "done parsing"
print "saving data"
data = {
'features': features,
'labels': labels,
'tracks': tracks, #(songindex, songid)
'track_titles': track_info
}
with open(targetfile, 'w') as f:
pickle.dump(data, f, pickle.HIGHEST_PROTOCOL)
print "data saved to %s" % targetfile
#SCRIPT
#parse_file(SOURCE_DATA_FILE_2,E_TARGET_DATA_FILE_2)
#parse_file(SOURCE_DATA_FILE,E_TARGET_DATA_FILE)
timbre_min = np.zeros((len(tracks),12))
timbre_max = np.zeros((len(tracks),12))
pitches_means = np.zeros((len(tracks),12))
pitches_vars = np.zeros((len(tracks),12))
pitches_median = np.zeros((len(tracks),12))
pitches_min = np.zeros((len(tracks),12))
pitches_max = np.zeros((len(tracks),12))
for idx,track in enumerate(tracks):
h5 = GETTERS.open_h5_file_read("../../msd_dense_subset/mood/"+track[2]+"/"+track[3]+"/"+track[4]+"/"+track+".h5") #fetch h5 file to allow faster preprocessing
keys[idx], modes[idx]= ut.get_key_feature(track,h5)
loudnesses[idx], loudnesses_var[idx], loudnesses_interval[idx] = ut.get_loudness(track,h5)
tempos[idx] = ut.get_tempo_feature(track,h5)
time_signatures[idx] = ut.get_time_signature(track,h5)
timbre_means[idx],timbre_vars[idx], timbre_median[idx], timbre_min[idx], timbre_max[idx] = ut.get_timbre(track,h5)
pitches_means[idx],pitches_vars[idx], pitches_median[idx], pitches_min[idx], pitches_max[idx]= ut.get_pitches(track,h5)
energies[idx] = ut.get_energy_feature(track)
h5.close()
#use binning for continious data
#problem: number of bins => freedman-driaconis rule
num_bins = 2* (stats.scoreatpercentile(loudnesses_interval, 75) - stats.scoreatpercentile(loudnesses_interval, 25))*len(loudnesses_interval)**(1/3)
bins = np.linspace(min(loudnesses_interval), max(loudnesses_interval),num=num_bins)
d_loudnesses_interval = np.digitize(loudnesses_interval, bins)
num_bins = 2* (stats.scoreatpercentile(loudnesses, 75) - stats.scoreatpercentile(loudnesses, 25))*len(loudnesses)**(1/3)
bins = np.linspace(min(loudnesses), max(loudnesses),num=100)
d_loudnesses = np.digitize(loudnesses, bins)
num_bins = 2* (stats.scoreatpercentile(tempos, 75) - stats.scoreatpercentile(tempos, 25))*len(tempos)**(1/3)
bins = np.linspace(min(tempos), max(tempos),num=100)
pitches_max = np.zeros((len(tracks), 12))
for idx, track in enumerate(tracks):
h5 = GETTERS.open_h5_file_read(
"../../msd_dense_subset/mood/" + track[2] + "/" + track[3] + "/" +
track[4] + "/" + track +
".h5") #fetch h5 file to allow faster preprocessing
keys[idx], modes[idx] = ut.get_key_feature(track, h5)
loudnesses[idx], loudnesses_var[idx], loudnesses_interval[
idx] = ut.get_loudness(track, h5)
tempos[idx] = ut.get_tempo_feature(track, h5)
time_signatures[idx] = ut.get_time_signature(track, h5)
timbre_means[idx], timbre_vars[idx], timbre_median[idx], timbre_min[
idx], timbre_max[idx] = ut.get_timbre(track, h5)
pitches_means[idx], pitches_vars[idx], pitches_median[idx], pitches_min[
idx], pitches_max[idx] = ut.get_pitches(track, h5)
energies[idx] = ut.get_energy_feature(track)
h5.close()
#use binning for continious data
#problem: number of bins => freedman-driaconis rule
num_bins = 2 * (stats.scoreatpercentile(loudnesses_interval, 75) -
stats.scoreatpercentile(loudnesses_interval, 25)
) * len(loudnesses_interval)**(1 / 3)
bins = np.linspace(min(loudnesses_interval),
max(loudnesses_interval),
num=num_bins)
d_loudnesses_interval = np.digitize(loudnesses_interval, bins)
num_bins = 2 * (stats.scoreatpercentile(loudnesses, 75) -
stats.scoreatpercentile(loudnesses, 25)) * len(loudnesses)**(