def extract_features(tracks_dir="tracks/", feat_dir="features/"):
utils.print_success("Extracting features")
tracks_fn = os.listdir(tracks_dir)
utils.create_dir(feat_dir)
feat_dir = utils.create_dir(feat_dir + "svmbff")
bextract = "bextract -mfcc -zcrs -ctd -rlf -flx -ws 1024 -as 898 -sv -fe "
for index, filename in enumerate(tracks_fn):
utils.print_progress_start(
str(index) + "/" + str(len(tracks_fn)) + " " + filename)
track_path = filename + ".mf"
with open(track_path, "w") as filep:
filep.write(tracks_dir + filename + "\n")
new_fn = filename.split(".")[0] + ".arff"
try:
os.system(bextract + track_path + " -w " + new_fn +
"> /dev/null 2>&1")
except:
utils.print_info(
"You have to make marsyas available systemwide, tips:")
utils.print_info(
"http://marsyas.info/doc/manual/marsyas-user/Step_002dby_002dstep-building-instructions.html#Step_002dby_002dstep-building-instructions"
)
utils.print_info("http://stackoverflow.com/a/21173918")
utils.print_error("Program exit")
# print(new_fn)
# print(feat_dir + " " + new_fn)
os.rename(new_fn, feat_dir + new_fn)
# os.rename("MARSYAS_EMPTY" + new_fn, feat_dir + new_fn)
os.system("rm " + track_path)
utils.print_progress_end()
os.system("rm bextract_single.mf")
def yaafe_feat_extraction(dir_tracks):
"""Description of yaafe_feat_extraction
yaafe.py -r 22050 -f "mfcc: MFCC blockSize=2048 stepSize=1024" audio_fn.txt
"""
utils.print_success("YAAFE features extraction (approx. 8 minutes)")
# Assert Python version
if sys.version_info.major != 2:
utils.print_error("Yaafe needs Python 2 environment")
# Assert folder exists
dir_tracks = utils.abs_path_dir(dir_tracks)
filelist = os.listdir(dir_tracks)
dir_feat = utils.create_dir(utils.create_dir("features") + "database1")
# dir_tmp = utils.create_dir("tmp")
# dir_yaafe = utils.create_dir(dir_tmp + "yaafe")
# fn_filelist = dir_yaafe + "filelist.txt"
dir_current = os.getcwd()
os.chdir(dir_tracks)
yaafe_cmd = 'yaafe -r 22050 -f "mfcc: MFCC blockSize=2048 stepSize=1024" '
yaafe_cmd += "--resample -b " + dir_feat + " "
for index, filen in enumerate(filelist):
utils.print_progress_start(str(index+1) + "/" + str(len(filelist)) + " " + filen)
os.system(yaafe_cmd + filen + "> /dev/null 2>&1")
utils.print_progress_end()
os.chdir(dir_current)
def run_kea_on_folds(folds_dir):
"""Description of run_kea_on_folds
Wrapper for kea on folds
"""
folds_dir = utils.abs_path_dir(folds_dir)
out_file = folds_dir + "/results.txt"
if os.path.exists(folds_dir + "/train_test.arff"):
train_file = folds_dir + "/train_test.arff"
test_file = train_file
run_kea(train_file, test_file, out_file)
else:
nb_folds = len([
name for name in os.listdir(folds_dir)
if os.path.isfile(os.path.join(folds_dir, name))
])
# Run on multiple train/test
for index in range(1, int(nb_folds / 2) + 1):
utils.print_progress_start("Train/Test on fold " + str(index))
train_file = folds_dir + "/train_" + str(index).zfill(2) + ".arff"
test_file = folds_dir + "/test_" + str(index).zfill(2) + ".arff"
out_file = folds_dir + "/results_" + str(index).zfill(2) + ".arff"
run_kea(train_file, test_file, out_file)
utils.print_progress_end()
utils.print_warning("TODO multiprocessing")
def preprocess_features(folder):
utils.print_success("Preprocessing train set")
folder = utils.abs_path_dir(folder)
filelist = os.listdir(folder)
nb_file = str(len(filelist))
for index, filename in enumerate(filelist):
utils.print_progress_start(str(index) + "/" + nb_file + " " + filename)
convert_feats_files(folder + filename)
utils.print_progress_end()
def generate_singing_voice_track(paths):
"""
@brief { function_description }
@param dir_audio The dir audio
@return { description_of_the_return_value }
"""
utils.print_success("Generating singing voice tracks")
for index, folder in enumerate(paths):
utils.print_progress_start(
str(index) + "/" + str(len(paths)) + " " + folder)
if os.path.isdir(folder) and os.path.exists(folder):
filelist = os.listdir(folder)
for filen in filelist:
nb_error = 0
if "-bv." in filen:
utils.print_error("Backing vocals file found in " + filen)
if filen.endswith('-ld.wav'):
song = filen
instru = filen.replace("-ld", "")
try:
song_samples, song_fs = sf.read(folder + "/" + song)
except RuntimeError as run_err:
error("RuntimeError", str(run_err))
nb_error += 1
try:
instru_samples, instru_fs = sf.read(folder + "/" +
instru)
except RuntimeError as run_err:
error("RuntimeError", str(run_err))
nb_error += 1
if nb_error == 0:
if song_fs != instru_fs:
error("SamplingFreq", filen)
elif len(instru_samples) != len(song_samples):
error("SampleSize", filen)
else:
voice_samples = song_samples - instru_samples
# print(instru)
# print(song)
# utils.print_error(len(song_samples))
# print(song_samples.shape)
# print(len(instru_samples))
# print(instru_samples.shape)
# print(len(voice_samples))
# print(voice_samples.shape)
sf.write(
folder + "/" + filen.replace("nbv-ld", "sv"),
voice_samples, song_fs)
with open("available.txt", "a") as filep:
filep.write(folder + "/" + filen + "\n")
utils.print_progress_end()
return "available.txt"
def match_feat_with_song_gt(dir_feat, dir_gts):
"""Description of match_feat_gt
Use groundtruth created by
http://www.mathieuramona.com/wp/data/jamendo/
associate to local features
csv 7041 lines yaafe
lab 326.973 sec ramona
Definition of YAAFE from
http://yaafe.sourceforge.net/features.html
"""
utils.print_success("Matching local feat to song/instru groundtruths")
dir_feat = utils.abs_path_dir(dir_feat)
dir_gts = utils.abs_path_dir(dir_gts)
block_size = 1024.
step_size = 512.
fech = 22050.
frame_size_ms = block_size / fech
filenames = [fn for fn in os.listdir(dir_gts)]
for index, filename in enumerate(filenames):
utils.print_progress_start(str(index) + "/" + str(len(filenames)) + " " + filename)
# gather groundtruths
groundtruths = []
with open(dir_gts + filename, "r") as filep:
for row in filep:
line = row.split(" ")
end = float(line[1])
if "no" in line[2]:
tag = ",i\n"
else:
tag = ",s\n"
groundtruths.append([end, tag])
gt_len = len(groundtruths)
overflow = False
gt_index = 0
cpt = 0
# Write features & groundtruths to file
str_to_write = ""
feat_fn = filename.split(".")[0]
feat_fn += ".wav.mfcc.csv"
with open(dir_feat + feat_fn, "r") as filep:
for index, line in enumerate(filep):
# todo cleanup
if gt_index < gt_len:
if frame_size_ms * index > groundtruths[gt_index][0]:
gt_index += 1
if gt_index < gt_len:
str_to_write += line[:-1] + groundtruths[gt_index][1]
with open(dir_feat + feat_fn, "w") as filep:
filep.write(str_to_write)
utils.print_progress_end()
def merge_arff(indir, outfilename):
"""Description of merge_arff
bextract program from Marsyas generate one output file per audio file
This function merge them all in one unique file
Check if analysed file are valid i.e. not empty
"""
utils.print_success("Preprocessing ARFFs")
indir = utils.abs_path_dir(indir)
filenames = os.listdir(indir)
outfn = open(outfilename, 'w')
cpt_invalid_fn = 0
# Write first lines of ARFF template file
for filename in filenames:
if os.path.isfile(indir + filename):
new_fn = validate_arff(indir + filename)
if new_fn:
with open(new_fn, 'r') as template:
nb_line = 74
for line in template:
if not nb_line:
break
nb_line -= 1
outfn.write(line)
break
else:
cpt_invalid_fn += 1
# Append all arff file to the output file
cur_file_num = 1
for filename in filenames:
if os.path.isfile(indir + filename):
new_fn = validate_arff(indir + filename)
if new_fn:
cur_file_num = cur_file_num + 1
utils.print_progress_start("Analysing file\t" +
str(cur_file_num))
fname = open(new_fn, 'r')
outfn.write("".join(fname.readlines()[74:77]))
fname.close()
else:
cpt_invalid_fn += 1
utils.print_progress_end()
outfn.close()
# os.system("rm " + indir + "*.arff")
if cpt_invalid_fn:
utils.print_warning(
str(cpt_invalid_fn) + " ARFF files with errors found")
return outfilename
def extract_features(dir_audio, dir_feat):
dir_audio = utils.abs_path_dir(dir_audio)
dir_feat = utils.abs_path_dir(dir_feat)
filelist = []
for elem in os.listdir(dir_audio):
if os.path.isfile(dir_audio + elem):
filelist.append(dir_audio + elem)
else:
for filename in os.listdir(dir_audio + elem):
if "ld.wav" in filename:
filelist.append(dir_audio + elem + "/" + filename)
# marsyas(dir_feat, filelist)
for index, filen in enumerate(filelist):
utils.print_progress_start(str(index+1) + "/" + str(len(filelist)) + " " + filen.split(os.sep)[-1])
utils.yaafe(filen)
essentia(dir_feat, filen)
utils.print_progress_end()
import os
import sys
sys.path.insert(0, './src/')
import utils
dir_tracks = "tracks/"
utils.print_success("YAAFE features extraction (approx. 8 minutes)")
# Assert Python version
if sys.version_info.major != 2:
utils.print_error("Yaafe needs Python 2 environment")
# Assert folder exists
dir_tracks = utils.abs_path_dir(dir_tracks)
filelist = os.listdir(dir_tracks)
dir_feat = utils.create_dir(utils.create_dir("features") + "database1")
# dir_tmp = utils.create_dir("tmp")
# dir_yaafe = utils.create_dir(dir_tmp + "yaafe")
# fn_filelist = dir_yaafe + "filelist.txt"
dir_current = os.getcwd()
os.chdir(dir_tracks)
yaafe_cmd = 'yaafe -r 22050 -f "mfcc: MFCC blockSize=2048 stepSize=1024" '
yaafe_cmd += "--resample -b " + dir_feat + " "
for index, filen in enumerate(filelist):
utils.print_progress_start(
str(index + 1) + "/" + str(len(filelist)) + " " + filen)
os.system(yaafe_cmd + filen + "> /dev/null 2>&1")
utils.print_progress_end()
os.chdir(dir_current)
def create_track_feat_testset(folder, infile, outfile, model_file, train=False):
"""Description of create_track_feat_testset
Need to read each test file
compute deltas on mfcc in the ram
predict and predict_proba
generate song and instru ngrams and histograms
Add the mean of mfcc+deltas
append 109 features vector in feat_track/feat_test.csv
"""
utils.print_success("Create track feat testset")
folder = utils.abs_path_dir(folder)
infile = utils.abs_path_file(infile)
clf = joblib.load(model_file)
track_gts = read_gts(infile, separator=",")
for index, filename in enumerate(track_gts):
utils.print_progress_start(str(index+1) + "/" + str(len(track_gts)) + " " + filename)
mfccs = []
mfccs_1 = []
extension = ""
if train:
extension = ""
else:
extension += "_audio_full_mono_22k"
extension += ".wav.mfcc.csv"
with open(folder + filename + extension, "r") as filep:
if train:
next(filep)
next(filep)
next(filep)
next(filep)
next(filep)
for line in filep:
if train:
line = line.split(",")
else:
line = line.split(" ")
mfccs_1.append(str2arr(line[:-1]))
# if train:
# mfccs.append(str2arr(line[:-1]))
# else:
# mfccs.append(str2arr(line[0:]))
mfccs = np.array(mfccs_1)
delta_mfcc = librosa.feature.delta(mfccs)
delta2_mfcc = librosa.feature.delta(mfccs, order=2)
tmp = np.append(mfccs, delta_mfcc, axis=1)
features = np.append(tmp, delta2_mfcc, axis=1)
preds_proba = clf.predict_proba(features)
# Histogramm
nb_hist_class = 10
numbers = column(preds_proba, 0)
hist_pred = np.histogram(numbers, nb_hist_class)
hist_pred_norm = hist_pred[0] / float(sum(hist_pred[0]))
ngram_threshold = 0.5
song_ngram_proba = ngram_proba(local_pred=numbers, threshold=ngram_threshold, above_threshold=True)
instru_ngram_proba = ngram_proba(local_pred=numbers, threshold=ngram_threshold, above_threshold=False)
preds = clf.predict(features)
song_ngram = ngram(preds, "s")
instru_ngram = ngram(preds, "i")
with open(outfile, "a") as filep:
filep.write(filename[:12] + "," +
arr2str(np.mean(mfccs, axis=0)) + "," +
arr2str(np.mean(delta_mfcc, axis=0)) + "," +
arr2str(np.mean(delta2_mfcc, axis=0)) + "," +
arr2str(hist_pred_norm) + "," +
song_ngram_proba + "," +
instru_ngram_proba + "," +
song_ngram + "," +
instru_ngram + "," +
track_gts[filename] + "\n")
utils.print_progress_end()
def process_local_feat(indir, file_gts_track, outdir_local, out_feat_global, train):
"""Description of process_local_feat
Add delta and double delta to MFCCs
"""
utils.print_success("Processing local features")
# Preprocess arg
indir = utils.abs_path_dir(indir)
file_gts_track = utils.abs_path_file(file_gts_track)
filelist = os.listdir(indir)
outdir_local = utils.abs_path_dir(outdir_local)
track_gts = {}
with open(file_gts_track, "r") as filep:
for line in filep:
line = line.split(",")
if train:
index = line[0]
else:
index = line[0] + ".wav.mfcc.csv"
track_gts[index] = line[1][:-1]
for index, filename in enumerate(filelist):
utils.print_progress_start(str(index) + "/" + str(len(filelist)) + " " + filename)
if filename in track_gts:
mfccs = []
groundtruths = []
with open(indir + filename, "r") as filep:
next(filep)
next(filep)
next(filep)
next(filep)
next(filep)
for line in filep:
line = line.split(",")
mfccs.append(str2arr(line[:-1]))
if train:
groundtruths.append(line[-1][:-1])
mfccs = np.array(mfccs)
delta_mfcc = librosa.feature.delta(mfccs)
delta2_mfcc = librosa.feature.delta(mfccs, order=2)
# Write local features in outdir_local
with open(outdir_local + filename, "w") as filep:
gt_to_write = ""
if "i" in track_gts[filename]:
gt_to_write = ",i"
elif "s" in track_gts[filename]:
# postpone frame groundtruth annotationa to another function later in the code
gt_to_write = ""
else:
utils.print_warning("bayle.py line 231 local frame groundtruth undefined")
if train:
for a, b, c, d in zip(mfccs, delta_mfcc, delta2_mfcc, groundtruths):
filep.write(arr2str(a) + "," + arr2str(b) + "," + arr2str(c) + "," + d + "\n")
else:
for a, b, c in zip(mfccs, delta_mfcc, delta2_mfcc):
filep.write(arr2str(a) + "," + arr2str(b) + "," + arr2str(c) + gt_to_write + "\n")
# # Write global features in out_feat_global
# with open(out_feat_global, "a") as filep:
# filep.write(filename + "," +
# arr2str(np.mean(mfccs, axis=0)) + "," +
# arr2str(np.mean(delta_mfcc, axis=0)) + "," +
# arr2str(np.mean(delta2_mfcc, axis=0)) + "," +
# track_gts[filename] + "\n")
utils.print_progress_end()
utils.print_success("Adding local groundtruths to Songs in Jamendo thanks to Ramona annotations")
match_feat_with_song_gt(dir_feat=outdir_local, dir_gts="groundtruths/frame_annot_jamendo_ramona/")
utils.print_success("Done")