def get_X(track_id,
use_mp3,
use_lyrics,
use_artist,
lyrics_d2v_model,
d_artist_vec,
music_datatype="mfcc"):
conn = MyConn()
rawmusic_path, vggish_embed_path, lyrics_path, artist = conn.query(
table="sub_tracks",
conditions={"track_id": track_id},
fetchall=False,
targets=[
"rawmusic_path", "vggish_embed_path", "lyrics_path", "artist"
])
vecs = []
if use_mp3:
if music_datatype == "mfcc":
music_vec = get_mfcc(rawmusic_path).ravel()
elif music_datatype == "vggish":
with open(vggish_embed_path, "rb") as f:
music_vec = pickle.load(f).detach().numpy()
vecs.append(music_vec)
if use_lyrics:
lyrics_vec = get_d2v_vector(lyrics_path, lyrics_d2v_model)
vecs.append(lyrics_vec)
if use_artist:
artist_vec = d_artist_vec[artist.lower().strip()]
vecs.append(artist_vec)
features_vec = concatenate_features(vecs)
return features_vec
def __getitem__(self, index):
'''
必须自定义，使用index获取成对数据
'''
if index<len(self.breakouts):
label = 1
id_ = self.breakouts[index]
track_id, beta = self.conn.query(targets=["track_id", "beta"], table="breakouts",
conditions={"id":id_})[0]
feature_words = self.conn.query(targets=["feature_words"], table="breakouts_feature_words_1",
conditions={"breakout_id":id_})[0][0].split()
else:
label = 0
index = index - len(self.breakouts)
id_ = self.no_breakouts[index]
track_id = id_.split('-')[0]
beta = 1
feature_words = self.conn.query(targets=["feature_words"], table="no_breakouts_feature_words_1",
conditions={"id":id_})[0][0].split()
rawmusic_path, lyrics_path = self.conn.query(
targets=["rawmusic_path", "lyrics_path"], conditions={"track_id": track_id})[0]
mfcc = torch.Tensor(get_mfcc(rawmusic_path))
lyrics_vec = torch.Tensor(get_d2v_vector(lyrics_path, self.d2v_model))
return label, beta, mfcc, lyrics_vec, feature_words
def get_X_y_embed(tracks_2_labels, conn, d2v_model, music_feature_extractor,
intrinsic_feature_embed):
# 构建.pkl文件的路径字典
dir1 = "/Volumes/nmusic/NetEase2020/data/breakouts_rawmusic"
dir2 = "/Volumes/nmusic/NetEase2020/data/no_breakouts_rawmusic"
tid_2_mp3path_d = get_tid_2_mp3path_d(dir1)
tid_2_mp3path_d2 = get_tid_2_mp3path_d(dir2)
for tid in tid_2_mp3path_d2:
if tid not in tid_2_mp3path_d:
tid_2_mp3path_d[tid] = tid_2_mp3path_d2[tid]
# print(len(tid_2_mp3path_d))
X = []
y = []
flag = 1
for tid, label in tracks_2_labels.items():
tid = str(tid)
try:
# 获取音频特征向量
mfcc = torch.tensor(get_mfcc(tid_2_mp3path_d[tid])).unsqueeze(0)
# 获取歌词特征向量
lyrics_path = conn.query(targets=["lyrics_path"],
conditions={"track_id": tibd})[0][0]
lyrics_vec = torch.tensor(
get_d2v_vector(
"/Volumes/nmusic/NetEase2020/data/" + lyrics_path,
d2v_model))
h1 = music_feature_extractor(mfcc).squeeze()
# print(h1.shape)
h2 = torch.cat((h1, lyrics_vec))
# print(h2.shape)
feature_vec = intrinsic_feature_embed(h2)
X.append(feature_vec)
y.append(label)
print(flag, tid)
flag += 1
except KeyboardInterrupt:
print("interrupted by user.")
sys.exit(1)
except:
print("ERROR", tid)
print(traceback.format_exc())
# continue
# break
return X, y
def get_mfcc(signal_rate=100, output_path=None, audio_paths=None):
"""
get mfcc
:param sr: sampling rate
:param output_path:
:param audio_paths:
:return:
"""
audio = audio_paths[0].split('.')[-1]
for audio_path in audio_paths:
name = re.findall('{}\/(.*)\.{}'.format(audio, audio), audio_path)[0]
mfcc_df = utils.get_mfcc(audio_path, signal_rate)
#mfcc_df = mfcc_df[(mfcc_df.T == 0).any() == False].reset_index(drop=True)
path = output_path
if not os.path.exists(path):
os.makedirs(path)
mfcc_df.T.to_csv(path + '/{0}.csv'.format(name), index=False)
def __getitem__(self, index):
'''
必须定义，使用index获取一条完整的数据
'''
tid, label = self.data[index]
rawmusic_path, vggish_embed_path, vggish_examples_path, mel_3seconds_groups_path, lyrics_path, artist, chorus_start = self.conn.query(
table="sub_tracks", conditions={"track_id":tid}, fetchall=False,
targets=["rawmusic_path", "vggish_embed_path", "vggish_examples_path", "mel_3seconds_groups_path", "lyrics_path", "artist", "chorus_start"])
if self.config.MUSIC_DATATYPE=="vggish":
with open(vggish_embed_path, "rb") as f:
music_vec = pickle.load(f)
elif self.config.MUSIC_DATATYPE=="mel":
music_vec = torch.Tensor(get_melspectrogram(rawmusic_path, config=self.config))
elif self.config.MUSIC_DATATYPE=="mfcc":
music_vec = torch.Tensor(get_mfcc(rawmusic_path, config=self.config))
elif self.config.MUSIC_DATATYPE=="vggish_examples":
with open(vggish_examples_path, "rb") as f:
music_vec = pickle.load(f)
elif self.config.MUSIC_DATATYPE=="mel_3seconds_groups":
with open(mel_3seconds_groups_path, "rb") as f:
music_vec = pickle.load(f)
lyrics_vec = torch.Tensor(get_d2v_vector(lyrics_path, self.d2v_model))
artist_vec = torch.Tensor(self.d_artist_vec[artist.lower().strip()])
# reviews_vec = torch.Tensor(get_reviews_vec(tid, breakout=label,
# w2v_model=self.w2v_model, key=self.config.REVIEWS_VEC_KEY)[:self.config.TOPK])
# reviews_topk_words = get_reviews_topk_words(tid, is_breakout=label, key="candidates")[:5]
reviews_vec = torch.Tensor(get_reviews_vec_with_freq(tid, breakout=label, w2v_model=self.w2v_model,
d_breakouts=self.d_breakouts_feature_words, d_no_breakouts=self.d_no_breakouts_feature_words,
d_pos_track_breakout=self.d_pos_track_breakout, with_freq=False)[:self.config.TOPK])
item_data = {
"track_id": tid,
"label": label,
"music_vec": music_vec,
"lyrics_vec": lyrics_vec,
"artist_vec": artist_vec,
"reviews_vec": reviews_vec,
# "reviews_topk_words": reviews_topk_words
}
return item_data