def get_query_data(self, song_id_list):
p = open(norm_param_pkl, "rb")
song_sv_dict = pickle.load(p)
p.close()
print "NUMBER OF SONGS IN QUERY: ", len(song_id_list)
if len(song_id_list) > 1:
timbre_result, rhythm_result = pardora_db_sqlite.get_song_features_from_query(
song_id_list, self.cursor)
t_feature_list = []
r_feature_list = []
for row in timbre_result:
feats = np.array(np.ndarray((row[0], row[1]), buffer=row[2]),
dtype=np.float32)
t_feature_list.append(feats)
timbre_features = np.array(np.concatenate(t_feature_list))
for row in rhythm_result:
feats = np.array(np.ndarray((row[0], row[1]), buffer=row[2]),
dtype=np.float32)
feats = feats.T
r_feature_list.append(feats)
rhythm_features = np.array(np.concatenate(r_feature_list))
print "INFO: Timbre features shape:", timbre_features.shape
print "INFO: Rhythm features shape:", rhythm_features.shape
query_timbre_sv = pardora_ubm.adapt_model(timbre_features,
self.timbre_ubm_params,
M)
query_rhythm_sv = pardora_ubm.adapt_model(rhythm_features,
self.rhythm_ubm_params,
M)
query_timbre_sv = msdtools.mcs_norm(query_timbre_sv,
song_sv_dict['t_sv_mean'])
query_rhythm_sv = msdtools.mcs_norm(query_rhythm_sv,
song_sv_dict['r_sv_mean'])
p_mean_t, p_sigma_t = msdtools.p_norm_params_single(
query_timbre_sv, song_sv_dict['t_sv_sample'].T)
p_mean_r, p_sigma_r = msdtools.p_norm_params_single(
query_rhythm_sv, song_sv_dict['r_sv_sample'].T)
query_dict = {}
query_dict['q_t_sv'] = query_timbre_sv
query_dict['q_r_sv'] = query_rhythm_sv
query_dict['p_mean_t'] = p_mean_t
query_dict['p_mean_r'] = p_mean_r
query_dict['p_sigma_t'] = p_sigma_t
query_dict['p_sigma_r'] = p_sigma_r
else:
query_dict = pardora_db_sqlite.get_song_sv_data(
song_id_list[0], self.cursor)
return query_dict
def get_query_data(self, song_id_list):
p = open(norm_param_pkl, "rb")
song_sv_dict = pickle.load(p)
p.close()
print "NUMBER OF SONGS IN QUERY: ", len(song_id_list)
if len(song_id_list) > 1:
timbre_result, rhythm_result = pardora_db.get_song_features_from_query(song_id_list, self.cursor)
t_feature_list = []
r_feature_list = []
for row in timbre_result:
feats = np.array(np.ndarray((row[0],row[1]), buffer=row[2]), dtype=np.float32)
t_feature_list.append(feats)
timbre_features = np.array(np.concatenate(t_feature_list))
for row in rhythm_result:
feats = np.array(np.ndarray((row[0],row[1]), buffer=row[2]), dtype=np.float32)
feats = feats.T
r_feature_list.append(feats)
rhythm_features = np.array(np.concatenate(r_feature_list))
print "INFO: Timbre features shape:", timbre_features.shape
print "INFO: Rhythm features shape:", rhythm_features.shape
query_timbre_sv = pardora_ubm.adapt_model(timbre_features, self.timbre_ubm_params, M)
query_rhythm_sv = pardora_ubm.adapt_model(rhythm_features, self.rhythm_ubm_params, M)
query_timbre_sv = msdtools.mcs_norm(query_timbre_sv, song_sv_dict['t_sv_mean'])
query_rhythm_sv = msdtools.mcs_norm(query_rhythm_sv, song_sv_dict['r_sv_mean'])
p_mean_t, p_sigma_t = msdtools.p_norm_params_single(query_timbre_sv, song_sv_dict['t_sv_sample'].T)
p_mean_r, p_sigma_r = msdtools.p_norm_params_single(query_rhythm_sv, song_sv_dict['r_sv_sample'].T)
query_dict = {}
query_dict['q_t_sv'] = query_timbre_sv
query_dict['q_r_sv'] = query_rhythm_sv
query_dict['p_mean_t'] = p_mean_t
query_dict['p_mean_r'] = p_mean_r
query_dict['p_sigma_t'] = p_sigma_t
query_dict['p_sigma_r'] = p_sigma_r
else:
query_dict = pardora_db.get_song_sv_data(song_id_list[0], self.cursor)
return query_dict
def compute_and_add_song_svs(timbre_ubm_params, rhythm_ubm_params, db_cursor):
print "............... Computing and Adding Supervectors To DB ...................."
p = open(song_id_pkl, "rb")
song_ids = pickle.load(p)
p.close()
song_id_chunks = chunks(song_ids, CHUNK_SIZE)
chunk_count = 0
t_mean_to_use = np.zeros(1)
t_sv = np.zeros(1)
r_mean_to_use = np.zeros(1)
r_sv = np.zeros(1)
all_songs_list = []
total_time = time.time()
for chunk in song_id_chunks:
song_id_list = []
timbre_sv_arr = []
rhythm_sv_arr = []
print "==== CHUNK: ", chunk_count, "===="
chunk_count += 1
chunk_time = time.time()
songs = pardora_db.get_rhythm_features_for_song_ids(chunk, db_cursor)
for s in songs:
if s[1] is not None and s[2] is not None:
feats = np.array(np.ndarray((s[1], s[2]), buffer=s[0]), dtype=np.float32)
feats_t = feats.T
rhythm_sv = pardora_ubm.adapt_model(feats_t, rhythm_ubm_params, M)
rhythm_sv_arr.append(rhythm_sv)
song_id_list.append(s[3])
print "INFO: Rhythm SV comp time: ", time.time() - st
st = time.time()
songs = pardora_db.get_timbre_features_for_song_ids(chunk, db_cursor)
for s in songs:
if s[3] in song_id_list:
feats = np.array(np.ndarray((s[1], s[2]), buffer=s[0]), dtype=np.float32)
timbre_sv = pardora_ubm.adapt_model(feats, timbre_ubm_params, M)
timbre_sv_arr.append(timbre_sv)
print "INFO: Timbre SV comp time: ", time.time() - st
st = time.time()
t_sv = np.vstack(timbre_sv_arr)
del timbre_sv_arr
r_sv = np.vstack(rhythm_sv_arr)
del rhythm_sv_arr
if chunk_count == 1:
t_mean_to_use = np.mean(t_sv, axis=0)
r_mean_to_use = np.mean(r_sv, axis=0)
t_sv = msdtools.mcs_norm(t_sv.T, t_mean_to_use).T
r_sv = msdtools.mcs_norm(r_sv.T, r_mean_to_use).T
print "INFO: MCS norm computation time: ", time.time() - st
st = time.time()
p_means_t = np.zeros(len(song_id_list))
p_sigmas_t = p_means_t.copy()
p_means_t = p_means_t.copy()
p_sigmas_t = p_means_t.copy()
st = time.time()
p_means_t, p_sigmas_t = msdtools.p_norm_params_chunk(t_sv.T, t_sv.T, NORM_CHUNK_SIZE)
p_means_r, p_sigmas_r = msdtools.p_norm_params_chunk(r_sv.T, r_sv.T, NORM_CHUNK_SIZE)
print "INFO: P-means computation time: ", time.time() - st
st = time.time()
idx = 0
for s_id in song_id_list:
t = np.array(t_sv[idx])
r = np.array(r_sv[idx])
if s_id not in all_songs_list:
pardora_db.add_sv_and_p_vals_to_db(
conn, c, s_id, t, r, p_means_t[idx], p_sigmas_t[idx], p_means_r[idx], p_sigmas_r[idx], db_cursor
)
all_songs_list.append(s_id)
idx += 1
conn.commit()
print "INFO: Database update time: ", time.time() - st
print "INFO: TOTAL CHUNK TIME:", time.time() - chunk_time
print "=============================================="
print "INFO: TOTAL TIME FOR SV COMP TIME: ", time.time() - total_time
print "=============================================="
d = {}
d["t_sv_mean"] = t_mean_to_use
d["t_sv_sample"] = t_sv
d["r_sv_mean"] = r_mean_to_use
d["r_sv_sample"] = r_sv
p = open(norm_param_pkl, "wb")
pickle.dump(d, p, True)
p.close()
return
def compute_and_add_song_svs(timbre_ubm_params, rhythm_ubm_params, db_cursor):
print "............... Computing and Adding Supervectors To DB ...................."
p = open(song_id_pkl, "rb")
song_ids = pickle.load(p)
p.close()
song_id_chunks = chunks(song_ids, CHUNK_SIZE)
chunk_count = 0
t_mean_to_use = np.zeros(1)
t_sv = np.zeros(1)
r_mean_to_use = np.zeros(1)
r_sv = np.zeros(1)
all_songs_list = []
total_time = time.time()
for chunk in song_id_chunks:
song_id_list = []
timbre_sv_arr = []
rhythm_sv_arr = []
print "==== CHUNK: ", chunk_count, "===="
chunk_count += 1
chunk_time = time.time()
songs = pardora_db.get_rhythm_features_for_song_ids(chunk, db_cursor)
for s in songs:
if s[1] is not None and s[2] is not None:
feats = np.array(np.ndarray((s[1], s[2]), buffer=s[0]),
dtype=np.float32)
feats_t = feats.T
rhythm_sv = pardora_ubm.adapt_model(feats_t, rhythm_ubm_params,
M)
rhythm_sv_arr.append(rhythm_sv)
song_id_list.append(s[3])
print "INFO: Rhythm SV comp time: ", time.time() - st
st = time.time()
songs = pardora_db.get_timbre_features_for_song_ids(chunk, db_cursor)
for s in songs:
if s[3] in song_id_list:
feats = np.array(np.ndarray((s[1], s[2]), buffer=s[0]),
dtype=np.float32)
timbre_sv = pardora_ubm.adapt_model(feats, timbre_ubm_params,
M)
timbre_sv_arr.append(timbre_sv)
print "INFO: Timbre SV comp time: ", time.time() - st
st = time.time()
t_sv = np.vstack(timbre_sv_arr)
del timbre_sv_arr
r_sv = np.vstack(rhythm_sv_arr)
del rhythm_sv_arr
if chunk_count == 1:
t_mean_to_use = np.mean(t_sv, axis=0)
r_mean_to_use = np.mean(r_sv, axis=0)
t_sv = msdtools.mcs_norm(t_sv.T, t_mean_to_use).T
r_sv = msdtools.mcs_norm(r_sv.T, r_mean_to_use).T
print "INFO: MCS norm computation time: ", time.time() - st
st = time.time()
p_means_t = np.zeros(len(song_id_list))
p_sigmas_t = p_means_t.copy()
p_means_t = p_means_t.copy()
p_sigmas_t = p_means_t.copy()
st = time.time()
p_means_t, p_sigmas_t = msdtools.p_norm_params_chunk(
t_sv.T, t_sv.T, NORM_CHUNK_SIZE)
p_means_r, p_sigmas_r = msdtools.p_norm_params_chunk(
r_sv.T, r_sv.T, NORM_CHUNK_SIZE)
print "INFO: P-means computation time: ", time.time() - st
st = time.time()
idx = 0
for s_id in song_id_list:
t = np.array(t_sv[idx])
r = np.array(r_sv[idx])
if s_id not in all_songs_list:
pardora_db.add_sv_and_p_vals_to_db(conn, c, s_id, t, r, \
p_means_t[idx], p_sigmas_t[idx],\
p_means_r[idx], p_sigmas_r[idx],\
db_cursor)
all_songs_list.append(s_id)
idx += 1
conn.commit()
print "INFO: Database update time: ", time.time() - st
print "INFO: TOTAL CHUNK TIME:", time.time() - chunk_time
print "=============================================="
print "INFO: TOTAL TIME FOR SV COMP TIME: ", time.time() - total_time
print "=============================================="
d = {}
d['t_sv_mean'] = t_mean_to_use
d['t_sv_sample'] = t_sv
d['r_sv_mean'] = r_mean_to_use
d['r_sv_sample'] = r_sv
p = open(norm_param_pkl, "wb")
pickle.dump(d, p, True)
p.close()
return
