2023-05-24 14:26:35 +02:00
|
|
|
|
"""
|
|
|
|
|
格式:直接cid为自带的index位;aid放不下了,通过字典来查,反正就5w个
|
|
|
|
|
"""
|
|
|
|
|
import faiss, numpy as np, os
|
2023-06-28 07:48:06 +02:00
|
|
|
|
from sklearn.cluster import MiniBatchKMeans
|
|
|
|
|
import traceback
|
|
|
|
|
from multiprocessing import cpu_count
|
2023-05-24 14:26:35 +02:00
|
|
|
|
|
|
|
|
|
# ###########如果是原始特征要先写save
|
2023-06-28 07:48:06 +02:00
|
|
|
|
n_cpu = 0
|
|
|
|
|
if n_cpu == 0:
|
|
|
|
|
n_cpu = cpu_count()
|
|
|
|
|
inp_root = r"./logs/anz/3_feature768"
|
2023-05-24 14:26:35 +02:00
|
|
|
|
npys = []
|
|
|
|
|
listdir_res = list(os.listdir(inp_root))
|
|
|
|
|
for name in sorted(listdir_res):
|
|
|
|
|
phone = np.load("%s/%s" % (inp_root, name))
|
|
|
|
|
npys.append(phone)
|
|
|
|
|
big_npy = np.concatenate(npys, 0)
|
|
|
|
|
big_npy_idx = np.arange(big_npy.shape[0])
|
|
|
|
|
np.random.shuffle(big_npy_idx)
|
|
|
|
|
big_npy = big_npy[big_npy_idx]
|
|
|
|
|
print(big_npy.shape) # (6196072, 192)#fp32#4.43G
|
2023-06-28 07:48:06 +02:00
|
|
|
|
if big_npy.shape[0] > 2e5:
|
|
|
|
|
# if(1):
|
|
|
|
|
info = "Trying doing kmeans %s shape to 10k centers." % big_npy.shape[0]
|
|
|
|
|
print(info)
|
|
|
|
|
try:
|
|
|
|
|
big_npy = (
|
|
|
|
|
MiniBatchKMeans(
|
|
|
|
|
n_clusters=10000,
|
|
|
|
|
verbose=True,
|
|
|
|
|
batch_size=256 * n_cpu,
|
|
|
|
|
compute_labels=False,
|
|
|
|
|
init="random",
|
|
|
|
|
)
|
|
|
|
|
.fit(big_npy)
|
|
|
|
|
.cluster_centers_
|
|
|
|
|
)
|
|
|
|
|
except:
|
|
|
|
|
info = traceback.format_exc()
|
|
|
|
|
print(info)
|
|
|
|
|
|
|
|
|
|
np.save("tools/infer/big_src_feature_mi.npy", big_npy)
|
2023-05-24 14:26:35 +02:00
|
|
|
|
|
|
|
|
|
##################train+add
|
|
|
|
|
# big_npy=np.load("/bili-coeus/jupyter/jupyterhub-liujing04/vits_ch/inference_f0/big_src_feature_mi.npy")
|
|
|
|
|
n_ivf = min(int(16 * np.sqrt(big_npy.shape[0])), big_npy.shape[0] // 39)
|
|
|
|
|
index = faiss.index_factory(768, "IVF%s,Flat" % n_ivf) # mi
|
|
|
|
|
print("training")
|
|
|
|
|
index_ivf = faiss.extract_index_ivf(index) #
|
|
|
|
|
index_ivf.nprobe = 1
|
|
|
|
|
index.train(big_npy)
|
|
|
|
|
faiss.write_index(
|
2023-06-28 07:48:06 +02:00
|
|
|
|
index, "tools/infer/trained_IVF%s_Flat_baseline_src_feat_v2.index" % (n_ivf)
|
2023-05-24 14:26:35 +02:00
|
|
|
|
)
|
|
|
|
|
print("adding")
|
|
|
|
|
batch_size_add = 8192
|
|
|
|
|
for i in range(0, big_npy.shape[0], batch_size_add):
|
|
|
|
|
index.add(big_npy[i : i + batch_size_add])
|
2023-06-28 07:48:06 +02:00
|
|
|
|
faiss.write_index(
|
|
|
|
|
index, "tools/infer/added_IVF%s_Flat_mi_baseline_src_feat.index" % (n_ivf)
|
|
|
|
|
)
|
2023-05-24 14:26:35 +02:00
|
|
|
|
"""
|
|
|
|
|
大小(都是FP32)
|
|
|
|
|
big_src_feature 2.95G
|
|
|
|
|
(3098036, 256)
|
|
|
|
|
big_emb 4.43G
|
|
|
|
|
(6196072, 192)
|
|
|
|
|
big_emb双倍是因为求特征要repeat后再加pitch
|
|
|
|
|
|
|
|
|
|
"""
|