2023-05-18 12:52:17 +00:00
|
|
|
|
import time,pdb
|
|
|
|
|
import tqdm
|
|
|
|
|
from time import time as ttime
|
2023-03-10 10:11:04 +00:00
|
|
|
|
import os
|
|
|
|
|
from pathlib import Path
|
|
|
|
|
import logging
|
|
|
|
|
import argparse
|
2023-05-18 13:09:47 +00:00
|
|
|
|
from kmeans import KMeansGPU
|
2023-03-10 10:11:04 +00:00
|
|
|
|
import torch
|
|
|
|
|
import numpy as np
|
2023-05-18 12:52:17 +00:00
|
|
|
|
from sklearn.cluster import KMeans,MiniBatchKMeans
|
|
|
|
|
|
2023-03-10 10:11:04 +00:00
|
|
|
|
logging.basicConfig(level=logging.INFO)
|
|
|
|
|
logger = logging.getLogger(__name__)
|
2023-05-18 12:52:17 +00:00
|
|
|
|
from time import time as ttime
|
|
|
|
|
import pynvml,torch
|
2023-03-10 10:11:04 +00:00
|
|
|
|
|
2023-05-18 12:52:17 +00:00
|
|
|
|
def train_cluster(in_dir, n_clusters, use_minibatch=True, verbose=False,use_gpu=False):#gpu_minibatch真拉,虽然库支持但是也不考虑
|
2023-03-10 10:11:04 +00:00
|
|
|
|
logger.info(f"Loading features from {in_dir}")
|
|
|
|
|
features = []
|
|
|
|
|
nums = 0
|
|
|
|
|
for path in tqdm.tqdm(in_dir.glob("*.soft.pt")):
|
2023-05-18 12:52:17 +00:00
|
|
|
|
# for name in os.listdir(in_dir):
|
|
|
|
|
# path="%s/%s"%(in_dir,name)
|
|
|
|
|
features.append(torch.load(path,map_location="cpu").squeeze(0).numpy().T)
|
2023-03-10 10:11:04 +00:00
|
|
|
|
# print(features[-1].shape)
|
|
|
|
|
features = np.concatenate(features, axis=0)
|
|
|
|
|
print(nums, features.nbytes/ 1024**2, "MB , shape:",features.shape, features.dtype)
|
|
|
|
|
features = features.astype(np.float32)
|
|
|
|
|
logger.info(f"Clustering features of shape: {features.shape}")
|
|
|
|
|
t = time.time()
|
2023-05-18 12:52:17 +00:00
|
|
|
|
if(use_gpu==False):
|
|
|
|
|
if use_minibatch:
|
|
|
|
|
kmeans = MiniBatchKMeans(n_clusters=n_clusters,verbose=verbose, batch_size=4096, max_iter=80).fit(features)
|
|
|
|
|
else:
|
|
|
|
|
kmeans = KMeans(n_clusters=n_clusters,verbose=verbose).fit(features)
|
2023-03-10 10:11:04 +00:00
|
|
|
|
else:
|
2023-05-18 12:52:17 +00:00
|
|
|
|
kmeans = KMeansGPU(n_clusters=n_clusters, mode='euclidean', verbose=2 if verbose else 0,max_iter=500,tol=1e-2)#
|
|
|
|
|
features=torch.from_numpy(features)#.to(device)
|
|
|
|
|
labels = kmeans.fit_predict(features)#
|
|
|
|
|
|
2023-03-10 10:11:04 +00:00
|
|
|
|
print(time.time()-t, "s")
|
|
|
|
|
|
|
|
|
|
x = {
|
2023-05-20 11:43:33 +00:00
|
|
|
|
"n_features_in_": kmeans.n_features_in_ if use_gpu==False else features.shape[1],
|
2023-05-18 12:52:17 +00:00
|
|
|
|
"_n_threads": kmeans._n_threads if use_gpu==False else 4,
|
|
|
|
|
"cluster_centers_": kmeans.cluster_centers_ if use_gpu==False else kmeans.centroids.cpu().numpy(),
|
2023-03-10 10:11:04 +00:00
|
|
|
|
}
|
|
|
|
|
print("end")
|
|
|
|
|
|
|
|
|
|
return x
|
|
|
|
|
|
|
|
|
|
if __name__ == "__main__":
|
|
|
|
|
parser = argparse.ArgumentParser()
|
|
|
|
|
parser.add_argument('--dataset', type=Path, default="./dataset/44k",
|
|
|
|
|
help='path of training data directory')
|
|
|
|
|
parser.add_argument('--output', type=Path, default="logs/44k",
|
|
|
|
|
help='path of model output directory')
|
2023-05-18 13:09:47 +00:00
|
|
|
|
parser.add_argument('--gpu',action='store_true', default=False ,
|
2023-05-18 12:52:17 +00:00
|
|
|
|
help='to use GPU')
|
|
|
|
|
|
2023-03-10 10:11:04 +00:00
|
|
|
|
|
|
|
|
|
args = parser.parse_args()
|
|
|
|
|
|
|
|
|
|
checkpoint_dir = args.output
|
|
|
|
|
dataset = args.dataset
|
2023-05-18 13:09:47 +00:00
|
|
|
|
use_gpu = args.gpu
|
2023-05-20 11:43:33 +00:00
|
|
|
|
n_clusters = 10000
|
2023-05-18 12:52:17 +00:00
|
|
|
|
|
2023-03-10 10:11:04 +00:00
|
|
|
|
ckpt = {}
|
|
|
|
|
for spk in os.listdir(dataset):
|
|
|
|
|
if os.path.isdir(dataset/spk):
|
|
|
|
|
print(f"train kmeans for {spk}...")
|
|
|
|
|
in_dir = dataset/spk
|
2023-05-18 12:52:17 +00:00
|
|
|
|
x = train_cluster(in_dir, n_clusters,use_minibatch=False,verbose=False,use_gpu=use_gpu)
|
2023-03-10 10:11:04 +00:00
|
|
|
|
ckpt[spk] = x
|
|
|
|
|
|
|
|
|
|
checkpoint_path = checkpoint_dir / f"kmeans_{n_clusters}.pt"
|
|
|
|
|
checkpoint_path.parent.mkdir(exist_ok=True, parents=True)
|
|
|
|
|
torch.save(
|
|
|
|
|
ckpt,
|
|
|
|
|
checkpoint_path,
|
|
|
|
|
)
|
2023-05-18 12:52:17 +00:00
|
|
|
|
|