From 7c754dc420996d2216d4d480a76f7603cc4cec4c Mon Sep 17 00:00:00 2001
From: ylzz1997 <ylzz1997@outlook.com>
Date: Sat, 20 May 2023 19:43:33 +0800
Subject: [PATCH] Debug Kmeans

---
 cluster/km_train.py      | 80 ----------------------------------------
 cluster/train_cluster.py |  4 +-
 2 files changed, 2 insertions(+), 82 deletions(-)
 delete mode 100644 cluster/km_train.py

diff --git a/cluster/km_train.py b/cluster/km_train.py
deleted file mode 100644
index 917b2da..0000000
--- a/cluster/km_train.py
+++ /dev/null
@@ -1,80 +0,0 @@
-import time,pdb
-import tqdm
-from time import time as ttime
-import os
-from pathlib import Path
-import logging
-import argparse
-from cluster.kmeans import KMeansGPU
-import torch
-import numpy as np
-from sklearn.cluster import KMeans,MiniBatchKMeans
-
-logging.basicConfig(level=logging.INFO)
-logger = logging.getLogger(__name__)
-from time import time as ttime
-import pynvml,torch
-
-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
-    logger.info(f"Loading features from {in_dir}")
-    features = []
-    nums = 0
-    for path in tqdm.tqdm(in_dir.glob("*.soft.pt")):
-    # 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)
-        # 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()
-    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)
-    else:
-            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)#
-
-    print(time.time()-t, "s")
-
-    x = {
-            "n_features_in_": kmeans.n_features_in_ if use_gpu==False else features.shape[0],
-            "_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(),
-    }
-    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')
-
-    args = parser.parse_args()
-
-    checkpoint_dir = args.output
-    dataset = args.dataset
-    n_clusters = 1000
-    
-    ckpt = {}
-    for spk in os.listdir(dataset):
-        if os.path.isdir(dataset/spk):
-            print(f"train kmeans for {spk}...")
-            in_dir = dataset/spk
-            x = train_cluster(in_dir, n_clusters,use_minibatch=False,verbose=False,use_gpu=True)
-            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,
-    )
-    
diff --git a/cluster/train_cluster.py b/cluster/train_cluster.py
index b8d4964..8644566 100644
--- a/cluster/train_cluster.py
+++ b/cluster/train_cluster.py
@@ -42,7 +42,7 @@ def train_cluster(in_dir, n_clusters, use_minibatch=True, verbose=False,use_gpu=
     print(time.time()-t, "s")
 
     x = {
-            "n_features_in_": kmeans.n_features_in_ if use_gpu==False else features.shape[0],
+            "n_features_in_": kmeans.n_features_in_ if use_gpu==False else features.shape[1],
             "_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(),
     }
@@ -65,7 +65,7 @@ if __name__ == "__main__":
     checkpoint_dir = args.output
     dataset = args.dataset
     use_gpu = args.gpu
-    n_clusters = 1000
+    n_clusters = 10000
     
     ckpt = {}
     for spk in os.listdir(dataset):