Spaces:
Sleeping
Sleeping
File size: 6,182 Bytes
7e4b981 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 |
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import argparse
import logging
import os
import time
import numpy as np
from sklearn.cluster import MiniBatchKMeans
import joblib
from examples.textless_nlp.gslm.speech2unit.pretrained.utils import (
get_and_dump_features,
get_features,
)
def get_logger():
log_format = "[%(asctime)s] [%(levelname)s]: %(message)s"
logging.basicConfig(format=log_format, level=logging.INFO)
logger = logging.getLogger(__name__)
return logger
def get_parser():
parser = argparse.ArgumentParser(
description="Learn K-means clustering over acoustic features."
)
# Features arguments
parser.add_argument(
"--in_features_path", type=str, default=None, help="Features file path"
)
parser.add_argument(
"--feature_type",
type=str,
choices=["logmel", "hubert", "w2v2", "cpc"],
default=None,
help="Acoustic feature type",
)
parser.add_argument(
"--manifest_path",
type=str,
default=None,
help="Manifest file containing the root dir and file names",
)
parser.add_argument(
"--out_features_path",
type=str,
default=None,
help="Features file path to write to",
)
parser.add_argument(
"--checkpoint_path",
type=str,
help="Pretrained acoustic model checkpoint",
)
parser.add_argument(
"--layer",
type=int,
help="The layer of the pretrained model to extract features from",
default=-1,
)
parser.add_argument(
"--sample_pct",
type=float,
help="Percent data to use for K-means training",
default=0.1,
)
# K-means arguments
parser.add_argument(
"--num_clusters", type=int, help="Nubmer of clusters", default=50
)
parser.add_argument("--init", default="k-means++")
parser.add_argument(
"--max_iter",
type=int,
help="Maximum number of iterations for K-means training",
default=150,
)
parser.add_argument(
"--batch_size",
type=int,
help="Batch size for K-means training",
default=10000,
)
parser.add_argument("--tol", default=0.0, type=float)
parser.add_argument("--max_no_improvement", default=100, type=int)
parser.add_argument("--n_init", default=20, type=int)
parser.add_argument("--reassignment_ratio", default=0.5, type=float)
parser.add_argument(
"--out_kmeans_model_path",
type=str,
required=True,
help="Path to save K-means model",
)
# Leftovers
parser.add_argument(
"--seed",
type=int,
help="Random seed to use for K-means training",
default=1369,
)
return parser
def get_kmeans_model(
n_clusters,
init,
max_iter,
batch_size,
tol,
max_no_improvement,
n_init,
reassignment_ratio,
random_state,
):
return MiniBatchKMeans(
n_clusters=n_clusters,
init=init,
max_iter=max_iter,
batch_size=batch_size,
tol=tol,
max_no_improvement=max_no_improvement,
n_init=n_init,
reassignment_ratio=reassignment_ratio,
random_state=random_state,
verbose=1,
compute_labels=True,
init_size=None,
)
def train_kmeans(kmeans_model, features_batch):
start_time = time.time()
kmeans_model.fit(features_batch)
time_taken = round((time.time() - start_time) // 60, 2)
return kmeans_model, time_taken
def main(args, logger):
# Features loading/extraction for K-means
if args.in_features_path:
# Feature loading
logger.info(f"Loading features from {args.in_features_path}...")
features_batch = np.load(args.in_features_path, allow_pickle=True)
else:
# Feature extraction
logger.info(f"Extracting {args.feature_type} acoustic features...")
features_batch = (
get_features(
feature_type=args.feature_type,
checkpoint_path=args.checkpoint_path,
layer=args.layer,
manifest_path=args.manifest_path,
sample_pct=args.sample_pct,
flatten=True,
)
if not args.out_features_path
else get_and_dump_features(
feature_type=args.feature_type,
checkpoint_path=args.checkpoint_path,
layer=args.layer,
manifest_path=args.manifest_path,
sample_pct=args.sample_pct,
flatten=True,
out_features_path=args.out_features_path,
)
)
if args.out_features_path:
logger.info(
f"Saved extracted features at {args.out_features_path}"
)
logger.info(f"Features shape = {features_batch.shape}\n")
# Learn and save K-means model
kmeans_model = get_kmeans_model(
n_clusters=args.num_clusters,
init=args.init,
max_iter=args.max_iter,
batch_size=args.batch_size,
tol=args.tol,
max_no_improvement=args.max_no_improvement,
n_init=args.n_init,
reassignment_ratio=args.reassignment_ratio,
random_state=args.seed,
)
logger.info("Starting k-means training...")
kmeans_model, time_taken = train_kmeans(
kmeans_model=kmeans_model, features_batch=features_batch
)
logger.info(f"...done k-means training in {time_taken} minutes")
inertia = -kmeans_model.score(features_batch) / len(features_batch)
logger.info(f"Total intertia: {round(inertia, 2)}\n")
logger.info(f"Saving k-means model to {args.out_kmeans_model_path}")
os.makedirs(os.path.dirname(args.out_kmeans_model_path), exist_ok=True)
joblib.dump(kmeans_model, open(args.out_kmeans_model_path, "wb"))
if __name__ == "__main__":
parser = get_parser()
args = parser.parse_args()
logger = get_logger()
logger.info(args)
main(args, logger)
|