File size: 8,662 Bytes
884b8f8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 | import os
# 如果在 "tests" 目录下运行，则切回上一层
pwd = os.getcwd()
if pwd.endswith("tests"):
os.chdir(os.path.dirname(pwd))
import time
import torch
import numpy as np
from marble.tasks.Chords1217._datamodule_v1 import Chords1217AudioTrain
from torch.utils.data import DataLoader
def benchmark_get_targets_single_clip(
jsonl_path: str,
sample_rate: int,
channels: int,
clip_seconds: float,
label_freq: int,
mode: str,
num_files: int = 20
):
"""
对 Chords1217AudioTrain 在指定 mode 下，测量前 num_files 个文件的第一个 slice (slice_idx=0)
调用 get_targets 的总耗时。仅供对比三种实现（python/numpy/numba）哪种最慢。
"""
print(f"\n--- Benchmark get_targets mode='{mode}' (仅前 {num_files} 个文件，每个文件的第 0 号 slice) ---")
# 1. 实例化 Dataset（这里只用 Train split）
dataset = Chords1217AudioTrain(
jsonl=jsonl_path,
sample_rate=sample_rate,
channels=channels,
clip_seconds=clip_seconds,
label_freq=label_freq,
mode=mode
)
# 假设每个文件至少有一个 slice，且 orig_sr == sample_rate，orig_clip_frames == clip_seconds * sample_rate
orig_sr = sample_rate
orig_clip_frames = int(clip_seconds * sample_rate)
t0 = time.perf_counter()
for file_idx in range(min(num_files, len(dataset.chords_meta))):
# 只测第 0 号 slice 的 get_targets
_ = dataset.get_targets(
file_idx=file_idx,
slice_idx=0,
orig_sr=orig_sr,
orig_clip_frames=orig_clip_frames
)
t1 = time.perf_counter()
total_time = t1 - t0
count = min(num_files, len(dataset.chords_meta))
print(f"Mode='{mode}'，总耗时：{total_time:.4f}s，平均每个文件第 0 号 slice get_targets 用时 {total_time/count:.4f}s")
def benchmark_get_targets_full_epoch(
jsonl_path: str,
sample_rate: int,
channels: int,
clip_seconds: float,
label_freq: int,
mode: str,
batch_size: int = 8,
num_workers: int = 4
):
"""
通过 DataLoader 对整个 train split 迭代一遍，测量整个 epoch 调用 get_targets + collate 的时间。
"""
print(f"\n--- Benchmark get_targets mode='{mode}' （使用 DataLoader 遍历整个 train split）---")
train_ds = Chords1217AudioTrain(
jsonl=jsonl_path,
sample_rate=sample_rate,
channels=channels,
clip_seconds=clip_seconds,
label_freq=label_freq,
mode=mode
)
loader = DataLoader(
train_ds,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
pin_memory=True,
prefetch_factor=2,
persistent_workers=True
)
t0 = time.perf_counter()
for batch in loader:
# batch 里会调用 get_targets(...) + 其他预处理（如音频 loading、transform 等）
pass
t1 = time.perf_counter()
print(f"Mode='{mode}'，整 train split 迭代一轮 get_targets 耗时：{t1-t0:.4f}s")
def benchmark_load_and_preprocess_single_clip(
jsonl_path: str,
sample_rate: int,
channels: int,
clip_seconds: float,
label_freq: int,
mode: str,
num_files: int = 20
):
"""
针对 Chords1217AudioTrain，在指定 mode 下，测量前 num_files 个文件的第一个 slice (slice_idx=0)
调用 _load_and_preprocess 的总耗时。仅供对比三种实现（python/numpy/numba）在纯加载+预处理上的差异。
"""
print(f"\n--- Benchmark load_and_preprocess mode='{mode}' (仅前 {num_files} 个文件，第 0 号 slice) ---")
# 1. 实例化 Dataset（这里只用 Train split）
dataset = Chords1217AudioTrain(
jsonl=jsonl_path,
sample_rate=sample_rate,
channels=channels,
clip_seconds=clip_seconds,
label_freq=label_freq,
mode=mode
)
# 假设 orig_sr == sample_rate，orig_clip_frames == clip_seconds * sample_rate
orig_sr = sample_rate
orig_clip_frames = int(clip_seconds * sample_rate)
t0 = time.perf_counter()
for file_idx in range(min(num_files, len(dataset.meta))):
info = dataset.meta[file_idx]
audio_path = info["audio_path"]
# 测量 _load_and_preprocess 的耗时
_ = dataset._load_and_preprocess(
path=audio_path,
slice_idx=0,
orig_sr=orig_sr,
orig_clip_frames=orig_clip_frames
)
t1 = time.perf_counter()
total_time = t1 - t0
count = min(num_files, len(dataset.meta))
print(f"Mode='{mode}'，总耗时：{total_time:.4f}s，平均每个文件第 0 号 slice _load_and_preprocess 用时 {total_time/count:.4f}s")
def benchmark_load_and_preprocess_full_epoch(
jsonl_path: str,
sample_rate: int,
channels: int,
clip_seconds: float,
label_freq: int,
mode: str,
batch_size: int = 8,
num_workers: int = 4
):
"""
通过 DataLoader 遍历整个 train split，但在 collate 阶段只收集 waveform，不调用 get_targets，
仅测 _load_and_preprocess 的耗时。
"""
print(f"\n--- Benchmark load_and_preprocess mode='{mode}' （使用 DataLoader 遍历 entire train split，仅 _load_and_preprocess） ---")
train_ds = Chords1217AudioTrain(
jsonl=jsonl_path,
sample_rate=sample_rate,
channels=channels,
clip_seconds=clip_seconds,
label_freq=label_freq,
mode=mode
)
# 定义一个只做加载 + 预处理的 collate_fn
def collate_waveforms(batch):
# batch 是 [(waveform, targets, audio_path), ...]
# waveform 已经包含了 _load_and_preprocess 的结果
waveforms = [item[0] for item in batch]
return torch.stack(waveforms, dim=0)
loader = DataLoader(
train_ds,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
collate_fn=collate_waveforms,
pin_memory=True,
prefetch_factor=2,
persistent_workers=True
)
t0 = time.perf_counter()
for _ in loader:
# 只负责跑 _load_and_preprocess，然后被 collate_fn 收集
pass
t1 = time.perf_counter()
print(f"Mode='{mode}'，整 train split 迭代一轮 _load_and_preprocess 耗时：{t1-t0:.4f}s")
if __name__ == '__main__':
# ====== 以下是需要根据实际情况修改的参数 ======
JSONL_PATH = 'data/Chords1217/Chords1217.train.jsonl' # 你的 Chords1217 metadata jsonl 文件
SAMPLE_RATE = 24000
CHANNELS = 1
CLIP_SECONDS = 15.0 # 每个 clip 时长
LABEL_FREQ = 75 # 每秒多少帧标签（对 get_targets 有用，对 load_and_preprocess 无影响）
NUM_FILES = 500 # 单文件测试时用到的文件数量
BATCH_SIZE = 64 # DataLoader 的 batch_size
NUM_WORKERS = 4 # DataLoader 的 num_workers
# -------------------------------
# 1) 先跑 get_targets 的基准测试
# -------------------------------
for m in ['python', 'numpy', 'numba']:
benchmark_get_targets_single_clip(
jsonl_path=JSONL_PATH,
sample_rate=SAMPLE_RATE,
channels=CHANNELS,
clip_seconds=CLIP_SECONDS,
label_freq=LABEL_FREQ,
mode=m,
num_files=NUM_FILES
)
# for m in ['python', 'numpy', 'numba']:
# benchmark_get_targets_full_epoch(
# jsonl_path=JSONL_PATH,
# sample_rate=SAMPLE_RATE,
# channels=CHANNELS,
# clip_seconds=CLIP_SECONDS,
# label_freq=LABEL_FREQ,
# mode=m,
# batch_size=BATCH_SIZE,
# num_workers=NUM_WORKERS
# )
# -----------------------------------------
# 2) 再跑 _load_and_preprocess 的基准测试
# -----------------------------------------
for m in ['python', 'numpy', 'numba']:
benchmark_load_and_preprocess_single_clip(
jsonl_path=JSONL_PATH,
sample_rate=SAMPLE_RATE,
channels=CHANNELS,
clip_seconds=CLIP_SECONDS,
label_freq=LABEL_FREQ,
mode=m,
num_files=NUM_FILES
)
# for m in ['python', 'numpy', 'numba']:
# benchmark_load_and_preprocess_full_epoch(
# jsonl_path=JSONL_PATH,
# sample_rate=SAMPLE_RATE,
# channels=CHANNELS,
# clip_seconds=CLIP_SECONDS,
# label_freq=LABEL_FREQ,
# mode=m,
# batch_size=BATCH_SIZE,
# num_workers=NUM_WORKERS
# )
|