TTS中英文内存使用过高，应该怎么解决

[mrg79433283]

`# -- coding:utf-8 -- import paddle import yaml import soundfile as sf from yacs.config import CfgNode from paddlespeech.t2s.frontend.mix_frontend import MixFrontend from paddlespeech.t2s.exps.syn_utils import get_am_inference from paddlespeech.t2s.exps.syn_utils import get_voc_inference from paddlespeech.t2s.exps.syn_utils import run_frontend

from memory_profiler import profile

sentence = """ 800多年前的今天，北京市丰台区的卢沟桥遭遇了一场猛烈的暴雨，但令人惊讶的是，这座古老的桥梁却安然无恙地屹立在这里，见证了时光的流转和历史的变迁。 卢沟桥是中国历史上著名的古桥之一，建于元代至正年间，距今已有800多年的历史。这座桥梁位于北京市丰台区卢沟镇境内，横跨于卢沟河之上，全长266.5米，宽9.3米，是一座典型的石拱桥。。 在800多年的历史中，卢沟桥经历了无数次风雨洗礼和战火炮烙，但每一次都能够奇迹般地幸免于难。其中最为著名的一次就是1937年的七七事变，日军侵华之际，卢沟桥成为了中日两军之间的战场。 在激战中，卢沟桥被毁坏殆尽，但在抗战胜利后经过修复后又恢复了原貌。而这次暴雨，则是近年来卢沟桥所遭遇的一次自然灾害。据当地居民介绍，这场暴雨是他们近几十年来见过的最大暴雨之一， 整个卢沟镇被淹没在水中，许多房屋和道路都被冲毁。但令人欣慰的是，卢沟桥并没有受到任何损害，仍然安然无恙地屹立在卢沟河之上。专家介绍说，卢沟桥之所以能够幸免于难， 主要是因为它在建造时就考虑到了防洪防涝的问题。在桥梁两侧设置了多个洪水泄洪口和堤坝，能够有效地控制卢沟河的水位和流量。此外，在修缮时也一直坚持使用传统工艺和材料， 保证了卢沟桥的稳固和耐久性。卢沟桥的安然无恙，不仅是一种幸运和奇迹，更是对我们传统文化和历史遗产的珍视和保护。我们应该更加重视和保护这些历史文化遗产，让它们能够继续传承下去， 成为我们民族文化的瑰宝和精神财富。 """

phones_dict = "./fastspeech2_mix_ckpt_1.2.0/phone_id_map.txt"

am_config_file = "./fastspeech2_mix_ckpt_1.2.0/default.yaml" am_ckpt = "./fastspeech2_mix_ckpt_1.2.0/snapshot_iter_99200.pdz" am_stat = "./fastspeech2_mix_ckpt_1.2.0/speech_stats.npy" speaker_dict = "./fastspeech2_mix_ckpt_1.2.0/speaker_id_map.txt"

voc_config_file = "./hifigan_aishell3_ckpt_0.2.0/default.yaml" voc_ckpt = "./hifigan_aishell3_ckpt_0.2.0/snapshot_iter_2500000.pdz" voc_stat = "./hifigan_aishell3_ckpt_0.2.0/feats_stats.npy"

voc_config_file = "./pwg_aishell3_ckpt_0.5/default.yaml"

voc_ckpt = "./pwg_aishell3_ckpt_0.5/snapshot_iter_1000000.pdz"

voc_stat = "./pwg_aishell3_ckpt_0.5/feats_stats.npy"

text frontend

@profile(precision=5)

def text_frontend(): frontend = MixFrontend(phone_vocab_path=phones_dict) print("frontend done!") return frontend

with open(am_config_file) as f: am_config = CfgNode(yaml.safe_load(f))

load AM

@profile(precision=5)

def load_am(): am_inference = get_am_inference( am="fastspeech2_mix", am_config=am_config, am_ckpt=am_ckpt, am_stat=am_stat, phones_dict=phones_dict, tones_dict=None, speaker_dict=speaker_dict) print("acoustic model done!") return am_inference

with open(voc_config_file) as f: voc_config = CfgNode(yaml.safe_load(f))

load Voc

@profile(precision=5)

def load_voc(): voc_inference = get_voc_inference( voc="hifigan_aishell3", voc_config=voc_config, voc_ckpt=voc_ckpt, voc_stat=voc_stat)

voc_inference = get_voc_inference(

#     voc="pwgan_aishell3",
#     voc_config=voc_config,
#     voc_ckpt=voc_ckpt,
#     voc_stat=voc_stat)
print("voc done!")
return voc_inference

get phone id

@profile(precision=5)

def get_phone_id(): frontend = text_frontend()

frontend_dict = run_frontend(
    frontend=frontend,
    text=sentence,
    merge_sentences=False,
    get_tone_ids=False,
    lang="mix")
phone_ids = frontend_dict['phone_ids']
return phone_ids

phone_ids = get_phone_id() am_inference = load_am() voc_inference = load_voc()

inference

flags = 0 print(f"len phone ids : {len(phone_ids)}") print(phone_ids) for i in range(len(phone_ids)): part_phone_ids = phone_ids[i] spk_id = 174 # baker:174, ljspeech:175, aishell3:0~173, vctk:176~282 spk_id = paddle.to_tensor(spk_id) tracemalloc.start() with paddle.no_grad(): mel = am_inference(part_phone_ids, spk_id) wav = voc_inference(mel)

if flags == 0:
    wav_all = wav
    flags = 1
else:
    wav_all = paddle.concat([wav_all, wav])
tracemalloc.stop()

print("infer successfully.")

save audio

wav = wav_all.numpy() sf.write("./out.wav", wav, am_config.fs) print("write successfully.") `

执行转译任务时，文字越多占用内存越高，1200多字算上加载模型大概能占用快到16G，且随着字数增多，内存仍在上涨。 想请问下，是我使用的问题吗，还是它本来就是这种表现

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.