note

• torch-model-archiver打包模型；利用torchserve加载前面打包的模型，并以grpc和http等接口往外提供推理服务
• 启动模型的api服务、curl命令发送http post请求，请求模型服务API；流程和TensorFlow serving流程大同小异

一、torchserve和archiver模块

• 模型部署需要用到两个模块
• torchserve用来模型部署
• torch-model-archiver打包模型

pip:
    - torch-workflow-archiver
    - torch-model-archiver 
    - torchserve 

二、Speech2Text Wav2Vec2模型部署

2.1 准备模型和自定义handler

• Wav2Vec2语音转文本的模型。这里我们为了简化流程从huggingface下载对应的模型，进行本地化利用torchserve部署
• hander将原始data进行转为模型输入所需的格式；nlp中很多任务可以直接用torchtext的text_classifier。

# 1. 导入huggingface模型
from transformers import AutoModelForCTC, AutoProcessor
import os

modelname = "facebook/wav2vec2-base-960h"
model = AutoModelForCTC.from_pretrained(modelname)
processor = AutoProcessor.from_pretrained(modelname)

modelpath = "model"
os.makedirs(modelpath, exist_ok=True)
model.save_pretrained(modelpath)
processor.save_pretrained(modelpath)

# 2. 自定义handler
import torch
import torchaudio
from transformers import AutoProcessor, AutoModelForCTC
import io


class Wav2VecHandler(object):
    def __init__(self):
        self._context = None
        self.initialized = False
        self.model = None
        self.processor = None
        self.device = None
        # Sampling rate for Wav2Vec model must be 16k
        self.expected_sampling_rate = 16_000

    def initialize(self, context):
        """Initialize properties and load model"""
        self._context = context
        self.initialized = True
        properties = context.system_properties

        # See https://pytorch.org/serve/custom_service.html#handling-model-execution-on-multiple-gpus
        self.device = torch.device("cuda:" + str(properties.get("gpu_id")) if torch.cuda.is_available() else "cpu")

        model_dir = properties.get("model_dir")
        self.processor = AutoProcessor.from_pretrained(model_dir)
        self.model = AutoModelForCTC.from_pretrained(model_dir)

    def handle(self, data, context):
        """Transform input to tensor, resample, run model and return transcribed text."""
        input = data[0].get("data")
        if input is None:
            input = data[0].get("body")
        
        # torchaudio.load accepts file like object, here `input` is bytes
        model_input, sample_rate = torchaudio.load(io.BytesIO(input), format="WAV")
        
        # Ensure sampling rate is the same as the trained model
        if sample_rate != self.expected_sampling_rate:
            model_input = torchaudio.functional.resample(model_input, sample_rate, self.expected_sampling_rate)
        
        model_input = self.processor(model_input, sampling_rate = self.expected_sampling_rate, return_tensors="pt").input_values[0]
        logits = self.model(model_input)[0]
        pred_ids = torch.argmax(logits, axis=-1)[0]
        output = self.processor.decode(pred_ids)

        return [output]

2.2 打包模型和启动模型api服务

• 可以直接在linux环境的terminal进行如下相关操作（打包模型、启动模型的api服务、curl命令发送http post请求，请求模型服务API）
• curl命令发送http post请求，请求模型服务API，如果遇到报错java.lang.NoSuchMethodError: java.nio.file.Files.readString(Ljava/nio/file/Path;)Ljava/lang/String;则应该是JRE没有安装或者需要升级：sudo apt install default-jre即可。
• curl那坨后正常会返回I HAD THAT CURIOSITY BESIDE ME AT THIS MOMENT%，测试数据是一段简单的sample.wav语音文件
• Waveform Audio File Format（WAVE，又或者是因为WAV后缀而被大众所知的），它采用RIFF（Resource Interchange File Format）文件格式结构。通常用来保存PCM格式的原始音频数据，所以通常被称为无损音频

# 打包部署模型文件, 把model部署到torchserve 
torch-model-archiver --model-name Wav2Vec2 --version 1.0 --serialized-file model/pytorch_model.bin --handler ./handler.py --extra-files "model/config.json,model/special_tokens_map.json,model/tokenizer_config.json,model/vocab.json,model/preprocessor_config.json" -f
mv Wav2Vec2.mar model_store

# 启动model服务, 加载前面打包的model, 并以grpc和http接口向外提供推理服务
torchserve --start --model-store model_store --models Wav2Vec2=Wav2Vec2.mar --ncs

# Once the server is running, let's try it with:
curl -X POST http://127.0.0.1:8080/predictions/Wav2Vec2 --data-binary '@./sample.wav' -H "Content-Type: audio/basic"

# 暂停torchserve serving
torchserve --stop

2.3 相关参数记录

torch-model-archiver：用来打包模型

• model-name: 设定部署的模型名称
• version: 设定部署的模型版本
• model-file: 定义模型结构的python文件
• serialized-file: 设定训练模型保存的pth文件
• export-path: 设定打包好的模型保存路径
• extra-files: 设定额外的文件，如label跟id映射的定义文件等，用于一并打包到模型压缩包中
• handler: 为一个处理器，用来指定模型推理预测前后的数据的处理问题；如 nlp模型中的文本分词和转换为id的步骤；以及分类问题中，模型预测结果映射为具体的label等数据处理功能

torch-model-archiver：用来打包模型
usage: torch-model-archiver [-h] --model-name MODEL_NAME
                            [--serialized-file SERIALIZED_FILE]
                            [--model-file MODEL_FILE] --handler HANDLER
                            [--extra-files EXTRA_FILES]
                            [--runtime {python,python2,python3}]
                            [--export-path EXPORT_PATH]
                            [--archive-format {tgz,no-archive,default}] [-f]
                            -v VERSION [-r REQUIREMENTS_FILE]

torchserve：该组件用来加载前面打包的模型，并以grpc和http等接口往外提供推理服务

• start 和 stop: 即推理服务的启动和停止；
• model-store: 打包模型存储的路径；
• models: 设定模型名称和模型文件名，如：MODEL_NAME=MODEL_PATH2 格式；
• no-config-snapshots: 即 --ncs，用来设置防止服务器存储配置快照文件；

Reference

[1] https://github.com/pytorch/serve
[2] Torch Model archiver for TorchServe
[3] https://github.com/pytorch/serve/tree/master/examples/speech2text_wav2vec2
[4] https://huggingface.co/docs/transformers/model_doc/wav2vec2
[5] https://github.com/pytorch/serve/tree/master/model-archiver
[6] pytorch 模型部署.nlper
[7] cURL - 学习/实践