aws sagemakerにて、学習済みの音響モデルをデプロイする方法がわかりません・・・。

Question

実現したいこと

aws sagemakerにて、学習済みの音響モデルをデプロイしたいと考えています。
以下の、比較的近い事例を参考に構築をすすめています。
https://dev.classmethod.jp/articles/realtime-inference-on-sagemaker-using-by-pose-estimation-model-pytorch-own-algorhythm/

困っていること

pythonの知識が乏しく、以下のソースコードをどのようにsagemaker指定の関数（input_fn,model_fn,predict_fn,output_fn）に落とし込む方法がわかりません・・・。

python,awsに詳しい方がいましたらご教授いただけますと幸いです。

該当のソースコード

TTS_inference.sh

model_tag="kan-bayashi/jsut_tacotron2_accent_with_pause"
train_config=""
model_file=""

vocoder_tag="parallel_wavegan/jsut_hifigan.v1"
vocoder_config=""
vocoder_file=""

prosodic="false"
fs=""

. ../utils/parse_options.sh

COMMAND="python tts_inference_with_accent.py "


pwg=`pip list | grep parallel`
if [ "$pwg" == "" ];
then
	pip install -U parallel_wavegan
fi

ip=`pip list | grep ipython`
if [ "$pwg" == "" ];
then
	pip install -U IPython
fi


if [ "$train_config" == "" ] && [ "$model_file" == "" ]
then
	COMMAND="${COMMAND}--model_tag \"${model_tag}\" "
else
	COMMAND="${COMMAND}--train_config \"${train_config}\" "
	COMMAND="${COMMAND}--model_file \"${model_file}\" "
fi

if [ "$vocoder_config" == "" ] && [ "$vocoder_file" == "" ]
then
	COMMAND="${COMMAND}--vocoder_tag \"${vocoder_tag}\" "
else
	COMMAND="${COMMAND}--vocoder_config \"${vocoder_config}\" "
	COMMAND="${COMMAND}--vocoder_file \"${vocoder_file}\" "
fi

if [ ! "$fs" == "" ]; then COMMAND="${COMMAND}--fs ${fs} "; fi

if [ "$prosodic" == "true" ]; then COMMAND="${COMMAND}-p"; fi

echo "${COMMAND}"
echo ""
echo ""

eval $COMMAND

tts_inference_with_accent.py

import os
import time
import torch
import pyopenjtalk
from espnet2.bin.tts_inference import Text2Speech
import matplotlib.pyplot as plt
from espnet2.tasks.tts import TTSTask
from espnet2.text.token_id_converter import TokenIDConverter
import numpy as np

import argparse
import text_processing as texp



prosodic=False

parser = argparse.ArgumentParser()

parser.add_argument("--model_tag")
parser.add_argument("--train_config")
parser.add_argument("--model_file")
parser.add_argument("--vocoder_tag")
parser.add_argument("--vocoder_config")
parser.add_argument("--vocoder_file")
parser.add_argument("-p", "--prosodic",help="Prosodic text input mode", action="store_true")
parser.add_argument("--fs",type=int,default=24000)

args = parser.parse_args()
print(args)

os.chdir('../')

# Case 2: Load the local model and the pretrained vocoder
print("download model = ",args.model_tag,"\n")
print("download vocoder = ",args.vocoder_tag,"\n")
print("モデルを読み込んでいます...\n")
if args.model_tag is not None :
    text2speech = Text2Speech.from_pretrained(
        model_tag=args.model_tag,
        vocoder_tag=args.vocoder_tag,
        device="cuda",
        )
elif args.vocoder_tag is not None :
    text2speech = Text2Speech.from_pretrained(
        train_config=args.train_config,
        model_file=args.model_file,
        vocoder_tag=args.vocoder_tag,
        device="cuda",
    )
else :
    text2speech = Text2Speech.from_pretrained(
        train_config=args.train_config,
        model_file=args.model_file,
        vocoder_config=args.vocoder_config,
        vocoder_file=args.vocoder_file,
        device="cuda",
    )

guide="セリフを入力してください" 
if args.prosodic :
    guide="アクセント句がスペースで区切られた韻律記号(^)付きのセリフをすべてひらがなで入力してください。(スペースや記号もすべて全角で)\n"
x=""
while(1):
    # decide the input sentence by yourself
    print(guide)

    x = "デモ原稿"
    if x == "exit" :
        break

    #model, train_args = TTSTask.build_model_from_file(
    #        args.train_config, args.model_file, "cuda"
    #        )
        
    token_id_converter = TokenIDConverter(
            token_list=text2speech.train_args.token_list,
            unk_symbol="<unk>",
            )

    text = x
    if args.prosodic :
        tokens = texp.a2p(x)
        text_ints = token_id_converter.tokens2ids(tokens)
        text = np.array(text_ints)
    else :
        print("\npyopenjtalk_accent_with_pauseによる解析結果：")
        print(texp.text2yomi(x),"\n")


    # synthesis
    with torch.no_grad():
        start = time.time()
        data = text2speech(text)
        wav = data["wav"]
        #print(text2speech.preprocess_fn("<dummy>",dict(text=x))["text"])
    rtf = (time.time() - start) / (len(wav) / text2speech.fs)
    print(f"RTF = {rtf:5f}")

    if not os.path.isdir("generated_wav"):
    	os.makedirs("generated_wav")

    if  args.model_tag is not None  :
        if "tacotron" in args.model_tag :
            mel = data['feat_gen_denorm'].cpu()
            plt.imshow(torch.t(mel).numpy(), 
                       aspect='auto', 
                       origin='bottom', 
                       interpolation='none', 
                       cmap='viridis'
                       )
            plt.savefig('generated_wav/'+x+'.png')
    else :
        if "tacotron" in args.model_file :
            mel = data['feat_gen_denorm'].cpu()
            plt.imshow(torch.t(mel).numpy(), 
                       aspect='auto', 
                       origin='bottom', 
                       interpolation='none', 
                       cmap='viridis'
                       )
            plt.savefig('generated_wav/'+x+'.png')


    # let us listen to generated samples
    from IPython.display import display, Audio
    import numpy as np
    #display(Audio(wav.view(-1).cpu().numpy(), rate=text2speech.fs))
    #Audio(wav.view(-1).cpu().numpy(), rate=text2speech.fs)
    np_wav=wav.view(-1).cpu().numpy()


    print("サンプリングレート",args.fs,"で出力します。")
    from scipy.io.wavfile import write
    samplerate = args.fs
    t = np.linspace(0., 1., samplerate)
    amplitude = np.iinfo(np.int16).max
    data = amplitude * np_wav/np.max(np.abs(np_wav))
    write("generated_wav/"+x+".wav", samplerate, data.astype(np.int16))
    print("\n\n\n")