main.py

from fastapi import FastAPI
import datetime
import torch
import os
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC, Wav2Vec2ProcessorWithLM, AutoConfig
from huggingface_hub import hf_hub_download
from fuzzywuzzy import fuzz
from utils import ffmpeg_read, query_dummy, query_raw, find_different

## config
API_TOKEN = os.environ["API_TOKEN"]
MODEL_PATH = os.environ["MODEL_PATH"]
PITCH_PATH = os.environ["PITCH_PATH"]

QUANTIZED_MODEL_PATH = hf_hub_download(repo_id=MODEL_PATH, filename='quantized_model.pt', token=API_TOKEN)
QUANTIZED_PITCH_MODEL_PATH = hf_hub_download(repo_id=PITCH_PATH, filename='quantized_model.pt', token=API_TOKEN)


## word preprocessor
processor_with_lm = Wav2Vec2ProcessorWithLM.from_pretrained(MODEL_PATH, use_auth_token=API_TOKEN)
processor = Wav2Vec2Processor.from_pretrained(MODEL_PATH, use_auth_token=API_TOKEN)

### quantized model
config = AutoConfig.from_pretrained(MODEL_PATH, use_auth_token=API_TOKEN)
dummy_model = Wav2Vec2ForCTC(config)
quantized_model = torch.quantization.quantize_dynamic(dummy_model, {torch.nn.Linear}, dtype=torch.qint8, inplace=True)
quantized_model.load_state_dict(torch.load(QUANTIZED_MODEL_PATH))

## pitch preprocessor 
processor_pitch = Wav2Vec2Processor.from_pretrained(PITCH_PATH, use_auth_token=API_TOKEN)

### quantized pitch mode
config = AutoConfig.from_pretrained(PITCH_PATH, use_auth_token=API_TOKEN)
dummy_pitch_model = Wav2Vec2ForCTC(config)
quantized_pitch_model = torch.quantization.quantize_dynamic(dummy_pitch_model, {torch.nn.Linear}, dtype=torch.qint8, inplace=True)
quantized_pitch_model.load_state_dict(torch.load(QUANTIZED_PITCH_MODEL_PATH))

app = FastAPI()

@app.get("/")
def read_root():
    return {"Message": "Application startup complete"}