rendering/universal_tts/examples/esd_inneremo/eval/eval_emo.py

import torch
import argparse
from tqdm import tqdm
from pathlib import Path
import numpy as np
import os
import torch.nn.functional as F
import soundfile as sf
from torchaudio.transforms import Resample
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
from funasr import AutoModel

def load_audio(manifest_path):
    import numpy as np
    generated_pathes, tgt_pathes, gt_texts= [], [], []
    with open(manifest_path, "r") as f:
        for ind, line in enumerate(f):
            if len(line.strip()) < 2:
                continue
            generated_path, tgt_path, gt_text = line.strip().split("\t")[:3]
            generated_pathes.append(generated_path)
            tgt_pathes.append(tgt_path)
            gt_texts.append(gt_text)
    return generated_pathes, tgt_pathes, gt_texts

def load_tsv(path):
    with open(path, "r") as rf:
        lines = rf.readlines()
    return lines
        
if __name__ == "__main__":
    parser = argparse.ArgumentParser(description='Inference')
    parser.add_argument('-t', '--tsv', type=str)
    parser.add_argument('-o', '--out_home', type=str)
    args = parser.parse_args()

    if os.path.exists(os.path.join(args.out_home, "emo.log")):
        with open(os.path.join(args.out_home, "emo.log"), "r") as rf:
            if len(rf.readlines()) > 100:
                exit()
    
    model = AutoModel(model="/workspace/echoloc/modelscope/iic/emotion2vec_plus_large/")

    generated_pathes, tgt_pathes, gt_texts = load_audio(args.tsv)
    simis = []
    with torch.no_grad():
        with open(os.path.join(args.out_home, "emo.log"), "w") as f:
            for index, (est_path, tgt_path, gt_text) in enumerate(tqdm(zip(generated_pathes, tgt_pathes, gt_texts))):
                try:
                    generated_emb = model.generate(est_path, granularity="utterance", extract_embedding=True, disable_pbar=True)[0]["feats"] # 1024
                    tgt_emb = model.generate(tgt_path, granularity="utterance", extract_embedding=True, disable_pbar=True)[0]["feats"] # 1024
                    simi = float(F.cosine_similarity(torch.FloatTensor([generated_emb]), torch.FloatTensor([tgt_emb])).item())
                except Exception as e:
                    simi = -1.0
                    print(e)
                simis.append(simi)
                print("%s %s %f"%(est_path, tgt_path, simi), file=f)
            print("------------------------------------------", file=f)
            simis = np.array(simis)
            print("with -1: emo2vec large:", np.mean(simis), file=f)
            print("without -1: emo2vec large:", np.mean(simis[simis != -1]), " -1 num:", len(simis[simis==-1]), file=f)