import whisper import re import sys import os, random, copy import numpy as np import torch import pandas as pd import torchaudio from tqdm.notebook import tqdm import collections, json import editdistance from whisper.normalizers import EnglishTextNormalizer from argparse import ArgumentParser from num2words import num2words sys.path.append('/home3/huyuchen/pytorch_workplace/jiwer') from jiwer import wer_embdiff import fasttext from huggingface_hub import hf_hub_download from pathlib import Path from typing import Optional from sentencepiece import SentencePieceProcessor, SentencePieceTrainer from sentence_transformers import SentenceTransformer from argparse import ArgumentParser from evaluate import load from lit_gpt.tokenizer import Tokenizer eval_wer = load("wer") normalizer = EnglishTextNormalizer() checkpoint_dir = Path('/home3/huyuchen/pytorch_workplace/wgpt/checkpoints/Llama-2-7b-hf') tokenizer = Tokenizer(checkpoint_dir) sbert_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2') def calculate_wer(all_hypo, all_refer): return eval_wer.compute(predictions=all_hypo, references=all_refer) def word_emb_diff(reference, hypothesis): output, edit_ops = wer_embdiff(reference, hypothesis) ref_words, hypo_words = output.references[0], output.hypotheses[0] emb_diffs = [] for op in edit_ops: if op.tag == 'replace': ref_word, hypo_word = ref_words[op.src_pos], hypo_words[op.dest_pos] elif op.tag == 'delete': ref_word, hypo_word = ref_words[op.src_pos], None elif op.tag == 'insert': ref_word, hypo_word = None, hypo_words[op.dest_pos] else: continue ref_emb = torch.from_numpy(sbert_model.encode([ref_word])[0]) if ref_word else torch.zeros([384]) hypo_emb = torch.from_numpy(sbert_model.encode([hypo_word])[0]) if hypo_word else torch.zeros([384]) emb_diff = ref_emb - hypo_emb emb_diffs.append(emb_diff) # print('word', hypo_emb.mean(), ref_emb.mean(), emb_diff.mean()) if len(emb_diffs) == 0: return torch.zeros([384]) else: return torch.stack(emb_diffs, dim=0).mean(dim=0) def sent_emb_diff(reference, hypothesis): embeddings = sbert_model.encode([reference, hypothesis]) ref_emb, hypo_emb = torch.from_numpy(embeddings[0]), torch.from_numpy(embeddings[1]) emb_diff = ref_emb - hypo_emb # print('sentence', hypo_emb.mean(), ref_emb.mean(), emb_diff.mean()) return emb_diff def generate_prompt(input1, input2): return ( f"Below is the best-hypotheses transcribed from speech recognition system. Please try to revise it using the words which are only included into other-hypothesis, and write the response for the true transcription.\n\n### Best-hypothesis:\n{input1}\n\n### Other-hypothesis:\n{input2}\n\n### Response:\n" ) DEVICE = "cuda" if torch.cuda.is_available() else "cpu" model = whisper.load_model('large-v2') f_noisy_wav = open(f'noisy_wav.scp', 'r') f_clean_wav = open(f'clean_wav.scp', 'r') f_text = open(f'text', 'r') id = 0 pt_file = [] all_hypo, all_refer = [], [] for line in f_noisy_wav.readlines(): utt_id, audio_path = line.strip().split()[0], line.strip().split()[1] clean_line = f_clean_wav.readline() clean_utt_id, clean_audio_path = clean_line.strip().split()[0], clean_line.strip().split()[1] assert clean_utt_id == utt_id, (line, clean_line) gt = ' '.join(f_text.readline().strip().split()[1:]) audio = whisper.load_audio(audio_path) audio = whisper.pad_or_trim(audio) mel = whisper.log_mel_spectrogram(audio).to(model.device) options = whisper.DecodingOptions(language='en', beam_size=50) texts, confidences = whisper.decode_score(model, mel, options) ## noisy audio feats audio_features = model.encoder(mel.unsqueeze(0))[0] ## clean audio feats clean_audio = whisper.load_audio(clean_audio_path) clean_audio = whisper.pad_or_trim(clean_audio) clean_mel = whisper.log_mel_spectrogram(clean_audio).to(model.device) clean_audio_features = model.encoder(clean_mel.unsqueeze(0))[0] input, score = [], [] for text, confidence in zip(texts, confidences): if len(input) < 5 and len(text) > 0 and text not in input: input.append(text) score.append(confidence) # print('before', input, score, len(input)) if len(input) < 5: options = whisper.DecodingOptions(language='en', temperature=1.2) for _ in range(5 - len(input)): result = whisper.decode(model, mel, options) text, condidence = result.text, result.avg_logprob if text in input: continue inserted = False for i in range(len(input)): if condidence > score[i]: input.insert(i, text) score.insert(i, condidence) inserted = True break if not inserted: input.append(text) score.append(condidence) # print('after ', input, score, len(input)) if len(input) < 5: num_to_add = 5 - len(input) for _ in range(num_to_add): rand_id = random.randint(0, len(input) - 1) rep_input, rep_score = copy.deepcopy(input[rand_id]), copy.deepcopy(score[rand_id]) input.insert(rand_id + 1, rep_input) score.insert(rand_id + 1, rep_score) for i in range(len(input)): try: text = normalizer(input[i]) text = re.sub(r"[-+]?\d*\.?\d+|\d+%?", lambda m: num2words(m.group()), text).replace('%', ' percent') except Exception: text = normalizer(input[i]) print(f'input exception: {text}') input[i] = text if len(text) > 0 else '' try: output = normalizer(gt) output = re.sub(r"[-+]?\d*\.?\d+|\d+%?", lambda m: num2words(m.group()), output).replace('%', ' percent') except Exception: output = normalizer(gt) print(f'output exception: {output}') output = output if len(output) > 0 else '' cur_wer = calculate_wer([input[0]], [output]) # calculate emb diff we_diffs, se_diffs = [], [] for i in range(5): for j in range(i + 1, 5): we_diffs.append(word_emb_diff(input[i], input[j])) se_diffs.append(sent_emb_diff(input[i], input[j])) we_diff = torch.stack(we_diffs, dim=0) # [10, 384] se_diff = torch.stack(se_diffs, dim=0) # [10, 384] emb_diff = torch.cat([we_diff, se_diff], dim=0) # [20, 384] # generate ids input1 = input[0] + '.' input2 = '. '.join(input[1:]) + '.' full_prompt = generate_prompt(input1, input2) full_prompt_and_response = full_prompt + output encoded_full_prompt = tokenizer.encode(full_prompt, max_length=1024) encoded_full_prompt_and_response = tokenizer.encode(full_prompt_and_response, eos=True, max_length=1024) labels = encoded_full_prompt_and_response.clone() labels[: len(encoded_full_prompt)] = -1 data = {"id": utt_id, "input_ids": encoded_full_prompt_and_response, "input_ids_no_response": encoded_full_prompt, "labels": labels, "input": input, 'ground_truth': output, "am_score": score, 'emb_diff': emb_diff, 'audio_features': noisy_audio_features, 'clean_audio_features': clean_audio_features} pt_file.append(data) # calculate wer id += 1 print(f'utterance {id}: wer = {cur_wer}, confidence = {score[0]}') all_hypo.append(input[0]) all_refer.append(output) torch.save(pt_file, f'/home3/huyuchen/pytorch_workplace/wllama/hypo_paradise_v2/train_rats.pt') f_noisy_wav.close() f_clean_wav.close() f_text.close() all_wer = calculate_wer(all_hypo, all_refer) print(f'all wer = {all_wer}')