Upload finetune_speech.py
Browse files- examples/finetune_speech.py +929 -0
examples/finetune_speech.py
ADDED
|
@@ -0,0 +1,929 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import argparse
|
| 2 |
+
import json
|
| 3 |
+
import os
|
| 4 |
+
from pathlib import Path
|
| 5 |
+
|
| 6 |
+
import numpy as np
|
| 7 |
+
import torch
|
| 8 |
+
import sacrebleu
|
| 9 |
+
|
| 10 |
+
from datasets import load_dataset
|
| 11 |
+
from torch.utils.data import Dataset, ConcatDataset
|
| 12 |
+
from tqdm import tqdm
|
| 13 |
+
from transformers import (
|
| 14 |
+
AutoProcessor,
|
| 15 |
+
AutoModel,
|
| 16 |
+
BatchFeature,
|
| 17 |
+
Trainer,
|
| 18 |
+
TrainingArguments,
|
| 19 |
+
StoppingCriteria,
|
| 20 |
+
StoppingCriteriaList,
|
| 21 |
+
)
|
| 22 |
+
from collections import defaultdict
|
| 23 |
+
|
| 24 |
+
import soundfile as sf
|
| 25 |
+
from datasets import Audio
|
| 26 |
+
import random
|
| 27 |
+
|
| 28 |
+
class MultipleTokenBatchStoppingCriteria(StoppingCriteria):
|
| 29 |
+
"""Stopping criteria capable of receiving multiple stop-tokens and handling batched inputs."""
|
| 30 |
+
|
| 31 |
+
def __init__(self, stop_tokens: torch.LongTensor, batch_size: int = 1) -> None:
|
| 32 |
+
"""Initialize the multiple token batch stopping criteria.
|
| 33 |
+
|
| 34 |
+
Args:
|
| 35 |
+
stop_tokens: Stop-tokens.
|
| 36 |
+
batch_size: Batch size.
|
| 37 |
+
|
| 38 |
+
"""
|
| 39 |
+
|
| 40 |
+
self.stop_tokens = stop_tokens
|
| 41 |
+
self.max_stop_tokens = stop_tokens.shape[-1]
|
| 42 |
+
self.stop_tokens_idx = torch.zeros(batch_size, dtype=torch.long, device=stop_tokens.device)
|
| 43 |
+
|
| 44 |
+
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
|
| 45 |
+
# Only gather the maximum number of inputs compatible with stop tokens
|
| 46 |
+
# and checks whether generated inputs are equal to `stop_tokens`
|
| 47 |
+
generated_inputs = torch.eq(input_ids[:, -self.max_stop_tokens :].unsqueeze(1), self.stop_tokens)
|
| 48 |
+
equal_generated_inputs = torch.all(generated_inputs, dim=2)
|
| 49 |
+
|
| 50 |
+
# Mark the position where a stop token has been produced for each input in the batch,
|
| 51 |
+
# but only if the corresponding entry is not already set
|
| 52 |
+
sequence_idx = torch.any(equal_generated_inputs, dim=1)
|
| 53 |
+
sequence_set_mask = self.stop_tokens_idx == 0
|
| 54 |
+
self.stop_tokens_idx[sequence_idx & sequence_set_mask] = input_ids.shape[-1]
|
| 55 |
+
|
| 56 |
+
return torch.all(self.stop_tokens_idx)
|
| 57 |
+
|
| 58 |
+
class BaseAudioDataset(Dataset):
|
| 59 |
+
def __init__(self, processor, split, sampling_rate=16000, debug=False):
|
| 60 |
+
self.processor = processor
|
| 61 |
+
self.training = "train" in split
|
| 62 |
+
self.debug = debug
|
| 63 |
+
self.sampling_rate = sampling_rate
|
| 64 |
+
self.name = ""
|
| 65 |
+
|
| 66 |
+
def set_dataset_name(self, name):
|
| 67 |
+
self.name = name
|
| 68 |
+
|
| 69 |
+
@staticmethod
|
| 70 |
+
def filter_corrupted_files(data, audio_field, text_fields, dataset_name, sampling_rate=16000, debug=True):
|
| 71 |
+
original_size = len(data)
|
| 72 |
+
|
| 73 |
+
data = data.cast_column(audio_field, Audio(decode=False))
|
| 74 |
+
|
| 75 |
+
def identify_corrupted_files(example):
|
| 76 |
+
try:
|
| 77 |
+
sf.read(example[audio_field]["path"])
|
| 78 |
+
|
| 79 |
+
for field in text_fields:
|
| 80 |
+
if field in example and example[field].replace('"', '') == "":
|
| 81 |
+
return False
|
| 82 |
+
return True
|
| 83 |
+
except Exception:
|
| 84 |
+
return False
|
| 85 |
+
|
| 86 |
+
data = data.filter(identify_corrupted_files, num_proc=16)
|
| 87 |
+
validated_size = len(data)
|
| 88 |
+
|
| 89 |
+
# Audio Decoding
|
| 90 |
+
data = data.cast_column(audio_field, Audio(sampling_rate=sampling_rate, decode=True))
|
| 91 |
+
|
| 92 |
+
if debug:
|
| 93 |
+
print(f"Dataset: {dataset_name}")
|
| 94 |
+
print(f"Original data nums: {original_size}")
|
| 95 |
+
print(f"After filtering data nums: {validated_size}")
|
| 96 |
+
print(f"Filtering ratio: {validated_size/original_size:.2%}")
|
| 97 |
+
|
| 98 |
+
return data
|
| 99 |
+
|
| 100 |
+
@staticmethod
|
| 101 |
+
def filter_by_audio_length(data, audio_field, min_sec=2, max_sec=20, debug=True):
|
| 102 |
+
original_size = len(data)
|
| 103 |
+
|
| 104 |
+
def filter_audio_by_length(example):
|
| 105 |
+
try:
|
| 106 |
+
audio = example[audio_field]['array']
|
| 107 |
+
channel = 1
|
| 108 |
+
if hasattr(audio, 'ndim') and audio.ndim > 1:
|
| 109 |
+
channel = audio.ndim
|
| 110 |
+
audio = audio.squeeze()
|
| 111 |
+
audio_length = len(audio) / example[audio_field]['sampling_rate'] / channel
|
| 112 |
+
return min_sec <= audio_length <= max_sec
|
| 113 |
+
except Exception as e:
|
| 114 |
+
if debug:
|
| 115 |
+
print(f"Error : {str(e)[:100]}... - sample excluded")
|
| 116 |
+
return False
|
| 117 |
+
|
| 118 |
+
data = data.filter(filter_audio_by_length, num_proc=16)
|
| 119 |
+
filtered_size = len(data)
|
| 120 |
+
|
| 121 |
+
if debug:
|
| 122 |
+
print(f"Before Length Filtering data nums: {original_size}")
|
| 123 |
+
print(f"After Length Filtering data nums: {filtered_size}")
|
| 124 |
+
print(f"Filtering ratio: {filtered_size/original_size:.2%}")
|
| 125 |
+
|
| 126 |
+
return data
|
| 127 |
+
|
| 128 |
+
def prepare_model_inputs(self, audio_array, instruction, answer_text):
|
| 129 |
+
user_message = {
|
| 130 |
+
'role': 'user',
|
| 131 |
+
'content': '<start_of_audio>' + instruction,
|
| 132 |
+
}
|
| 133 |
+
prompt = self.processor.tokenizer.apply_chat_template(
|
| 134 |
+
[user_message], tokenize=False, add_generation_prompt=True, add_bos=True
|
| 135 |
+
)
|
| 136 |
+
|
| 137 |
+
inputs = self.processor(
|
| 138 |
+
text=prompt,
|
| 139 |
+
audio=[audio_array],
|
| 140 |
+
add_special_tokens=False,
|
| 141 |
+
return_tensors='pt'
|
| 142 |
+
)
|
| 143 |
+
|
| 144 |
+
answer = f"{answer_text}{ANSWER_SUFFIX}"
|
| 145 |
+
answer_ids = self.processor.tokenizer(answer, add_special_tokens=False, return_tensors='pt').input_ids
|
| 146 |
+
|
| 147 |
+
if self.debug:
|
| 148 |
+
self.debug = False
|
| 149 |
+
task_type = 'AST' if hasattr(self, 'ast') and self.ast else 'ASR'
|
| 150 |
+
lang_info = f" - {self.lang}" if hasattr(self, 'lang') else ""
|
| 151 |
+
print(f"{task_type}{lang_info}\nPROMPT: {prompt}\nINPUT: {self.processor.decode(inputs.input_ids[0], skip_special_tokens=False)}\nANSWER: {self.processor.decode(answer_ids[0], skip_special_tokens=False)}\n")
|
| 152 |
+
print(f"INPUT_MODE: {inputs.input_modes[0].item()}")
|
| 153 |
+
|
| 154 |
+
if self.training:
|
| 155 |
+
input_ids = torch.cat([inputs.input_ids, answer_ids], dim=1)
|
| 156 |
+
labels = torch.full_like(input_ids, _IGNORE_INDEX)
|
| 157 |
+
labels[:, -answer_ids.shape[1]:] = answer_ids
|
| 158 |
+
padding = torch.zeros((inputs.token_type_ids.shape[0], answer_ids.shape[1]))
|
| 159 |
+
token_type_ids = torch.cat([inputs.token_type_ids, padding], dim=1)
|
| 160 |
+
else:
|
| 161 |
+
input_ids = inputs.input_ids
|
| 162 |
+
labels = answer_ids
|
| 163 |
+
token_type_ids = inputs.token_type_ids
|
| 164 |
+
|
| 165 |
+
return {
|
| 166 |
+
'input_ids': input_ids,
|
| 167 |
+
'labels': labels,
|
| 168 |
+
'token_type_ids': token_type_ids,
|
| 169 |
+
'input_audio_embeds': inputs.input_audio_embeds,
|
| 170 |
+
'audio_embed_sizes': inputs.audio_embed_sizes,
|
| 171 |
+
'input_modes': inputs.input_modes,
|
| 172 |
+
}
|
| 173 |
+
|
| 174 |
+
# CoVoST2 Dataset Class
|
| 175 |
+
class CoVoSTDataset(BaseAudioDataset):
|
| 176 |
+
def __init__(self, processor, data_dir, split, ast=False,
|
| 177 |
+
lang=("en_ko", "Korean"), sampling_rate=16000, debug=False):
|
| 178 |
+
super().__init__(processor, split, sampling_rate, debug)
|
| 179 |
+
|
| 180 |
+
self.set_dataset_name("CoVoST")
|
| 181 |
+
self.ast = ast
|
| 182 |
+
self.lang = lang[0]
|
| 183 |
+
|
| 184 |
+
self.data = load_dataset("junnei/covost2",
|
| 185 |
+
lang[0],
|
| 186 |
+
data_dir=data_dir,
|
| 187 |
+
split=split,
|
| 188 |
+
trust_remote_code=True
|
| 189 |
+
)
|
| 190 |
+
|
| 191 |
+
text_fields = ["sentence", "translation"] if ast else ["sentence"]
|
| 192 |
+
self.data = self.filter_corrupted_files(self.data, "audio", text_fields, "CoVoST")
|
| 193 |
+
|
| 194 |
+
# (Optional) Audio length Filtering
|
| 195 |
+
self.data = self.filter_by_audio_length(self.data, "audio")
|
| 196 |
+
|
| 197 |
+
# Instruction Setting
|
| 198 |
+
self.instruction = random.choice(INSTRUCTION["ast"]).format(lang[1]) if ast else random.choice(INSTRUCTION["asr"])
|
| 199 |
+
|
| 200 |
+
def __len__(self):
|
| 201 |
+
return len(self.data)
|
| 202 |
+
|
| 203 |
+
def __getitem__(self, idx):
|
| 204 |
+
data = self.data[idx]
|
| 205 |
+
|
| 206 |
+
if self.ast:
|
| 207 |
+
answer_text = data["translation"]
|
| 208 |
+
else:
|
| 209 |
+
answer_text = data["sentence"].replace('"', '')
|
| 210 |
+
|
| 211 |
+
return self.prepare_model_inputs(
|
| 212 |
+
data["audio"]["array"],
|
| 213 |
+
self.instruction,
|
| 214 |
+
answer_text
|
| 215 |
+
)
|
| 216 |
+
|
| 217 |
+
# Zeroth Korean Dataset Class
|
| 218 |
+
class ZerothKoreanDataset(BaseAudioDataset):
|
| 219 |
+
def __init__(self, processor, split, sampling_rate=16000, debug=False):
|
| 220 |
+
super().__init__(processor, split, sampling_rate, debug)
|
| 221 |
+
|
| 222 |
+
self.set_dataset_name("Zeroth")
|
| 223 |
+
# only ASR
|
| 224 |
+
self.ast = False
|
| 225 |
+
self.lang = "ko"
|
| 226 |
+
|
| 227 |
+
# load dataset
|
| 228 |
+
self.data = load_dataset("Bingsu/zeroth-korean",
|
| 229 |
+
split=split,
|
| 230 |
+
trust_remote_code=True
|
| 231 |
+
)
|
| 232 |
+
|
| 233 |
+
# (Optional) Audio length Filtering
|
| 234 |
+
self.data = self.filter_by_audio_length(self.data, "audio")
|
| 235 |
+
|
| 236 |
+
# Instruction Setting
|
| 237 |
+
self.instruction = random.choice(INSTRUCTION["asr"])
|
| 238 |
+
|
| 239 |
+
def __len__(self):
|
| 240 |
+
return len(self.data)
|
| 241 |
+
|
| 242 |
+
def __getitem__(self, idx):
|
| 243 |
+
data = self.data[idx]
|
| 244 |
+
|
| 245 |
+
# Zeroth Korean is only for ASR
|
| 246 |
+
answer_text = data["text"].replace('"', '')
|
| 247 |
+
|
| 248 |
+
return self.prepare_model_inputs(
|
| 249 |
+
data["audio"]["array"],
|
| 250 |
+
self.instruction,
|
| 251 |
+
answer_text
|
| 252 |
+
)
|
| 253 |
+
|
| 254 |
+
# Libri Speech Dataset Class
|
| 255 |
+
class LibriSpeechDataset(BaseAudioDataset):
|
| 256 |
+
def __init__(self, processor, subset, split, sampling_rate=16000, debug=False):
|
| 257 |
+
super().__init__(processor, split, sampling_rate, debug)
|
| 258 |
+
|
| 259 |
+
self.set_dataset_name(f"LibriSpeech_{subset}")
|
| 260 |
+
# only ASR
|
| 261 |
+
self.ast = False
|
| 262 |
+
self.lang = "en"
|
| 263 |
+
|
| 264 |
+
# load dataset
|
| 265 |
+
self.data = load_dataset("fixie-ai/librispeech_asr",
|
| 266 |
+
subset,
|
| 267 |
+
split=split,
|
| 268 |
+
trust_remote_code=True
|
| 269 |
+
)
|
| 270 |
+
|
| 271 |
+
# (Optional) Audio length Filtering
|
| 272 |
+
self.data = self.filter_by_audio_length(self.data, "audio")
|
| 273 |
+
|
| 274 |
+
# Instruction Setting
|
| 275 |
+
self.instruction = random.choice(INSTRUCTION["asr"])
|
| 276 |
+
|
| 277 |
+
def __len__(self):
|
| 278 |
+
return len(self.data)
|
| 279 |
+
|
| 280 |
+
def __getitem__(self, idx):
|
| 281 |
+
data = self.data[idx]
|
| 282 |
+
|
| 283 |
+
# Libri Speech is only for ASR
|
| 284 |
+
answer_text = data["text"].replace('"', '')
|
| 285 |
+
|
| 286 |
+
return self.prepare_model_inputs(
|
| 287 |
+
data["audio"]["array"],
|
| 288 |
+
self.instruction,
|
| 289 |
+
answer_text
|
| 290 |
+
)
|
| 291 |
+
|
| 292 |
+
# Fleurs Dataset Class
|
| 293 |
+
class FleursDataset(BaseAudioDataset):
|
| 294 |
+
def __init__(self, processor, split, source_lang, target_lang=None,
|
| 295 |
+
mode="asr", sampling_rate=16000, debug=False):
|
| 296 |
+
super().__init__(processor, split, sampling_rate, debug)
|
| 297 |
+
|
| 298 |
+
self.set_dataset_name("Fleurs")
|
| 299 |
+
# Mode Setting (ASR or AST)
|
| 300 |
+
if mode not in ["asr", "ast"]:
|
| 301 |
+
raise ValueError("mode must be 'asr' or 'ast'.")
|
| 302 |
+
|
| 303 |
+
self.mode = mode
|
| 304 |
+
self.ast = (mode == "ast")
|
| 305 |
+
self.source_lang = source_lang
|
| 306 |
+
|
| 307 |
+
# Language name mapping (expand if needed)
|
| 308 |
+
self.lang_names = {
|
| 309 |
+
'en_us': 'English', 'ko_kr': 'Korean'
|
| 310 |
+
}
|
| 311 |
+
|
| 312 |
+
# load dataset - source language dataset
|
| 313 |
+
self.data = load_dataset("google/fleurs",
|
| 314 |
+
source_lang,
|
| 315 |
+
split=split,
|
| 316 |
+
trust_remote_code=True
|
| 317 |
+
)
|
| 318 |
+
|
| 319 |
+
# (Optional) Audio length Filtering
|
| 320 |
+
self.data = self.filter_by_audio_length(self.data, "audio")
|
| 321 |
+
|
| 322 |
+
# When AST mode, load target language dataset.
|
| 323 |
+
if self.ast:
|
| 324 |
+
if target_lang is None:
|
| 325 |
+
raise ValueError("AST mode requires target_lang.")
|
| 326 |
+
|
| 327 |
+
self.target_lang = target_lang
|
| 328 |
+
self.lang = f"{source_lang}_{target_lang}"
|
| 329 |
+
|
| 330 |
+
# load dataset - target language dataset (for translation)
|
| 331 |
+
target_data = load_dataset("google/fleurs",
|
| 332 |
+
target_lang,
|
| 333 |
+
split=split,
|
| 334 |
+
trust_remote_code=True
|
| 335 |
+
)
|
| 336 |
+
|
| 337 |
+
source_dict = {item['id']: item for item in self.data}
|
| 338 |
+
target_dict = {item['id']: item for item in target_data}
|
| 339 |
+
|
| 340 |
+
# only Common ID, add translation fields
|
| 341 |
+
common_ids = set(source_dict.keys()) & set(target_dict.keys())
|
| 342 |
+
print(f"FLEURS AST Common data filtering: {len(self.data)} -> {len(common_ids)}")
|
| 343 |
+
self.data = [
|
| 344 |
+
{**source_dict[id], 'translation': target_dict[id]['transcription']}
|
| 345 |
+
for id in common_ids
|
| 346 |
+
]
|
| 347 |
+
|
| 348 |
+
# Instruction Setting - use target language name
|
| 349 |
+
target_lang_name = self.lang_names.get(target_lang, target_lang.capitalize())
|
| 350 |
+
self.instruction = random.choice(INSTRUCTION["ast"]).format(target_lang_name)
|
| 351 |
+
else:
|
| 352 |
+
# ASR mode
|
| 353 |
+
self.lang = source_lang
|
| 354 |
+
self.instruction = random.choice(INSTRUCTION["asr"])
|
| 355 |
+
|
| 356 |
+
if self.debug:
|
| 357 |
+
print(f"FLEURS dataset loaded: {self.mode.upper()} mode")
|
| 358 |
+
print(f"source lang: {source_lang} ({self.lang_names.get(source_lang, source_lang)})")
|
| 359 |
+
if self.ast:
|
| 360 |
+
print(f"target lang: {target_lang} ({self.lang_names.get(target_lang, target_lang)})")
|
| 361 |
+
print(f"dataset size: {len(self.data)}")
|
| 362 |
+
|
| 363 |
+
def __len__(self):
|
| 364 |
+
return len(self.data)
|
| 365 |
+
|
| 366 |
+
def __getitem__(self, idx):
|
| 367 |
+
data = self.data[idx]
|
| 368 |
+
audio_array = data["audio"]["array"]
|
| 369 |
+
|
| 370 |
+
if self.ast:
|
| 371 |
+
answer_text = data["translation"]
|
| 372 |
+
else:
|
| 373 |
+
answer_text = data["transcription"]
|
| 374 |
+
|
| 375 |
+
return self.prepare_model_inputs(
|
| 376 |
+
audio_array,
|
| 377 |
+
self.instruction,
|
| 378 |
+
answer_text
|
| 379 |
+
)
|
| 380 |
+
|
| 381 |
+
def covost_collate_fn(batch):
|
| 382 |
+
input_ids_list = []
|
| 383 |
+
labels_list = []
|
| 384 |
+
token_type_ids_list = []
|
| 385 |
+
input_audio_embeds_list = []
|
| 386 |
+
audio_embed_sizes_list = []
|
| 387 |
+
audio_attention_mask_list = []
|
| 388 |
+
input_modes_list = []
|
| 389 |
+
for inputs in batch:
|
| 390 |
+
input_ids_list.append(inputs['input_ids'][0])
|
| 391 |
+
labels_list.append(inputs['labels'][0])
|
| 392 |
+
token_type_ids_list.append(inputs['token_type_ids'][0])
|
| 393 |
+
input_audio_embeds_list.append(inputs['input_audio_embeds'])
|
| 394 |
+
audio_embed_sizes_list.append(inputs['audio_embed_sizes'])
|
| 395 |
+
audio_attention_mask_list.append(
|
| 396 |
+
inputs['input_audio_embeds'].new_full((inputs['input_audio_embeds'].size(1),), True, dtype=torch.bool)
|
| 397 |
+
)
|
| 398 |
+
input_modes_list.append(inputs['input_modes'])
|
| 399 |
+
|
| 400 |
+
try:
|
| 401 |
+
token_type_ids = pad_sequence(token_type_ids_list, padding_side='left', padding_value=0)
|
| 402 |
+
input_ids = pad_sequence(input_ids_list, padding_side='left', padding_value=0)
|
| 403 |
+
labels = pad_sequence(labels_list, padding_side='left', padding_value=0)
|
| 404 |
+
audio_attention_mask = (
|
| 405 |
+
pad_sequence(audio_attention_mask_list, padding_side='left', padding_value=False)
|
| 406 |
+
if len(audio_attention_mask_list) > 1
|
| 407 |
+
else None
|
| 408 |
+
)
|
| 409 |
+
except Exception as e:
|
| 410 |
+
print(e)
|
| 411 |
+
print(input_ids_list)
|
| 412 |
+
print(labels_list)
|
| 413 |
+
raise
|
| 414 |
+
attention_mask = (input_ids != 0).long()
|
| 415 |
+
input_audio_embeds = cat_with_pad(input_audio_embeds_list, dim=0)
|
| 416 |
+
audio_embed_sizes = torch.cat(audio_embed_sizes_list)
|
| 417 |
+
input_modes = torch.cat(input_modes_list)
|
| 418 |
+
|
| 419 |
+
return BatchFeature(
|
| 420 |
+
{
|
| 421 |
+
'input_ids': input_ids,
|
| 422 |
+
'labels': labels,
|
| 423 |
+
'token_type_ids': token_type_ids,
|
| 424 |
+
'attention_mask': attention_mask,
|
| 425 |
+
'input_audio_embeds': input_audio_embeds,
|
| 426 |
+
'audio_embed_sizes': audio_embed_sizes,
|
| 427 |
+
'audio_attention_mask': audio_attention_mask,
|
| 428 |
+
'input_modes': input_modes,
|
| 429 |
+
}
|
| 430 |
+
)
|
| 431 |
+
|
| 432 |
+
def pad_sequence(sequences, padding_side='left', padding_value=0):
|
| 433 |
+
"""
|
| 434 |
+
Pad a list of sequences to the same length.
|
| 435 |
+
sequences: list of tensors in [seq_len, *] shape
|
| 436 |
+
"""
|
| 437 |
+
assert padding_side in ['right', 'left']
|
| 438 |
+
max_size = sequences[0].size()
|
| 439 |
+
trailing_dims = max_size[1:]
|
| 440 |
+
max_len = max(len(seq) for seq in sequences)
|
| 441 |
+
batch_size = len(sequences)
|
| 442 |
+
output = sequences[0].new_full((batch_size, max_len) + trailing_dims, padding_value)
|
| 443 |
+
for i, seq in enumerate(sequences):
|
| 444 |
+
length = seq.size(0)
|
| 445 |
+
if padding_side == 'right':
|
| 446 |
+
output.data[i, :length] = seq
|
| 447 |
+
else:
|
| 448 |
+
output.data[i, -length:] = seq
|
| 449 |
+
return output
|
| 450 |
+
|
| 451 |
+
def cat_with_pad(tensors, dim, padding_value=0):
|
| 452 |
+
"""
|
| 453 |
+
cat along dim, while pad to max for all other dims
|
| 454 |
+
"""
|
| 455 |
+
ndim = tensors[0].dim()
|
| 456 |
+
assert all(
|
| 457 |
+
t.dim() == ndim for t in tensors[1:]
|
| 458 |
+
), 'All tensors must have the same number of dimensions'
|
| 459 |
+
|
| 460 |
+
out_size = [max(t.shape[i] for t in tensors) for i in range(ndim)]
|
| 461 |
+
out_size[dim] = sum(t.shape[dim] for t in tensors)
|
| 462 |
+
output = tensors[0].new_full(out_size, padding_value)
|
| 463 |
+
|
| 464 |
+
index = 0
|
| 465 |
+
for t in tensors:
|
| 466 |
+
# Create a slice list where every dimension except dim is full slice
|
| 467 |
+
slices = [slice(0, t.shape[d]) for d in range(ndim)]
|
| 468 |
+
# Update only the concat dimension slice
|
| 469 |
+
slices[dim] = slice(index, index + t.shape[dim])
|
| 470 |
+
|
| 471 |
+
output[slices] = t
|
| 472 |
+
index += t.shape[dim]
|
| 473 |
+
|
| 474 |
+
return output
|
| 475 |
+
|
| 476 |
+
def count_parameters_by_module(model):
|
| 477 |
+
# dictionary for parameters number by modules
|
| 478 |
+
module_params = defaultdict(lambda: {"total": 0, "trainable": 0})
|
| 479 |
+
|
| 480 |
+
# all params
|
| 481 |
+
total_params = 0
|
| 482 |
+
total_trainable_params = 0
|
| 483 |
+
|
| 484 |
+
# Check Embedding Token masks
|
| 485 |
+
embedding_masks = {}
|
| 486 |
+
for name, param in model.named_parameters():
|
| 487 |
+
if 'embed_tokens.weight' in name and hasattr(param, '_backward_hooks') and param._backward_hooks:
|
| 488 |
+
# check if params has embedding_grad_mask_hook
|
| 489 |
+
for hook_id, hook_fn in param._backward_hooks.items():
|
| 490 |
+
if hook_fn.__code__.co_name == 'embedding_grad_mask_hook':
|
| 491 |
+
# Accessing mask variables in the closure of hook functions
|
| 492 |
+
for cell in hook_fn.__closure__ or []:
|
| 493 |
+
if isinstance(cell.cell_contents, torch.Tensor) and cell.cell_contents.dtype == torch.bool:
|
| 494 |
+
# check mask tensor
|
| 495 |
+
embedding_masks[name] = ~cell.cell_contents # True : Trainable
|
| 496 |
+
|
| 497 |
+
# Count params by modules
|
| 498 |
+
for name, param in model.named_parameters():
|
| 499 |
+
# extracts top module_name
|
| 500 |
+
module_name = name.split('.')[0]
|
| 501 |
+
param_count = param.numel()
|
| 502 |
+
|
| 503 |
+
module_params[module_name]["total"] += param_count
|
| 504 |
+
total_params += param_count
|
| 505 |
+
|
| 506 |
+
if param.requires_grad:
|
| 507 |
+
# Only count for real trainable params. (with masks)
|
| 508 |
+
if name in embedding_masks:
|
| 509 |
+
trainable_count = embedding_masks[name].sum().item()
|
| 510 |
+
module_params[module_name]["trainable"] += trainable_count
|
| 511 |
+
total_trainable_params += trainable_count
|
| 512 |
+
else:
|
| 513 |
+
module_params[module_name]["trainable"] += param_count
|
| 514 |
+
total_trainable_params += param_count
|
| 515 |
+
|
| 516 |
+
print(f"All Params: {total_params:,}")
|
| 517 |
+
print(f"Trainable Params: {total_trainable_params:,} ({total_trainable_params/total_params*100:.2f}%)")
|
| 518 |
+
print("\nParams by Module:")
|
| 519 |
+
|
| 520 |
+
for module_name, counts in sorted(module_params.items()):
|
| 521 |
+
trainable_percentage = counts["trainable"] / counts["total"] * 100 if counts["total"] > 0 else 0
|
| 522 |
+
total_percentage = counts["total"] / total_params * 100
|
| 523 |
+
|
| 524 |
+
print(f"- {module_name}:")
|
| 525 |
+
print(f" Total: {counts['total']:,} ({total_percentage:.2f}% of model)")
|
| 526 |
+
print(f" Trainable: {counts['trainable']:,} ({trainable_percentage:.2f}% of module)")
|
| 527 |
+
|
| 528 |
+
return module_params
|
| 529 |
+
|
| 530 |
+
def create_model(model_name_or_path, revision="main", use_flash_attention = False):
|
| 531 |
+
model = AutoModel.from_pretrained(
|
| 532 |
+
model_name_or_path,
|
| 533 |
+
revision=revision,
|
| 534 |
+
torch_dtype=torch.bfloat16,
|
| 535 |
+
device_map="auto",
|
| 536 |
+
attn_implementation="flash_attention_2" if use_flash_attention else "eager",
|
| 537 |
+
trust_remote_code=True,
|
| 538 |
+
)
|
| 539 |
+
|
| 540 |
+
# Set use_cache to False after model loaded
|
| 541 |
+
model.config.use_cache = False
|
| 542 |
+
|
| 543 |
+
# Freeze all parameters
|
| 544 |
+
for param in model.parameters():
|
| 545 |
+
param.requires_grad = False
|
| 546 |
+
|
| 547 |
+
model.set_lora_adapter('speech')
|
| 548 |
+
model.to(torch.bfloat16)
|
| 549 |
+
|
| 550 |
+
# (Optional) unfreeze audio_tower parameters
|
| 551 |
+
#for param in model.audio_tower.parameters():
|
| 552 |
+
# param.requires_grad = True
|
| 553 |
+
|
| 554 |
+
# Only unfreeze audio_projector parameters
|
| 555 |
+
for param in model.audio_projector.parameters():
|
| 556 |
+
param.requires_grad = True
|
| 557 |
+
|
| 558 |
+
# (Optional) unfreeze audio embed_tokens
|
| 559 |
+
train_embed = True
|
| 560 |
+
if train_embed:
|
| 561 |
+
embed_tokens = model.language_model.model.model.embed_tokens
|
| 562 |
+
|
| 563 |
+
embed_tokens.weight.requires_grad = False
|
| 564 |
+
|
| 565 |
+
# Added Speech token IDs (only this tokens be trainable)
|
| 566 |
+
trainable_token_ids = [256001, 256002]
|
| 567 |
+
|
| 568 |
+
embed_tokens.weight.requires_grad = True
|
| 569 |
+
mask = torch.ones_like(embed_tokens.weight, dtype=torch.bool)
|
| 570 |
+
mask[trainable_token_ids] = False # Trainable Tokens are False (unfreeze), else True (freeze)
|
| 571 |
+
|
| 572 |
+
# backward hook, with gradient masking
|
| 573 |
+
def embedding_grad_mask_hook(grad):
|
| 574 |
+
return grad.masked_fill(mask, 0)
|
| 575 |
+
|
| 576 |
+
embed_tokens.weight.register_hook(embedding_grad_mask_hook)
|
| 577 |
+
|
| 578 |
+
model.language_model.model.model.embed_tokens = embed_tokens
|
| 579 |
+
|
| 580 |
+
count_parameters_by_module(model)
|
| 581 |
+
|
| 582 |
+
return model
|
| 583 |
+
|
| 584 |
+
@torch.no_grad()
|
| 585 |
+
def evaluate(model, processor, eval_dataset, save_path=None, disable_tqdm=False, eval_batch_size=1):
|
| 586 |
+
model.eval()
|
| 587 |
+
all_generated_texts = []
|
| 588 |
+
all_labels = []
|
| 589 |
+
|
| 590 |
+
eval_dataloader = torch.utils.data.DataLoader(
|
| 591 |
+
eval_dataset,
|
| 592 |
+
batch_size=eval_batch_size,
|
| 593 |
+
collate_fn=covost_collate_fn,
|
| 594 |
+
shuffle=False,
|
| 595 |
+
drop_last=False,
|
| 596 |
+
num_workers=8,
|
| 597 |
+
prefetch_factor=2,
|
| 598 |
+
pin_memory=True,
|
| 599 |
+
)
|
| 600 |
+
stop_tokens = [processor.tokenizer.eos_token]
|
| 601 |
+
stop_tokens_ids = processor.tokenizer(stop_tokens, add_special_tokens=False, padding="longest", return_tensors="pt")["input_ids"]
|
| 602 |
+
stop_tokens_ids = stop_tokens_ids.to('cuda')
|
| 603 |
+
|
| 604 |
+
for inputs in tqdm(
|
| 605 |
+
eval_dataloader, disable= disable_tqdm, desc='running eval'
|
| 606 |
+
):
|
| 607 |
+
stopping_criteria=StoppingCriteriaList([MultipleTokenBatchStoppingCriteria(stop_tokens_ids, batch_size=inputs.input_ids.size(0))])
|
| 608 |
+
inputs = inputs.to('cuda').to(model.dtype)
|
| 609 |
+
generated_ids = model.generate(
|
| 610 |
+
**inputs, eos_token_id=processor.tokenizer.eos_token_id, max_new_tokens=64,
|
| 611 |
+
stopping_criteria=stopping_criteria,
|
| 612 |
+
)
|
| 613 |
+
|
| 614 |
+
stop_tokens_idx = stopping_criteria[0].stop_tokens_idx.reshape(inputs.input_ids.size(0), -1)[:, 0]
|
| 615 |
+
|
| 616 |
+
stop_tokens_idx = torch.where(
|
| 617 |
+
stop_tokens_idx > 0,
|
| 618 |
+
stop_tokens_idx - stop_tokens_ids.shape[-1],
|
| 619 |
+
generated_ids.shape[-1],
|
| 620 |
+
)
|
| 621 |
+
generated_text = [
|
| 622 |
+
processor.decode(_pred_ids[inputs["input_ids"].shape[1] : _stop_tokens_idx], skip_special_tokens=True, clean_up_tokenization_spaces=False)
|
| 623 |
+
for _pred_ids, _stop_tokens_idx in zip(generated_ids, stop_tokens_idx)
|
| 624 |
+
]
|
| 625 |
+
all_generated_texts.extend(generated_text)
|
| 626 |
+
labels = [processor.decode(_label_ids[_label_ids != 0]).removesuffix(ANSWER_SUFFIX) for _label_ids in inputs["labels"]]
|
| 627 |
+
all_labels.extend(labels)
|
| 628 |
+
|
| 629 |
+
assert len(all_generated_texts) == len(all_labels)
|
| 630 |
+
bleu = sacrebleu.corpus_bleu(all_generated_texts, [all_labels])
|
| 631 |
+
print(bleu)
|
| 632 |
+
if save_path:
|
| 633 |
+
with open(save_path, 'w') as f:
|
| 634 |
+
save_dict = {
|
| 635 |
+
'all_generated_texts': all_generated_texts,
|
| 636 |
+
'all_labels': all_labels,
|
| 637 |
+
'score': bleu.score,
|
| 638 |
+
}
|
| 639 |
+
json.dump(save_dict, f)
|
| 640 |
+
|
| 641 |
+
return bleu.score
|
| 642 |
+
|
| 643 |
+
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
| 644 |
+
|
| 645 |
+
INSTRUCTION = {
|
| 646 |
+
"ast": [
|
| 647 |
+
"Translate the audio to {0}.",
|
| 648 |
+
"Translate the audio clip into {0}.",
|
| 649 |
+
"Based on the attached audio, generate a comprehensive {0} translation of the spoken content.",
|
| 650 |
+
"Translate the provided audio file into {0}.",
|
| 651 |
+
"Convert the audio speech to {0} text.",
|
| 652 |
+
"Write an {0} translation of the audio file.",
|
| 653 |
+
"Translate spoken words from the audio into {0}.",
|
| 654 |
+
"Create an {0} version of the audio content.",
|
| 655 |
+
"Produce an accurate {0} translation of the audio.",
|
| 656 |
+
"Extract speech from the audio and translate it to {0}.",
|
| 657 |
+
"Turn the audio into readable {0} text.",
|
| 658 |
+
"Write all spoken content from the audio in {0}.",
|
| 659 |
+
"Generate an {0} translation of the speech in the file.",
|
| 660 |
+
"Convert the recording into {0} text.",
|
| 661 |
+
"Accurately translate the audio recording to {0}.",
|
| 662 |
+
"Write down dialogue from the given audio in {0}.",
|
| 663 |
+
"Translate all speech in this audio file to {0}.",
|
| 664 |
+
"Create an accurate {0} version of the speech.",
|
| 665 |
+
"Perform a complete {0} translation of the audio."
|
| 666 |
+
],
|
| 667 |
+
"asr": [
|
| 668 |
+
"Transcribe the audio clip into text.",
|
| 669 |
+
"Based on the attached audio, generate a comprehensive text transcription of the spoken content.",
|
| 670 |
+
"Transcribe the provided audio file into text.",
|
| 671 |
+
"Convert the audio speech to text.",
|
| 672 |
+
"Write a transcript of the audio file.",
|
| 673 |
+
"Transcribe spoken words from the audio.",
|
| 674 |
+
"Create a text version of the audio content.",
|
| 675 |
+
"Produce a verbatim transcript of the audio.",
|
| 676 |
+
"Extract and transcribe speech from the audio.",
|
| 677 |
+
"Turn the audio into readable text.",
|
| 678 |
+
"Write all spoken words from the audio.",
|
| 679 |
+
"Generate a transcript of the speech in the file.",
|
| 680 |
+
"Convert the recording into a text transcript.",
|
| 681 |
+
"Accurately transcribe the audio recording.",
|
| 682 |
+
"Write down dialogue from the given audio.",
|
| 683 |
+
"Transcribe all speech in this audio file.",
|
| 684 |
+
"Create an accurate text version of the speech.",
|
| 685 |
+
"Perform a complete transcription of the audio."
|
| 686 |
+
],
|
| 687 |
+
}
|
| 688 |
+
|
| 689 |
+
ANSWER_SUFFIX = "<end_of_turn>"
|
| 690 |
+
_IGNORE_INDEX = -100
|
| 691 |
+
|
| 692 |
+
model_name_or_path = 'junnei/gemma-3-4b-it-speech'
|
| 693 |
+
use_flash_attention = True
|
| 694 |
+
|
| 695 |
+
output_dir = '/workspace/output'
|
| 696 |
+
batch_size = 128
|
| 697 |
+
batch_size_per_gpu = 32
|
| 698 |
+
learning_rate = 4.0e-5 # 1.0e-4 for fine-tuning
|
| 699 |
+
wd = 0.01
|
| 700 |
+
num_train_epochs = 5
|
| 701 |
+
|
| 702 |
+
revision = "main" #"v1.0"
|
| 703 |
+
|
| 704 |
+
processor = AutoProcessor.from_pretrained(
|
| 705 |
+
model_name_or_path,
|
| 706 |
+
revision=revision,
|
| 707 |
+
trust_remote_code=True,
|
| 708 |
+
)
|
| 709 |
+
|
| 710 |
+
model = create_model(
|
| 711 |
+
model_name_or_path,
|
| 712 |
+
revision=revision,
|
| 713 |
+
use_flash_attention=use_flash_attention,
|
| 714 |
+
)
|
| 715 |
+
|
| 716 |
+
train_datasets = []
|
| 717 |
+
|
| 718 |
+
# Covost ASR mode (English -> English text)
|
| 719 |
+
covost_asr_dataset = CoVoSTDataset(
|
| 720 |
+
processor=processor,
|
| 721 |
+
data_dir="/workspace/CommonVoice/EN",
|
| 722 |
+
split="train",
|
| 723 |
+
ast=False,
|
| 724 |
+
lang=("en_ko", "Korean")
|
| 725 |
+
)
|
| 726 |
+
train_datasets.append(covost_asr_dataset)
|
| 727 |
+
|
| 728 |
+
# Covost AST mode (English -> Korean text)
|
| 729 |
+
covost_dataset = CoVoSTDataset(
|
| 730 |
+
processor=processor,
|
| 731 |
+
data_dir="/workspace/CommonVoice/EN",
|
| 732 |
+
split="train",
|
| 733 |
+
ast=True,
|
| 734 |
+
lang=("en_ko", "Korean")
|
| 735 |
+
)
|
| 736 |
+
train_datasets.append(covost_dataset)
|
| 737 |
+
|
| 738 |
+
# Libri Speech Clean ASR mode (English -> English text)
|
| 739 |
+
libri_speech_clean = LibriSpeechDataset(
|
| 740 |
+
processor=processor,
|
| 741 |
+
subset="clean",
|
| 742 |
+
split="train.360"
|
| 743 |
+
)
|
| 744 |
+
train_datasets.append(libri_speech_clean)
|
| 745 |
+
|
| 746 |
+
# Libri Speech Other ASR mode (English -> English text)
|
| 747 |
+
libri_speech_other = LibriSpeechDataset(
|
| 748 |
+
processor=processor,
|
| 749 |
+
subset="other",
|
| 750 |
+
split="train.500"
|
| 751 |
+
)
|
| 752 |
+
train_datasets.append(libri_speech_other)
|
| 753 |
+
|
| 754 |
+
# Fleurs ASR mode (English -> English text)
|
| 755 |
+
en_asr_fleurs = FleursDataset(
|
| 756 |
+
processor=processor,
|
| 757 |
+
split="train",
|
| 758 |
+
source_lang="en_us", # English
|
| 759 |
+
mode="asr"
|
| 760 |
+
)
|
| 761 |
+
train_datasets.append(en_asr_fleurs)
|
| 762 |
+
|
| 763 |
+
# Fleurs AST mode (English -> Korean text)
|
| 764 |
+
en_ko_ast_fleurs = FleursDataset(
|
| 765 |
+
processor=processor,
|
| 766 |
+
split="train",
|
| 767 |
+
source_lang="en_us", # English
|
| 768 |
+
target_lang="ko_kr", # Korean
|
| 769 |
+
mode="ast"
|
| 770 |
+
)
|
| 771 |
+
train_datasets.append(en_ko_ast_fleurs)
|
| 772 |
+
|
| 773 |
+
# Covost ASR mode (Korean -> Korean text)
|
| 774 |
+
covost_ko_asr_dataset = CoVoSTDataset(
|
| 775 |
+
processor=processor,
|
| 776 |
+
data_dir="/workspace/CommonVoice/ko",
|
| 777 |
+
split="train",
|
| 778 |
+
ast=False,
|
| 779 |
+
lang=("ko_en", "English")
|
| 780 |
+
)
|
| 781 |
+
train_datasets.append(covost_ko_asr_dataset)
|
| 782 |
+
|
| 783 |
+
# Covost AST mode (Korean -> English text)
|
| 784 |
+
covost_ko_dataset = CoVoSTDataset(
|
| 785 |
+
processor=processor,
|
| 786 |
+
data_dir="/workspace/CommonVoice/ko",
|
| 787 |
+
split="train",
|
| 788 |
+
ast=True,
|
| 789 |
+
lang=("ko_en", "English")
|
| 790 |
+
)
|
| 791 |
+
train_datasets.append(covost_ko_dataset)
|
| 792 |
+
|
| 793 |
+
# Zeroth ASR mode (Korean -> Korean text)
|
| 794 |
+
ko_asr_zeroth = ZerothKoreanDataset(
|
| 795 |
+
processor=processor,
|
| 796 |
+
split="train"
|
| 797 |
+
)
|
| 798 |
+
train_datasets.append(ko_asr_zeroth)
|
| 799 |
+
|
| 800 |
+
# Fleurs ASR mode (Korean -> Korean text)
|
| 801 |
+
ko_asr_fleurs = FleursDataset(
|
| 802 |
+
processor=processor,
|
| 803 |
+
split="train",
|
| 804 |
+
source_lang="ko_kr", # Korean
|
| 805 |
+
mode="asr"
|
| 806 |
+
)
|
| 807 |
+
train_datasets.append(ko_asr_fleurs)
|
| 808 |
+
|
| 809 |
+
# Fleurs AST mode (Korean -> English text)
|
| 810 |
+
ko_en_ast_fleurs = FleursDataset(
|
| 811 |
+
processor=processor,
|
| 812 |
+
split="train",
|
| 813 |
+
source_lang="ko_kr", # Korean
|
| 814 |
+
target_lang="en_us", # English
|
| 815 |
+
mode="ast"
|
| 816 |
+
)
|
| 817 |
+
train_datasets.append(ko_en_ast_fleurs)
|
| 818 |
+
|
| 819 |
+
print("Count Num of Datasets", len(train_datasets))
|
| 820 |
+
print([len(dataset) for dataset in train_datasets])
|
| 821 |
+
|
| 822 |
+
# ConcatDataset
|
| 823 |
+
train_dataset = ConcatDataset(train_datasets) if len(train_datasets) > 1 else train_datasets[0]
|
| 824 |
+
print("Count Length of Datas", len(train_dataset))
|
| 825 |
+
|
| 826 |
+
# Check GPUs
|
| 827 |
+
num_gpus = torch.cuda.device_count()
|
| 828 |
+
print(f'training on {num_gpus} GPUs')
|
| 829 |
+
|
| 830 |
+
assert (
|
| 831 |
+
batch_size % (num_gpus * batch_size_per_gpu) == 0
|
| 832 |
+
), 'Batch size must be divisible by the number of GPUs'
|
| 833 |
+
gradient_accumulation_steps = batch_size // (num_gpus * batch_size_per_gpu)
|
| 834 |
+
|
| 835 |
+
# hard coded training args
|
| 836 |
+
training_args = TrainingArguments(
|
| 837 |
+
num_train_epochs=num_train_epochs,
|
| 838 |
+
per_device_train_batch_size=batch_size_per_gpu,
|
| 839 |
+
gradient_checkpointing=True,
|
| 840 |
+
gradient_checkpointing_kwargs={'use_reentrant': False},
|
| 841 |
+
gradient_accumulation_steps=gradient_accumulation_steps,
|
| 842 |
+
optim='adamw_torch',
|
| 843 |
+
adam_beta1=0.9,
|
| 844 |
+
adam_beta2=0.95,
|
| 845 |
+
adam_epsilon=1e-7,
|
| 846 |
+
learning_rate=learning_rate,
|
| 847 |
+
weight_decay=wd,
|
| 848 |
+
max_grad_norm=1.0,
|
| 849 |
+
lr_scheduler_type='cosine',
|
| 850 |
+
warmup_steps=50,
|
| 851 |
+
logging_steps=50,
|
| 852 |
+
output_dir=output_dir,
|
| 853 |
+
save_strategy='no',
|
| 854 |
+
save_total_limit=10,
|
| 855 |
+
save_only_model=True,
|
| 856 |
+
bf16=True,
|
| 857 |
+
fp16=False,
|
| 858 |
+
remove_unused_columns=False,
|
| 859 |
+
report_to='none',
|
| 860 |
+
deepspeed=None,
|
| 861 |
+
disable_tqdm=False,
|
| 862 |
+
dataloader_num_workers=4,
|
| 863 |
+
ddp_find_unused_parameters=True,
|
| 864 |
+
)
|
| 865 |
+
|
| 866 |
+
out_path = Path(training_args.output_dir)
|
| 867 |
+
out_path.mkdir(parents=True, exist_ok=True)
|
| 868 |
+
|
| 869 |
+
# create optimizer only for trainable params
|
| 870 |
+
optimizer = torch.optim.AdamW(
|
| 871 |
+
filter(lambda p: p.requires_grad, model.parameters()),
|
| 872 |
+
lr=learning_rate,
|
| 873 |
+
weight_decay=wd,
|
| 874 |
+
betas=(0.9, 0.95),
|
| 875 |
+
eps=1e-7,
|
| 876 |
+
)
|
| 877 |
+
|
| 878 |
+
# Trainer Setting
|
| 879 |
+
trainer = Trainer(
|
| 880 |
+
model=model,
|
| 881 |
+
args=training_args,
|
| 882 |
+
data_collator=covost_collate_fn,
|
| 883 |
+
train_dataset=train_dataset,
|
| 884 |
+
optimizers=(optimizer, None),
|
| 885 |
+
)
|
| 886 |
+
|
| 887 |
+
trainer.train()
|
| 888 |
+
|
| 889 |
+
import shutil
|
| 890 |
+
|
| 891 |
+
# setting output dir
|
| 892 |
+
output_dir = "/workspace/output"
|
| 893 |
+
|
| 894 |
+
# 1. Save LoRA Adapter
|
| 895 |
+
model.language_model.model.save_pretrained(output_dir)
|
| 896 |
+
|
| 897 |
+
# 1-1. Delete Markdown file
|
| 898 |
+
markdown_file = os.path.join(output_dir, "README.md")
|
| 899 |
+
if os.path.exists(markdown_file):
|
| 900 |
+
os.remove(markdown_file)
|
| 901 |
+
|
| 902 |
+
# 2. Save entire model
|
| 903 |
+
model.save_pretrained(output_dir)
|
| 904 |
+
|
| 905 |
+
# 3. Cleanup Memory
|
| 906 |
+
del model
|
| 907 |
+
del trainer
|
| 908 |
+
__import__('gc').collect()
|
| 909 |
+
torch.cuda.empty_cache()
|
| 910 |
+
|
| 911 |
+
from huggingface_hub import HfApi, login, create_repo, Repository, upload_folder
|
| 912 |
+
|
| 913 |
+
upload_dir = "/workspace/upload"
|
| 914 |
+
|
| 915 |
+
# 4. Clone Repo
|
| 916 |
+
repo_id = "junnei/gemma-3-4b-it-speech"
|
| 917 |
+
branch_name = "main" # 새 브랜치 이름
|
| 918 |
+
|
| 919 |
+
repo = Repository(local_dir=upload_dir, clone_from = repo_id)
|
| 920 |
+
repo.git_checkout(branch_name, create_branch_ok=True)
|
| 921 |
+
|
| 922 |
+
# 4-1. Move Trained model to Repo
|
| 923 |
+
for item in os.listdir(output_dir):
|
| 924 |
+
s = os.path.join(output_dir, item)
|
| 925 |
+
d = os.path.join(upload_dir, item)
|
| 926 |
+
if os.path.isdir(s):
|
| 927 |
+
shutil.copytree(s, d, dirs_exist_ok=True)
|
| 928 |
+
else:
|
| 929 |
+
shutil.copy2(s, d)
|