Upload eval_model.py
Browse files- eval_model.py +153 -0
eval_model.py
ADDED
@@ -0,0 +1,153 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoTokenizer, GPT2LMHeadModel
|
2 |
+
from datasets import load_dataset, Dataset, DatasetDict
|
3 |
+
import random
|
4 |
+
import string
|
5 |
+
import torch
|
6 |
+
|
7 |
+
from torchmetrics.text import WordErrorRate, CharErrorRate
|
8 |
+
|
9 |
+
wer = WordErrorRate()
|
10 |
+
cer = CharErrorRate()
|
11 |
+
|
12 |
+
def process(text):
|
13 |
+
|
14 |
+
# Lower case every letter
|
15 |
+
text = text.lower()
|
16 |
+
|
17 |
+
# Remove punctuation
|
18 |
+
punctuation_to_remove = string.punctuation.replace("'", "")
|
19 |
+
translation_table = str.maketrans('', '', punctuation_to_remove)
|
20 |
+
text = text.translate(translation_table)
|
21 |
+
|
22 |
+
# Remove whitespaces from front and behind
|
23 |
+
while text[0] == ' ' or text[-1] == ' ':
|
24 |
+
if text[0] == ' ':
|
25 |
+
text = text[1:]
|
26 |
+
if text[-1] == ' ':
|
27 |
+
text = text[:-1]
|
28 |
+
|
29 |
+
return text
|
30 |
+
|
31 |
+
import jiwer
|
32 |
+
from edit_distance import SequenceMatcher
|
33 |
+
def correct_text(text):
|
34 |
+
transforms = jiwer.Compose(
|
35 |
+
[
|
36 |
+
jiwer.ExpandCommonEnglishContractions(),
|
37 |
+
jiwer.ToLowerCase(),
|
38 |
+
jiwer.RemoveMultipleSpaces(),
|
39 |
+
jiwer.Strip(),
|
40 |
+
jiwer.RemovePunctuation(),
|
41 |
+
jiwer.ReduceToListOfListOfWords(),
|
42 |
+
]
|
43 |
+
)
|
44 |
+
return transforms(text)
|
45 |
+
|
46 |
+
def align_gt_asr(gt, asr):
|
47 |
+
sm = SequenceMatcher(a=gt, b=asr)
|
48 |
+
best_path = []
|
49 |
+
opcodes = sm.get_opcodes()
|
50 |
+
for tag, i1, i2, j1, j2 in opcodes:
|
51 |
+
if tag == "delete":
|
52 |
+
for i in range(i1, i2):
|
53 |
+
best_path.append([gt[i], ""])
|
54 |
+
if tag == "replace" or tag == "equal":
|
55 |
+
for i, j in zip(range(i1, i2), range(j1, j2)):
|
56 |
+
best_path.append([gt[i], asr[j]])
|
57 |
+
if tag == "insert":
|
58 |
+
for j in range(j1, j2):
|
59 |
+
best_path.append(["", asr[j]])
|
60 |
+
return best_path
|
61 |
+
|
62 |
+
dtype = torch.float16
|
63 |
+
|
64 |
+
dataset_name = "./../libripseech_tokenized"
|
65 |
+
dataset = DatasetDict.load_from_disk(dataset_name)
|
66 |
+
|
67 |
+
with open("./../prompting/blist/all_rare_words.txt") as fin:
|
68 |
+
rarewords = [process(word.strip()) for word in fin]
|
69 |
+
|
70 |
+
tokenizer = AutoTokenizer.from_pretrained("./../tokenizer")
|
71 |
+
tokenizer.pad_token_id = 0
|
72 |
+
tokenizer.pad_token = "<|padding|>"
|
73 |
+
tokenizer.padding_side = "left"
|
74 |
+
|
75 |
+
# Adding new tokens for introducing prompts
|
76 |
+
tokenizer.add_tokens(["<|startofprompt|>", "<|sepofprompt|>", "<|endofprompt|>"])
|
77 |
+
sot_token = tokenizer.encode("<|startoftranscript|>")[0]
|
78 |
+
eot_token = tokenizer.encode("<|endoftranscript|>")[0]
|
79 |
+
|
80 |
+
from math import ceil
|
81 |
+
from tqdm import tqdm
|
82 |
+
|
83 |
+
val_bs = 32
|
84 |
+
n_bwords = 25
|
85 |
+
context_length = 2048
|
86 |
+
|
87 |
+
def prepare(element):
|
88 |
+
|
89 |
+
# Add audio
|
90 |
+
audio_tkns = element["audio_tokens"]
|
91 |
+
data = "".join([f"<|audio:{tkn}|>" for tkn in audio_tkns])
|
92 |
+
|
93 |
+
# sample context words and mix with the biasing list
|
94 |
+
b_words = element["b_words"]
|
95 |
+
if n_bwords > len(b_words):
|
96 |
+
context = b_words + random.sample(rarewords, n_bwords - len(b_words))
|
97 |
+
else:
|
98 |
+
context = random.sample(b_words, n_bwords)
|
99 |
+
random.shuffle(context)
|
100 |
+
|
101 |
+
# add the context words
|
102 |
+
data += "<|startofprompt|>" + "<|sepofprompt|>".join(context) + "<|endofprompt|>"
|
103 |
+
|
104 |
+
# Add text
|
105 |
+
data += "<|startoftranscript|>"
|
106 |
+
|
107 |
+
return {"data": data, "context": context}
|
108 |
+
|
109 |
+
@torch.no_grad()
|
110 |
+
def evaluate_model(model):
|
111 |
+
|
112 |
+
transcripts = []
|
113 |
+
|
114 |
+
processed_data = dataset["test.clean"].map(prepare)
|
115 |
+
data = processed_data["data"]
|
116 |
+
|
117 |
+
for idx in tqdm(range(ceil(len(data)/val_bs))):
|
118 |
+
|
119 |
+
outputs = tokenizer(data[idx * val_bs: (idx + 1) * val_bs], truncation=False, max_length=None, padding=True, return_tensors="pt").to(model.device)
|
120 |
+
input_ids = outputs["input_ids"]
|
121 |
+
par = input_ids.shape[-1]
|
122 |
+
|
123 |
+
generations = model.generate(
|
124 |
+
input_ids,
|
125 |
+
max_new_tokens=context_length - par - 1,
|
126 |
+
eos_token_id = eot_token
|
127 |
+
)
|
128 |
+
transcripts += tokenizer.batch_decode(generations[:, par:], skip_special_tokens=True)
|
129 |
+
|
130 |
+
bias_word_cnt = 0
|
131 |
+
normal_word_cnt = 0
|
132 |
+
u_wer = 0.0
|
133 |
+
b_wer = 0.0
|
134 |
+
pred_list = correct_text(transcripts)
|
135 |
+
text_list = correct_text(processed_data["text"])
|
136 |
+
prompt_list = processed_data["context"]
|
137 |
+
for a, b, c in zip(pred_list, text_list, prompt_list):
|
138 |
+
aligned_pair = align_gt_asr(b, a)
|
139 |
+
for gt_word, asr_word in aligned_pair:
|
140 |
+
if gt_word in c or asr_word in c:
|
141 |
+
if gt_word != asr_word:
|
142 |
+
b_wer += 1.0
|
143 |
+
if gt_word in c:
|
144 |
+
bias_word_cnt += 1
|
145 |
+
else:
|
146 |
+
if gt_word != asr_word:
|
147 |
+
u_wer += 1.0
|
148 |
+
if gt_word != "":
|
149 |
+
normal_word_cnt += 1
|
150 |
+
u_wer = u_wer / normal_word_cnt * 100
|
151 |
+
b_wer = b_wer / bias_word_cnt * 100
|
152 |
+
|
153 |
+
return wer(transcripts, processed_data["text"]).item() * 100, cer(transcripts, processed_data["text"]).item() * 100, b_wer, u_wer
|