-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathcsasr_st.py
186 lines (156 loc) · 8.07 KB
/
csasr_st.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
import re, random
import torch
from collections import Counter
import csv, os
import pandas as pd
from tqdm import tqdm
import numpy as np
import regex
from config import MyParser
import whisper
punc = [",", "?", "¿", ".", "!", "¡", ";", "；", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞","؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。", "、", "﹂", "﹁", "‧", "～", "﹏", "，", "｛", "｝", "（", "）", "［", "］", "【", "】", "‥", "〽", "『", "』", "〝", "〟", "⟨", "⟩", "〜", "：", "！", "？", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"]
if __name__ == "__main__":
torch.cuda.empty_cache()
args = MyParser().parse_args()
print(args)
# seed everything
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# both dataset and post-processing are dataset specific, so all done in ${dataset}.py
if args.dataset == "ascend":
from data.ascend import get_dataloader, calc_metrics
elif args.dataset == "seame":
from data.seame import get_dataloader, calc_metrics
elif args.dataset == "covost2":
from data.covost2 import get_dataloader
if args.language == "zh":
from data.covost2 import calc_metrics_zh as calc_metrics
else:
from data.covost2 import calc_metrics_ar as calc_metrics
elif args.dataset == "libritrans":
from data.libritrans import get_dataloader, calc_metrics
elif args.dataset == "mustcv1":
from data.mustcv1 import get_dataloader, calc_metrics
###################################
tokenizer, data_loader = get_dataloader(args)
model = whisper.load_model(args.model)
model.eval()
model.cuda()
if args.logit_mask != "0":
def construct(lang, path):
local_tokenizer = whisper.tokenizer.get_tokenizer(multilingual=True, language=lang, task="transcribe")
counter = Counter()
if args.dataset == "covost2":
data = pd.read_csv(path, sep="\t", header=0, encoding="utf-8", escapechar="\\", quoting=csv.QUOTE_NONE, na_filter=False)
elif args.dataset == "libritrans" or args.dataset == "mustcv1":
with open(path, "r") as ff:
all_trans = [l for l in ff.readlines()]
data = {'translation': all_trans}
for text in data['translation']:
tokens = local_tokenizer.encode(text.strip())
counter.update(tokens)
del data
return counter
if args.language == "zh":
lang_in = "zh-CN"
else:
lang_in = args.language
if args.dataset == "covost2":
path=f"{args.dataset_dir}/metadata/covost_v2.en_{lang_in}.train.tsv"
elif args.dataset == "libritrans":
path=f"{args.dataset_dir}/train/train.fr"
elif args.dataset == "mustcv1":
path = f"{args.dataset_dir}/en-{lang_in}/data/train/txt/train.{lang_in}"
if not os.path.isfile(path):
path = path.replace("/data/scratch/", "/data3/scratch/") # handle rtx path
if not os.path.isfile(path):
path = path.replace("/data3/scratch/", "/scratch/cluster/")
# construct vocab
counter = construct(args.language, path)
# only allow the most frequent tokens
n_vocab = model.dims.n_vocab
cap_p = getattr(args, "vocab_cap", 0.7)
cap_n = round(len(counter)*cap_p)
constraint_ind = [item[0] for item in counter.most_common(cap_n)]
special_inds = list(tokenizer.tokenizer.get_added_vocab().values())
constraint_ind += special_inds # add the indices of the special tokens
# redo constraint for zh ar and ru as we can constrain the output script
if args.language == "zh" or args.language == "ar" or args.language == "ru":
lang2range = {"zh": r"[\u4e00-\ufaff]", "ar": r"[\u0600-\u06ff]"}
constraint_ind = []
for i in range(n_vocab):
decoding_res = tokenizer.decode(i)
if args.language == "ru":
constraint = regex.findall(r'\p{Cyrillic}+', decoding_res)
else:
constraint_reg_range = lang2range[args.language]
constraint = re.findall(constraint_reg_range, decoding_res, re.UNICODE)
if len(decoding_res) > 0 and len(constraint) > 0:
constraint_ind.append(i)
constraint_ind += list(tokenizer.tokenizer.get_added_vocab().values()) # add the indices of the special tokens
# # control whether outputting punctuations
punc2ind = {}
for p in punc:
punc2ind[p] = tokenizer.encode(p)
pind = np.unique(list(punc2ind.values())).tolist()
for p in pind:
constraint_ind += p
constraint_ind = np.unique(constraint_ind).tolist()
logit_mask = torch.ones((1, n_vocab)) * -1000000.
logit_mask[:, constraint_ind] = 0.0
print(f"allowed vocab: {args.language} scripts")
print(f"total vocab size: {n_vocab}, allowed vocab size: {len(constraint_ind)}")
else:
logit_mask = None
refs = []
preds = []
single_preds = []
prompts = []
for i, b in enumerate(tqdm(data_loader)):
input_mels = b["input_mels"].half().cuda()
raw_texts = b['raw_text']
with torch.no_grad():
# for input_mel, label in zip(input_mels, labels):
for input_mel, raw_text in zip(input_mels, raw_texts):
if args.code_switching != "0":
main_lang, second_lang = args.code_switching.split("-")
_, probs = whisper.detect_language(model, input_mel)
max_lang = max(probs, key=probs.get)
prob = probs[max_lang]
if max_lang == main_lang:
lang = main_lang
elif max_lang == second_lang:
lang = second_lang
else: # Whisper language identification is not working well, assigning main_lang as the language
lang = main_lang
options = whisper.DecodingOptions(task=args.task, language=lang, without_timestamps=True, beam_size=args.beam_size, block_ngrams=args.block_ngrams, concat_lang_token=args.code_switching if (args.concat_lang_token != 0 and prob < args.single_lang_threshold) else "0", logit_mask=logit_mask)
else:
options = whisper.DecodingOptions(task=args.task, language=args.language, without_timestamps=True, beam_size=args.beam_size, block_ngrams=args.block_ngrams, concat_lang_token="0", logit_mask=logit_mask)
with torch.no_grad():
results = whisper.decode(model, input_mel, options)
preds.append(results.text)
ref = raw_text
refs.append(ref)
inference_metrics, (wer_list, processed_preds, processed_refs) = calc_metrics()(refs, preds)
print("results:", inference_metrics)
print("results:", inference_metrics)
# in the case of speech translation, the metric is actually BLUE score
if args.topk > 0:
import numpy as np
inds = np.argsort(wer_list)[::-1]
for ind in inds[:args.topk]:
print("-"*10)
print("wer/mer: ", wer_list[ind])
print("ref: ", processed_refs[ind])
print("pred: ", processed_preds[ind])
# print("prompt: ", prompts[ind])
else:
for j, (k, v) in enumerate(zip(processed_refs, processed_preds)):
if j % 100 == 0:
print("-"*10)
print("ref: ", k)
print("pred: ", v)
print("results:", inference_metrics)
print("results:", inference_metrics)