streaming
Browse files
app.py
CHANGED
@@ -2,13 +2,16 @@ import gradio as gr
|
|
2 |
from nltk.tokenize.treebank import TreebankWordDetokenizer
|
3 |
from somajo import SoMaJo
|
4 |
|
5 |
-
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM, AutoModelForSeq2SeqLM
|
|
|
6 |
from datasets import Dataset
|
7 |
from transformers.pipelines.pt_utils import KeyDataset
|
8 |
from hybrid_textnorm.lexicon import Lexicon
|
9 |
from hybrid_textnorm.normalization import predict_type_normalization, reranked_normalization, prior_normalization
|
10 |
from hybrid_textnorm.preprocess import recombine_tokens, german_transliterate
|
11 |
from tqdm import tqdm
|
|
|
|
|
12 |
|
13 |
text_tokenizer = SoMaJo("de_CMC", split_camel_case=True)
|
14 |
lexicon_dataset_name = 'aehrm/dtaec-lexicon'
|
@@ -20,78 +23,72 @@ def predict(input_str, model_name, progress=gr.Progress()):
|
|
20 |
tokenized_sentences = list(text_tokenizer.tokenize_text([input_str]))
|
21 |
|
22 |
if model_name == 'type normalizer':
|
23 |
-
|
24 |
elif model_name == 'type normalizer + lm':
|
25 |
-
|
26 |
elif model_name == 'transnormer':
|
27 |
-
|
28 |
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
|
34 |
def predict_transnormer(tokenized_sentences, progress):
|
35 |
model_name = 'ybracke/transnormer-19c-beta-v02'
|
36 |
|
37 |
-
progress(0, desc='loading model')
|
38 |
-
|
|
|
|
|
39 |
|
40 |
raw_sentences = []
|
41 |
for tokenized_sent in tokenized_sentences:
|
42 |
sent = ''.join(tok.text + (' ' if tok.space_after else '') for tok in tokenized_sent)
|
43 |
-
raw_sentences.append(sent)
|
44 |
|
|
|
45 |
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
return output_sentences
|
54 |
|
55 |
|
56 |
|
57 |
def predict_only_type_transformer(tokenized_sentences, progress):
|
58 |
type_model_name = 'aehrm/dtaec-type-normalizer'
|
59 |
-
progress(0, desc='loading model')
|
60 |
-
|
61 |
-
|
62 |
|
63 |
transliterated_sentences = []
|
64 |
for sentence in tokenized_sentences:
|
65 |
-
|
66 |
|
67 |
-
|
68 |
-
|
|
|
|
|
|
|
69 |
|
70 |
-
progress(0, desc='running normalization')
|
71 |
-
ds = KeyDataset(Dataset.from_dict(dict(types=list(oov_types))), "types")
|
72 |
-
for in_type, out in zip(ds, progress.tqdm(pipe(ds))):
|
73 |
-
oov_normalizations[in_type] = out[0]['generated_text']
|
74 |
-
|
75 |
-
output_sentences = []
|
76 |
-
for sent in transliterated_sentences:
|
77 |
output_sent = []
|
78 |
-
for t in
|
79 |
if t in train_lexicon.keys():
|
80 |
output_sent.append(train_lexicon[t].most_common(1)[0][0])
|
81 |
-
elif t in
|
82 |
-
output_sent.append(
|
83 |
else:
|
84 |
raise ValueError()
|
85 |
|
86 |
-
|
87 |
-
|
88 |
-
return output_sentences
|
89 |
|
90 |
def predict_type_transformer_with_lm(tokenized_sentences, progress):
|
91 |
type_model_name = 'aehrm/dtaec-type-normalizer'
|
92 |
language_model_name = 'dbmdz/german-gpt2'
|
93 |
|
94 |
-
progress(0, desc='loading model')
|
95 |
type_model_tokenizer = AutoTokenizer.from_pretrained(type_model_name)
|
96 |
type_model = AutoModelForSeq2SeqLM.from_pretrained(type_model_name)
|
97 |
language_model_tokenizer = AutoTokenizer.from_pretrained(language_model_name)
|
@@ -99,25 +96,18 @@ def predict_type_transformer_with_lm(tokenized_sentences, progress):
|
|
99 |
if 'pad_token' not in language_model_tokenizer.special_tokens_map:
|
100 |
language_model_tokenizer.add_special_tokens({'pad_token': '<pad>'})
|
101 |
|
102 |
-
transliterated_sentences = []
|
103 |
-
for sentence in tokenized_sentences:
|
104 |
-
transliterated_sentences.append([german_transliterate(tok.text) for tok in sentence])
|
105 |
-
|
106 |
-
oov_types = set(tok for sent in transliterated_sentences for tok in sent) - train_lexicon.keys()
|
107 |
oov_replacement_probabilities = {}
|
|
|
|
|
|
|
108 |
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
|
113 |
-
|
114 |
-
progress(0, desc='running LM re-ranking')
|
115 |
-
for hist_sent in progress.tqdm(transliterated_sentences):
|
116 |
-
predictions = reranked_normalization(hist_sent, train_lexicon, oov_replacement_probabilities, language_model_tokenizer, language_model, batch_size=1)
|
117 |
best_pred, _, _, _ = predictions[0]
|
118 |
-
|
119 |
-
|
120 |
-
return output_sentences
|
121 |
|
122 |
|
123 |
gradio_app = gr.Interface(
|
|
|
2 |
from nltk.tokenize.treebank import TreebankWordDetokenizer
|
3 |
from somajo import SoMaJo
|
4 |
|
5 |
+
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM, AutoModelForSeq2SeqLM, TextStreamer, TextIteratorStreamer
|
6 |
+
from threading import Thread
|
7 |
from datasets import Dataset
|
8 |
from transformers.pipelines.pt_utils import KeyDataset
|
9 |
from hybrid_textnorm.lexicon import Lexicon
|
10 |
from hybrid_textnorm.normalization import predict_type_normalization, reranked_normalization, prior_normalization
|
11 |
from hybrid_textnorm.preprocess import recombine_tokens, german_transliterate
|
12 |
from tqdm import tqdm
|
13 |
+
import re
|
14 |
+
from collections import Counter
|
15 |
|
16 |
text_tokenizer = SoMaJo("de_CMC", split_camel_case=True)
|
17 |
lexicon_dataset_name = 'aehrm/dtaec-lexicon'
|
|
|
23 |
tokenized_sentences = list(text_tokenizer.tokenize_text([input_str]))
|
24 |
|
25 |
if model_name == 'type normalizer':
|
26 |
+
stream = predict_only_type_transformer(tokenized_sentences, progress)
|
27 |
elif model_name == 'type normalizer + lm':
|
28 |
+
stream = predict_type_transformer_with_lm(tokenized_sentences, progress)
|
29 |
elif model_name == 'transnormer':
|
30 |
+
stream = predict_transnormer(tokenized_sentences, progress)
|
31 |
|
32 |
+
accumulated = ""
|
33 |
+
for out in stream:
|
34 |
+
accumulated += out
|
35 |
+
yield accumulated
|
36 |
|
37 |
def predict_transnormer(tokenized_sentences, progress):
|
38 |
model_name = 'ybracke/transnormer-19c-beta-v02'
|
39 |
|
40 |
+
#progress(0, desc='loading model')
|
41 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
42 |
+
streamer = TextIteratorStreamer(tokenizer)
|
43 |
+
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
|
44 |
|
45 |
raw_sentences = []
|
46 |
for tokenized_sent in tokenized_sentences:
|
47 |
sent = ''.join(tok.text + (' ' if tok.space_after else '') for tok in tokenized_sent)
|
|
|
48 |
|
49 |
+
inputs = tokenizer([sent], return_tensors='pt')
|
50 |
|
51 |
+
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1000, num_beams=1)
|
52 |
+
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
53 |
+
thread.start()
|
54 |
+
for new_text in streamer:
|
55 |
+
yield re.sub(r'(<pad>|</s>)', '', new_text)
|
56 |
+
yield '\n'
|
|
|
|
|
57 |
|
58 |
|
59 |
|
60 |
def predict_only_type_transformer(tokenized_sentences, progress):
|
61 |
type_model_name = 'aehrm/dtaec-type-normalizer'
|
62 |
+
#progress(0, desc='loading model')
|
63 |
+
type_model_tokenizer = AutoTokenizer.from_pretrained(type_model_name)
|
64 |
+
type_model = AutoModelForSeq2SeqLM.from_pretrained(type_model_name)
|
65 |
|
66 |
transliterated_sentences = []
|
67 |
for sentence in tokenized_sentences:
|
68 |
+
transliterated = [german_transliterate(tok.text) for tok in sentence]
|
69 |
|
70 |
+
oov_replacement_probabilities = {}
|
71 |
+
oov_types = set(transliterated) - train_lexicon.keys() - oov_replacement_probabilities.keys()
|
72 |
+
#print('oov:', oov_types)
|
73 |
+
for input_type, probas in predict_type_normalization(oov_types, type_model_tokenizer, type_model, batch_size=8):
|
74 |
+
oov_replacement_probabilities[input_type] = probas
|
75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
76 |
output_sent = []
|
77 |
+
for t in transliterated:
|
78 |
if t in train_lexicon.keys():
|
79 |
output_sent.append(train_lexicon[t].most_common(1)[0][0])
|
80 |
+
elif t in oov_replacement_probabilities.keys():
|
81 |
+
output_sent.append(Counter(dict(oov_replacement_probabilities[t])).most_common(1)[0][0])
|
82 |
else:
|
83 |
raise ValueError()
|
84 |
|
85 |
+
yield detok.detokenize(recombine_tokens(output_sent)) + '\n'
|
|
|
|
|
86 |
|
87 |
def predict_type_transformer_with_lm(tokenized_sentences, progress):
|
88 |
type_model_name = 'aehrm/dtaec-type-normalizer'
|
89 |
language_model_name = 'dbmdz/german-gpt2'
|
90 |
|
91 |
+
#progress(0, desc='loading model')
|
92 |
type_model_tokenizer = AutoTokenizer.from_pretrained(type_model_name)
|
93 |
type_model = AutoModelForSeq2SeqLM.from_pretrained(type_model_name)
|
94 |
language_model_tokenizer = AutoTokenizer.from_pretrained(language_model_name)
|
|
|
96 |
if 'pad_token' not in language_model_tokenizer.special_tokens_map:
|
97 |
language_model_tokenizer.add_special_tokens({'pad_token': '<pad>'})
|
98 |
|
|
|
|
|
|
|
|
|
|
|
99 |
oov_replacement_probabilities = {}
|
100 |
+
for sentence in tokenized_sentences:
|
101 |
+
transliterated = [german_transliterate(tok.text) for tok in sentence]
|
102 |
+
oov_types = set(transliterated) - train_lexicon.keys() - oov_replacement_probabilities.keys()
|
103 |
|
104 |
+
#print('oov:', oov_types)
|
105 |
+
for input_type, probas in predict_type_normalization(oov_types, type_model_tokenizer, type_model, batch_size=8):
|
106 |
+
oov_replacement_probabilities[input_type] = probas
|
107 |
|
108 |
+
predictions = reranked_normalization(transliterated, train_lexicon, oov_replacement_probabilities, language_model_tokenizer, language_model, batch_size=1)
|
|
|
|
|
|
|
109 |
best_pred, _, _, _ = predictions[0]
|
110 |
+
yield detok.detokenize(recombine_tokens(best_pred)) + '\n'
|
|
|
|
|
111 |
|
112 |
|
113 |
gradio_app = gr.Interface(
|