Spaces:
Running
Running
Added model selection feature
Browse files
app.py
CHANGED
@@ -1,6 +1,5 @@
|
|
1 |
import logging
|
2 |
import os
|
3 |
-
import tempfile
|
4 |
from pathlib import Path
|
5 |
from typing import List, Tuple
|
6 |
|
@@ -8,6 +7,7 @@ import gradio as gr
|
|
8 |
import pandas as pd
|
9 |
import spacy
|
10 |
import torch
|
|
|
11 |
from transformers import AutoModelForTokenClassification, AutoTokenizer
|
12 |
|
13 |
from preprocessing import expand_contractions
|
@@ -17,27 +17,56 @@ try:
|
|
17 |
except Exception:
|
18 |
os.system("python -m spacy download pt_core_news_sm")
|
19 |
nlp = spacy.load("pt_core_news_sm")
|
20 |
-
|
21 |
-
|
22 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
logger = logging.getLogger()
|
24 |
logger.setLevel(logging.DEBUG)
|
25 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
|
27 |
-
def predict(text,
|
28 |
-
doc =
|
29 |
-
tokens = [token.text for token in doc]
|
30 |
|
31 |
logger.info("Starting predictions for sentence: {}".format(text))
|
|
|
32 |
|
33 |
-
input_tokens = tokenizer(
|
34 |
tokens,
|
35 |
return_tensors="pt",
|
36 |
is_split_into_words=True,
|
37 |
return_offsets_mapping=True,
|
38 |
return_special_tokens_mask=True,
|
39 |
)
|
40 |
-
output = model(input_tokens["input_ids"])
|
41 |
|
42 |
i_token = 0
|
43 |
labels = []
|
@@ -49,7 +78,7 @@ def predict(text, nlp, logger=None) -> Tuple[List[str], List[str]]:
|
|
49 |
):
|
50 |
if is_special_token or off[0] > 0:
|
51 |
continue
|
52 |
-
label = model.config.__dict__["id2label"][int(pred.argmax(axis=-1))]
|
53 |
if logger is not None:
|
54 |
logger.info("{}, {}, {}".format(off, tokens[i_token], label))
|
55 |
labels.append(label)
|
@@ -63,7 +92,11 @@ def predict(text, nlp, logger=None) -> Tuple[List[str], List[str]]:
|
|
63 |
|
64 |
def text_analysis(text):
|
65 |
text = expand_contractions(text)
|
66 |
-
tokens, labels, scores = predict(text,
|
|
|
|
|
|
|
|
|
67 |
pos_count = pd.DataFrame(
|
68 |
{
|
69 |
"token": tokens,
|
@@ -99,7 +132,7 @@ def batch_analysis(input_file):
|
|
99 |
sent = expand_contractions(sent)
|
100 |
conllu_output.append("# sent_id = {}-{}\n".format(name, i + 1))
|
101 |
conllu_output.append("# text = {}\n".format(sent))
|
102 |
-
tokens, labels, scores = predict(sent,
|
103 |
for j, (token, label) in enumerate(zip(tokens, labels)):
|
104 |
conllu_output.append(
|
105 |
"{}\t{}\t_\t{}".format(j + 1, token, label) + "\t_" * 5 + "\n"
|
@@ -119,6 +152,8 @@ bottom_html = open("bottom.html").read()
|
|
119 |
|
120 |
with gr.Blocks(css=css) as demo:
|
121 |
gr.HTML(top_html)
|
|
|
|
|
122 |
with gr.Tab("Single sentence"):
|
123 |
text = gr.Textbox(placeholder="Enter your text here...", label="Input")
|
124 |
examples = gr.Examples(
|
@@ -167,4 +202,4 @@ with gr.Blocks(css=css) as demo:
|
|
167 |
gr.HTML(bottom_html)
|
168 |
|
169 |
|
170 |
-
demo.launch(debug=True
|
|
|
1 |
import logging
|
2 |
import os
|
|
|
3 |
from pathlib import Path
|
4 |
from typing import List, Tuple
|
5 |
|
|
|
7 |
import pandas as pd
|
8 |
import spacy
|
9 |
import torch
|
10 |
+
from dante_tokenizer import DanteTokenizer
|
11 |
from transformers import AutoModelForTokenClassification, AutoTokenizer
|
12 |
|
13 |
from preprocessing import expand_contractions
|
|
|
17 |
except Exception:
|
18 |
os.system("python -m spacy download pt_core_news_sm")
|
19 |
nlp = spacy.load("pt_core_news_sm")
|
20 |
+
dt_tokenizer = DanteTokenizer()
|
21 |
+
|
22 |
+
default_model = "News"
|
23 |
+
model_choices = {
|
24 |
+
"News": "Emanuel/porttagger-news-base",
|
25 |
+
"Tweets": "Emanuel/porttagger-tweets-base",
|
26 |
+
"Oil and Gas": "Emanuel/porttagger-oilgas-base",
|
27 |
+
"Multigenre": "Emanuel/porttagger-base",
|
28 |
+
}
|
29 |
+
pre_tokenizers = {
|
30 |
+
"News": nlp,
|
31 |
+
"Tweets": dt_tokenizer.tokenize,
|
32 |
+
"Oil and Gas": nlp,
|
33 |
+
"Multigenre": nlp,
|
34 |
+
}
|
35 |
logger = logging.getLogger()
|
36 |
logger.setLevel(logging.DEBUG)
|
37 |
|
38 |
+
class MyApp:
|
39 |
+
def __init__(self) -> None:
|
40 |
+
self.model = None
|
41 |
+
self.tokenizer = None
|
42 |
+
self.pre_tokenizer = None
|
43 |
+
self.load_model()
|
44 |
+
|
45 |
+
def load_model(self, model_name: str = default_model):
|
46 |
+
if model_name not in model_choices.keys():
|
47 |
+
logger.error("Selected model is not supported, resetting to the default model.")
|
48 |
+
model_name = default_model
|
49 |
+
self.model = AutoModelForTokenClassification.from_pretrained(model_choices[model_name])
|
50 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_choices[model_name])
|
51 |
+
self.pre_tokenizer = pre_tokenizers[model_name]
|
52 |
+
|
53 |
+
myapp = MyApp()
|
54 |
|
55 |
+
def predict(text, logger=None) -> Tuple[List[str], List[str]]:
|
56 |
+
doc = myapp.pre_tokenizer(text)
|
57 |
+
tokens = [token.text if not isinstance(token, str) else token for token in doc]
|
58 |
|
59 |
logger.info("Starting predictions for sentence: {}".format(text))
|
60 |
+
print("Using model {}".format(myapp.model.config.__dict__["_name_or_path"]))
|
61 |
|
62 |
+
input_tokens = myapp.tokenizer(
|
63 |
tokens,
|
64 |
return_tensors="pt",
|
65 |
is_split_into_words=True,
|
66 |
return_offsets_mapping=True,
|
67 |
return_special_tokens_mask=True,
|
68 |
)
|
69 |
+
output = myapp.model(input_tokens["input_ids"])
|
70 |
|
71 |
i_token = 0
|
72 |
labels = []
|
|
|
78 |
):
|
79 |
if is_special_token or off[0] > 0:
|
80 |
continue
|
81 |
+
label = myapp.model.config.__dict__["id2label"][int(pred.argmax(axis=-1))]
|
82 |
if logger is not None:
|
83 |
logger.info("{}, {}, {}".format(off, tokens[i_token], label))
|
84 |
labels.append(label)
|
|
|
92 |
|
93 |
def text_analysis(text):
|
94 |
text = expand_contractions(text)
|
95 |
+
tokens, labels, scores = predict(text, logger)
|
96 |
+
if len(labels) != len(tokens):
|
97 |
+
m = len(tokens) - len(labels)
|
98 |
+
labels += [None] * m
|
99 |
+
scores += [0] * m
|
100 |
pos_count = pd.DataFrame(
|
101 |
{
|
102 |
"token": tokens,
|
|
|
132 |
sent = expand_contractions(sent)
|
133 |
conllu_output.append("# sent_id = {}-{}\n".format(name, i + 1))
|
134 |
conllu_output.append("# text = {}\n".format(sent))
|
135 |
+
tokens, labels, scores = predict(sent, logger)
|
136 |
for j, (token, label) in enumerate(zip(tokens, labels)):
|
137 |
conllu_output.append(
|
138 |
"{}\t{}\t_\t{}".format(j + 1, token, label) + "\t_" * 5 + "\n"
|
|
|
152 |
|
153 |
with gr.Blocks(css=css) as demo:
|
154 |
gr.HTML(top_html)
|
155 |
+
select_model = gr.Dropdown(choices=list(model_choices.keys()), label="Tagger model", value=default_model)
|
156 |
+
select_model.change(myapp.load_model, inputs=[select_model])
|
157 |
with gr.Tab("Single sentence"):
|
158 |
text = gr.Textbox(placeholder="Enter your text here...", label="Input")
|
159 |
examples = gr.Examples(
|
|
|
202 |
gr.HTML(bottom_html)
|
203 |
|
204 |
|
205 |
+
demo.launch(debug=True)
|