Spaces:
Running
Running
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
import gradio as gr
|
4 |
+
import time
|
5 |
+
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
|
6 |
+
from flores200_codes import flores_codes
|
7 |
+
|
8 |
+
|
9 |
+
def load_models():
|
10 |
+
# build model and tokenizer
|
11 |
+
model_name_dict = {
|
12 |
+
#'nllb-distilled-600M': 'facebook/nllb-200-distilled-600M',
|
13 |
+
#'nllb-1.3B': 'facebook/nllb-200-1.3B',
|
14 |
+
#'nllb-distilled-1.3B': 'facebook/nllb-200-distilled-1.3B',
|
15 |
+
#'nllb-3.3B': 'facebook/nllb-200-3.3B',
|
16 |
+
'nllb-moe-54b': 'facebook/nllb-moe-54b',
|
17 |
+
}
|
18 |
+
|
19 |
+
model_dict = {}
|
20 |
+
|
21 |
+
for call_name, real_name in model_name_dict.items():
|
22 |
+
print('\tLoading model: %s' % call_name)
|
23 |
+
model = AutoModelForSeq2SeqLM.from_pretrained(real_name)
|
24 |
+
tokenizer = AutoTokenizer.from_pretrained(real_name)
|
25 |
+
model_dict[call_name+'_model'] = model
|
26 |
+
model_dict[call_name+'_tokenizer'] = tokenizer
|
27 |
+
|
28 |
+
return model_dict
|
29 |
+
|
30 |
+
|
31 |
+
def translation(source, target, text):
|
32 |
+
if len(model_dict) == 2:
|
33 |
+
model_name = 'nllb-distilled-600M'
|
34 |
+
|
35 |
+
start_time = time.time()
|
36 |
+
source = flores_codes[source]
|
37 |
+
target = flores_codes[target]
|
38 |
+
|
39 |
+
model = model_dict[model_name + '_model']
|
40 |
+
tokenizer = model_dict[model_name + '_tokenizer']
|
41 |
+
|
42 |
+
translator = pipeline('translation', model=model, tokenizer=tokenizer, src_lang=source, tgt_lang=target)
|
43 |
+
output = translator(text, max_length=400)
|
44 |
+
|
45 |
+
end_time = time.time()
|
46 |
+
|
47 |
+
output = output[0]['translation_text']
|
48 |
+
result = {'inference_time': end_time - start_time,
|
49 |
+
'source': source,
|
50 |
+
'target': target,
|
51 |
+
'result': output}
|
52 |
+
return result
|
53 |
+
|
54 |
+
|
55 |
+
if __name__ == '__main__':
|
56 |
+
print('\tinit models')
|
57 |
+
|
58 |
+
global model_dict
|
59 |
+
|
60 |
+
model_dict = load_models()
|
61 |
+
|
62 |
+
# define gradio demo
|
63 |
+
lang_codes = list(flores_codes.keys())
|
64 |
+
#inputs = [gr.inputs.Radio(['nllb-distilled-600M', 'nllb-1.3B', 'nllb-distilled-1.3B'], label='NLLB Model'),
|
65 |
+
inputs = [gr.inputs.Dropdown(lang_codes, default='English', label='Source'),
|
66 |
+
gr.inputs.Dropdown(lang_codes, default='Korean', label='Target'),
|
67 |
+
gr.inputs.Textbox(lines=5, label="Input text"),
|
68 |
+
]
|
69 |
+
|
70 |
+
outputs = gr.outputs.JSON()
|
71 |
+
|
72 |
+
title = "NLLB distilled 600M demo"
|
73 |
+
|
74 |
+
demo_status = "Demo is running on CPU"
|
75 |
+
description = f"Details: https://github.com/facebookresearch/fairseq/tree/nllb. {demo_status}"
|
76 |
+
examples = [
|
77 |
+
['English', 'Korean', 'Hi. nice to meet you']
|
78 |
+
]
|
79 |
+
|
80 |
+
gr.Interface(translation,
|
81 |
+
inputs,
|
82 |
+
outputs,
|
83 |
+
title=title,
|
84 |
+
description=description,
|
85 |
+
).launch()
|