using translator
Browse files
app.py
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# app.py
|
| 2 |
+
|
| 3 |
+
import gradio as gr
|
| 4 |
+
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
|
| 5 |
+
|
| 6 |
+
# Translation models for English→English correction
|
| 7 |
+
MODEL_OPTIONS = {
|
| 8 |
+
"Helsinki-NLP/opus-mt-en-en (light, CPU-friendly)": "Helsinki-NLP/opus-mt-en-en",
|
| 9 |
+
"facebook/mbart-large-50-many-to-many-mmt (heavier)": "facebook/mbart-large-50-many-to-many-mmt"
|
| 10 |
+
}
|
| 11 |
+
|
| 12 |
+
# Cache loaded pipelines
|
| 13 |
+
loaded_pipelines = {}
|
| 14 |
+
|
| 15 |
+
def get_pipeline(model_id: str):
|
| 16 |
+
if model_id not in loaded_pipelines:
|
| 17 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 18 |
+
model = AutoModelForSeq2SeqLM.from_pretrained(
|
| 19 |
+
model_id,
|
| 20 |
+
low_cpu_mem_usage=True,
|
| 21 |
+
torch_dtype="auto"
|
| 22 |
+
)
|
| 23 |
+
pipe = pipeline("translation", model=model, tokenizer=tokenizer, device=-1)
|
| 24 |
+
# Warm-up
|
| 25 |
+
_ = pipe("This is a test.", max_length=32)
|
| 26 |
+
loaded_pipelines[model_id] = pipe
|
| 27 |
+
return loaded_pipelines[model_id]
|
| 28 |
+
|
| 29 |
+
def polish(sentence: str, model_choice: str) -> str:
|
| 30 |
+
model_id = MODEL_OPTIONS[model_choice]
|
| 31 |
+
translator = get_pipeline(model_id)
|
| 32 |
+
|
| 33 |
+
# For mbart we need to set language codes
|
| 34 |
+
if "mbart" in model_id:
|
| 35 |
+
inputs = translator.tokenizer(sentence, return_tensors="pt")
|
| 36 |
+
inputs["forced_bos_token_id"] = translator.tokenizer.lang_code_to_id["en_XX"]
|
| 37 |
+
out = translator.model.generate(**inputs, max_length=128, num_beams=4)
|
| 38 |
+
text = translator.tokenizer.decode(out[0], skip_special_tokens=True)
|
| 39 |
+
else:
|
| 40 |
+
out = translator(sentence, max_length=128)
|
| 41 |
+
text = out[0]["translation_text"]
|
| 42 |
+
|
| 43 |
+
return text.strip()
|
| 44 |
+
|
| 45 |
+
# Gradio interface
|
| 46 |
+
demo = gr.Interface(
|
| 47 |
+
fn=polish,
|
| 48 |
+
inputs=[
|
| 49 |
+
gr.Textbox(lines=2, placeholder="Enter a sentence to correct..."),
|
| 50 |
+
gr.Dropdown(choices=list(MODEL_OPTIONS.keys()),
|
| 51 |
+
value="Helsinki-NLP/opus-mt-en-en (light, CPU-friendly)",
|
| 52 |
+
label="Choose Model")
|
| 53 |
+
],
|
| 54 |
+
outputs=gr.Textbox(label="Corrected English"),
|
| 55 |
+
title="English→English Grammar Polisher",
|
| 56 |
+
description="Uses translation models (Helsinki-NLP opus-mt-en-en and facebook mbart-large-50) to rewrite English sentences into fluent, corrected English."
|
| 57 |
+
)
|
| 58 |
+
|
| 59 |
+
if __name__ == "__main__":
|
| 60 |
+
demo.launch()
|