NLP-model / app.py
Ahmed-AI-Engineer's picture
Use native ZeroGPU Gradio API
46141e0
Raw History Blame Contribute Delete
2.67 kB
import spaces
import gradio as gr
from transformers import (
AutoTokenizer,
AutoModelForSequenceClassification,
TextClassificationPipeline,
)
MODEL_REPO = "Ahmed-AI-Engineer/smartassist-nlp-intent-model"
CONFIDENCE_THRESHOLD = 0.60
print(f"Loading model from: {MODEL_REPO}")
# -----------------------------------
# Load tokenizer
# -----------------------------------
tokenizer = AutoTokenizer.from_pretrained(
MODEL_REPO
)
# -----------------------------------
# Load model
# -----------------------------------
model = AutoModelForSequenceClassification.from_pretrained(
MODEL_REPO
)
# IMPORTANT:
# ZeroGPU supports CUDA placement during startup
# through its CUDA emulation layer.
model = model.to("cuda")
model.eval()
print("Model loaded successfully.")
# -----------------------------------
# Create pipeline
# -----------------------------------
classifier = TextClassificationPipeline(
model=model,
tokenizer=tokenizer,
device=0
)
# -----------------------------------
# Prediction function
# -----------------------------------
@spaces.GPU(duration=30)
def predict_intent(text: str):
text = text.strip()
if not text:
return {
"intent": "",
"confidence": 0.0,
"needs_confirmation": True,
"raw_text": text,
}
result = classifier(text)[0]
confidence = float(result["score"])
return {
"intent": result["label"],
"confidence": confidence,
"needs_confirmation": confidence < CONFIDENCE_THRESHOLD,
"raw_text": text,
}
# -----------------------------------
# Gradio UI
# -----------------------------------
with gr.Blocks(
title="SmartAssist NLP API"
) as demo:
gr.Markdown(
"""
# SmartAssist NLP API
Intent classification for SmartAssist home automation commands.
Supports:
- English
- Urdu
- Roman Urdu
"""
)
text_input = gr.Textbox(
label="Voice command text",
placeholder="Example: light on karo"
)
predict_button = gr.Button(
"Predict",
variant="primary"
)
prediction_output = gr.JSON(
label="Prediction"
)
predict_button.click(
fn=predict_intent,
inputs=text_input,
outputs=prediction_output,
api_name="predict"
)
text_input.submit(
fn=predict_intent,
inputs=text_input,
outputs=prediction_output,
api_name=False
)
# -----------------------------------
# Launch Space
# -----------------------------------
if __name__ == "__main__":
demo.launch()