app.py

import gradio as gr
import numpy as np
import torch
from transformers import (
    AutoTokenizer,
    AutoModel,
    AutoModelForSequenceClassification
)
from scipy.special import softmax

# ==============================
# LOAD MODELS ONCE (GLOBAL)
# ==============================

bert_model_name = "bert-base-uncased"
tokenizer = AutoTokenizer.from_pretrained(bert_model_name)
bert_model = AutoModel.from_pretrained(bert_model_name)
bert_model.eval()

sentiment_model_name = "cardiffnlp/twitter-roberta-base-sentiment"
sentiment_tokenizer = AutoTokenizer.from_pretrained(sentiment_model_name)
sentiment_model = AutoModelForSequenceClassification.from_pretrained(sentiment_model_name)
sentiment_model.eval()


# ==============================
# CORE FUNCTION (NO PANDAS)
# Input: list of lists
# Output: list of lists
# ==============================

def nlp_encode_sentence(values):

    feature_rows = []

    for row in values:
        sentence = row[0]   # first column

        inputs = tokenizer(
            sentence,
            return_tensors="pt",
            truncation=True,
            padding=True
        )

        with torch.no_grad():
            outputs = bert_model(**inputs)

        cls_embedding = outputs.last_hidden_state[:, 0, :].squeeze().numpy()

        embedding_mean = float(np.mean(cls_embedding))
        embedding_median = float(np.median(cls_embedding))
        embedding_std = float(np.std(cls_embedding))
        embedding_min = float(np.min(cls_embedding))
        embedding_max = float(np.max(cls_embedding))

        sentiment_inputs = sentiment_tokenizer(
            sentence,
            return_tensors="pt",
            truncation=True,
            padding=True
        )

        with torch.no_grad():
            sentiment_outputs = sentiment_model(**sentiment_inputs)

        probs = softmax(sentiment_outputs.logits.numpy()[0])
        sentiment_score = float(probs[2] - probs[0])

        feature_rows.append([
            embedding_mean,
            embedding_median,
            embedding_std,
            embedding_min,
            embedding_max,
            sentiment_score
        ])

    return feature_rows


# ==============================
# GRADIO APP
# ==============================

with gr.Blocks() as demo:
    gr.Markdown("### NLP Encoder")

    input_data = gr.Dataframe(
        headers=["value"],
        datatype=["str"],
        type="array"
    )

    output_data = gr.Dataframe(
        headers=[
            "embedding_mean",
            "embedding_median",
            "embedding_std",
            "embedding_min",
            "embedding_max",
            "sentiment_score"
        ],
        type="array"
    )

    btn = gr.Button("Run")

    btn.click(
        fn=nlp_encode_sentence,
        inputs=input_data,
        outputs=output_data
    )

if __name__ == "__main__":
    demo.launch()