app.py

import gradio as gr
import torch
from transformers import DebertaV2Model, DebertaV2Config, AutoTokenizer, PreTrainedModel
from transformers.models.deberta.modeling_deberta import ContextPooler
from transformers import pipeline
import torch.nn as nn

# Model cards and thresholds
BASE_MODEL = "microsoft/mdeberta-v3-base"
SENT_SUBJ_MODEL = "MatteoFasulo/mdeberta-v3-base-subjectivity-sentiment-multilingual-no-arabic"
SUBJ_ONLY_MODEL = "MatteoFasulo/mdeberta-v3-base-subjectivity-multilingual-no-arabic"
THRESHOLD = 0.65

# Custom model for subjectivity (+ optional sentiment features)
class CustomModel(PreTrainedModel):
    config_class = DebertaV2Config

    def __init__(self, config, sentiment_dim=0, num_labels=2, *args, **kwargs):
        super().__init__(config, *args, **kwargs)
        self.deberta = DebertaV2Model(config)
        self.pooler = ContextPooler(config)
        output_dim = self.pooler.output_dim
        self.dropout = nn.Dropout(0.1)
        self.classifier = nn.Linear(output_dim + sentiment_dim, num_labels)

    def forward(self, input_ids, attention_mask=None, token_type_ids=None,
                positive=None, neutral=None, negative=None):
        outputs = self.deberta(input_ids=input_ids, attention_mask=attention_mask)
        pooled = self.pooler(outputs[0])
        if positive is not None and neutral is not None and negative is not None:
            sent_feats = torch.stack((positive, neutral, negative), dim=1)
            combined = torch.cat((pooled, sent_feats), dim=1)
        else:
            combined = pooled
        logits = self.classifier(self.dropout(combined))
        return logits

# Load tokenizer and model helper
def load_models():
    # Tokenizer shared
    tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
    # Sentiment+Subjectivity model
    cfg1 = DebertaV2Config.from_pretrained(
        SENT_SUBJ_MODEL,
        num_labels=2,
        id2label={0: 'OBJ', 1: 'SUBJ'},
        label2id={'OBJ': 0, 'SUBJ': 1},
        output_attentions=False,
        output_hidden_states=False
    )
    model1 = CustomModel(config=cfg1, sentiment_dim=3)
    model1 = model1.from_pretrained(SENT_SUBJ_MODEL)
    # Subjectivity-only model
    cfg2 = DebertaV2Config.from_pretrained(
        SUBJ_ONLY_MODEL,
        num_labels=2,
        id2label={0: 'OBJ', 1: 'SUBJ'},
        label2id={'OBJ': 0, 'SUBJ': 1},
        output_attentions=False,
        output_hidden_states=False
    )
    model2 = CustomModel(config=cfg2, sentiment_dim=0)
    model2 = model2.from_pretrained(SUBJ_ONLY_MODEL)
    return tokenizer, model1, model2

# Sentiment pipeline
sentiment_pipe = pipeline(
    "sentiment-analysis",
    model="cardiffnlp/twitter-xlm-roberta-base-sentiment",
    tokenizer="cardiffnlp/twitter-xlm-roberta-base-sentiment",
    top_k=None
)

def get_sentiment_scores(text: str):
    results = sentiment_pipe(text)[0]
    return {lbl: score for lbl, score in [(list(d.keys())[0], list(d.values())[0]) for d in results]}

# Prediction function
# Caches models on first call
tokenizer, model_sent_subj, model_subj_only = None, None, None

def predict_subjectivity(text):
    global tokenizer, model_sent_subj, model_subj_only
    if tokenizer is None:
        tokenizer, model_sent_subj, model_subj_only = load_models()

    # Tokenize input
    inputs = tokenizer(text, padding=True, truncation=True, max_length=256, return_tensors='pt')

    # Sentiment + subjectivity model inference
    sent_scores = get_sentiment_scores(text)
    pos, neu, neg = sent_scores['positive'], sent_scores['neutral'], sent_scores['negative']
    logits1 = model_sent_subj(
        input_ids=inputs['input_ids'],
        attention_mask=inputs.get('attention_mask'),
        positive=torch.tensor([pos]),
        neutral=torch.tensor([neu]),
        negative=torch.tensor([neg])
    )
    probs1 = torch.softmax(logits1, dim=1)[0]

    # Subjectivity-only model inference
    logits2 = model_subj_only(
        input_ids=inputs['input_ids'],
        attention_mask=inputs.get('attention_mask')
    )
    probs2 = torch.softmax(logits2, dim=1)[0]

    # Formatting
    output = []
    output.append("Sentiment Scores (sent-subj model):")
    output.append(f"- Positive: {pos:.2%}")
    output.append(f"- Neutral: {neu:.2%}")
    output.append(f"- Negative: {neg:.2%}\n")

    output.append(f"Subjectivity (with sentiment) - OBJ: {probs1[0]:.2%}, SUBJ: {probs1[1]:.2%}")
    output.append(f"Subjectivity (text only)   - OBJ: {probs2[0]:.2%}, SUBJ: {probs2[1]:.2%}")

    return "\n".join(output)

# Build Gradio interface
demo = gr.Interface(
    fn=predict_subjectivity,
    inputs=gr.Textbox(
        label='Input sentence',
        placeholder='Enter a sentence from a news article',
        info='Paste a sentence from a news article to determine subjectivity'
    ),
    outputs=gr.Textbox(
        label='Results',
        info='Sentiment & dual-model subjectivity probabilities'
    ),
    title='Dual-Model Subjectivity Detection',
    description='Outputs sentiment scores and class probabilities from two subjectivity models.'
)

demo.launch()