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 definitions
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"

# -- Custom model builder
from functools import partial

def build_custom_model(sentiment_dim=0):
    class CustomModel(PreTrainedModel):
        config_class = DebertaV2Config
        def __init__(self, config, *args, **kwargs):
            super().__init__(config, *args, **kwargs)
            self.deberta = DebertaV2Model(config)
            self.pooler = ContextPooler(config)
            self.dropout = nn.Dropout(0.1)
            hidden_dim = self.pooler.output_dim + sentiment_dim
            self.classifier = nn.Linear(hidden_dim, config.num_labels)
        def forward(self, input_ids, attention_mask=None, **sent_kwargs):
            x = self.deberta(input_ids=input_ids, attention_mask=attention_mask)[0]
            pooled = self.pooler(x)
            if sentiment_dim:
                sent_feats = torch.stack((sent_kwargs['positive'], sent_kwargs['neutral'], sent_kwargs['negative']), dim=1)
                pooled = torch.cat((pooled, sent_feats), dim=1)
            return self.classifier(self.dropout(pooled))
    return CustomModel

# -- Load models and tokenizer
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

# sentiment+subjectivity
cfg1 = DebertaV2Config.from_pretrained(SENT_SUBJ_MODEL, num_labels=2, id2label={0:'OBJ',1:'SUBJ'}, label2id={'OBJ':0,'SUBJ':1})
Model1Cls = build_custom_model(sentiment_dim=3)
model1 = Model1Cls.from_pretrained(SENT_SUBJ_MODEL, config=cfg1, ignore_mismatched_sizes=True)

# subjectivity-only
cfg2 = DebertaV2Config.from_pretrained(SUBJ_ONLY_MODEL, num_labels=2, id2label={0:'OBJ',1:'SUBJ'}, label2id={'OBJ':0,'SUBJ':1})
Model2Cls = build_custom_model(sentiment_dim=0)
model2 = Model2Cls.from_pretrained(SUBJ_ONLY_MODEL, config=cfg2)

# 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):
    out = sentiment_pipe(text)[0]
    return {list(d.keys())[0]: list(d.values())[0] for d in out}

# -- Prediction logic
def analyze(text):
    # Tokenize
    inputs = tokenizer(text, truncation=True, padding=True, max_length=256, return_tensors='pt')
    # Sentiment
    scores = get_sentiment_scores(text)
    pos, neu, neg = scores['positive'], scores['neutral'], scores['negative']
    # Model1
    logits1 = model1(input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, positive=torch.tensor([pos]), neutral=torch.tensor([neu]), negative=torch.tensor([neg]))
    p1 = torch.softmax(logits1, dim=1)[0]
    # Model2
    logits2 = model2(input_ids=inputs.input_ids, attention_mask=inputs.attention_mask)
    p2 = torch.softmax(logits2, dim=1)[0]
    # Build results
    return {
        'Positive': f"{pos:.2%}", 'Neutral': f"{neu:.2%}", 'Negative': f"{neg:.2%}",
        'Sent-Subj OBJ': f"{p1[0]:.2%}", 'Sent-Subj SUBJ': f"{p1[1]:.2%}",
        'TextOnly OBJ': f"{p2[0]:.2%}", 'TextOnly SUBJ': f"{p2[1]:.2%}"
    }

# -- Build Gradio Dashboard with Blocks
dark_theme = gr.themes.Dark()

with gr.Blocks(theme=dark_theme, css="""
#result_table td { padding: 8px; font-size: 1rem; }
#header { text-align: center; font-size: 2rem; font-weight: bold; margin-bottom: 10px; }
""") as demo:
    gr.Markdown("<div id='header'>🚀 Advanced Subjectivity & Sentiment Dashboard 🚀</div>")
    with gr.Row():
        txt = gr.Textbox(label="Enter text to analyze", placeholder="Paste news sentence here...", lines=2)
        btn = gr.Button("Analyze 🔍", variant="primary")
    with gr.Tabs():
        with gr.TabItem("Overview 📊"):
            chart = gr.BarPlot(x="category", y="value", label="Results", elem_id="result_chart")
        with gr.TabItem("Raw Scores 📋"):
            table = gr.Dataframe(headers=["Metric", "Value"], datatype=["str","str"], interactive=False, elem_id="result_table")
        with gr.TabItem("About ℹ️"):
            gr.Markdown("This dashboard uses two DeBERTa-based models (with and without sentiment integration) to detect subjectivity, alongside sentiment scores from an XLM-RoBERTa model.")
            gr.Markdown("**Threshold** for subjective classification is adjustable in code (default: 0.65). Feel free to fork and customize! 🚀")
    # Link inputs to outputs
    btn.click(fn=analyze, inputs=txt, outputs=[chart, table])
    # Add confetti effect on button click
    btn.js_on_event("click", {
        "type": "confetti",
        "props": {"particleCount": 100, "spread": 60}
    })

# -- Launch
demo.queue().launch(server_name="0.0.0.0", share=True)