llm/mentalBERT.py

import torch
from transformers import RobertaTokenizer, RobertaForSequenceClassification
import gradio as gr

# Load the tokenizer and models
tokenizer = RobertaTokenizer.from_pretrained("mental/mental-roberta-base")
sentiment_model = RobertaForSequenceClassification.from_pretrained("mental/mental-roberta-base")
emotion_model = RobertaForSequenceClassification.from_pretrained("j-hartmann/emotion-english-distilroberta-base")

# Define the labels
sentiment_labels = ["negative", "positive"]
emotion_labels = ["anger", "disgust", "fear", "joy", "neutral", "sadness", "surprise"]

def analyze_text(text):
    try:
        # Tokenize the input text
        inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512)
        
        # Get sentiment model outputs
        sentiment_outputs = sentiment_model(**inputs)
        sentiment_logits = sentiment_outputs.logits
        sentiment_probs = torch.nn.functional.softmax(sentiment_logits, dim=-1)
        
        # Debugging: Print logits and probs shapes
        print("Sentiment logits shape:", sentiment_logits.shape)
        print("Sentiment logits:", sentiment_logits)
        print("Sentiment probs shape:", sentiment_probs.shape)
        print("Sentiment probs:", sentiment_probs)
        
        # Get the highest probability and corresponding label for sentiment
        max_sentiment_prob, max_sentiment_index = torch.max(sentiment_probs, dim=1)
        sentiment = sentiment_labels[max_sentiment_index.item()]
        
        # Get emotion model outputs
        emotion_outputs = emotion_model(**inputs)
        emotion_logits = emotion_outputs.logits
        emotion_probs = torch.nn.functional.softmax(emotion_logits, dim=-1)
        
        # Debugging: Print logits and probs shapes
        print("Emotion logits shape:", emotion_logits.shape)
        print("Emotion logits:", emotion_logits)
        print("Emotion probs shape:", emotion_probs.shape)
        print("Emotion probs:", emotion_probs)
        
        # Get the highest probability and corresponding label for emotion
        max_emotion_prob, max_emotion_index = torch.max(emotion_probs, dim=1)
        emotion = emotion_labels[max_emotion_index.item()]
        
        return sentiment, f"{max_sentiment_prob.item():.4f}", emotion, f"{max_emotion_prob.item():.4f}"
    except Exception as e:
        print("Error:", str(e))
        return "Error", "N/A", "Error", "N/A"

# Define the Gradio interface
interface = gr.Interface(
    fn=analyze_text,
    inputs=gr.Textbox(
        lines=5, 
        placeholder="Enter text here...", 
        value="I don’t know a lot but what I do know is, we don’t start off very big and we all try to make each other smaller."
    ),
    outputs=[
        gr.Textbox(label="Detected Sentiment"),
        gr.Textbox(label="Sentiment Confidence Score"),
        gr.Textbox(label="Detected Emotion"),
        gr.Textbox(label="Emotion Confidence Score")
    ],
    title="Sentiment and Emotion Analysis: Detecting Positive/Negative Sentiment and Specific Emotions",
    description="Enter a piece of text to detect overall sentiment (positive or negative) and specific emotions (anger, disgust, fear, joy, neutral, sadness, surprise)."
)

# Launch the interface
interface.launch()