app.py

import gradio as gr
import time
from transformers import pipeline, AutoTokenizer, AutoModelForSequenceClassification

model_name = "ethanrom/a2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

pretrained_model_name = "roberta-large-mnli"
pretrained_tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name)
pretrained_model = pipeline("zero-shot-classification", model=pretrained_model_name, tokenizer=pretrained_tokenizer)
candidate_labels = ["negative", "positive", "no impact", "mixed"]

accuracy_scores = {"Fine-tuned": 0.0, "Pretrained": 0.0}
model_sizes = {"Fine-tuned": 0, "Pretrained": 0}
inference_times = {"Fine-tuned": 0.0, "Pretrained": 0.0}

def predict_sentiment(text_input, model_selection):
    global accuracy_scores, model_sizes, inference_times
    
    start_time = time.time()
    if model_selection == "Fine-tuned":
        inputs = tokenizer.encode_plus(text_input, return_tensors='pt')
        outputs = model(**inputs)
        logits = outputs.logits.detach().cpu().numpy()[0]
        predicted_class = int(logits.argmax())
        accuracy_scores[model_selection] += 1 if candidate_labels[predicted_class] == "positive" else 0
        model_sizes[model_selection] = model.num_parameters()
    else:
        result = pretrained_model(text_input, candidate_labels)
        predicted_class = result["labels"][0]
        accuracy_scores[model_selection] += 1 if predicted_class == 1 else 0
        model_sizes[model_selection] = pretrained_model.model.num_parameters()
    end_time = time.time()
    inference_times[model_selection] = end_time - start_time

    return candidate_labels[predicted_class]

def accuracy(model_selection):
    return accuracy_scores[model_selection]/10

def model_size(model_selection):
    return str(model_sizes[model_selection]//(1024*1024)) + " MB"

def inference_time(model_selection):
    return str(inference_times[model_selection]*1000) + " ms"

inputs = [
    gr.inputs.Textbox("Enter text"),
    gr.inputs.Dropdown(["Pretrained", "Fine-tuned"], label="Select model"),
]

outputs = [
    gr.outputs.Textbox(label="Predicted Sentiment"),
    gr.outputs.Label(label="Accuracy:"),
    gr.outputs.Label(label="Model Size:"),
    gr.outputs.Label(label="Inference Time:")
]

gr.Interface(fn=predict_sentiment, inputs=inputs, outputs=outputs, 
             title="Sentiment Analysis", description="Compare the output of two models",
             live=True,
             examples=[["on us lift up the light", "Fine-tuned"], ["max laid his hand upon the old man's arm", "Pretrained"]]
            ).launch();