import gradio as gr
from transformers import BartForSequenceClassification, BartTokenizer


# model = pipeline("text-generation")

# following https://joeddav.github.io/blog/2020/05/29/ZSL.html
tokenizer_bart = BartTokenizer.from_pretrained('facebook/bart-large-mnli')
model_bart_sq = BartForSequenceClassification.from_pretrained('facebook/bart-large-mnli')


def zs(premise,hypothesis):
    input_ids = tokenizer_bart.encode(premise, hypothesis, return_tensors='pt')
    logits = model_bart_sq(input_ids)[0]
    entail_contradiction_logits = logits[:,[0,1,2]]
    probs = entail_contradiction_logits.softmax(dim=1)
    contra_prob = round(probs[:,0].item() * 100,2)
    neut_prob = round(probs[:,1].item() * 100,2)
    entail_prob = round(probs[:,2].item() * 100,2)
    return contra_prob, neut_prob, entail_prob

# gr.Interface(fn=zs, inputs=["text", "text"], outputs=["text","text","text"]).launch()


with gr.Blocks() as demo:
    with gr.Row():
        premise = gr.Textbox(label="Premise")
        hypothesis = gr.Textbox(label="Hypothesis")
    with gr.Row():
        greet_btn = gr.Button("Compute")
    with gr.Row():
        entailment = gr.Textbox(label="Entailment Probability")
        contradiction = gr.Textbox(label="Contradiction Probability")
        neutral = gr.Textbox(label="Neutral Probability")
        greet_btn.click(fn=zs, inputs=[premise,hypothesis], outputs=[contradiction,neutral,entailment])

demo.launch()