import gradio as gr
import torch
from transformers import GPT2Tokenizer, GPT2LMHeadModel, TextDataset, DataCollatorForLanguageModeling
from transformers import Trainer, TrainingArguments

__checkpoint = "gpt2"
__tokenizer = GPT2Tokenizer.from_pretrained(__checkpoint)
__model = GPT2LMHeadModel.from_pretrained(__checkpoint)
# Create a Data collator object
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False, return_tensors="pt")
def queryGPT(question):
    return generate_response(__model, __tokenizer, question)

def generate_response(__model, __tokenizer, prompt, max_length=200):

    input_ids = tokenizer.encode(prompt, return_tensors="pt")      # 'pt' for returning pytorch tensor

    # Create the attention mask and pad token id
    attention_mask = torch.ones_like(input_ids)
    pad_token_id = __tokenizer.eos_token_id

    output = __model.generate(
        input_ids,
        max_length=max_length,
        num_return_sequences=1,
        attention_mask=attention_mask,
        pad_token_id=pad_token_id
    )

    return __tokenizer.decode(output[0], skip_special_tokens=True)
    
with gr.Blocks() as demo:

    txt_input = gr.Textbox(label="Input Question", lines=2)
    txt_output = gr.Textbox(value="", label="Answer")
    btn = gr.Button(value="Submit")
    btn.click(queryGPT, inputs=[txt_input], outputs=[txt_output])
if __name__ == "__main__":
    demo.launch()