import gradio as gr

import nltk
import string
from transformers import GPT2LMHeadModel, GPT2Tokenizer, GenerationConfig, set_seed
import random

nltk.download('punkt')

response_length = 200

sentence_detector = nltk.data.load('tokenizers/punkt/english.pickle')

tokenizer = GPT2Tokenizer.from_pretrained("gpt2-medium")
tokenizer.truncation_side = 'right'

model = GPT2LMHeadModel.from_pretrained('checkpoint-10000')
generation_config = GenerationConfig.from_pretrained('gpt2-medium')
generation_config.max_new_tokens = response_length
generation_config.pad_token_id = generation_config.eos_token_id


outputs = []


def generate_response(new_prompt):
    print('a')
    global outputs
    print('b')
    story_so_far = "\n".join(outputs[:int(1024 / response_length + 1)])
    print('c')
    set_seed(random.randint(0, 4000000000))
    inputs = tokenizer.encode(story_so_far + '\n' + new_prompt if story_so_far else new_prompt,
                              return_tensors='pt', truncation=True,
                              max_length=1024 - response_length)
    print('d')
    output = model.generate(inputs, do_sample=True, generation_config=generation_config)
    print('e')
    response = clean_paragraph(tokenizer.batch_decode(output)[0][((len(story_so_far) + 1) if story_so_far else 0):])
    print('f')
    outputs.append(response)
    print('g')
    return ((story_so_far + '\n' if story_so_far else '') + response).replace('\n', '\n\n')

def undo():
    global outputs
    print(outputs)
    outputs = outputs[:-1]
    print(outputs)
    return "\n".join(outputs).replace('\n', '\n\n')

def clean_paragraph(entry):
    paragraphs = entry.split('\n')

    for i in range(len(paragraphs)):
        split_sentences = nltk.tokenize.sent_tokenize(paragraphs[i], language='english')
        if i == len(paragraphs) - 1 and split_sentences[:1][-1] not in string.punctuation:
            paragraphs[i] = " ".join(split_sentences[:-1])

    return capitalize_first_char("\n".join(paragraphs))

def reset():
    global outputs
    outputs = []
    return None

def capitalize_first_char(entry):
    for i in range(len(entry)):
        if entry[i].isalpha():
            return entry[:i] + entry[i].upper() + entry[i + 1:]
    return entry

with gr.Blocks() as demo:
    story = gr.Textbox(interactive=False, lines=20)
    story.style(show_copy_button=True)

    prompt = gr.Textbox(placeholder="Continue the story here!", lines=3, max_lines=3)

    with gr.Row():
        gen_button = gr.Button('Generate')
        undo_button = gr.Button("Undo")
        res_button = gr.Button("Reset")

    prompt.submit(generate_response, prompt, story, scroll_to_output=True)
    gen_button.click(generate_response, prompt, story, scroll_to_output=True)
    undo_button.click(undo, [], story, scroll_to_output=True)
    res_button.click(reset, [], story, scroll_to_output=True)

demo.launch(inbrowser=True)