import gradio as gr
import random
import os
from huggingface_hub import login
from transformers import pipeline
from transformers import GPT2Tokenizer, GPT2LMHeadModel
login(os.environ["HF_TOKEN"])
#https://huggingface.co/facebook/opt-1.3b
#generator = pipeline('text-generation', model="microsoft/DialoGPT-medium")
tokenizer = GPT2Tokenizer.from_pretrained('microsoft/DialoGPT-medium')
original_model = GPT2LMHeadModel.from_pretrained('microsoft/DialoGPT-medium')
fine_tuned_model = GPT2LMHeadModel.from_pretrained('zmbfeng/FineTune-1')

def create_response_original(input_str, 
                             num_beams, 
                             num_return_sequences,
                             temperature,
                             repetition_penalty,
                             top_p,
                             top_k,
                             do_sample):
  print("num_beams=" + str(num_beams))
  print("num_return_sequences" + str(num_return_sequences))
  print("top_p" + str(top_p))
  print("top_k" + str(top_k))
  print("repetition_penalty" + str(repetition_penalty))
  print("temperature" + str(temperature))
  print("do_sample" + str(do_sample))

  #output_raw= generator(input_str)
  """print (output_raw)"""

  #output_str = output_raw[0]['generated_text']
  #output_str = output_str.replace("\n", "")
  #output_str = output_str.replace(input_str, "")
  #output_str = tokenizer.decode(model.generate(**tokenizer("What are John West's hobbies?"+tokenizer.eos_token,return_tensors="pt",max_length=200))[0])
  # output_str = tokenizer.decode(original_model.generate(**tokenizer(input_str+tokenizer.eos_token,return_tensors="pt",max_length=200),
  #                                                      num_beams=num_beams,
  #                                                      num_return_sequences=num_return_sequences)[0])

  input_ids = tokenizer.encode(input_str + tokenizer.eos_token, return_tensors="pt")
  output_ids = fine_tuned_model.generate(input_ids,do_sample=True, max_length=100, temperature=0.2, top_p=0.9, repetition_penalty=1.5,num_return_sequences=6)                                
  outputs = []                                 
  for output_id in output_ids:
    output = tokenizer.decode(output_id, skip_special_tokens=True)
    outputs.append(output)
  return outputs

def create_response_fine_tuned(input_str):
  #output_raw= generator(input_str)
  """print (output_raw)"""

  #output_str = output_raw[0]['generated_text']
  #output_str = output_str.replace("\n", "")
  #output_str = output_str.replace(input_str, "")
  #output_str = tokenizer.decode(model.generate(**tokenizer("What are John West's hobbies?"+tokenizer.eos_token,return_tensors="pt",max_length=200))[0])
  output_str = tokenizer.decode(fine_tuned_model.generate(**tokenizer(input_str+tokenizer.eos_token,return_tensors="pt",max_length=200))[0])
  return (output_str)
interface1 = gr.Interface(fn=create_response_original, 
    title="original",
    description="original language model, no fine tuning",
    examples=[ 
      ["What is death?"], # The first example
      ["One of the best teachers in all of life turns out to be what?"], # The second example
      ["what is your most meaningful relationship?"], # The third example
      ["What actually gives life meaning?"]
    ],
    inputs=[
    gr.Textbox(label="input text here", lines=3),
    gr.Number(label="num_beams (integer) explores the specified number of possible outputs and selects the most " +
                "likely ones (specified in num_beams)", value=7),
    gr.Number(label="num_return_sequences (integer)  the number of outputs selected from num_beams possible output",
                  value=5),
        gr.Number(
            label="temperature (decimal) controls the creativity or randomness of the output. A higher temperature" +
                  " (e.g., 0.9) results in more diverse and creative output, while a lower temperature (e.g., 0.2)" +
                  " makes the output more deterministic and focused",
            value=0.2),
        gr.Number(label="repetition_penalty (decimal) penalizes words that have already appeared in the output, " +
                        "making them less likely to be generated again. A higher repetition_penalty (e.g., 1.5) results" +
                        "in more varied and non-repetitive output.",
                  value=1.5),
        gr.Number(label="top_p (decimal) the model will only consider the words that have a high enough probability" +
                        " to reach a certain threshold",
                  value=0.9),
        gr.Number(label="top_k (integer) The number of highest probability vocabulary word will be considered" +
                        "This means that only the tokens with the highest probabilities are considered for sampling" +
                        "This reduces the diversity of the generated sequences, "+
                        "but also makes them more likely to be coherent and fluent.",
                  value=50),
        gr.Checkbox(label="do_sample. If is set to False, num_return_sequences must be 1 because the generate function will use greedy decoding, " +
                          "which means that it will select the word with the highest probability at each step. " +
                          "This results in a deterministic and fluent output, but it might also lack diversity and creativity" +
                          "If is set to True, the generate function will use stochastic sampling, which means that it will randomly" +
                          " select a word from the probability distribution at each step. This results in a more diverse and creative" +
                          " output, but it might also introduce errors and inconsistencies ", value=True)
    ], outputs="list")
interface2 = gr.Interface(fn=create_response_fine_tuned, inputs="text", outputs="text", title="Fine Tuned")
demo = gr.TabbedInterface([interface1, interface2], ["Original", "Fine Tuned"])
# with gr.Blocks() as demo:
#   with gr.Row():
#
demo.launch()