import gradio as gr
import random
import os
import copy
import torch
from huggingface_hub import login
from transformers import pipeline
from transformers import GPT2Tokenizer, GPT2LMHeadModel,set_seed
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')
untethered_model = GPT2LMHeadModel.from_pretrained('zmbfeng/untethered_20240225_epochs_500')
untethered_paraphrased_model = GPT2LMHeadModel.from_pretrained('zmbfeng/untethered_20240227_epochs_350')
default_num_return_sequences=5
default_temperature=0.5
default_repetition_penalty=1.5
default_top_p=2
default_top_k=50
default_do_sample=True
default_seed=42
def create_response(input_str, 
                             # num_beams, 
                             num_return_sequences,
                             temperature,
                             repetition_penalty,
                             top_p,
                             top_k,
                             do_sample,
                             seed,   
                             model_name):
  print("input_str="+input_str)
  print("model_name="+str(model_name))
  # num_beams = int(num_beams)
  # print("num_beams=" + str(num_beams))
  num_return_sequences=int(num_return_sequences)                               
  print("num_return_sequences" + str(num_return_sequences))
  print("top_p" + str(top_p))
  top_k=int(top_k)
  print("top_k" + str(top_k))
  print("repetition_penalty" + str(repetition_penalty))
  print("temperature" + str(temperature))
  print("do_sample" + str(do_sample))
  if not do_sample:
      num_beams = 1
      print("num_beams=" + str(num_beams))
  seed=int(seed)
  print("seed" + str(seed))     
  encoded = tokenizer.encode_plus(input_str + tokenizer.eos_token, return_tensors="pt")
  input_ids = encoded["input_ids"]
  #attention_mask = encoded["attention_mask"]

  if seed != -1:
    set_seed(seed)    
  if model_name == "original_model":      
    output = original_model.generate(input_ids,do_sample=do_sample, max_length=100, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty,num_return_sequences=num_return_sequences,return_dict_in_generate=True, output_scores=True )
    transition_scores = original_model.compute_transition_scores(output.sequences, output.scores,
                                                                 normalize_logits=False)

  elif model_name == "untethered_model":
    output = untethered_model.generate(input_ids, do_sample=do_sample, max_length=100, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty,num_return_sequences=num_return_sequences,return_dict_in_generate=True, output_scores=True )
    transition_scores = untethered_model.compute_transition_scores(output.sequences, output.scores,
                                                               normalize_logits=False)
  elif model_name == "untethered_paraphrased_model":
    output = untethered_paraphrased_model.generate(input_ids, do_sample=do_sample,  max_length=100, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty,num_return_sequences=num_return_sequences,return_dict_in_generate=True, output_scores=True )
    transition_scores = untethered_paraphrased_model.compute_transition_scores(output.sequences, output.scores,
                                                               normalize_logits=False)
  else:
    output = original_model.generate(input_ids,do_sample=do_sample, max_length=100, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty,num_return_sequences=num_return_sequences,return_dict_in_generate=True, output_scores=True )
    transition_scores = original_model.compute_transition_scores(output.sequences, output.scores,
                                                                 normalize_logits=False)
  score_list = []
  for scores in transition_scores:
      # exp_tensor_list = torch.exp(scores)
      # print(scores)
      # print(exp_tensor_list)
      sum_exp_tensor_list = torch.sum(scores)
      len_exp_tensor_list = torch.numel(scores)
      average_exp_tensor_list = sum_exp_tensor_list / len_exp_tensor_list
      print(average_exp_tensor_list)
      score_list.append(average_exp_tensor_list.item())
  outputs = model_name+" generated <br>"
  for index, output_id in enumerate(output.sequences):
      sentence = tokenizer.decode(output_id, skip_special_tokens=True)
      #print(sentence + " score:" + str(score_list[index]))
      outputs = outputs + sentence +  " score:" + str(score_list[index]) +"<br/>"
  # for output_id in output_ids:
  #   output = tokenizer.decode(output_id, skip_special_tokens=True)
  #   outputs=outputs+output+"<br/>"
  return outputs


common_examples_string="<br/>Sample Inputs:<br/>What is death?<br/>One of the best teachers in all of life turns out to be what?<br/>what is your most meaningful relationship?\nWhat actually gives life meaning?<br/>"
common_examples=[ 
      ["What is death?",default_num_return_sequences,default_temperature,default_repetition_penalty,default_top_p,default_top_k,default_do_sample], # The first example
      ["One of the best teachers in all of life turns out to be what?",default_num_return_sequences,default_temperature,default_repetition_penalty,default_top_p,default_top_k,default_do_sample], # The second example
      ["what is your most meaningful relationship?",default_num_return_sequences,default_temperature,default_repetition_penalty,default_top_p,default_top_k,default_do_sample], # The third example
      ["What actually gives life meaning?",default_num_return_sequences,default_temperature,default_repetition_penalty,default_top_p,default_top_k,default_do_sample]
    ]
examples = copy.deepcopy(common_examples)
print(examples)
for example in examples:
    example.append("original_model")
print(examples)   
interface_original = gr.Interface(fn=create_response, 
    title="original",
    description="original language model, no fine tuning"+common_examples_string,
    #examples=examples,
    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=default_num_return_sequences),
        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=default_temperature),
        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=default_repetition_penalty),
        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=default_top_p),
        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=default_top_k),
        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=default_do_sample),
        gr.Number(
            label="seed (integer) random seed, set to -1 to use a random seed everytime",
                value=default_seed),
        gr.Textbox(label="model", lines=3, value="original_model",visible=False)
    ],
    outputs="html"
    )
examples = copy.deepcopy(common_examples)
print(examples)
for example in examples:
    example.append("untethered_model")
print(examples)


interface_untethered_model = gr.Interface(fn=create_response, 
    title="untethered model",
    description="language model fine tuned with'The Untethered Soul' chapter 17"+common_examples_string,
    #examples=examples,
    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=default_num_return_sequences),
        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=default_temperature),
        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=default_repetition_penalty),
        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=default_top_p),
        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=default_top_k),
        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=default_do_sample),
        gr.Number(
            label="seed (integer) random seed, set to -1 to use a random seed everytime",
                value=default_seed),
        gr.Textbox(label="model", lines=3, value="untethered_model",visible=False)
    ],
    outputs="html"
    )

examples = copy.deepcopy(common_examples)
print(examples)
for example in examples:
    example.append("untethered_paraphrased_model")
print(examples)
interface_untethered_paraphrased_model = gr.Interface(fn=create_response, 
    title="untethered paraphrased_model",
    description="language model fine tuned with'The Untethered Soul' chapter 17 paraphrased"+common_examples_string,
    #examples=examples,
    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=default_num_return_sequences),
        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=default_temperature),
        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=default_repetition_penalty),
        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=default_top_p),
        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=default_top_k),
        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=default_do_sample),
        gr.Number(
            label="seed (integer) random seed, set to -1 to use a random seed everytime",
                value=default_seed),
        gr.Textbox(label="model", lines=3, value="untethered_paraphrased_model",visible=False)
    ],
    outputs= "html"
    )



demo = gr.TabbedInterface([interface_original, interface_untethered_model, interface_untethered_paraphrased_model], ["Original", "Untethered", "Untethered paraphrased"])

demo.launch()