import logging
logging.basicConfig(level='ERROR')
import numpy as np
from pathlib import Path
import openai
import torch
import zlib
import statistics
from torch.utils.data import DataLoader
from transformers import AutoTokenizer, AutoModelForCausalLM
from tqdm import tqdm
import math
import numpy as np
from datasets import load_dataset
from options import Options
from ipdb import set_trace as bp
from eval import *
from utils import evaluate_model
from analyze import analyze_data
import argparse
import os
import sys
import gc
import pickle

models = {}

def save_data(filename, data):
    with open(filename, 'wb') as filehandle:
        # store the data as binary data stream
        pickle.dump(data, filehandle)

def load_data(filename):
    with open(filename, 'rb') as filehandle:
        # read the data as binary data stream
        loaded_data = pickle.load(filehandle)
        
    return loaded_data

def load_model(name1):
    if name1 not in models:
        model1 = AutoModelForCausalLM.from_pretrained(name1, return_dict=True, device_map='auto')
        model1.eval()
        tokenizer1 = AutoTokenizer.from_pretrained(name1)
        tokenizer1.pad_token = tokenizer1.eos_token
        models[name1] = model1
        models[name1 + "_tokenizer"] = tokenizer1
    return models[name1], models[name1 + "_tokenizer"]

def calculatePerplexity(sentence, model, tokenizer, gpu):
    """
    exp(loss)
    """
    input_ids = torch.tensor(tokenizer.encode(sentence)).unsqueeze(0)
    input_ids = input_ids.to(gpu)
    with torch.no_grad():
        outputs = model(input_ids, labels=input_ids)
    loss, logits = outputs[:2]
    
    '''
    extract logits:
    '''
    # Apply softmax to the logits to get probabilities
    probabilities = torch.nn.functional.log_softmax(logits, dim=-1)
    # probabilities = torch.nn.functional.softmax(logits, dim=-1)
    all_prob = []
    input_ids_processed = input_ids[0][1:]

    for i, token_id in enumerate(input_ids_processed):
        probability = probabilities[0, i, token_id].item()
        all_prob.append(probability)
    return torch.exp(loss).item(), all_prob, loss.item()

def sample_generation(sentence, model, tokenizer, args,data_name):
    half_sentence_index = math.ceil(len(sentence.split())*args['prefix_length'])

    if half_sentence_index > 0:
        prefix = " ".join(sentence.split()[:half_sentence_index])
    else:
        prefix = '<|startoftext|> '
    
    input_ids = torch.tensor(tokenizer.encode(prefix)).unsqueeze(0)
    input_ids = input_ids.to(model.device)

    output = model.generate(input_ids, max_new_tokens=(len(sentence.split())-half_sentence_index), min_new_tokens=1, num_return_sequences=int(args['num_z']), pad_token_id=tokenizer.eos_token_id, **args['generate_args'])
    # print(output)
    complete_generated_text = tokenizer.batch_decode(output, skip_special_tokens=True)
 
    return complete_generated_text
    

def RMIA_1(text,target_loss,ref_loss,model1,tokenizer1,ratio_gen,neighbors_dl):
    target_losses_z = evaluate_model(model1,tokenizer1,neighbors_dl)
    result = torch.count_nonzero(target_losses_z < target_loss).item() / len(target_losses_z)
    return result

def get_neighbors(text,ref_loss,model2,tokenizer2,ratio_gen,data_name):
    cur_args = {'prefix_length': ratio_gen, 'num_z': 50, 'generate_args': {'do_sample': True}}
    neighbors = sample_generation(text, model2, tokenizer2, cur_args,data_name)
    neighbors_dl = DataLoader(neighbors, batch_size=32, shuffle=False)
    return neighbors_dl

def evaluate_data(test_data, col_name, target_model, ref_model, ratio_gen, data_name):
    global model1,model2,tokenizer1,tokenizer2
    print(f"all data size: {len(test_data)}")
    random.seed(0)
    random.shuffle(test_data)
    test_data = test_data[:100]
    
    inference2_pass = None
    neighbors_dls = None
    ref_model_clean = ref_model.replace("/","-")
    data_name_clean = data_name.replace("/","-")
    os.makedirs(os.path.join(f"saves/{ref_model_clean}",f"{data_name_clean}"),exist_ok=True)
    try:
        inference2_pass = load_data(f'saves/{ref_model_clean}/{data_name_clean}/inference2_pass.txt')
        neighbors_dls = load_data(f'saves/{ref_model_clean}/{data_name_clean}/neighbors_dls.txt')
    except:
        ### MODEL 2 likelihoods
        model2, tokenizer2 = load_model(ref_model)
        inference2_pass = [] #0: p_ref, #1: all_prob_ref, #2: p_ref_likelihood
        for ex in tqdm(test_data): 
            text = ex[col_name]
            new_ex = inference_model2(model2, tokenizer2, text)
            inference2_pass.append(new_ex)
        # Invariant. Doesn't take in model1 so I'm good 

        ### Neighbors:
        neighbors_dls = []
        counter = 0
        for ex in tqdm(test_data):
            text = ex[col_name]
            new_ex = get_neighbors(text,inference2_pass[counter][2],model2,tokenizer2,ratio_gen,data_name)
            counter = counter + 1
            neighbors_dls.append(new_ex)

        del models[ref_model]
        del models[ref_model + "_tokenizer"]
        model2.cpu()
        del model2
        del tokenizer2
        gc.collect()
        torch.cuda.empty_cache()

        # Because it uses temp it is not invariant, however taking a snapshot in time should be just fine.
        save_data(f'saves/{ref_model_clean}/{data_name_clean}/inference2_pass.txt',inference2_pass)
        save_data(f'saves/{ref_model_clean}/{data_name_clean}/neighbors_dls.txt',neighbors_dls)
        print("Saved ref data, exiting.")

    ### MODEL 1 likelihoods
    model1, tokenizer1 = load_model(target_model)
    inference1_pass = [] #0: p1, #1: all_prob, #2: p1_likelihood, #3: p_lower, #4: p_lower_likelihood
    for ex in tqdm(test_data):
        text = ex[col_name]
        new_ex = inference_model1(model1,tokenizer1,text)
        inference1_pass.append(new_ex)

    ### RIMA results
    counter = 0
    results = []
    for ex in tqdm(test_data):
        text = ex[col_name]
        new_ex = RMIA_1(text,inference1_pass[counter][2],inference2_pass[counter][2],model1,tokenizer1,ratio_gen,neighbors_dls[counter])
        counter = counter + 1
        results.append(new_ex)
    
    del models[target_model]
    del models[target_model + "_tokenizer"]
    model1.cpu()
    del model1
    del tokenizer1
    gc.collect()
    torch.cuda.empty_cache()
    
    ### Inference ex
    all_output = []
    counter = 0
    for ex in tqdm(test_data):
        text = ex[col_name]
        pred = {}
        pred["minkprob_w/_ref"] = results[counter]
        pred["ppl"] = inference1_pass[counter][0]
        pred["ppl/Ref_ppl (calibrate PPL to the reference model)"] = inference1_pass[counter][2]-inference2_pass[counter][2]
        pred["ppl/lowercase_ppl"] = -(np.log(inference1_pass[counter][3]) / np.log(inference1_pass[counter][0])).item()
        zlib_entropy = len(zlib.compress(bytes(text, 'utf-8')))
        pred["ppl/zlib"] = np.log(inference1_pass[counter][0])/zlib_entropy
        ex["pred"] = pred 
        counter = counter + 1
        all_output.append(ex)
    return all_output

def inference_model1 (model1, tokenizer1, text):
    p1, all_prob, p1_likelihood = calculatePerplexity(text, model1, tokenizer1, gpu=model1.device)
    p_lower, _, p_lower_likelihood = calculatePerplexity(text.lower(), model1, tokenizer1, gpu=model1.device)
    return [p1, all_prob, p1_likelihood, p_lower, p_lower_likelihood]

def inference_model2 (model2, tokenizer2, text):
    p_ref, all_prob_ref, p_ref_likelihood = calculatePerplexity(text, model2, tokenizer2, gpu=model2.device)
    return [p_ref,all_prob_ref,p_ref_likelihood]

def main(target_model,ref_model,output_dir,data,length,key_name,ratio_gen):
    output_dir = f"{output_dir}/{target_model}_{ref_model}/{key_name}"
    Path(output_dir).mkdir(parents=True, exist_ok=True)
    # load model and data
    data_name = data
    if "jsonl" in data:
        data = load_jsonl(f"{data}")
    elif data == "truthful_qa": 
        # bp()
        dataset = load_dataset(data, "multiple_choice", split="validation")
        data = convert_huggingface_data_to_list_dic(dataset)
        data = process_truthful_qa(data)
    elif data == "cais/mmlu":
        dataset = load_dataset(data, "all", split="test")
        data = convert_huggingface_data_to_list_dic(dataset)
        data = process_mmlu(data)
    elif data == "ai2_arc":
        dataset = load_dataset(data, "ARC-Challenge", split="test")
        data = convert_huggingface_data_to_list_dic(dataset)
        data = process_arc(data)
    elif data == "gsm8k": 
        dataset = load_dataset(data, "main", split="test")
        data = convert_huggingface_data_to_list_dic(dataset)
        data = process_gsm8k(data)
    elif data == "Rowan/hellaswag": 
        dataset = load_dataset(data, "default", split="validation")
        # We use validation since labels for the test set are not available?
        data = convert_huggingface_data_to_list_dic(dataset)
        data = process_hellaswag(data)
    elif data == "winogrande": 
        dataset = load_dataset(data,"winogrande_debiased", split="validation")
        data = convert_huggingface_data_to_list_dic(dataset)
        data = process_winogrande(data)

    #model1, model2, tokenizer1, tokenizer2 = load_model(target_model, ref_model)

    all_output = evaluate_data(data,key_name, target_model, ref_model,ratio_gen,data_name)
    dump_jsonl(all_output, f"{output_dir}/all_output.jsonl")
    return analyze_data(all_output)
    # fig_fpr_tpr(all_output, output_dir)