code/main.py

import os
import sys
import argparse
import logging
import time
import json

#os.environ["CUDA_VISIBLE_DEVICES"] = "1"

from peft import PeftModel,LoraConfig, get_peft_model, prepare_model_for_kbit_training
from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments, BitsAndBytesConfig
from datasets import load_dataset,Dataset
import torch

from utils.warp import Warp,WarpLJP
from utils.dataset import DataCollatorForReward
from utils.trainer import PRGTrainer
from tree.base import Tree,Node,I_policy
from configs.hyperparametric import Reward_config,Tree_config
from model.logitsprocessor import OutputControlLogitsProcessor,RewardControlLogitsProcessor
from tree.asts import AST
from utils.model_generate import generate_string,generate_score

tree_config = Tree_config().to_dict()
#reward_config = Reward_config().to_dict()

import torch
#torch.cuda.set_device(0)


#TASKS = ['ecthr_a','ecthr_b']
TASKS = ['ljp',]


def get_args():
    parser = argparse.ArgumentParser()

    ## ___datasets___
    #parser.add_argument('--data_path',default='lex_glue',type=str, help='Path containing dataset')
    parser.add_argument('--train_path',default='',type=str, help='Path containing dataset')
    parser.add_argument('--eval_path',default='',type=str, help='Path containing dataset')
    parser.add_argument('--test_path',default='',type=str, help='Path containing dataset')
    parser.add_argument('--dataset',default='ljp',type=str, help='Dataset of choice in data_path')
    parser.add_argument('--save_data_path',default='',type=str, help='The path used to save dataset')
    parser.add_argument('--output_path',default='',type=str, help='The path used to save outputs')
    parser.add_argument('--sample_path',default='',type=str, help='The path used to samples')
    parser.add_argument('--control_file',default='./codekey_proofread.txt',type=str, help='The path used to output control')

    ## ___model___
    parser.add_argument('--generate_model_path',default='',type=str, help='Path containing model')
    parser.add_argument('--reward_model_path',default='',type=str, help='Path containing model')
    parser.add_argument('--reward_save_path',default='./output/reward',type=str, help='Path containing model')
    parser.add_argument('--reward_lora_path',default='',type=str,)
    parser.add_argument('--per_device_train_batch_size',default=2,type=int)
    parser.add_argument('--gradient_accumulation_steps',default=2,type=int)
    parser.add_argument('--learning_rate',default=1e-3,type=float)
    parser.add_argument('--num_train_epochs',default=10,type=int)
    parser.add_argument('--logging_steps',default=200,type=int)
    parser.add_argument('--save_strategy',default='epoch',type=str,)
    parser.add_argument('--fp16',action='store_true',default=True,)
    parser.add_argument('--optim',default='paged_adamw_8bit',type=str,)
    
    parser.add_argument('--lora_rank',default=64,type=int)
    parser.add_argument('--lora_alpha',default=16,type=int)
    parser.add_argument('--lora_dropout',default=0.1,type=float)

    ## ___pipline___
    parser.add_argument('--do_train',action='store_true',default=False, help='Training or not')
    parser.add_argument('--do_test',action='store_true',default=True, help='Eval or not')

    ## ___parameter___
    parser.add_argument('--budget',default=20,type=int, help='iterations of search')
    parser.add_argument('--reward_funcation',default='leaf',type=str,choices=['random','reward','leaf'], help='iterations of search')
    parser.add_argument('--iteration',default=3,type=int, help='iterations of sample')

    ## ___special___
    parser.add_argument('--ljp_mode',default='p',type=str,choices=['p','pd','pdf'])
    parser.add_argument('--logits_control',action='store_true',default=False, help='Training or not')
    parser.add_argument('--add_reward',action='store_true',default=False,)
    parser.add_argument('--inference_mode',default='zeroshot',type=str,choices=['zeroshot','fewshot','cot'])



    return parser.parse_args()

def get_logger(path='./'):
    log_path = os.path.join(path,"log_%s.txt"%(time.strftime("%Y-%m-%d-%H-%M-%S",time.localtime())))
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        handlers=[logging.StreamHandler(sys.stdout),
                  logging.FileHandler(log_path)],
    )
    logger = logging.getLogger(__name__)
    logger.setLevel(level=logging.DEBUG)
    return logger

def load_data(args):
    if os.path.isdir(args.train_path):
        save_path = os.path.join(args.save_data_path,args.dataset)
        if not os.path.exists(save_path):
            dataset = load_dataset(path=args.train_path,name=args.dataset)
            dataset.save_to_disk(save_path)
        else:
            dataset = load_dataset(save_path)
    if os.path.isfile(args.train_path):
        data_files = {mode:path for mode,path
                      in zip(['train','validation','test'],[args.train_path,args.eval_path,args.test_path])
                      if path}
        dataset = load_dataset('json',data_files=data_files)
    return dataset

def load_samples(sample_path,):
    path_list = os.listdir(sample_path)
    samples = []
    for path in path_list:
        path = os.path.join(sample_path,path)
        with open(path,'r') as f:
            for l in f.readlines():
                sample = json.loads(l)
                samples.append(sample)
    return samples



def train(args,warp,dataset,):
    # init TrainingArgument
    training_args = TrainingArguments(
        output_dir=os.path.join(args.reward_save_path,'reward_%s'%(time.strftime("%Y-%m-%d-%H-%M-%S",time.localtime()))),
        per_device_train_batch_size=args.per_device_train_batch_size,
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        learning_rate=args.learning_rate,
        num_train_epochs=args.num_train_epochs,
        logging_steps=args.logging_steps,
        save_strategy=args.save_strategy,
        fp16=args.fp16,
        optim=args.optim,
        remove_unused_columns=False
        )
    
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.float16,
        bnb_4bit_use_double_quant=True,
        bnb_4bit_quant_type="nf4",
        llm_int8_threshold=6.0,
        llm_int8_has_fp16_weight=False, 
        )

    peft_config = LoraConfig(
        r=args.lora_rank,
        lora_alpha=args.lora_alpha,
        lora_dropout=args.lora_dropout,
        target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
        bias="none",
        task_type="CAUSAL_LM"
        )

    # init model
    if not warp.reward_model:
        warp.load_reward_model(bnb_config=bnb_config)

    model = warp.reward_model
    tokenizer = warp.reward_tokenizer

    model = prepare_model_for_kbit_training(model)
    model = get_peft_model(model, peft_config)

    # init collator
    collator = DataCollatorForReward(tokenizer=tokenizer)
    logits_processor = RewardControlLogitsProcessor(tokenizer=tokenizer)

    # init dataset
    trainset = dataset['0']

    # init trainer
    trainer = PRGTrainer(
            tokenizer=tokenizer,
            model=model,
            args=training_args,
            train_dataset=trainset,
            data_collator=collator,
            logits_processor=logits_processor
        )
    
    # clean memory 
    warp.generate_model = None
    torch.cuda.empty_cache()

    # training
    logger.info('start training..')
    for i,trainset in dataset.items():
        if i != '0':
            trainer.train_dataset = trainset
        
        trainer.train()

    logger.info('training end..')
    warp.reward_model = model
    warp.reward_tokenizer = tokenizer

def evaluate(args,warp,dataset,):
    if warp.generate_model == None:
        warp.load_generate_model()
    if args.add_reward:
        if args.reward_lora_path != '':
            bnb_config = BitsAndBytesConfig(load_in_4bit=True,
                                            bnb_4bit_compute_dtype=torch.float16,
                                            bnb_4bit_use_double_quant=True,
                                            bnb_4bit_quant_type="nf4",
                                            llm_int8_threshold=6.0,
                                            llm_int8_has_fp16_weight=False,
                                            )
            warp.load_reward_model(bnb_config=bnb_config)
            warp.reward_model = PeftModel.from_pretrained(warp.reward_model, args.reward_lora_path) 
        else:
            warp.load_reward_model()
        rewarder = {'model':warp.reward_model,'tokenizer':warp.reward_tokenizer,}
        if args.logits_control:
            rewarder['reward_processor'] = RewardControlLogitsProcessor(tokenizer=rewarder['tokenizer'])
    else:
        rewarder = {}

    model = warp.generate_model
    tokenizer = warp.generate_tokenizer
    
    def get_response(x,a,tokenizer,model,rewarder={}):
        if rewarder == {}:
            inputs = warp.prompt_to_crime(x,a,bos=tokenizer.bos_token,eos=tokenizer.eos_token)
            if args.logits_control:
                outputs = generate_string(inputs,tokenizer=tokenizer,model=model,logits_processor=warp.logits_processor)
            else:
                outputs = generate_string(inputs,tokenizer=tokenizer,model=model,)

            response = warp.step_from_response(outputs)
            return response
        if 'reward_processor' not in rewarder.keys():
            rewarder['reward_processor'] = None
        
        for i in range(tree_config['branch']):
            inputs = warp.prompt_to_crime(x,a,bos=tokenizer.bos_token,eos=tokenizer.eos_token)
            if args.logits_control:
                outputs = generate_string(inputs,tokenizer=tokenizer,model=model,logits_processor=warp.logits_processor)
            else:
                outputs = generate_string(inputs,tokenizer=tokenizer,model=model,)
            response = warp.step_from_response(outputs)
            
            thought = warp.prompt_to_value(x,a+response,bos=rewarder['tokenizer'].bos_token,eos=rewarder['tokenizer'].eos_token)
            reward = generate_score(thought,
                                    tokenizer=rewarder['tokenizer'],
                                    model=rewarder['model'],
                                    reward_processor=rewarder['reward_processor']
                                    )
            r = warp.value_from_response(reward)
            if '拒绝' in r:
                continue
            else:
                break
        return response
    
    logger.info('start eval..')
    time_start = time.time()
    reward_control = 'rewardcontrol' if args.add_reward else 'un-reward'
    task_name = args.test_path.split('/')[-2]
    save_path = os.path.join(args.output_path,'eval',"%s_%s_%s_%s.json"%(task_name,args.inference_mode,reward_control,time.strftime("%Y-%m-%d-%H-%M-%S",time.localtime())))
    
    preds = []
    for i,data in enumerate(dataset):
        x,a,y = warp.processing_single(data)
        if args.inference_mode == 'fewshot':
            x = '这是一个例子：根据案情描述和已有步骤仅给出一个推理。如果是结论则直接输出<e></e>，例如<e>盗窃罪</e>。如果是步骤则直接输出<p></p>，例如<p>步骤1：…</p>\n案情描述:2013年下半年至2015年10月26日，被告人张和菊利用担任山东泰开电力建设工程有限公司、山东泰开国际工程技术有限公司现金出纳的职务便利，多次将公司的资金共计4472572.91元挪出，用于其在深圳石油化工交易所、天津渤海商品交易所的投资交易，已全部亏损。2015年10月26日，张和菊从公司提取现金26万元后，携款潜逃至济南市长清区租房处藏匿。2015年11月6日，张和菊被公安机关抓获。\n已有推理步骤：\n<e>挪用资金罪</e>\n这是问题:\n根据案情描述和已有步骤仅给出一个推理。如果是结论则直接输出<e></e>，例如<e>盗窃罪</e>。如果是步骤则直接输出<p></p>，例如<p>步骤1：…</p>\n案情描述:'+x
        elif args.inference_mode == 'cot':
            x = '一步步思考并回答，' + x
        y = y['crime']
        a = ''
        for d in range(tree_config['max_depth']):
            try:
                response = get_response(x,a,tokenizer,model,rewarder=rewarder)
            except Exception as E:
                response = ''     
            if '<e>' in response:
                break
            if '<p>' in response:
                a += response
        y_ = response
        preds.append({'x':x,'y':y,'pred':y_})

        if i % args.logging_steps == 0:
            logger.info('{x}'.format(x=str({'x':x,'y':y,'pred':y_})))
    
    logger.info('eval: save...')
    with open(save_path,'w') as file:
        for l in preds:
            line = json.dumps(l,ensure_ascii=False)
            file.write(line)
            file.write('\n')
    time_end = time.time()
    logger.info('reward_eval : {x} '.format(x=args.reward_model_path))
    logger.info('save_eval : {x} '.format(x=save_path))
    logger.info('running time: {x}'.format(x=time_end-time_start))
    logger.info('eval: fin')

def sample(args,warp):

    # load datset
    logger.info('load datset..')
    dataset = load_data(args)

    logger.info('start sampling..')
    samples = {}
    for iter in range(args.iteration):
        logger.info('iter_{x}'.format(x=iter))
        
        train_samples = []

        sample_path = 'branch{b}_deep{d}_budget{g}_iter{i}'.format(b=tree_config['branch'],d=tree_config['max_depth'],g=args.budget,i=iter)
        if args.sample_path != '' and sample_path in os.listdir(args.sample_path):
            sample_path = os.path.join(args.sample_path,sample_path)
            train_samples += load_samples(sample_path)

        else:
            save_path = os.path.join(args.output_path,'data',
                             'branch{b}_deep{d}_budget{g}_iter{i}'.format(b=tree_config['branch'],d=tree_config['max_depth'],g=args.budget,i=iter),
                             )
            os.makedirs(save_path) if not os.path.exists(save_path) else None
            for i,sample in enumerate(dataset['train']):

                time_start = time.time()
  
                tree_of_sample = Tree(sample=sample,warp=warp) 
                tree_of_sample.monte_carlo_tree_search(budget=args.budget,reward_funcation=args.reward_funcation)
                #train_samples += tree_of_sample.sample(attribute='positive')
        
                save_path = os.path.join(args.output_path,'data',
                                         'branch{b}_deep{d}_budget{g}_iter{i}'.format(b=tree_config['branch'],d=tree_config['max_depth'],g=args.budget,i=iter),
                                         'samples' + time.strftime("-%Y-%m-%d-%H:%M:%S", time.localtime()) + '.json')
                train_samples += tree_of_sample.save(path=save_path)

                time_end = time.time()
                logger.info('\nrunning time: {x}'.format(x=time_end-time_start))
                _example = tree_of_sample.root.x[:50] if len(tree_of_sample.root.x) > 50 else tree_of_sample.root.x
                logger.info('{i}-th sample: {x}'.format(i=i,x=_example))
        
        train_samples = Dataset.from_list(train_samples)
        samples[str(iter)] = train_samples
    
    return samples


def run(args):
    
    # create framework
    if args.dataset in ['ljp']:
        warp = WarpLJP(args=args)

    warp.load_generate_model()

    # load training data
    #

    # data collection
    if args.do_train:
        trainsets = sample(args=args,warp=warp)
        #raise ValueError
        train(args,warp,trainsets)
    # test
    if args.do_test:
        datasets = load_data(args)
        evaluate(args,warp,datasets['test'])



if __name__ == "__main__":
    args = get_args()
    logger = get_logger(args.output_path)

    loginfo = '\n'.join(['{k}: {v}'.format(k=k,v=v) for k,v in vars(args).items()])
    logger.info(loginfo)

    try:
        run(args)
    except Exception as E:
        logger.exception('{x}'.format(x=E))