train.py

from __future__ import absolute_import, division, print_function

import sys
print(sys.path)
sys.path.append("/home/fzus/lyh/DiCos/models/")

import nltk

import argparse
import json
import logging
import math
import os
import random
import datetime
import torch.nn as nn
import spacy

import numpy as np
import torch
import torch.nn.functional as F
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler, TensorDataset)
import megengine as mge
import megengine.autodiff as ad
import gc
from utils.lion import *

# 可视化相关
import visdom
vis = visdom.Visdom(env='train')
opt = {
    'xlable': 'step',
    'ylabel': 'loss_value',
    'title': 'mean_loss'
}
opt_lr = {
    'xlable': 'step',
    'ylabel': 'lr',
    'title': 'learning rate'
}
loss_window = vis.line(
    X = [0],
    Y = [0],
    opts=opt
)
lr_window = vis.line(
    X=[0],
    Y=[0],
    opts=opt_lr
)


mge.dtr.enable()

torch.autograd.set_detect_anomaly(True)

try:
    from torch.utils.tensorboard import SummaryWriter
except:
    from tensorboardX import SummaryWriter

writer = SummaryWriter('/home/fzus/lyh/DiCoS/log/' + datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))

from tqdm import tqdm
from models.generator import Generator
from utils import helper
from utils.data_kb_util import *
from utils.data_utils import prepare_dataset, MultiWozDataset
from utils.constant import track_slots, ansvocab, slot_map, n_slot, TURN_SPLIT, TEST_TURN_SPLIT
from utils.data_utils import make_slot_meta, domain2id, OP_SET, make_turn_label, postprocessing
from evaluation import op_evaluation, joint_evaluation
# from transformers.configuration_albert import AlbertConfig
from transformers import AlbertConfig
# from transformers.tokenization_albert import AlbertTokenizer
from transformers import AlbertTokenizer
from transformers.optimization import AdamW
# from pytorch_transformers import AdamW, WarmupLinearSchedule
from transformers import get_linear_schedule_with_warmup
from utils.logger import get_logger

import sys
import csv

os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0, 1"

logger_trainInfo = get_logger("train_logger", "./saved_models/logger.log")
logger_trainInfo.info("")

csv.field_size_limit(sys.maxsize)
logger = logging.getLogger(__name__)

nlp = spacy.load('en_core_web_trf')
f = open("/home/fzus/lyh/entity_en.json", 'r', encoding='UTF-8')
entities_dict = json.load(f)
entities_dict = DataDict(entities_dict)


def set_seed(args):
    random.seed(args.seed)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    if args.n_gpu > 0:
        torch.cuda.manual_seed_all(args.seed)


'''compute joint operation scores based on logits of two stages 
'''


def compute_jointscore(start_scores, end_scores, gen_scores, pred_ops, ans_vocab, slot_mask):
    seq_lens = start_scores.shape[-1]
    joint_score = start_scores.unsqueeze(-2) + end_scores.unsqueeze(-1)
    triu_mask = np.triu(np.ones((joint_score.size(-1), joint_score.size(-1))))
    triu_mask[0, 1:] = 0
    triu_mask = (torch.Tensor(triu_mask) == 0).bool()
    joint_score = joint_score.masked_fill(triu_mask.unsqueeze(0).unsqueeze(0).cuda(), -1e9).masked_fill(
        slot_mask.unsqueeze(1).unsqueeze(-2) == 0, -1e9)
    joint_score = F.softmax(joint_score.view(joint_score.size(0), joint_score.size(1), -1),
                            dim=-1).view(joint_score.size(0), joint_score.size(1), seq_lens, -1)

    score_diff = (joint_score[:, :, 0, 0] - joint_score[:, :, 1:, 1:].max(dim=-1)[0].max(dim=-1)[
        0])
    score_noans = pred_ops[:, :, -1] - pred_ops[:, :, 0]
    slot_ans_count = (ans_vocab.sum(-1) != 0).sum(dim=-1) - 2
    ans_idx = torch.where(slot_ans_count < 0, torch.zeros_like(slot_ans_count), slot_ans_count)
    neg_ans_mask = torch.cat((torch.linspace(0, ans_vocab.size(0) - 1,
                                             ans_vocab.size(0)).unsqueeze(0).long(),
                              ans_idx.unsqueeze(0)),
                             dim=0)
    neg_ans_mask = torch.sparse_coo_tensor(neg_ans_mask, torch.ones(ans_vocab.size(0)),
                                           (ans_vocab.size(0),
                                            ans_vocab.size(1))).to_dense().cuda()
    score_neg = gen_scores.masked_fill(neg_ans_mask.unsqueeze(0) == 0, -1e9).max(dim=-1)[0]
    score_has = gen_scores.masked_fill(neg_ans_mask.unsqueeze(0) == 1, -1e9).max(dim=-1)[0]
    cate_score_diff = score_neg - score_has
    score_diffs = score_diff.view(-1).cpu().detach().numpy().tolist()
    cate_score_diffs = cate_score_diff.view(-1).cpu().detach().numpy().tolist()
    score_noanses = score_noans.view(-1).cpu().detach().numpy().tolist()
    return score_diffs, cate_score_diffs, score_noanses


def saveOperationLogits(model, device, dataset, save_path, turn):
    score_ext_map = {}
    model.eval()
    for batch in tqdm(dataset, desc="Evaluating"):
        batch = [b.to(device) if not isinstance(b, int) and not isinstance(b, list) else b for b in batch]
        input_ids, input_mask, slot_mask, segment_ids, state_position_ids, op_ids, pred_ops, domain_ids, gen_ids, start_position, end_position, max_value, max_update, slot_ans_ids, start_idx, end_idx, sid = batch
        batch_size = input_ids.shape[0]
        seq_lens = input_ids.shape[1]
        start_logits, end_logits, has_ans, gen_scores, _, _, _ = model(input_ids=input_ids,
                                                                       token_type_ids=segment_ids,
                                                                       state_positions=state_position_ids,
                                                                       attention_mask=input_mask,
                                                                       slot_mask=slot_mask,
                                                                       max_value=max_value,
                                                                       op_ids=op_ids,
                                                                       max_update=max_update)

        score_ext = has_ans.cpu().detach().numpy().tolist()
        for i, sd in enumerate(score_ext):
            score_ext_map[sid[i]] = sd
    with open(os.path.join(save_path, "cls_score_test_turn{}.json".format(turn)), "w") as writer:
        writer.write(json.dumps(score_ext_map, indent=4) + "\n")


# ============================CL 损失计算=======================================

def sim(z1: torch.Tensor, z2: torch.Tensor):
    z1 = F.normalize(z1)
    z2 = F.normalize(z2)
    return torch.mm(z1, z2.t())

def semi_loss(self, z1: torch.Tensor, z2: torch.Tensor):  # 损失输入为两个特征矩阵
    f = lambda x: torch.exp(x / self.tau)
    refl_sim = f(sim(z1, z1))
    between_sim = f(sim(z1, z2))

    return -torch.log(
        between_sim.diag()
        / (refl_sim.sum(1) + between_sim.sum(1) - refl_sim.diag()))

def batched_semi_loss(z1: torch.Tensor, z2: torch.Tensor,
                        batch_size: int): # 输入特征是把几个batch的特征拼到一起
    # Space complexity: O(BN) (semi_loss: O(N^2))
    z1 = z1.transpose(0, 1)
    z2 = z2.transpose(0, 1)
    z1 = z1.reshape(z1.shape[0]*z1.shape[1], -1)
    z2 = z2.reshape(z2.shape[0]*z2.shape[1], -1)

    device = z1.device
    num_nodes = z1.size(0)
    num_batches = (num_nodes - 1) // batch_size + 1
    f = lambda x: torch.exp(x / 0.5)
    indices = torch.arange(0, num_nodes).to(device)
    losses = []

    for i in range(num_batches):
        mask = indices[i * batch_size:(i + 1) * batch_size]
        refl_sim = f(sim(z1[mask], z1))  # [B, N]
        between_sim = f(sim(z1[mask], z2))  # [B, N]

        losses.append(-torch.log(
            between_sim[:, i * batch_size:(i + 1) * batch_size].diag()
            / (refl_sim.sum(1) + between_sim.sum(1)
                - refl_sim[:, i * batch_size:(i + 1) * batch_size].diag())))

    return torch.cat(losses)


def cl_loss(z1: torch.Tensor, z2: torch.Tensor,
            mean: bool = True, batch_size: int = 0):  # 对比损失
    h1 = z1.detach()
    h2 = z2.detach()

    if batch_size == 0:
        l1 = semi_loss(h1, h2)
        l2 = semi_loss(h2, h1)
    else:
        l1 = batched_semi_loss(h1, h2, batch_size)
        l2 = batched_semi_loss(h2, h1, batch_size)

    ret = (l1 + l2) * 0.5
    ret = ret.mean() if mean else ret.sum()

    return ret
# ===============================================================

def compute_span_loss(gen_ids, input_ids, fuse_score, start_scores, end_scores, generate_turn, sample_mask,
                      generate_mask):
    loss = 0
    for i in range(start_scores.shape[0]):
        if i >= 2:
            select_mask = [generate_turn[i][t] in fuse_score[i].argsort(dim=-1, descending=True)[:, :2][t] for t in
                           range(30)]
            select_mask = torch.from_numpy(np.array(select_mask)).cuda()
            batch_mask = sample_mask[i] & ((generate_mask[i] == 0) | select_mask)
        else:
            batch_mask = sample_mask[i]
        start_idx = [-1 for i in range(n_slot)]
        end_idx = [-1 for i in range(n_slot)]
        for ti in range(n_slot):
            if not batch_mask[ti]:
                continue
            value = gen_ids[i][ti][0]
            value = value[:-1] if isinstance(value, list) else [value]
            batch_input = input_ids[i][ti].cpu().detach().numpy().tolist()
            for text_idx in range(len(input_ids[i][ti]) - len(value)):
                if batch_input[text_idx: text_idx + len(value)] == value:
                    start_idx[ti] = text_idx
                    end_idx[ti] = text_idx + len(value) - 1
                    break
        start_idx = torch.from_numpy(np.array(start_idx)).cuda()
        end_idx = torch.from_numpy(np.array(end_idx)).cuda()
        loss += masked_cross_entropy_for_value(start_scores[i].contiguous(),
                                               start_idx.contiguous(),
                                               sample_mask=batch_mask,
                                               pad_idx=-1
                                               )
        batch_loss = masked_cross_entropy_for_value(end_scores[i].contiguous(),
                                                    end_idx.contiguous(),
                                                    sample_mask=batch_mask,
                                                    pad_idx=-1
                                                    )
        loss += batch_loss
    loss /= start_scores.shape[0]
    return loss


def masked_cross_entropy_for_value(logits, target, sample_mask=None, slot_mask=None, pad_idx=-1):
    mask = logits.eq(0)
    pad_mask = target.ne(pad_idx)
    target = target.masked_fill(target < 0, 0)
    sample_mask = pad_mask & sample_mask if sample_mask is not None else pad_mask
    sample_mask = slot_mask & sample_mask if slot_mask is not None else sample_mask
    target = target.masked_fill(sample_mask == 0, 0)
    logits = logits.masked_fill(mask, 1)
    logits_flat = logits.view(-1, logits.size(-1))
    log_probs_flat = torch.log(logits_flat)
    target_flat = target.view(-1, 1)
    losses_flat = -torch.gather(log_probs_flat, dim=1, index=target_flat)
    losses = losses_flat.view(*target.size())
    # if mask is not None:
    sample_num = sample_mask.sum().float()
    losses = losses * sample_mask.float()
    loss = (losses.sum() / sample_num) if sample_num != 0 else losses.sum()
    return loss


# [SLOT], [NULL], [EOS]
def addSpecialTokens(tokenizer, specialtokens):
    special_key = "additional_special_tokens"
    tokenizer.add_special_tokens({special_key: specialtokens})


def fixontology(ontology, turn, tokenizer):
    ans_vocab = []
    esm_ans_vocab = []
    esm_ans = ansvocab
    slot_mm = np.zeros((len(slot_map), len(esm_ans)))
    max_anses_length = 0
    max_anses = 0
    for i, k in enumerate(ontology.keys()):
        if k in track_slots:
            s = ontology[k]
            s['name'] = k
            if not s['type']:
                s['db'] = []
            slot_mm[i][slot_map[s['name']]] = 1
            ans_vocab.append(s)
    for si in esm_ans:
        slot_anses = []
        for ans in si:
            enc_ans = tokenizer.encode(ans)
            max_anses_length = max(max_anses_length, len(ans))
            slot_anses.append(enc_ans)
        max_anses = max(max_anses, len(slot_anses))
        esm_ans_vocab.append(slot_anses)
    for s in esm_ans_vocab:
        for ans in s:
            gap = max_anses_length - len(ans)
            ans += [0] * gap
        gap = max_anses - len(s)
        s += [[0] * max_anses_length] * gap
    esm_ans_vocab = np.array(esm_ans_vocab)
    ans_vocab_tensor = torch.from_numpy(esm_ans_vocab)
    slot_mm = torch.from_numpy(slot_mm).float()
    return ans_vocab, slot_mm, ans_vocab_tensor


def mask_ans_vocab(ontology, slot_meta, tokenizer):
    ans_vocab = []
    max_anses = 0
    max_anses_length = 0
    change_k = []
    cate_mask = []
    for k in ontology:
        if (' range' in k['name']) or (' at' in k['name']) or (' by' in k['name']):
            change_k.append(k)
    for key in change_k:
        new_k = key['name'].replace(' ', '')
        key['name'] = new_k
    for s in ontology:
        cate_mask.append(s['type'])
        v_list = s['db']
        slot_anses = []
        for v in v_list:
            ans = tokenizer.encode(v)
            max_anses_length = max(max_anses_length, len(ans))
            slot_anses.append(ans)
        max_anses = max(max_anses, len(slot_anses))
        ans_vocab.append(slot_anses)
    for s in ans_vocab:
        for ans in s:
            gap = max_anses_length - len(ans)
            ans += [0] * gap
        gap = max_anses - len(s)
        s += [[0] * max_anses_length] * gap
    ans_vocab = np.array(ans_vocab)
    ans_vocab_tensor = torch.from_numpy(ans_vocab)
    return ans_vocab_tensor, ans_vocab, cate_mask


def main():
    parser = argparse.ArgumentParser()
    ## Required parameters
    parser.add_argument("--model_type", default='albert', type=str,
                        help="Model type selected in the list: ")
    parser.add_argument("--model_name_or_path", default='pretrained_models/albert_large/', type=str,
                        help="Path to pre-trained model or shortcut name selected in the list: ")
    parser.add_argument("--output_dir", default="saved_models/", type=str,
                        help="The output directory where the model predictions and checkpoints will be written.")
    ## Other parameters
    parser.add_argument("--config_name", default="", type=str,
                        help="Pretrained config name or path if not the same as model_name")
    parser.add_argument("--tokenizer_name", default="", type=str,
                        help="Pretrained tokenizer name or path if not the same as model_name")
    parser.add_argument("--cache_dir", default="", type=str,
                        help="Where do you want to store the pre-trained models downloaded from s3")
    parser.add_argument("--do_train", default=True, action='store_true',
                        help="Whether to run training.")
    parser.add_argument("--evaluate_during_training", action='store_true',
                        help="Rul evaluation during training at each logging step.")
    parser.add_argument("--do_lower_case", action='store_true',
                        help="Set this flag if you are using an uncased model.")

    parser.add_argument("--per_gpu_train_batch_size", default=16, type=int,
                        help="Batch size per GPU/CPU for training.")
    parser.add_argument("--per_gpu_eval_batch_size", default=32, type=int,
                        help="Batch size per GPU/CPU for evaluation.")
    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
                        help="Number of updates steps to accumulate before performing a backward/update pass.")
    parser.add_argument("--learning_rate", default=3e-5, type=float,
                        help="The initial learning rate for Adam.")
    parser.add_argument("--weight_decay", default=0.1, type=float,
                        help="Weight decay if we apply some.")
    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
                        help="Epsilon for Adam optimizer.")
    parser.add_argument("--max_grad_norm", default=5.0, type=float,
                        help="Max gradient norm.")
    parser.add_argument("--num_train_epochs", default=3.0, type=float,
                        help="Total number of training epochs to perform.")
    parser.add_argument("--max_steps", default=-1, type=int,
                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
    parser.add_argument("--warmup_steps", default=0, type=int,
                        help="Linear warmup over warmup_steps.")

    parser.add_argument('--logging_steps', type=int, default=50,
                        help="Log every X updates steps.")
    parser.add_argument('--save_steps', type=int, default=50,
                        help="Save checkpoint every X updates steps.")
    parser.add_argument("--eval_all_checkpoints", action='store_true',
                        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
    parser.add_argument("--no_cuda", action='store_true',
                        help="Avoid using CUDA when available")
    parser.add_argument('--overwrite_output_dir', default=True, action='store_true',
                        help="Overwrite the content of the output directory")
    parser.add_argument('--overwrite_cache', action='store_true',
                        help="Overwrite the cached training and evaluation sets")
    parser.add_argument('--seed', type=int, default=42,
                        help="random seed for initialization")

    parser.add_argument('--fp16', action='store_true',
                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
    parser.add_argument('--fp16_opt_level', type=str, default='O1',
                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
                             "See details at https://nvidia.github.io/apex/amp.html")

    parser.add_argument("--local_rank", type=int, default=-1,
                        help="For distributed training: local_rank")  # DST params
    # 数据集
    parser.add_argument("--data_root", default='data/mwz2.2/', type=str) # 2.2
    parser.add_argument("--train_data", default='train_dials.json', type=str)
    parser.add_argument("--dev_data", default='test_dials.json', type=str)
    parser.add_argument("--test_data", default='test_dials.json', type=str)

    parser.add_argument("--ontology_data", default='schema.json', type=str)
    parser.add_argument("--vocab_path", default='assets/vocab.txt', type=str)
    parser.add_argument("--save_dir", default='saved_models', type=str)
    parser.add_argument("--load_model", default=False, action='store_true')
    parser.add_argument("--load_ckpt_epoch", default='checkpoint_epoch_5996_with_ke.bin', type=str)
    parser.add_argument("--load_test_op_data_path", default='cls_score_test_state_update_predictor_output.json',
                        type=str)
    parser.add_argument("--random_seed", default=42, type=int)
    parser.add_argument("--num_workers", default=4, type=int)
    parser.add_argument("--batch_size", default=4, type=int)    # real batch size
    parser.add_argument("--enc_warmup", default=0.01, type=float)
    parser.add_argument("--dec_warmup", default=0.01, type=float)
    parser.add_argument("--enc_lr", default=5e-6, type=float) # 5e-6
    parser.add_argument("--base_lr", default=2e-4, type=float) # 1e-4 2e-4
    parser.add_argument("--n_epochs", default=30, type=int) # 训练代数 10
    parser.add_argument("--eval_epoch", default=1, type=int)
    parser.add_argument("--eval_step", default=5, type=int)
    parser.add_argument("--turn", default=2, type=int)
    parser.add_argument("--op_code", default="2", type=str)
    parser.add_argument("--slot_token", default="[SLOT]", type=str)
    parser.add_argument("--dropout", default=0.0, type=float) # 0.0
    parser.add_argument("--hidden_dropout_prob", default=0.0, type=float)
    parser.add_argument("--attention_probs_dropout_prob", default=0.1, type=float)
    parser.add_argument("--decoder_teacher_forcing", default=0.5, type=float)
    parser.add_argument("--word_dropout", default=0.1, type=float)
    parser.add_argument("--not_shuffle_state", default=True, action='store_true')

    parser.add_argument("--n_history", default=3, type=int)
    parser.add_argument("--max_seq_length", default=256, type=int)
    parser.add_argument("--sketch_weight", default=0.55, type=float)
    parser.add_argument("--answer_weight", default=0.6, type=float)
    parser.add_argument("--generation_weight", default=0.2, type=float)
    parser.add_argument("--extraction_weight", default=0.1, type=float)
    parser.add_argument("--msg", default=None, type=str)
    
    parser.add_argument("--clearning", default=False, type=bool)
    parser.add_argument("--m_gcn", default=True, type=bool)
    parser.add_argument("--syn_guide", default=True, type=bool)
    parser.add_argument("--knowledge_enhance", default=True, type=bool)
    parser.add_argument("--use_clr", default=False, type=bool)
    parser.add_argument("--use_lion_opt", default=False, type=bool)
    
    
    

    args = parser.parse_args()

    if os.path.exists(args.output_dir) and os.listdir(
            args.output_dir) and args.do_train and not args.overwrite_output_dir:
        raise ValueError(
            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
                args.output_dir))

    # Setup CUDA, GPU & distributed training
    if args.local_rank == -1 or args.no_cuda:
        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
        args.n_gpu = torch.cuda.device_count()
    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
        torch.cuda.set_device(args.local_rank)
        device = torch.device("cuda", args.local_rank)
        torch.distributed.init_process_group(backend='nccl')
        args.n_gpu = 1
    args.device = device

    # Setup logging
    logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
                        datefmt='%m/%d/%Y %H:%M:%S',
                        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
                   args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)

    # Set seed
    set_seed(args)

    def worker_init_fn(worker_id):
        np.random.seed(args.random_seed + worker_id)

    # Prepare GLUE task
    n_gpu = 0
    if torch.cuda.is_available():
        n_gpu = torch.cuda.device_count()

    # Load pretrained model and tokenizer
    if args.local_rank not in [-1, 0]:
        # Make sure only the first process in distributed training will download model & vocab
        torch.distributed.barrier()

    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
    turn = args.turn

    ontology = json.load(open(args.data_root + args.ontology_data))

    criterion = nn.MSELoss()

    _, slot_meta = make_slot_meta(ontology)

    

    with torch.cuda.device(0):
        op2id = OP_SET[args.op_code]  # 槽位选择的标签

        rng = random.random(args.random_seed)
        print(op2id)
        logger_trainInfo.info(op2id)
        tokenizer = AlbertTokenizer.from_pretrained(args.model_name_or_path + "spiece.model")
        addSpecialTokens(tokenizer, ['[SLOT]', '[NULL]', '[EOS]', '[dontcare]', '[negans]', '[noans]', '[TURN]'])
        turn_id = tokenizer.encode('[TURN]')[1]
        args.vocab_size = len(tokenizer)  # 词表大小
        ontology, slot_mm, esm_ans_vocab = fixontology(slot_meta, turn, tokenizer)
        ans_vocab, ans_vocab_nd, cate_mask = mask_ans_vocab(ontology, slot_meta, tokenizer)

        # nltk 相关配置
        

        if turn == 2:
            train_op_data_path = None
            test_op_data_path = args.data_root + "cls_score_test_state_update_predictor_output.json"
            isfilter = True

        model = Generator(args, len(op2id), len(domain2id), op2id['update'], esm_ans_vocab, slot_mm, turn=turn,
                          turn_id=turn_id, tokenizer=tokenizer)
        train_data_raw, _, _ = prepare_dataset(data_path=args.data_root + args.train_data,
                                               tokenizer=tokenizer,
                                               slot_meta=slot_meta,
                                               n_history=args.n_history,
                                               max_seq_length=args.max_seq_length,
                                               op_code=args.op_code,
                                               slot_ans=ontology,
                                               turn=turn,
                                               op_data_path=None,
                                               isfilter=isfilter,
                                               if_train=True
                                               )
        train_data = MultiWozDataset(train_data_raw,
                                     tokenizer,
                                     slot_meta,
                                     args.max_seq_length,
                                     ontology,
                                     args.word_dropout,
                                     turn=turn)
        print("# train examples %d" % len(train_data_raw))
        logger_trainInfo.info("# train examples %d" % len(train_data_raw))

        dev_data_raw, idmap, _ = prepare_dataset(data_path=args.data_root + args.dev_data,
                                                 tokenizer=tokenizer,
                                                 slot_meta=slot_meta,
                                                 n_history=args.n_history,
                                                 max_seq_length=args.max_seq_length,
                                                 op_code=args.op_code,
                                                 turn=turn,
                                                 slot_ans=ontology,
                                                 op_data_path=test_op_data_path,
                                                 isfilter=False,
                                                 if_train=False)
        dev_data = MultiWozDataset(dev_data_raw,
                                   tokenizer,
                                   slot_meta,
                                   args.max_seq_length,
                                   ontology,
                                   word_dropout=0,
                                   turn=turn)
        print("# dev examples %d" % len(dev_data_raw))
        logger_trainInfo.info("# dev examples %d" % len(dev_data_raw))

        train_sampler = RandomSampler(train_data)
        train_dataloader = DataLoader(train_data,
                                      sampler=train_sampler,
                                      batch_size=args.batch_size,
                                      collate_fn=train_data.collate_fn,
                                      num_workers=args.num_workers,
                                      worker_init_fn=worker_init_fn)

        dev_sampler = RandomSampler(dev_data)
        dev_dataloader = DataLoader(dev_data,
                                    sampler=dev_sampler,
                                    batch_size=args.batch_size,
                                    collate_fn=dev_data.collate_fn,
                                    num_workers=args.num_workers,
                                    worker_init_fn=worker_init_fn)

        if args.load_model:
            # load best
            # checkpoint = torch.load(os.path.join(args.save_dir, 'model_best_turn{}.bin'.format(turn)))
            # model.load_state_dict(checkpoint['model'])
            # load inter
            checkpoint = torch.load(os.path.join(args.save_dir, args.load_ckpt_epoch))
            model.load_state_dict(checkpoint)

        model.to(args.device)
        logger.info("Training/evaluation parameters %s", args)
        if args.do_train:
            num_train_steps = int(len(train_dataloader) / args.batch_size * args.n_epochs)
            bert_params_ids = list(map(id, model.albert.parameters()))
            no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
            enc_param_optimizer = list(model.named_parameters())

            enc_optimizer_grouped_parameters = [
                {'params': [p for n, p in enc_param_optimizer if
                            (id(p) in bert_params_ids and not any(nd in n for nd in no_decay))], 'weight_decay': 0.01,
                 'lr': args.enc_lr},
                {'params': [p for n, p in enc_param_optimizer if
                            id(p) in bert_params_ids and any(nd in n for nd in no_decay)], 'weight_decay': 0.0,
                 'lr': args.enc_lr},
                {'params': [p for n, p in enc_param_optimizer if
                            id(p) not in bert_params_ids and not any(nd in n for nd in no_decay)], 'weight_decay': 0.01,
                 'lr': args.base_lr},
                {'params': [p for n, p in enc_param_optimizer if
                            id(p) not in bert_params_ids and any(nd in n for nd in no_decay)], 'weight_decay': 0.0,
                 'lr': args.base_lr}]

            if args.use_lion_opt:
                enc_optimizer = Lion(enc_optimizer_grouped_parameters, lr=args.base_lr)
            else:
                enc_optimizer = AdamW(enc_optimizer_grouped_parameters, lr=args.base_lr)
            # cl_optimizer = AdamW()

            
            # CLR
            if args.use_clr:
                enc_scheduler = torch.optim.lr_scheduler.CyclicLR(enc_optimizer, base_lr=1e-4, max_lr=4e-4, step_size_up=2620)
            else:
                enc_scheduler = get_linear_schedule_with_warmup(enc_optimizer,
                                                            num_warmup_steps=int(num_train_steps * args.enc_warmup),
                                                            num_training_steps=num_train_steps)  # 线性warmup

            best_score = {'epoch': 0, 'overall_jga': 0, 'cate_jga': 0, 'noncate_jga': 0}
            file_logger = helper.FileLogger(args.save_dir + '/log.txt',
                                            header="# epoch\tstep\ttrain_loss\tbest_jointacc\tbest_catejointacc\tbest_noncatejointacc\tnow_jointacc\tnow_catejointacc\tnow_noncatejointacc")
            model.train()
            loss = 0
            sketchy_weight, answer_weight, generation_weight, extraction_weight = args.sketch_weight, args.answer_weight, args.generation_weight, args.extraction_weight
            verify_weight = 1 - sketchy_weight
            global_step = 0
            for epoch in range(args.n_epochs):
                batch_loss = []
                for step, batch in enumerate(tqdm(train_dataloader)):
                    global_step += 1
                    batch = [b.to(device) if not isinstance(b, int) and not isinstance(b, list) else b for b in batch]
                    input_ids, input_mask, slot_mask, segment_ids, state_position_ids, op_ids, pred_ops, domain_ids, gen_ids, start_position, end_position, max_value, max_update, slot_ans_ids, start_idx, end_idx, position_ids, sample_mm, generate_turn, generate_mask, ref_slot, gold_ans_label, sid, update_current_mm, slot_all_connect, update_mm, slot_domain_connect = batch
                    if input_ids.numel() == 0 or sample_mm.numel() == 0:
                        continue

                    assert len(input_ids) <= TURN_SPLIT  # train在之前已经切分了，这里不应该有任何false的情况
                    sample_mask = (pred_ops.argmax(dim=-1) == 0) if turn == 1 else (op_ids == 0)
                    start_logits, end_logits, gen_scores, _, _, _, fuse_score, input_ids, ctl_loss = model(input_ids=input_ids,
                                                                                                  token_type_ids=segment_ids,
                                                                                                  state_positions=state_position_ids,
                                                                                                  attention_mask=input_mask,
                                                                                                  slot_mask=slot_mask,
                                                                                                  first_edge_mask=update_current_mm,
                                                                                                  second_edge_mask=slot_all_connect,
                                                                                                  third_edge_mask=update_mm,
                                                                                                  fourth_edge_mask=slot_domain_connect,
                                                                                                  max_value=max_value,
                                                                                                  op_ids=op_ids,
                                                                                                  max_update=max_update,
                                                                                                  position_ids=position_ids,
                                                                                                  sample_mm=sample_mm) 

                    if turn == 2:
                        loss_selector = masked_cross_entropy_for_value(fuse_score.contiguous(),
                                                                       generate_turn.contiguous(),
                                                                       sample_mask=sample_mask & (generate_mask == 1)
                                                                       )
                        loss_classifier = masked_cross_entropy_for_value(gen_scores.contiguous(),
                                                                         slot_ans_ids.contiguous(),
                                                                         sample_mask=sample_mask
                                                                         ) # 分类损失
                        loss_extractor = compute_span_loss(gen_ids, input_ids, fuse_score, start_logits, end_logits,
                                                           generate_turn, sample_mask, generate_mask)   # 跨度提取损失
                        # todo: 这里新增一个图对比学习的损失
                        loss_grace = ctl_loss
                        # loss_grace = cl_loss(slot_node, another_slot_node, True, slot_node.shape[0]) # 对比损失
                        # loss_grace = criterion(slot_node.contiguous(), another_slot_node.contiguous())

                        # 加上对比学习损失 cl_loss
                        loss = loss_classifier + loss_extractor  + loss_grace
                        loss = math.log(input_ids.shape[0], 8) * loss

                        loss.backward(retain_graph=False)         # 损失反向传播
                        for name, par in model.named_parameters():
                            if par.requires_grad and par.grad is not None:
                                if torch.sum(torch.isnan(par.grad)) != 0:
                                    model.zero_grad()

                    batch_loss.append(loss.item())
                    del batch
                    # gc.collect()
                    torch.cuda.empty_cache() # 37 - 31
                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
                    enc_optimizer.step()
                    enc_scheduler.step()
                    model.zero_grad()

                    vis.line(X=[global_step], Y=[np.mean(batch_loss)], win=loss_window, opts=opt, update='append')
                    vis.line(X=[global_step], Y=[enc_optimizer.state_dict()['param_groups'][0]['lr']], win=lr_window, opts=opt_lr, update='append')
                    
                    if step % 100 == 0:
                        writer.add_scalar('train_mean_loss', np.mean(batch_loss), step)
                        writer.add_scalar('lr', enc_optimizer.state_dict()['param_groups'][0]['lr'], step)


                        print("[%d/%d] [%d/%d] mean_loss : %.3f" \
                              % (epoch + 1, args.n_epochs, step,
                                 len(train_dataloader), np.mean(batch_loss),), end='\t')
                        logger_trainInfo.debug("[%d/%d] [%d/%d] mean_loss : %.3f" \
                                               % (epoch + 1, args.n_epochs, step,
                                                  len(train_dataloader), np.mean(batch_loss),))
                        print(datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
                        logger_trainInfo.debug(datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S'))

                        file_logger.log(
                            "Epoch {}\t Step {}\t loss {:.6f}\t cl_loss {:.6f}\t time {}".format(epoch, step, loss, loss_grace,
                                                                                datetime.datetime.now().strftime(
                                                                                    '%Y-%m-%d %H:%M:%S')))

                        logger_trainInfo.info("Epoch {}\t Step {}\t loss {:.6f}\t cl_loss {:.6f}\t time {}".format(epoch, step, loss, loss_grace,
                                                                                                  datetime.datetime.now().strftime(
                                                                                                      '%Y-%m-%d %H:%M:%S')))
                        # torch.cuda.empty_cache()
                        del batch_loss
                        batch_loss = []
                    # del batch_loss, batch, input_ids
                if True:
                    with torch.no_grad(): # 测试无需梯度
                        joint_acc, catejoint_acc, noncatejoint_acc = evaluate(dev_dataloader, model, device, ans_vocab_nd,
                                                                            cate_mask, turn=2, tokenizer=tokenizer,
                                                                            ontology=ontology)
                    if joint_acc > best_score['overall_jga']:
                        best_score['epoch'] = epoch
                        best_score['overall_jga'] = joint_acc
                        best_score['cate_jga'] = catejoint_acc
                        best_score['noncate_jga'] = noncatejoint_acc
                        saved_name = 'model_best_turn' + str(turn) + '.bin'
                        save_path = os.path.join(args.save_dir, saved_name)
                        model_to_save = model.module if hasattr(model, 'module') else model
                        params = {
                            'model': model_to_save.state_dict(),
                            'optimizer': enc_optimizer.state_dict(),
                            'scheduler': enc_scheduler.state_dict(),
                            'args': args
                        }
                        print("开保存最佳模型。。。。。。。")
                        torch.save(params, save_path)

                    file_logger.log(
                        "{}\t{}\t{:.6f}\t{:.4f}\t{:.4f}\t{:.4f}\t{:.4f}\t{:.4f}\t{:.4f}".format(epoch, step, loss,
                                                                                                best_score[
                                                                                                    'overall_jga'],
                                                                                                best_score['cate_jga'],
                                                                                                best_score[
                                                                                                    'noncate_jga'],
                                                                                                joint_acc,
                                                                                                catejoint_acc,
                                                                                                noncatejoint_acc))
                    logger_trainInfo.warning(
                        "{}\t{}\t{:.6f}\t{:.4f}\t{:.4f}\t{:.4f}\t{:.4f}\t{:.4f}\t{:.4f}".format(epoch, step, loss,
                                                                                                best_score[
                                                                                                    'overall_jga'],
                                                                                                best_score['cate_jga'],
                                                                                                best_score[
                                                                                                    'noncate_jga'],
                                                                                                joint_acc,
                                                                                                catejoint_acc,
                                                                                                noncatejoint_acc))
                    print(
                        "Current: Epoch_{}\tstep_{}\tloss_{:.6f}\tJointAcc: {:.4f}\tCategorical-JointAcc: {:.4f}\tnon-Categorical-JointAcc: {:.4f}".format(
                            epoch, step, loss, joint_acc, catejoint_acc, noncatejoint_acc))
                    logger_trainInfo.warning(
                        "Current: Epoch_{}\tstep_{}\tloss_{:.6f}\tJointAcc: {:.4f}\tCategorical-JointAcc: {:.4f}\tnon-Categorical-JointAcc: {:.4f}".format(
                            epoch, step, loss, joint_acc, catejoint_acc, noncatejoint_acc))
                    print(
                        "Best: Epoch_{}\tJointAcc: {:.4f}\tCategorical-JointAcc: {:.4f}\tnon-Categorical-JointAcc: {:.4f}".format(
                            best_score['epoch'], best_score['overall_jga'], best_score['cate_jga'],
                            best_score['noncate_jga']))
                    logger_trainInfo.warning(
                        "Best: Epoch_{}\tJointAcc: {:.4f}\tCategorical-JointAcc: {:.4f}\tnon-Categorical-JointAcc: {:.4f}".format(
                            best_score['epoch'], best_score['overall_jga'], best_score['cate_jga'],
                            best_score['noncate_jga']))
                del loss

                # 每个epoch保存模型
                model_to_save = model.module if hasattr(model, 'module') else model
                save_path = os.path.join(args.save_dir, 'checkpoint_epoch_' + str(epoch) + '.bin')
                torch.save(model_to_save.state_dict(), save_path)

        else:
            with torch.no_grad(): # 测试无需梯度
                        joint_acc, catejoint_acc, noncatejoint_acc = evaluate(dev_dataloader, model, device, ans_vocab_nd,
                                                                            cate_mask, turn=2, tokenizer=tokenizer,
                                                                            ontology=ontology)
            print(
                "Test Result:\tJointAcc: {:.4f}\tCategorical-JointAcc: {:.4f}\tnon-Categorical-JointAcc: {:.4f}".format(
                    joint_acc, catejoint_acc, noncatejoint_acc))
            logger_trainInfo.warning(
                "Test Result:\tJointAcc: {:.4f}\tCategorical-JointAcc: {:.4f}\tnon-Categorical-JointAcc: {:.4f}".format(
                    joint_acc, catejoint_acc, noncatejoint_acc))


def evaluate(dev_dataloader, model, device, ans_vocab_nd, cate_mask, turn=2, tokenizer=None, ontology=None):
    model.eval()
    start_predictions = []
    start_ids = []
    end_predictions = []
    end_ids = []
    has_ans_predictions = []
    has_ans_labels = []
    gen_predictions = []
    gen_labels = []
    score_diffs = []
    cate_score_diffs = []
    score_noanses = []
    all_input_ids = []
    sample_ids = []
    ref_scores = []
    fuse_scores = []
    gen_correct = 0
    gen_guess = 0
    catecorrect = 0.0
    noncatecorrect = 0.0
    cate_slot_correct = 0.0
    nocate_slot_correct = 0.0
    domain_joint = {"hotel": 0, "train": 0, "attraction": 0, "taxi": 0, "restaurant": 0}
    joint_correct = 0
    update_guess = 0
    update_correct = 0
    update_gold = 0
    samples = 0
    gen_id_labels = []
    gold_ans_labels = []

    for step, batch in enumerate(tqdm(dev_dataloader)):
        batch = [b.to(device) if not isinstance(b, int) and not isinstance(b, list) else b for b in
                 batch]
        input_ids, input_mask, slot_mask, segment_ids, state_position_ids, op_ids, pred_ops, domain_ids, gen_ids, start_position, end_position, max_value, max_update, slot_ans_ids, start_idx, end_idx, position_ids, sample_mm, generate_turn, generate_mask, ref_slot, gold_ans_label, sid, update_current_mm, slot_all_connect, update_mm, slot_domain_connect = batch
        if input_ids.numel() == 0 or sample_mm.numel() == 0:
            continue

        if turn == 2:
            has_ans_predictions += pred_ops.argmax(dim=-1).cpu().detach().numpy().tolist()
            start_ids += start_idx.cpu().detach().numpy().tolist()
            end_ids += end_idx.cpu().detach().numpy().tolist()
            has_ans_labels += op_ids.cpu().detach().numpy().tolist()
            gen_labels += slot_ans_ids.cpu().detach().numpy().tolist()
            gen_id_labels += gen_ids
            gold_ans_labels += gold_ans_label
            sample_ids += sid

            # assert len(input_ids) <= TEST_TURN_SPLIT  # test的之前没切分，在这里应该有False的情况出现
            # 测试样本在这里切分
            if len(input_ids) <= TEST_TURN_SPLIT:
                with torch.no_grad():
                    start_logits, end_logits, gen_scores, _, _, _, fuse_score, input_ids, _ = model(input_ids=input_ids,
                                                                                              token_type_ids=segment_ids,
                                                                                              state_positions=state_position_ids,
                                                                                              attention_mask=input_mask,
                                                                                              slot_mask=slot_mask,
                                                                                              first_edge_mask=update_current_mm,
                                                                                              second_edge_mask=slot_all_connect,
                                                                                              third_edge_mask=update_mm,
                                                                                              fourth_edge_mask=slot_domain_connect,
                                                                                              max_value=max_value,
                                                                                              op_ids=op_ids,
                                                                                              max_update=max_update,
                                                                                              position_ids=position_ids,
                                                                                              sample_mm=sample_mm)
                start_predictions += start_logits.argmax(dim=-1).cpu().detach().numpy().tolist()
                end_predictions += end_logits.argmax(dim=-1).cpu().detach().numpy().tolist()
                gen_predictions += gen_scores.argmax(dim=-1).cpu().detach().numpy().tolist()
                fuse_scores += fuse_score.argmax(dim=-1).cpu().detach().numpy().tolist()
                all_input_ids += input_ids[:, :, 1:].cpu().detach().numpy().tolist()

                del start_logits, end_logits, gen_scores, _, fuse_score, input_ids,
                torch.cuda.empty_cache()

            else:
                tmp_input_ids = [input_ids[:TEST_TURN_SPLIT, :], input_ids[TEST_TURN_SPLIT:, :]]
                tmp_segment_ids = [segment_ids[:TEST_TURN_SPLIT, :], segment_ids[TEST_TURN_SPLIT:, :]]
                tmp_state_position_ids = [state_position_ids[:TEST_TURN_SPLIT, :],
                                          state_position_ids[TEST_TURN_SPLIT:, :]]
                tmp_input_mask = [input_mask[:TEST_TURN_SPLIT, :], input_mask[TEST_TURN_SPLIT:, :]]
                tmp_slot_mask = [slot_mask[:TEST_TURN_SPLIT, :], slot_mask[TEST_TURN_SPLIT:, :]]
                tmp_update_current_mm = [update_current_mm[:TEST_TURN_SPLIT, :, :TEST_TURN_SPLIT],
                                         update_current_mm[TEST_TURN_SPLIT:, :, TEST_TURN_SPLIT:]]
                tmp_slot_all_connect = [slot_all_connect[:TEST_TURN_SPLIT, :, :],
                                        slot_all_connect[TEST_TURN_SPLIT:, :, :]]
                tmp_update_mm = [update_mm[:TEST_TURN_SPLIT, :, :TEST_TURN_SPLIT],
                                 update_mm[TEST_TURN_SPLIT:, :, TEST_TURN_SPLIT:]]
                tmp_slot_domain_connect = [slot_domain_connect[:TEST_TURN_SPLIT, :, :],
                                           slot_domain_connect[TEST_TURN_SPLIT:, :, :]]
                tmp_max_value = [max_value, max_value]
                tmp_op_ids = [op_ids[:TEST_TURN_SPLIT, :], op_ids[TEST_TURN_SPLIT:, :]]
                tmp_max_update = [max_update, max_update]
                tmp_position_ids = [position_ids[:TEST_TURN_SPLIT, :], position_ids[TEST_TURN_SPLIT:, :]]
                tmp_sample_mm = [sample_mm[:TEST_TURN_SPLIT, :, :], sample_mm[TEST_TURN_SPLIT:, :, :]]

                for cnt in range(2):
                    with torch.no_grad():
                        start_logits, end_logits, gen_scores, _, _, _, fuse_score, input_ids, _ = model(
                        input_ids=tmp_input_ids[cnt],
                        token_type_ids=tmp_segment_ids[cnt],
                        state_positions=tmp_state_position_ids[cnt],
                        attention_mask=tmp_input_mask[cnt],
                        slot_mask=tmp_slot_mask[cnt],
                        first_edge_mask=tmp_update_current_mm[cnt],
                        second_edge_mask=tmp_slot_all_connect[cnt],
                        third_edge_mask=tmp_update_mm[cnt],
                        fourth_edge_mask=tmp_slot_domain_connect[cnt],
                        max_value=tmp_max_value[cnt],
                        op_ids=tmp_op_ids[cnt],
                        max_update=tmp_max_update[cnt],
                        position_ids=tmp_position_ids[cnt],
                        sample_mm=tmp_sample_mm[cnt])
                    start_predictions += start_logits.argmax(dim=-1).cpu().detach().numpy().tolist()
                    end_predictions += end_logits.argmax(dim=-1).cpu().detach().numpy().tolist()
                    gen_predictions += gen_scores.argmax(dim=-1).cpu().detach().numpy().tolist()
                    fuse_scores += fuse_score.argmax(dim=-1).cpu().detach().numpy().tolist()
                    all_input_ids += input_ids[:, :, 1:].cpu().detach().numpy().tolist()

                    del start_logits, end_logits, gen_scores, _, fuse_score, input_ids,
                    torch.cuda.empty_cache()

        if (step + 1) % 50 == 0:
            if turn == 2:
                gen_acc, op_acc, opguess, opgold, opcorrect, gen_correct, gen_guess, cate_slot_correct_b, nocate_slot_correct_b, catecorrect_b, noncatecorrect_b, domain_correct_b, joint_cor, sample_l = joint_evaluation(
                    start_predictions,
                    end_predictions,
                    gen_predictions,
                    has_ans_predictions,
                    start_ids, end_ids,
                    gen_labels, gen_id_labels,
                    has_ans_labels,
                    all_input_ids,
                    ans_vocab_nd, ref_scores,
                    fuse_scores,
                    gold_ans_labels,
                    tokenizer=tokenizer,
                    score_diffs=None,
                    cate_score_diffs=None,
                    score_noanses=None,
                    sketchy_weight=None,
                    verify_weight=None,
                    sid=sample_ids,
                    catemask=cate_mask,
                    ontology=ontology)
                gen_correct += gen_correct
                gen_guess += gen_guess
                joint_correct += joint_cor
                samples += sample_l
                update_gold += opgold
                update_guess += opguess
                update_correct += opcorrect
                catecorrect += catecorrect_b
                noncatecorrect += noncatecorrect_b
                cate_slot_correct += cate_slot_correct_b
                nocate_slot_correct += nocate_slot_correct_b
                for key in domain_joint.keys():
                    domain_joint[key] += domain_correct_b[key]

            start_predictions = []
            start_ids = []
            end_predictions = []
            end_ids = []
            has_ans_predictions = []
            has_ans_labels = []
            gen_predictions = []
            gen_labels = []
            score_diffs = []
            cate_score_diffs = []
            score_noanses = []
            all_input_ids = []
            sample_ids = []
            ref_scores = []
            fuse_scores = []
            gen_id_labels = []
            gold_ans_labels = []
    # print(gen_correct / gen_guess)
    # print(update_correct/update_gold)
    # print(update_correct/update_guess)
    # print(cate_slot_correct/(samples * n_slot))
    # print(nocate_slot_correct/(samples * n_slot))
    # print(catecorrect/samples)
    # print(noncatecorrect/samples)
    # for key in domain_joint.keys():
    #     print("{}, {}".format(key, domain_joint[key]/samples))
    # print(joint_correct/samples)
    return joint_correct / samples, catecorrect / samples, noncatecorrect / samples


if __name__ == "__main__":
    main()