# coding: UTF-8
# Copyright 2022 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import argparse
import time
import math
import os, sys
import itertools
import numpy as np

import torch
import torch.nn as nn
import torch.optim as optim
import torch.multiprocessing as mp
from data_utils import get_lm_corpus
from mem_transformer import MemTransformerLM
from utils.exp_utils import create_exp_dir
from utils.data_parallel import BalancedDataParallel
from apex import amp
import torch.distributed as dist
import apex
import warnings


parser = argparse.ArgumentParser(description='PyTorch Transformer Language Model')
parser.add_argument('--data', type=str, default='../data/enwik8',
                    help='location of the data corpus')
parser.add_argument('--dataset', type=str, default='enwik8',
                    choices=['wt103', 'lm1b', 'enwik8', 'text8'],
                    help='dataset name')
parser.add_argument('--n_layer', type=int, default=12,
                    help='number of total layers')
parser.add_argument('--n_head', type=int, default=8,
                    help='number of heads')
parser.add_argument('--d_head', type=int, default=64,
                    help='head dimension')
parser.add_argument('--d_embed', type=int, default=-1,
                    help='embedding dimension')
parser.add_argument('--d_model', type=int, default=512,
                    help='model dimension')
parser.add_argument('--d_inner', type=int, default=2048,
                    help='inner dimension in FF')
parser.add_argument('--dropout', type=float, default=0.1,
                    help='global dropout rate')
parser.add_argument('--dropatt', type=float, default=0.0,
                    help='attention probability dropout rate')
parser.add_argument('--init', default='normal', type=str,
                    help='parameter initializer to use.')
parser.add_argument('--emb_init', default='normal', type=str,
                    help='parameter initializer to use.')
parser.add_argument('--init_range', type=float, default=0.1,
                    help='parameters initialized by U(-init_range, init_range)')
parser.add_argument('--emb_init_range', type=float, default=0.01,
                    help='parameters initialized by U(-init_range, init_range)')
parser.add_argument('--init_std', type=float, default=0.02,
                    help='parameters initialized by N(0, init_std)')
parser.add_argument('--proj_init_std', type=float, default=0.01,
                    help='parameters initialized by N(0, init_std)')
parser.add_argument('--optim', default='adam', type=str,
                    choices=['adam', 'sgd', 'adagrad'],
                    help='optimizer to use.')
parser.add_argument('--lr', type=float, default=0.00025,
                    help='initial learning rate (0.00025|5 for adam|sgd)')
parser.add_argument('--mom', type=float, default=0.0,
                    help='momentum for sgd')
parser.add_argument('--scheduler', default='cosine', type=str,
                    choices=['cosine', 'inv_sqrt', 'dev_perf', 'constant'],
                    help='lr scheduler to use.')
parser.add_argument('--warmup_step', type=int, default=0,
                    help='upper epoch limit')
parser.add_argument('--decay_rate', type=float, default=0.5,
                    help='decay factor when ReduceLROnPlateau is used')
parser.add_argument('--lr_min', type=float, default=0.0,
                    help='minimum learning rate during annealing')
parser.add_argument('--clip', type=float, default=0.25,         # 源码中 clip 的 default=0.25
                    help='gradient clipping')
parser.add_argument('--clip_nonemb', action='store_true',
                    help='only clip the gradient of non-embedding params')
parser.add_argument('--max_step', type=int, default=400000,
                    help='upper epoch limit')
parser.add_argument('--batch_size', type=int, default=22,
                    help='batch size')
parser.add_argument('--batch_chunk', type=int, default=1,
                    help='split batch into chunks to save memory')
parser.add_argument('--tgt_len', type=int, default=512,
                    help='number of tokens to predict')
parser.add_argument('--eval_tgt_len', type=int, default=128,
                    help='number of tokens to predict for evaluation')
parser.add_argument('--ext_len', type=int, default=0,
                    help='length of the extended context')
parser.add_argument('--mem_len', type=int, default=512,
                    help='length of the retained previous heads')
parser.add_argument('--not_tied', action='store_true',
                    help='do not tie the word embedding and softmax weights')
parser.add_argument('--seed', type=int, default=1111,
                    help='random seed')
# parser.add_argument('--npu', default=True, help='use NPU')
parser.add_argument('--adaptive', action='store_true',
                    help='use adaptive softmax')
parser.add_argument('--div_val', type=int, default=1,
                    help='divident value for adapative input and softmax')
parser.add_argument('--pre_lnorm', action='store_true',
                    help='apply LayerNorm to the input instead of the output')
parser.add_argument('--varlen', action='store_true',
                    help='use variable length')
parser.add_argument('--multi_gpu', action='store_true',
                    help='use multiple GPU')
parser.add_argument('--log-interval', type=int, default=200,
                    help='report interval')
parser.add_argument('--eval-interval', type=int, default=4000,
                    help='evaluation interval')
parser.add_argument('--work_dir', default='LM-TFM', type=str,
                    help='experiment directory.')
parser.add_argument('--restart', action='store_true',
                    help='restart training from the saved checkpoint')
parser.add_argument('--restart_dir', type=str, default='',
                    help='restart dir')
parser.add_argument('--debug', action='store_true',
                    help='run in debug mode (do not create exp dir)')
parser.add_argument('--same_length', action='store_true',
                    help='use the same attn length for all tokens')
parser.add_argument('--attn_type', type=int, default=0,
                    help='attention type. 0 for ours, 1 for Shaw et al,'
                    '2 for Vaswani et al, 3 for Al Rfou et al.')
parser.add_argument('--clamp_len', type=int, default=-1,
                    help='use the same pos embeddings after clamp_len')
parser.add_argument('--eta_min', type=float, default=0.0,
                    help='min learning rate for cosine scheduler')
parser.add_argument('--gpu0_bsz', type=int, default=-1,
                    help='batch size on gpu 0')
parser.add_argument('--max_eval_steps', type=int, default=-1,
                    help='max eval steps')
parser.add_argument('--sample_softmax', type=int, default=-1,
                    help='number of samples in sampled softmax')
parser.add_argument('--patience', type=int, default=0,
                    help='patience')
parser.add_argument('--finetune_v2', action='store_true',
                    help='finetune v2')
parser.add_argument('--finetune_v3', action='store_true',
                    help='finetune v3')
parser.add_argument('--static-loss-scale', type=float, default=128.0,
                    help='Static loss scale, positive power of 2 values can '
                    'improve fp16 convergence.')
parser.add_argument('--dynamic-loss-scale', action='store_true',
                    help='Use dynamic loss scaling.  If supplied, this argument'
                    ' supersedes --static-loss-scale.')
#edit this for 8p
parser.add_argument('--dist-backend', type=str, default='hccl')
parser.add_argument('--world-size', type=int, default=-1)
parser.add_argument('--rank', type=int, default=-1)
parser.add_argument('--local_rank', type=int, default=0)
parser.add_argument('--addr', type=str, default='127.0.0.1')
parser.add_argument('--device_num', type=int, default=-1)
parser.add_argument('--workers', type=int, default=32)
parser.add_argument('--device-list', default='', type=str)
parser.add_argument('--dist-url', type=str, default='tcp://127.0.0.1:50000')
parser.add_argument('--device', type=str, default='npu')
parser.add_argument('--gpu', default=None, type=int, help='GPU id to use.')
parser.add_argument('--multiprocessing-distributed', action='store_true',
                    help='Use multi-processing distributed training to launch '
                         'N processes per node, which has N GPUs. This is the '
                         'fastest way to use PyTorch for either single node or '
                         'multi node data parallel training')
warnings.filterwarnings('ignore')
#############end#################

def main():
    args = parser.parse_args()
    args.tied = not args.not_tied
    torch.manual_seed(args.seed)

    global train_step, train_loss, best_val_loss, eval_start_time, log_start_time
    ##############################
    # edit this for 8p
    os.environ['MASTER_ADDR'] = args.addr
    os.environ['MASTER_PORT'] = '29888'
    os.environ['LOCAL_DEVICE_ID'] = str(0)
    print("+++++++++++++++++++++++++++LOCAL_DEVICE_ID:", os.environ['LOCAL_DEVICE_ID'])
    if args.dist_url == "env://" and args.world_size == -1:
        args.world_size = int(os.environ["WORLD_SIZE"])
    args.distributed = args.world_size > 1 or args.multiprocessing_distributed
    if args.device_list != '':
        ngpus_per_node = len(args.device_list.split(','))
    elif args.device_num != -1:
        ngpus_per_node = args.device_num
    elif args.device == 'npu':
        ngpus_per_node = int(os.environ["RANK_SIZE"])
    else:
        ngpus_per_node = torch.cuda.device_count()
    if args.multiprocessing_distributed:
        args.world_size = ngpus_per_node * args.world_size
        if args.device == 'npu':
           main_worker(args.local_rank, ngpus_per_node,args) 
    else:
        main_worker(args.gpu, ngpus_per_node, args)
    ##############################


def main_worker(gpu, ngpus_per_node, args):

    global train_step, train_loss, best_val_loss, eval_start_time, log_start_time
    if args.d_embed < 0:
        args.d_embed = args.d_model

    assert args.ext_len >= 0, 'extended context length must be non-negative'
    assert args.batch_size % args.batch_chunk == 0

    args.work_dir = '{}-{}'.format(args.work_dir, args.dataset)
    args.work_dir = os.path.join(args.work_dir, time.strftime('%Y%m%d-%H%M%S'))
    logging = create_exp_dir(args.work_dir,
        scripts_to_save=['train.py', 'mem_transformer.py'], debug=args.debug)

    if args.device_list != '':
        args.gpu = int(args.device_list.split(',')[gpu])
    else:
        args.gpu = gpu

    print("[npu id:", args.gpu, "]", "++++++++++++++++ before set LOCAL_DEVICE_ID:", os.environ['LOCAL_DEVICE_ID'])
    os.environ['LOCAL_DEVICE_ID'] = str(args.gpu)
    print("[npu id:", args.gpu, "]", "++++++++++++++++ LOCAL_DEVICE_ID:", os.environ['LOCAL_DEVICE_ID'])

    if args.gpu is not None:
        print("[npu id:", args.gpu, "]", "Use GPU: {} for training".format(args.gpu))

    if args.distributed:
        if args.dist_url == "env://" and args.rank == -1:
            args.rank = int(os.environ["RANK"])
        if args.multiprocessing_distributed:
            args.rank = args.rank * ngpus_per_node + gpu

        if args.device == 'npu':
            dist.init_process_group(backend=args.dist_backend,
                                    world_size=args.world_size, rank=args.rank)
        else:
            dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
                                    world_size=args.world_size, rank=args.rank)

    loc = 'npu:{}'.format(args.gpu)
    torch.npu.set_device(loc)

    args.workers = int((args.workers + ngpus_per_node - 1) / ngpus_per_node)

    print("[npu id:", args.gpu, "]", "===============main_worker()=================")
    print("[npu id:", args.gpu, "]", args)
    print("[npu id:", args.gpu, "]", "===============main_worker()=================")


    ###############################################################################
    # Load data
    ###############################################################################
    corpus = get_lm_corpus(args.data, args.dataset)
    ntokens = len(corpus.vocab)
    args.n_token = ntokens

    eval_batch_size = 10
    tr_iter = corpus.get_iterator('train', args.batch_size, args.tgt_len,
        device=loc, ext_len=args.ext_len)
    va_iter = corpus.get_iterator('valid', eval_batch_size, args.eval_tgt_len,
        device=loc, ext_len=args.ext_len)
    te_iter = corpus.get_iterator('test.py', eval_batch_size, args.eval_tgt_len,
        device=loc, ext_len=args.ext_len)

    # adaptive softmax / embedding
    cutoffs, tie_projs = [], [False]
    if args.adaptive:
        assert args.dataset in ['wt103', 'lm1b']
        if args.dataset == 'wt103':
            cutoffs = [20000, 40000, 200000]
            tie_projs += [True] * len(cutoffs)
        elif args.dataset == 'lm1b':
            cutoffs = [60000, 100000, 640000]
            tie_projs += [False] * len(cutoffs)

    ###############################################################################
    # Build the model
    ###############################################################################
    def init_weight(weight):
        if args.init == 'uniform':
            nn.init.uniform_(weight, -args.init_range, args.init_range)
        elif args.init == 'normal':
            nn.init.normal_(weight, 0.0, args.init_std)

    def init_bias(bias):
        nn.init.constant_(bias, 0.0)

    def weights_init(m):
        classname = m.__class__.__name__
        if classname.find('Linear') != -1:
            if hasattr(m, 'weight') and m.weight is not None:
                init_weight(m.weight)
            if hasattr(m, 'bias') and m.bias is not None:
                init_bias(m.bias)
        elif classname.find('AdaptiveEmbedding') != -1:
            if hasattr(m, 'emb_projs'):
                for i in range(len(m.emb_projs)):
                    if m.emb_projs[i] is not None:
                        nn.init.normal_(m.emb_projs[i], 0.0, args.proj_init_std)
        elif classname.find('Embedding') != -1:
            if hasattr(m, 'weight'):
                init_weight(m.weight)
        elif classname.find('ProjectedAdaptiveLogSoftmax') != -1:
            if hasattr(m, 'cluster_weight') and m.cluster_weight is not None:
                init_weight(m.cluster_weight)
            if hasattr(m, 'cluster_bias') and m.cluster_bias is not None:
                init_bias(m.cluster_bias)
            if hasattr(m, 'out_projs'):
                for i in range(len(m.out_projs)):
                    if m.out_projs[i] is not None:
                        nn.init.normal_(m.out_projs[i], 0.0, args.proj_init_std)
        elif classname.find('LayerNorm') != -1:
            if hasattr(m, 'weight'):
                nn.init.normal_(m.weight, 1.0, args.init_std)
            if hasattr(m, 'bias') and m.bias is not None:
                init_bias(m.bias)
        elif classname.find('TransformerLM') != -1:
            if hasattr(m, 'r_emb'):
                init_weight(m.r_emb)
            if hasattr(m, 'r_w_bias'):
                init_weight(m.r_w_bias)
            if hasattr(m, 'r_r_bias'):
                init_weight(m.r_r_bias)
            if hasattr(m, 'r_bias'):
                init_bias(m.r_bias)

    def update_dropout(m):
        classname = m.__class__.__name__
        if classname.find('Dropout') != -1:
            if hasattr(m, 'p'):
                m.p = args.dropout

    def update_dropatt(m):
        if hasattr(m, 'dropatt'):
            m.dropatt.p = args.dropatt

    if args.restart:
        with open(os.path.join(args.restart_dir, 'model.pt'), 'rb') as f:
            model = MemTransformerLM(ntokens, args.n_layer, args.n_head, args.d_model,
                                     args.d_head, args.d_inner, args.dropout, args.dropatt,
                                     tie_weight=args.tied, d_embed=args.d_embed, div_val=args.div_val,
                                     tie_projs=tie_projs, pre_lnorm=args.pre_lnorm, tgt_len=args.tgt_len,
                                     ext_len=args.ext_len, mem_len=args.mem_len, cutoffs=cutoffs,
                                     same_length=args.same_length, attn_type=args.attn_type,
                                     clamp_len=args.clamp_len, sample_softmax=args.sample_softmax)
            model.apply(weights_init)
            model.word_emb.apply(weights_init)
            model = model.to(loc)
            ckpt = torch.load(f, map_location=loc)
            model.load_state_dict(ckpt)
        model.apply(update_dropout)
        model.apply(update_dropatt)
    else:
        model = MemTransformerLM(ntokens, args.n_layer, args.n_head, args.d_model,
            args.d_head, args.d_inner, args.dropout, args.dropatt,
            tie_weight=args.tied, d_embed=args.d_embed, div_val=args.div_val,
            tie_projs=tie_projs, pre_lnorm=args.pre_lnorm, tgt_len=args.tgt_len,
            ext_len=args.ext_len, mem_len=args.mem_len, cutoffs=cutoffs,
            same_length=args.same_length, attn_type=args.attn_type,
            clamp_len=args.clamp_len, sample_softmax=args.sample_softmax)
        model.apply(weights_init)
        model.word_emb.apply(weights_init) # ensure embedding init is not overridden by out_layer in case of weight sharing



    args.n_all_param = sum([p.nelement() for p in model.parameters()])
    args.n_nonemb_param = sum([p.nelement() for p in model.layers.parameters()])



    #### optimizer
    if args.optim.lower() == 'sgd':
        if args.sample_softmax > 0:
            dense_params, sparse_params = [], []
            for param in model.parameters():
                if param.size() == model.word_emb.weight.size():
                    sparse_params.append(param)
                else:
                    dense_params.append(param)
            optimizer_sparse = optim.SGD(sparse_params, lr=args.lr * 2)
            optimizer = optim.SGD(dense_params, lr=args.lr, momentum=args.mom)
        else:
            optimizer = optim.SGD(model.parameters(), lr=args.lr,
                momentum=args.mom)
    elif args.optim.lower() == 'adam':
        if args.sample_softmax > 0:
            dense_params, sparse_params = [], []
            for param in model.parameters():
                if param.size() == model.word_emb.weight.size():
                    sparse_params.append(param)
                else:
                    dense_params.append(param)
            optimizer_sparse = optim.SparseAdam(sparse_params, lr=args.lr)
            optimizer = optim.Adam(dense_params, lr=args.lr)
        else:
            #optimizer = optim.Adam(model.parameters(), lr=args.lr)
            optimizer = apex.optimizers.NpuFusedAdam(model.parameters(), lr=args.lr)
    elif args.optim.lower() == 'adagrad':
        optimizer = optim.Adagrad(model.parameters(), lr=args.lr)

    model = model.to(loc)
    ###################################################################################################
    opt_level = "O2"
    model, optimizer = amp.initialize(model, optimizer, opt_level=opt_level, loss_scale=128.0, combine_grad=True)
    ###################################################################################################

    if args.multi_gpu:

        if args.gpu0_bsz >= 0:
            para_model = BalancedDataParallel(args.gpu0_bsz // args.batch_chunk,
                                              model, dim=1).to(loc)
        else:
            para_model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu], broadcast_buffers=False)
    else:
        para_model = model.to(loc)

    #### scheduler
    if args.scheduler == 'cosine':
        scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer,
            args.max_step, eta_min=args.eta_min)
        if args.sample_softmax > 0:
            scheduler_sparse = optim.lr_scheduler.CosineAnnealingLR(optimizer_sparse,
                args.max_step, eta_min=args.eta_min)
    elif args.scheduler == 'inv_sqrt':
        def lr_lambda(step):
            if step == 0 and args.warmup_step == 0:
                return 1.
            else:
                return 1. / (step ** 0.5) if step > args.warmup_step \
                       else step / (args.warmup_step ** 1.5)
        scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lr_lambda)
    elif args.scheduler == 'dev_perf':
        scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
            factor=args.decay_rate, patience=args.patience, min_lr=args.lr_min)
        if args.sample_softmax > 0:
            scheduler_sparse = optim.lr_scheduler.ReduceLROnPlateau(optimizer_sparse,
                factor=args.decay_rate, patience=args.patience, min_lr=args.lr_min)
    elif args.scheduler == 'constant':
        pass


    if args.restart:
        if os.path.exists(os.path.join(args.restart_dir, 'optimizer.pt')):
            with open(os.path.join(args.restart_dir, 'optimizer.pt'), 'rb') as f:
                opt_state_dict = torch.load(f, map_location=loc)
                optimizer.load_state_dict(opt_state_dict)
        else:
            print('Optimizer was not saved. Start from scratch.')

    logging('=' * 100)
    for k, v in args.__dict__.items():
        logging('    - {} : {}'.format(k, v))
    logging('=' * 100)
    logging('#params = {}'.format(args.n_all_param))
    logging('#non emb params = {}'.format(args.n_nonemb_param))

    ###############################################################################
    # Training code
    ###############################################################################

    def evaluate(eval_iter):
        model.eval()
        if args.mem_len == 0:
            model.reset_length(args.eval_tgt_len,
                               args.ext_len+args.tgt_len-args.eval_tgt_len, args.mem_len)
        else:
            model.reset_length(args.eval_tgt_len,
                               args.ext_len, args.mem_len+args.tgt_len-args.eval_tgt_len)

        # Evaluation
        total_len, total_loss = 0, 0.
        with torch.no_grad():
            mems = tuple()
            for i, (data, target, seq_len) in enumerate(eval_iter):
                if args.max_eval_steps > 0 and i >= args.max_eval_steps:
                    break
                ret = model(data, target, *mems)
                loss, mems = ret[0], ret[1:]
                loss = loss.mean()
                total_loss += seq_len * loss.float().item()
                total_len += seq_len

        model.reset_length(args.tgt_len, args.ext_len, args.mem_len)
        model.train()
        return total_loss / total_len


    def train():
        # Turn on training mode which enables dropout.
        global train_step, train_loss, best_val_loss, eval_start_time, log_start_time

        model.train()
        if args.batch_chunk > 1:
            mems = [tuple() for _ in range(args.batch_chunk)]
        else:
            mems = tuple()
        train_iter = tr_iter.get_varlen_iter() if args.varlen else tr_iter
        for batch, (data, target, seq_len) in enumerate(train_iter):
            model.zero_grad()
            if args.batch_chunk > 1:
                data_chunks = torch.chunk(data, args.batch_chunk, 1)
                target_chunks = torch.chunk(target, args.batch_chunk, 1)
                for i in range(args.batch_chunk):
                    data_i = data_chunks[i].contiguous()
                    target_i = target_chunks[i].contiguous()
                    ret = para_model(data_i, target_i, *mems[i])
                    loss, mems[i] = ret[0], ret[1:]
                    loss = loss.float().mean().type_as(loss) / args.batch_chunk
                    ####################################################################
                    with amp.scale_loss(loss, optimizer) as scaled_loss:
                        scaled_loss.backward()
                    ####################################################################
                    with torch.no_grad():
                        train_loss += loss.float().bool().item()
            else:
               ret = para_model(data, target, *mems)
               loss, mems = ret[0], ret[1:]
               loss = loss.float().mean().type_as(loss)
               ####################################################
               with torch.no_grad():
                   train_loss += loss.float().item()
               ###################################################################
               with amp.scale_loss(loss, optimizer) as scaled_loss:
                   scaled_loss.backward()


            optimizer.step()
            if args.sample_softmax > 0:
                optimizer_sparse.step()

            # step-wise learning rate annealing
            train_step += 1
            if args.scheduler in ['cosine', 'constant', 'dev_perf']:
                # linear warmup stage
                if train_step < args.warmup_step:
                    curr_lr = args.lr * train_step / args.warmup_step
                    optimizer.param_groups[0]['lr'] = curr_lr
                    if args.sample_softmax > 0:
                        optimizer_sparse.param_groups[0]['lr'] = curr_lr * 2
                else:
                    if args.scheduler == 'cosine':
                        scheduler.step(train_step)
                        if args.sample_softmax > 0:
                            scheduler_sparse.step(train_step)
            elif args.scheduler == 'inv_sqrt':
                scheduler.step(train_step)

            if train_step % args.log_interval == 0:
                cur_loss = train_loss / args.log_interval
                elapsed = time.time() - log_start_time
                log_str = '| epoch {:3d} step {:>8d} | {:>6d} batches | lr {:.3g} ' \
                          '| ms/batch {:5.2f} | loss {:5.2f} | fps {:.2f}'.format(
                    epoch, train_step, batch+1, optimizer.param_groups[0]['lr'],
                    elapsed * 1000 / args.log_interval, cur_loss, args.log_interval*args.batch_size*args.tgt_len*8/elapsed)
                if args.dataset in ['enwik8', 'text8']:
                    log_str += ' | bpc {:9.5f}'.format(cur_loss / math.log(2))
                else:
                    log_str += ' | ppl {:9.3f}'.format(math.exp(cur_loss))
                logging(log_str)
                train_loss = 0
                log_start_time = time.time()

            if train_step % args.eval_interval == 0:
                print('train_step is :', train_step)
                print('ars.eval_interval is :', args.eval_interval)
                print(train_step % args.eval_interval)
                print('*'*50)
                ts = time.time()
                val_loss = evaluate(va_iter)
                print('evaluation use time {} s'.format(time.time()-ts))
                logging('-' * 100)
                log_str = '| Eval {:3d} at step {:>8d} | time: {:5.2f}s ' \
                          '| valid loss {:5.2f}'.format(
                    train_step // args.eval_interval, train_step,
                    (time.time() - ts), val_loss)
                if args.dataset in ['enwik8', 'text8']:
                    log_str += ' | bpc {:9.5f}'.format(val_loss / math.log(2))
                else:
                    log_str += ' | valid ppl {:9.3f}'.format(math.exp(val_loss))
                logging(log_str)
                logging('-' * 100)
                # Save the model if the validation loss is the best we've seen so far.
                if not best_val_loss or val_loss < best_val_loss:
                    if not args.debug:
                        with open('model.pt', 'wb') as f:
                            torch.save(model.state_dict(), f)
                        with open('optimizer.pt', 'wb') as f:
                            torch.save(optimizer.state_dict(), f)
                    best_val_loss = val_loss

                # dev-performance based learning rate annealing
                if args.scheduler == 'dev_perf':
                    scheduler.step(val_loss)
                    if args.sample_softmax > 0:
                        scheduler_sparse.step(val_loss)

                eval_start_time = time.time()

            if train_step == args.max_step:
                sys.exit()    

    # At any point you can hit Ctrl + C to break out of training early.
    try:
        for epoch in itertools.count(start=1):
            train()
            if train_step == args.max_step:
                logging('-' * 100)
                logging('End of training')
                sys.exit() 
    except KeyboardInterrupt:
        logging('-' * 100)
        logging('Exiting from training early')

    # # Load the best saved model.
    # with open('model.pt', 'rb') as f:
    #     model.load_state_dict(torch.load(f, map_location=loc))
    # para_model = model.to(loc)

    # # Run on test data.
    # test_loss = evaluate(te_iter)
    # logging('=' * 100)
    # if args.dataset in ['enwik8', 'text8']:
    #     logging('| End of training | test loss {:5.2f} | test bpc {:9.5f}'.format(
    #         test_loss, test_loss / math.log(2)))
    # else:
    #     logging('| End of training | test loss {:5.2f} | test ppl {:9.3f}'.format(
    #         test_loss, math.exp(test_loss)))
    # logging('=' * 100)


if __name__ == '__main__':
    global train_step, train_loss, best_val_loss, eval_start_time, log_start_time
    train_step = 0
    train_loss = 0
    best_val_loss = None
    log_start_time = time.time()
    eval_start_time = time.time()
    main()