task_run.py

# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# 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.
"""
Fine-tuning the library models for language modeling on a text file (GPT, GPT-2, BERT, RoBERTa).
GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa are fine-tuned
using a masked language modeling (MLM) loss.
"""

from __future__ import absolute_import
import os
import sys
import pickle
import torch
import json
# import bleu_sum
import random
import logging
import bleu
import argparse
import numpy as np
from io import open
from itertools import cycle
import torch.nn as nn
from model import Seq2Seq
from tqdm import tqdm, trange
# from bleu import _bleu
from torch.utils.data import DataLoader, Dataset, SequentialSampler, RandomSampler,TensorDataset
from torch.utils.data.distributed import DistributedSampler
from transformers import (WEIGHTS_NAME, AdamW, get_linear_schedule_with_warmup,
                          RobertaConfig, RobertaModel, RobertaTokenizer)
MODEL_CLASSES = {'roberta': (RobertaConfig, RobertaModel, RobertaTokenizer)}

torch.multiprocessing.set_sharing_strategy('file_system')

logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
                    datefmt = '%m/%d/%Y %H:%M:%S',
                    level = logging.INFO)
from parser import DFG_python,DFG_java,DFG_ruby,DFG_go,DFG_php,DFG_javascript
from parser import (remove_comments_and_docstrings,
                   tree_to_token_index,
                   index_to_code_token,
                   tree_to_variable_index)
from tree_sitter import Language, Parser

import torch.multiprocessing
torch.multiprocessing.set_sharing_strategy('file_system')

from adapter_utils import AdapterUtils
logger = logging.getLogger(__name__)

adapter_utils = AdapterUtils()

dfg_function={
    'python':DFG_python,
    'java':DFG_java,
    'ruby':DFG_ruby,
    'go':DFG_go,
    'php':DFG_php,
    'javascript':DFG_javascript
}

#load parsers
parsers={}        
for lang in dfg_function:
    LANGUAGE = Language('parser/my-languages.so', lang)
    parser = Parser()
    parser.set_language(LANGUAGE) 
    parser = [parser,dfg_function[lang]]    
    parsers[lang]= parser

#remove comments, tokenize code and extract dataflow     
def extract_dataflow(code, parser,lang):
    #remove comments
    try:
        code=remove_comments_and_docstrings(code,lang)
    except:
        pass    
    #obtain dataflow
    if lang=="php":
        code="<?php"+code+"?>"    
    try:
        tree = parser[0].parse(bytes(code,'utf8'))    
        root_node = tree.root_node  
        tokens_index=tree_to_token_index(root_node)     
        code=code.split('\n')
        code_tokens=[index_to_code_token(x,code) for x in tokens_index]  
        index_to_code={}
        for idx,(index,code) in enumerate(zip(tokens_index,code_tokens)):
            index_to_code[index]=(idx,code)  
        try:
            DFG,_=parser[1](root_node,index_to_code,{}) 
        except:
            DFG=[]
        DFG=sorted(DFG,key=lambda x:x[1])
        indexs=set()
        for d in DFG:
            if len(d[-1])!=0:
                indexs.add(d[1])
            for x in d[-1]:
                indexs.add(x)
        new_DFG=[]
        for d in DFG:
            if d[1] in indexs:
                new_DFG.append(d)
        dfg=new_DFG
    except:
        dfg=[]
    return code_tokens,dfg


class Example(object):
    """A single training/test example."""
    def __init__(self,
                 idx,
                 source,
                 target,
                 ):
        self.idx = idx
        self.source = source
        self.target = target

def read_examples(filename):
    """Read examples from filename."""
    examples=[]
    with open(filename,encoding="utf-8") as f:
        for idx, line in enumerate(f):
            line=line.strip()
            js=json.loads(line)
            if 'idx' not in js:
                js['idx']=idx
            code=' '.join(js['code_tokens']).replace('\n',' ')
            code=' '.join(code.strip().split())
            nl=' '.join(js['docstring_tokens']).replace('\n','')
            nl=' '.join(nl.strip().split())            
            examples.append(
                Example(
                        idx = idx,
                        source=code,
                        target = nl,
                        ) 
            )
    return examples


class InputFeatures(object):
    """A single training/test features for a example."""
    def __init__(self,
                 example_id,
                 source_ids,
                 position_idx,
                 dfg_to_code,
                 dfg_to_dfg,                 
                 target_ids,
                 source_mask,
                 target_mask,

    ):
        self.example_id = example_id
        self.source_ids = source_ids
        self.position_idx = position_idx
        self.dfg_to_code = dfg_to_code
        self.dfg_to_dfg = dfg_to_dfg
        self.target_ids = target_ids
        self.source_mask = source_mask
        self.target_mask = target_mask       
        

def convert_examples_to_features(examples, tokenizer, args,stage=None):
    features = []
    for example_index, example in enumerate(tqdm(examples,total=len(examples))):
        ##extract data flow
        code_tokens,dfg=extract_dataflow(example.source,parsers[args.lang],args.lang)
        code_tokens=[tokenizer.tokenize('@ '+x)[1:] if idx!=0 else tokenizer.tokenize(x) for idx,x in enumerate(code_tokens)]
        ori2cur_pos={}
        ori2cur_pos[-1]=(0,0)
        for i in range(len(code_tokens)):
            ori2cur_pos[i]=(ori2cur_pos[i-1][1],ori2cur_pos[i-1][1]+len(code_tokens[i]))    
        code_tokens=[y for x in code_tokens for y in x]  
        
        #truncating
        code_tokens=code_tokens[:args.max_source_length-3]
        source_tokens =[tokenizer.cls_token]+code_tokens+[tokenizer.sep_token]
        source_ids =  tokenizer.convert_tokens_to_ids(source_tokens)
        position_idx = [i+tokenizer.pad_token_id + 1 for i in range(len(source_tokens))]
        dfg=dfg[:args.max_source_length-len(source_tokens)]
        source_tokens+=[x[0] for x in dfg]
        position_idx+=[0 for x in dfg]
        source_ids+=[tokenizer.unk_token_id for x in dfg]
        padding_length=args.max_source_length-len(source_ids)
        position_idx+=[tokenizer.pad_token_id]*padding_length
        source_ids+=[tokenizer.pad_token_id]*padding_length      
        source_mask = [1] * (len(source_tokens))
        source_mask+=[0]*padding_length        
        
        #reindex
        reverse_index={}
        for idx,x in enumerate(dfg):
            reverse_index[x[1]]=idx
        for idx,x in enumerate(dfg):
            dfg[idx]=x[:-1]+([reverse_index[i] for i in x[-1] if i in reverse_index],)    
        dfg_to_dfg=[x[-1] for x in dfg]
        dfg_to_code=[ori2cur_pos[x[1]] for x in dfg]
        length=len([tokenizer.cls_token])
        dfg_to_code=[(x[0]+length,x[1]+length) for x in dfg_to_code]        

        #target
        if stage=="test":
            target_tokens = tokenizer.tokenize("None")
        else:
            target_tokens = tokenizer.tokenize(example.target)[:args.max_target_length-2]
        target_tokens = [tokenizer.cls_token]+target_tokens+[tokenizer.sep_token]            
        target_ids = tokenizer.convert_tokens_to_ids(target_tokens)
        target_mask = [1] *len(target_ids)
        padding_length = args.max_target_length - len(target_ids)
        target_ids+=[tokenizer.pad_token_id]*padding_length
        target_mask+=[0]*padding_length   
   
        if example_index < 5:
            if stage=='train':
                logger.info("*** Example ***")
                logger.info("source_tokens: {}".format([x.replace('\u0120','_') for x in source_tokens]))
                logger.info("source_ids: {}".format(' '.join(map(str, source_ids))))
                logger.info("source_mask: {}".format(' '.join(map(str, source_mask))))
                logger.info("position_idx: {}".format(position_idx))
                logger.info("dfg_to_code: {}".format(' '.join(map(str, dfg_to_code))))
                logger.info("dfg_to_dfg: {}".format(' '.join(map(str, dfg_to_dfg))))
                
                logger.info("target_tokens: {}".format([x.replace('\u0120','_') for x in target_tokens]))
                logger.info("target_ids: {}".format(' '.join(map(str, target_ids))))
                logger.info("target_mask: {}".format(' '.join(map(str, target_mask))))
       
        features.append(
            InputFeatures(
                 example_index,
                 source_ids,
                 position_idx,
                 dfg_to_code,
                 dfg_to_dfg,
                 target_ids,
                 source_mask,
                 target_mask,
            )
        )
    return features


class TextDataset(Dataset):
    def __init__(self, examples, args,file_path=None):
        self.examples = examples
        self.args=args

    def __len__(self):
        return len(self.examples)
    
    def __getitem__(self, item):
        #calculate graph-guided masked function
        attn_mask=np.zeros((self.args.max_source_length,self.args.max_source_length),dtype=np.bool)
        #calculate begin index of node and max length of input
        node_index=sum([i>1 for i in self.examples[item].position_idx])
        max_length=sum([i!=1 for i in self.examples[item].position_idx])
        #sequence can attend to sequence
        attn_mask[:node_index,:node_index]=True
        #special tokens attend to all tokens
        for idx,i in enumerate(self.examples[item].source_ids):
            if i in [0,2]:
                attn_mask[idx,:max_length]=True
        #nodes attend to code tokens that are identified from
        for idx,(a,b) in enumerate(self.examples[item].dfg_to_code):
            if a<node_index and b<node_index:
                attn_mask[idx+node_index,a:b]=True
                attn_mask[a:b,idx+node_index]=True
        #nodes attend to adjacent nodes         
        for idx,nodes in enumerate(self.examples[item].dfg_to_dfg):
            for a in nodes:
                if a+node_index<len(self.examples[item].position_idx):
                    attn_mask[idx+node_index,a+node_index]=True  
                    
        return (torch.tensor(self.examples[item].source_ids),
                torch.tensor(self.examples[item].source_mask),
                torch.tensor(self.examples[item].position_idx),
                torch.tensor(attn_mask), 
                torch.tensor(self.examples[item].target_ids),
                torch.tensor(self.examples[item].target_mask),)
    
def set_seed(seed=42):
    random.seed(seed)
    os.environ['PYHTONHASHSEED'] = str(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.backends.cudnn.deterministic = True
        
def main():
    parser = argparse.ArgumentParser()

    ## Required parameters  
    parser.add_argument("--model_type", default=None, type=str, required=True,
                        help="Model type: e.g. roberta")
    
    parser.add_argument("--lang", default=None, type=str, required=True,
                        help="lang")
    
    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
                        help="Path to pre-trained model: e.g. roberta-base" )   
    parser.add_argument("--output_dir", default=None, type=str, required=True,
                        help="The output directory where the model predictions and checkpoints will be written.")
    parser.add_argument("--load_model_path", default=None, type=str, 
                        help="Path to trained model: Should contain the .bin files" )    
    ## Other parameters
    parser.add_argument("--train_filename", default=None, type=str, 
                        help="The train filename. Should contain the .jsonl files for this task.")
    parser.add_argument("--dev_filename", default=None, type=str, 
                        help="The dev filename. Should contain the .jsonl files for this task.")
    parser.add_argument("--test_filename", default=None, type=str, 
                        help="The test filename. Should contain the .jsonl files for this task.")  
    
    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("--max_source_length", default=64, type=int,
                        help="The maximum total source sequence length after tokenization. Sequences longer "
                             "than this will be truncated, sequences shorter will be padded.")
    parser.add_argument("--max_target_length", default=32, type=int,
                        help="The maximum total target sequence length after tokenization. Sequences longer "
                             "than this will be truncated, sequences shorter will be padded.")
    
    parser.add_argument("--do_train", action='store_true',
                        help="Whether to run training.")
    parser.add_argument("--do_eval", action='store_true',
                        help="Whether to run eval on the dev set.")
    parser.add_argument("--do_test", action='store_true',
                        help="Whether to run eval on the dev set.")
    parser.add_argument("--do_lower_case", action='store_true',
                        help="Set this flag if you are using an uncased model.")
    parser.add_argument("--no_cuda", action='store_true',
                        help="Avoid using CUDA when available") 
    
    parser.add_argument("--train_batch_size", default=8, type=int,
                        help="Batch size per GPU/CPU for training.")
    parser.add_argument("--eval_batch_size", default=8, 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=5e-5, type=float,
                        help="The initial learning rate for Adam.")
    parser.add_argument("--beam_size", default=10, type=int,
                        help="beam size for beam search")    
    parser.add_argument("--weight_decay", default=0.0, type=float,
                        help="Weight deay 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=1.0, type=float,
                        help="Max gradient norm.")
    parser.add_argument("--num_train_epochs", default=3, type=int,
                        help="Total number of training epochs to perform.")
    parser.add_argument("--eval_steps", default=-1, type=int,
                        help="")
    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("--train_steps", default=-1, type=int,
                        help="")
    parser.add_argument("--warmup_steps", default=0, type=int,
                        help="Linear warmup over warmup_steps.")
    parser.add_argument("--local_rank", type=int, default=-1,
                        help="For distributed training: local_rank")   
    parser.add_argument('--seed', type=int, default=42,
                        help="random seed for initialization")
    
    parser.add_argument('--train_adapter_fusion', action='store_true', help="should call when adapters need to be trained")
    parser.add_argument('--load_adapter_fusion', action='store_true', help="should call when adapters need to be trained")
    parser.add_argument('--adapter_fusion_path', default="", type=str, help="should call when adapters need to be trained")

    # print arguments
    args = parser.parse_args()
    logger.info(args)

    # Setup CUDA, GPU
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    args.n_gpu = torch.cuda.device_count()
    args.device = device
    
    # Set seed
    set_seed(args.seed)
    
    # make dir if output_dir not exist
    if os.path.exists(args.output_dir) is False:
        os.makedirs(args.output_dir)
        
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,do_lower_case=args.do_lower_case)
    config = config_class.from_pretrained(args.model_name_or_path)

    
    #budild model
    encoder = model_class.from_pretrained(args.model_name_or_path,config=config)    

    adapter_utils.add_new_adapter(encoder,f'adapter_{args.lang}',adapter_config='lora')

    decoder_layer = nn.TransformerDecoderLayer(d_model=config.hidden_size, nhead=config.num_attention_heads)
    decoder = nn.TransformerDecoder(decoder_layer, num_layers=6)




    model=Seq2Seq(encoder=encoder,decoder=decoder,config=config,
                  beam_size=args.beam_size,max_length=args.max_target_length,
                  sos_id=tokenizer.cls_token_id,eos_id=tokenizer.sep_token_id)
    
    if args.load_model_path is not None:
        logger.info("reload model from {}".format(args.load_model_path))
        model.load_state_dict(torch.load(args.load_model_path))
        
    model.to(device)

    # try:
    #     from apex.parallel import DistributedDataParallel as DDP
    # except ImportError:
    #     raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")


    if args.n_gpu > 1:
        # multi-gpu training
        model = torch.nn.DataParallel(model)

    if args.do_train:
        # Prepare training data loader

        lang = 'multilingual'
        args.train_filename=f"data/code2nl/CodeSearchNet/{lang}/train.jsonl" 
        args.train_foldername=f"data/code2nl/CodeSearchNet/{lang}"
        prefix=args.train_filename.split('/')[-1][:-6]
        # THIS_DIR = os.path.dirname(args.train_filename)
        # UP_DIR = os.path.dirname(os.path.dirname(THIS_DIR))

        
        example_cache_file=args.train_foldername+'/'+prefix+'_examples.pkl'
        feature_cache_file=args.train_foldername+'/'+prefix+'_features.pkl'

        if os.path.exists(feature_cache_file):
            logger.info("***** loading features from cache *****")
            train_features=pickle.load(open(feature_cache_file,'rb'))
            train_examples=pickle.load(open(example_cache_file,'rb'))
        else:
            logger.info("***** start creating features *****")
            train_examples = read_examples(args.train_filename)
            train_features = convert_examples_to_features(train_examples, tokenizer,args,stage='train')
            pickle.dump(train_examples,open(example_cache_file,'wb'))
            pickle.dump(train_features,open(feature_cache_file,'wb'))



        # train_examples = read_examples(args.train_filename)
        # train_features = convert_examples_to_features(train_examples, tokenizer,args,stage='train')
        train_data = TextDataset(train_features,args)
        train_sampler = RandomSampler(train_data)
        train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size//args.gradient_accumulation_steps,num_workers=0)

        num_train_optimization_steps =  args.train_steps

        # Prepare optimizer and schedule (linear warmup and decay)
        no_decay = ['bias', 'LayerNorm.weight']
        optimizer_grouped_parameters = [
            {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
             'weight_decay': args.weight_decay},
            {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]
        optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
        # scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=len(train_dataloader)*args.num_train_epochs*0.1,num_training_steps=len(train_dataloader)*args.num_train_epochs)
        scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps,
                                                    num_training_steps=num_train_optimization_steps)
        #Start training
        logger.info("***** Running training *****")
        logger.info("  Num examples = %d", len(train_examples))
        logger.info("  Batch size = %d", args.train_batch_size)
        logger.info("  Num epoch = %d", num_train_optimization_steps*args.train_batch_size//len(train_examples))
        
        model.train()
        dev_dataset={}
        nb_tr_examples, nb_tr_steps,tr_loss,global_step,best_bleu,best_loss = 0, 0,0,0,0,1e6 

        bar = tqdm(range(num_train_optimization_steps),total=num_train_optimization_steps)
        train_dataloader=cycle(train_dataloader)
        eval_flag = True
        for step in bar:
            batch = next(train_dataloader)
            batch = tuple(t.to(device) for t in batch)
            source_ids,source_mask,position_idx,att_mask,target_ids,target_mask = batch
            loss,_,_ = model(source_ids,source_mask,position_idx,att_mask,target_ids,target_mask)

            if args.n_gpu > 1:
                loss = loss.mean() # mean() to average on multi-gpu.
            if args.gradient_accumulation_steps > 1:
                loss = loss / args.gradient_accumulation_steps
            tr_loss += loss.item()
            train_loss=round(tr_loss*args.gradient_accumulation_steps/(nb_tr_steps+1),4)
            bar.set_description(" loss {}".format(train_loss))
            nb_tr_examples += source_ids.size(0)
            nb_tr_steps += 1
            loss.backward()

            if (nb_tr_steps + 1) % args.gradient_accumulation_steps == 0:
                #Update parameters
                optimizer.step()
                optimizer.zero_grad()
                scheduler.step()
                global_step += 1
                eval_flag = True

            if args.do_eval and ((global_step + 1) %args.eval_steps == 0) and eval_flag:
                #Eval model with dev dataset
                tr_loss = 0
                nb_tr_examples, nb_tr_steps = 0, 0                     
                eval_flag=False    
                if 'dev_loss' in dev_dataset:
                    eval_examples,eval_data=dev_dataset['dev_loss']
                else:
                    eval_examples = read_examples(args.dev_filename)
                    eval_features = convert_examples_to_features(eval_examples, tokenizer, args,stage='dev')
                    eval_data = TextDataset(eval_features,args)
                    dev_dataset['dev_loss']=eval_examples,eval_data
                                      

                eval_sampler = SequentialSampler(eval_data)
                eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size,num_workers=0)

                logger.info("\n***** Running evaluation *****")
                logger.info("  Num examples = %d", len(eval_examples))
                logger.info("  Batch size = %d", args.eval_batch_size)

                #Start Evaling model
                model.eval()
                eval_loss,tokens_num = 0,0
                for batch in eval_dataloader:
                    batch = tuple(t.to(device) for t in batch)
                    source_ids,source_mask,position_idx,att_mask,target_ids,target_mask = batch
                    with torch.no_grad():
                        _,loss,num = model(source_ids,source_mask,position_idx,att_mask,target_ids,target_mask) 
 
                    eval_loss += loss.sum().item()
                    tokens_num += num.sum().item()
                #Pring loss of dev dataset    
                model.train()
                eval_loss = eval_loss / tokens_num
                result = {'eval_ppl': round(np.exp(eval_loss),5),
                          'global_step': global_step+1,
                          'train_loss': round(train_loss,5)}
                for key in sorted(result.keys()):
                    logger.info("  %s = %s", key, str(result[key]))
                logger.info("  "+"*"*20)   

                #save last checkpoint
                last_output_dir = os.path.join(args.output_dir, 'checkpoint-last')
                if not os.path.exists(last_output_dir):
                    os.makedirs(last_output_dir)
                model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self
                output_model_file = os.path.join(last_output_dir, "pytorch_model.bin")
                torch.save(model_to_save.state_dict(), output_model_file)                    
                if eval_loss<best_loss:
                    logger.info("  Best ppl:%s",round(np.exp(eval_loss),5))
                    logger.info("  "+"*"*20)
                    best_loss=eval_loss
                    # Save best checkpoint for best ppl
                    output_dir = os.path.join(args.output_dir, 'checkpoint-best-ppl')
                    if not os.path.exists(output_dir):
                        os.makedirs(output_dir)
                    model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self

                    output_model_file = os.path.join(output_dir, "pytorch_model.bin")
                    torch.save(model_to_save.state_dict(), output_model_file)  


                #Calculate bleu  
                if 'dev_bleu' in dev_dataset:
                    eval_examples,eval_data=dev_dataset['dev_bleu']
                else:

                    eval_examples = read_examples(args.dev_filename)
                    eval_examples = random.sample(eval_examples,min(1000,len(eval_examples)))
                    eval_features = convert_examples_to_features(eval_examples, tokenizer, args,stage='test')
                    eval_data = TextDataset(eval_features,args)
                    dev_dataset['dev_bleu']=eval_examples,eval_data

                eval_sampler = SequentialSampler(eval_data)
                eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size,num_workers=0)
                model.eval() 
                p=[]
                for batch in eval_dataloader:
                    batch = tuple(t.to(device) for t in batch)
                    source_ids,source_mask,position_idx,att_mask,target_ids,target_mask = batch                 
                    with torch.no_grad():
                        preds = model(source_ids,source_mask,position_idx,att_mask)  
                        for pred in preds:
                            t=pred[0].cpu().numpy()
                            t=list(t)
                            if 0 in t:
                                t=t[:t.index(0)]
                            text = tokenizer.decode(t,clean_up_tokenization_spaces=False)
                            p.append(text)
                model.train()
                predictions=[]
                accs=[]
                with open(os.path.join(args.output_dir,"dev.output"),'w') as f, open(os.path.join(args.output_dir,"dev.gold"),'w') as f1:
                    for ref,gold in zip(p,eval_examples):
                        predictions.append(str(gold.idx)+'\t'+ref)
                        f.write(str(gold.idx)+'\t'+ref+'\n')
                        f1.write(str(gold.idx)+'\t'+gold.target+'\n')     

                (goldMap, predictionMap) = bleu.computeMaps(predictions, os.path.join(args.output_dir, "dev.gold")) 
                dev_bleu=round(bleu.bleuFromMaps(goldMap, predictionMap)[0],2)


                # dev_bleu=round(_bleu(os.path.join(args.output_dir, "dev.gold"), os.path.join(args.output_dir, "dev.output")),2)
                # (goldMap, predictionMap) = bleu_sum.computeMaps(os.path.join(args.output_dir, "dev.output"), os.path.join(args.output_dir, "dev.gold")) 
                # dev_bleu=round(bleu_sum.bleuFromMaps(goldMap, predictionMap)[0],2)
                logger.info("  %s = %s "%("bleu-4",str(dev_bleu)))
                logger.info("  "+"*"*20)      
                if dev_bleu>best_bleu:
                    logger.info("  Best BLEU+xMatch:%s",dev_bleu)
                    logger.info("  "+"*"*20)
                    best_bleu=dev_bleu
                    # Save best checkpoint for best bleu
                    output_dir = os.path.join(args.output_dir, 'checkpoint-best-bleu')
                    if not os.path.exists(output_dir):
                        os.makedirs(output_dir)
                    model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self
                    adapter_utils.save_adapter(encoder,f'adapter_{args.lang}',output_dir)
                    output_model_file = os.path.join(output_dir, "pytorch_model.bin")
                    torch.save(model_to_save.state_dict(), output_model_file)
               
    if args.do_test:
        files=[]
        if args.dev_filename is not None:
            files.append(args.dev_filename)
        if args.test_filename is not None:
            files.append(args.test_filename)
        for idx,file in enumerate(files):   
            logger.info("Test file: {}".format(file))
            eval_examples = read_examples(file)
            eval_features = convert_examples_to_features(eval_examples, tokenizer, args,stage='test')
            eval_data = TextDataset(eval_features,args)   

            # Calculate bleu
            eval_sampler = SequentialSampler(eval_data)
            eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size,num_workers=0)

            model.eval() 
            p=[]
            for batch in tqdm(eval_dataloader,total=len(eval_dataloader)):
                batch = tuple(t.to(device) for t in batch)
                source_ids,source_mask,position_idx,att_mask,target_ids,target_mask = batch                      
                with torch.no_grad():
                    preds = model(source_ids,source_mask,position_idx,att_mask)  
                    for pred in preds:
                        t=pred[0].cpu().numpy()
                        t=list(t)
                        if 0 in t:
                            t=t[:t.index(0)]
                        text = tokenizer.decode(t,clean_up_tokenization_spaces=False)
                        p.append(text)
            model.train()
            predictions=[]
            with open(os.path.join(args.output_dir,"test_{}.output_lora_task".format(str(idx))),'w') as f, open(os.path.join(args.output_dir,"test_{}.gold_lora_task".format(str(idx))),'w') as f1:
                for ref,gold in zip(p,eval_examples):
                    predictions.append(str(gold.idx)+'\t'+ref)
                    f.write(str(gold.idx)+'\t'+ref+'\n')
                    f1.write(str(gold.idx)+'\t'+gold.target+'\n')     
            (goldMap, predictionMap) = bleu.computeMaps(predictions, os.path.join(args.output_dir, "test_{}.gold_lora_task".format(str(idx)))) 
            dev_bleu=round(bleu.bleuFromMaps(goldMap, predictionMap)[0],2)

            # dev_bleu=round(_bleu(os.path.join(args.output_dir, "dev.gold"), os.path.join(args.output_dir, "dev.output")),2)
            # (goldMap, predictionMap) = bleu.computeMaps(predictions, os.path.join(args.output_dir, "test_{}.gold".format(idx))) 
            # dev_bleu=round(bleu.bleuFromMaps(goldMap, predictionMap)[0],2)
            # (goldMap, predictionMap) = bleu_sum.computeMaps(os.path.join(args.output_dir, "dev.output"), os.path.join(args.output_dir, "dev.gold")) 
            # dev_bleu=round(bleu_sum.bleuFromMaps(goldMap, predictionMap)[0],2)
            logger.info("  %s = %s "%("bleu-4",str(dev_bleu)))
            logger.info("  "+"*"*20)  
            
if __name__ == "__main__":
    main()