unknown

add files

df55b07 over 2 years ago

23.5 kB

# 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
from bleu import _bleu
import pickle
import torch
import json
import random
import logging
import argparse
import numpy as np
from io import open
from itertools import cycle
import torch.nn as nn
from tqdm import tqdm, trange
from torch.nn.utils.rnn import pad_sequence
from fuzzywuzzy import fuzz
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, T5ForConditionalGeneration, AutoTokenizer)

logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
                    datefmt = '%m/%d/%Y %H:%M:%S',
                    level = logging.INFO)
logger = logging.getLogger(__name__)
divide_number = 3
# 

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

            examples.append(
                Example(
                        idx = idx,
                        source=" ".join(js['natrual_language']),
                        ts_v = ",".join(js['TS_V_token']),
                        target = " ".join(js["ground_truth"][1:-1]),
                        ) 
            )
            
    return examples


class InputFeatures(object):
    """A single training/test features for a example."""
    def __init__(self,
                 example_id,
                 source_ids, source_mask,
                 target_ids, target_mask
    ):
        self.example_id = example_id
        self.source_ids = source_ids
        self.source_mask = source_mask
        self.target_ids = target_ids
        self.target_mask = target_mask      
        
def convert_examples_to_features(examples, tokenizer, args,stage=None):
    features = []
    for example_index, example in enumerate(examples):
        #source

        source_ids = torch.LongTensor(tokenizer.encode(example.source + tokenizer.pad_token + example.ts_v, 
                add_special_tokens=True, max_length=args.max_source_length, truncation=True))
        
        source_mask = torch.ones_like(source_ids)
        #target
        if stage=="test":
            target_tokens = tokenizer.tokenize("None")
        else:
            target_tokens = tokenizer.tokenize(example.target)[:args.max_target_length-2]
        
        target_ids = torch.LongTensor(tokenizer.encode(example.target, 
                add_special_tokens=True, max_length=args.max_target_length, truncation=True))
        target_mask = torch.ones_like(target_ids)
   

       
        features.append(
            InputFeatures(
                 example_index,
                 source_ids, source_mask,
                 target_ids, target_mask
            )
        )
    return features



def set_seed(seed=20240124):
    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_name_or_path", default=None, type=str, required=True,
                        help="Path to pre-trained model: e.g. roberta-base" )
    parser.add_argument("--load_model_path", default=None, type=str,
                        help="Path to trained model" )
    parser.add_argument("--output_dir", default=None, type=str, required=True,
                        help="The output directory where the model predictions and checkpoints will be written.")   
  
    ## 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("--max_source_length", default=256, 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=512, 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("--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('--seed', type=int, default=20240124,
                        help="random seed for initialization")
    
    # print arguments
    args = parser.parse_args()
    # set log
    logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
                    datefmt='%m/%d/%Y %H:%M:%S',level=logging.INFO )
    # set device
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    args.n_gpu = torch.cuda.device_count()
    args.device = device
    logger.info("device: %s, n_gpu: %s",device, args.n_gpu)
    
    # 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)

    # build model
    tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
    is_trust = False 
    if "codet5p-220m" in args.model_name_or_path:
        is_trust = False
    else:
        is_trust = True
    model = T5ForConditionalGeneration.from_pretrained(args.model_name_or_path) 
    logger.info("Training/evaluation parameters %s", args)
    if args.load_model_path is not None:
        logger.info("reload model from {}".format(args.load_model_path + "/pytorch_model.bin"))
        model.load_state_dict(torch.load(args.load_model_path + "/pytorch_model.bin"))
    model.to(args.device)   
    
    if args.n_gpu > 1:
        # multi-gpu training
        model = torch.nn.DataParallel(model)

    if args.do_train:
        # Prepare training data loader
        train_examples = read_examples(args.train_filename)
        train_features = convert_examples_to_features(train_examples, tokenizer,args,stage='train')
        all_source_ids = pad_sequence([f.source_ids for f in train_features], batch_first=True, padding_value=tokenizer.pad_token_id)
        all_source_mask = pad_sequence([f.source_mask for f in train_features], batch_first=True, padding_value=0)
        all_target_ids = pad_sequence([f.target_ids for f in train_features], batch_first=True, padding_value=tokenizer.pad_token_id)
        all_target_mask = pad_sequence([f.target_mask for f in train_features], batch_first=True, padding_value=0)
        train_data = TensorDataset(all_source_ids,all_source_mask,all_target_ids,all_target_mask)
        train_sampler = RandomSampler(train_data)
        train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size // args.gradient_accumulation_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=int(len(train_dataloader)*args.num_train_epochs*0.1),
                                                    num_training_steps=len(train_dataloader)*args.num_train_epochs)
    
        #Start training
        logger.info("***** Running training *****")
        logger.info("  Num examples = %d", len(train_examples))
        logger.info("  Batch size = %d", args.train_batch_size * args.gradient_accumulation_steps)
        logger.info("  Num epoch = %d", args.num_train_epochs)
        

        model.train()
        patience, best_score, losses, dev_dataset = 0, 0, [], {}
        for epoch in range(args.num_train_epochs):
            for idx,batch in enumerate(train_dataloader):
                batch = tuple(t.to(device) for t in batch)
                source_ids,source_mask,target_ids,target_mask = batch
                # loss,_,_ = model(source_ids=source_ids,target_ids=target_ids)

                loss = model(input_ids=source_ids, attention_mask=source_mask.gt(0), 
                                  labels=target_ids, decoder_attention_mask=target_mask.gt(0)).loss
                
                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
                    
                losses.append(loss.item())
                loss.backward()
                if len(losses) % args.gradient_accumulation_steps == 0:
                    #Update parameters
                    optimizer.step()
                    optimizer.zero_grad()
                    scheduler.step()
                    if len(losses) // args.gradient_accumulation_steps % 100 == 0:
                        logger.info("epoch {} step {} loss {}".format(epoch,
                                                     len(losses)//args.gradient_accumulation_steps,
                                                     round(np.mean(losses[-100*args.gradient_accumulation_steps:]),4)))
            if args.do_eval:
                #Eval model with dev dataset                   
                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')
                    all_source_ids = pad_sequence([f.source_ids for f in eval_features], batch_first=True, padding_value=tokenizer.pad_token_id)
                    all_source_mask = pad_sequence([f.source_mask for f in eval_features], batch_first=True, padding_value=0)
                    all_target_ids = pad_sequence([f.target_ids for f in eval_features], batch_first=True, padding_value=tokenizer.pad_token_id)
                    all_target_mask = pad_sequence([f.target_mask for f in eval_features], batch_first=True, padding_value=0)
                    eval_data = TensorDataset(all_source_ids,all_source_mask,all_target_ids,all_target_mask)
                    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)

                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,target_ids,target_mask = batch
                    with torch.no_grad():
                        loss = model(input_ids=source_ids, attention_mask=source_mask, 
                                    labels=target_ids, decoder_attention_mask=target_mask).loss

                        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
                        eval_loss += loss.item()
                        tokens_num += 1
                #Pring loss of dev dataset    
                model.train()
                eval_loss = eval_loss / tokens_num
                result = {'eval_ppl': round(np.exp(eval_loss),5)}
                for key in sorted(result.keys()):
                    logger.info("  %s = %s", key, str(result[key]))
                logger.info("  "+"*"*20)   

                #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)
                    eval_features = convert_examples_to_features(eval_examples, tokenizer, args,stage='test')
                    all_source_ids = pad_sequence([f.source_ids for f in eval_features], batch_first=True, padding_value=tokenizer.pad_token_id)
                    all_source_mask = pad_sequence([f.source_mask for f in eval_features], batch_first=True, padding_value=0)
                    all_target_ids = pad_sequence([f.target_ids for f in eval_features], batch_first=True, padding_value=tokenizer.pad_token_id)
                    all_target_mask = pad_sequence([f.target_mask for f in eval_features], batch_first=True, padding_value=0)
                    eval_data = TensorDataset(all_source_ids,all_source_mask,all_target_ids,all_target_mask)
                    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)

                model.eval() 
                p=[]
                for batch in eval_dataloader:
                    batch = tuple(t.to(device) for t in batch)
                    source_ids,source_mask,target_ids,target_mask = batch              
                    with torch.no_grad():
                        preds = model.module.generate(source_ids, attention_mask=source_mask, use_cache=True, 
                                num_beams=args.beam_size, max_new_tokens =args.max_target_length)
                            
                        # convert ids to text
                        for pred in preds:
                            # print(pred)
                            text = tokenizer.decode(pred, skip_special_tokens=True, clean_up_tokenization_spaces=False)
                            p.append(text)

                model.train()
                predictions = []
                res_list = []
                EM = []
                is_gened = False
                with open(args.output_dir+"/dev.output",'w') as f, open(args.output_dir+"/dev.gold",'w') as f1:
                    for ref,gold in zip(p,eval_examples):
                        predictions.append(ref)
                        if len(ref) > 0:
                            is_gened = True
                        f.write(ref+'\n')
                        f1.write(gold.target+'\n')     
                        EM.append(ref.split()==gold.target.split())   
                        res_list.append([ref,gold.target])
                if is_gened:
                    dev_bleu = _bleu(args.output_dir+"/dev.gold", args.output_dir+"/dev.output")
                else:
                    dev_bleu = 0
                logger.info("  %s = %s "%("Epoch",str(epoch)))
                logger.info("  %s = %s "%("bleu-4",str(dev_bleu)))
                logger.info("  %s = %s "%("EM",str(round(np.mean(EM)*100,2))))
                logger.info("  "+"*"*20)    
                dev_score = (dev_bleu+round(np.mean(EM)*100,2))
                if dev_score>best_score:
                    best_score=dev_score
                    # Save best checkpoint for best bleu
                    output_dir = args.output_dir
                    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)
                    patience = 0
                else:
                    patience += 1
                    if patience == 3:
                        break
                output_dir = args.output_dir
                with open(output_dir + "/last_training_result.jsonl", 'w') as wf:
                    for line in res_list:
                        dic = {}
                        dic["Pred"] = line[0]
                        dic["GT"] = line[1]
                        wf.write(json.dumps(dic))
                        wf.write("\n")

                logger.info("  Best score:%s",best_score)
                logger.info("  "+"*"*20)
    if args.do_test:
        res_list = []
        
        if args.load_model_path is not None:
            checkpoint_prefix = 'pytorch_model.bin'
            output_dir = os.path.join(args.output_dir, checkpoint_prefix)  
            model_to_load = model.module if hasattr(model, 'module') else model  
            model_to_load.load_state_dict(torch.load(output_dir))
        
        

        eval_examples = read_examples(args.test_filename)
        eval_features = convert_examples_to_features(eval_examples, tokenizer, args,stage='test')
        all_source_ids = pad_sequence([f.source_ids for f in eval_features], batch_first=True, padding_value=tokenizer.pad_token_id)
        all_source_mask = pad_sequence([f.source_mask for f in eval_features], batch_first=True, padding_value=0)
        all_target_ids = pad_sequence([f.target_ids for f in eval_features], batch_first=True, padding_value=tokenizer.pad_token_id)
        all_target_mask = pad_sequence([f.target_mask for f in eval_features], batch_first=True, padding_value=0)
        eval_data = TensorDataset(all_source_ids,all_source_mask,all_target_ids,all_target_mask)

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

        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, _, _  = batch 
            with torch.no_grad():
                preds = model.module.generate(source_ids, attention_mask=source_mask, use_cache=True, 
                                        num_beams=args.beam_size, max_new_tokens =args.max_target_length)
            for pred in preds:
                # print(pred)
                text = tokenizer.decode(pred, skip_special_tokens=True, clean_up_tokenization_spaces=False)
                p.append(text)

        predictions=[]
        EM = []
        edit_dis = 0
        cnt = 0
        with open(args.output_dir+"/test.output",'w') as f, open(args.output_dir+"/test.gold",'w') as f1:
            for ref,gold in zip(p,eval_examples):
                res_list.append([ref,gold.target])
                predictions.append(ref)
                f.write(ref+'\n')
                f1.write(gold.target+'\n')     
                EM.append(ref.split()==gold.target.split())
                edit_dis += fuzz.ratio(ref, gold.target)
                cnt += 1

        dev_bleu = _bleu(args.output_dir+"/test.gold", args.output_dir+"/test.output")   
        logger.info("  %s = %s "%("bleu-4",str(dev_bleu)))
        logger.info("  %s = %s "%("EM",str(round(np.mean(EM)*100,2))))
        logger.info("  %s = %s "%("Edit Distance",str(round(float(edit_dis)/cnt,2))))
        logger.info("  "+"*"*20)


        with open(args.output_dir + "/last_training_result.jsonl", 'w') as wf:
            for line in res_list:
                dic = {}
                dic["Pred"] = line[0]
                dic["GT"] = line[1]
                wf.write(json.dumps(dic))
                wf.write("\n")
                
if __name__ == "__main__":
    main()

	# 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
	from bleu import _bleu
	import pickle
	import torch
	import json
	import random
	import logging
	import argparse
	import numpy as np
	from io import open
	from itertools import cycle
	import torch.nn as nn
	from tqdm import tqdm, trange
	from torch.nn.utils.rnn import pad_sequence
	from fuzzywuzzy import fuzz
	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, T5ForConditionalGeneration, AutoTokenizer)

	logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
	datefmt = '%m/%d/%Y %H:%M:%S',
	level = logging.INFO)
	logger = logging.getLogger(__name__)
	divide_number = 3
	#

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

	examples.append(
	Example(
	idx = idx,
	source=" ".join(js['natrual_language']),
	ts_v = ",".join(js['TS_V_token']),
	target = " ".join(js["ground_truth"][1:-1]),
	)
	)

	return examples


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

	def convert_examples_to_features(examples, tokenizer, args,stage=None):
	features = []
	for example_index, example in enumerate(examples):
	#source

	source_ids = torch.LongTensor(tokenizer.encode(example.source + tokenizer.pad_token + example.ts_v,
	add_special_tokens=True, max_length=args.max_source_length, truncation=True))

	source_mask = torch.ones_like(source_ids)
	#target
	if stage=="test":
	target_tokens = tokenizer.tokenize("None")
	else:
	target_tokens = tokenizer.tokenize(example.target)[:args.max_target_length-2]

	target_ids = torch.LongTensor(tokenizer.encode(example.target,
	add_special_tokens=True, max_length=args.max_target_length, truncation=True))
	target_mask = torch.ones_like(target_ids)



	features.append(
	InputFeatures(
	example_index,
	source_ids, source_mask,
	target_ids, target_mask
	)
	)
	return features



	def set_seed(seed=20240124):
	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_name_or_path", default=None, type=str, required=True,
	help="Path to pre-trained model: e.g. roberta-base" )
	parser.add_argument("--load_model_path", default=None, type=str,
	help="Path to trained model" )
	parser.add_argument("--output_dir", default=None, type=str, required=True,
	help="The output directory where the model predictions and checkpoints will be written.")

	## 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("--max_source_length", default=256, 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=512, 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("--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('--seed', type=int, default=20240124,
	help="random seed for initialization")

	# print arguments
	args = parser.parse_args()
	# set log
	logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
	datefmt='%m/%d/%Y %H:%M:%S',level=logging.INFO )
	# set device
	device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
	args.n_gpu = torch.cuda.device_count()
	args.device = device
	logger.info("device: %s, n_gpu: %s",device, args.n_gpu)

	# 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)

	# build model
	tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
	is_trust = False
	if "codet5p-220m" in args.model_name_or_path:
	is_trust = False
	else:
	is_trust = True
	model = T5ForConditionalGeneration.from_pretrained(args.model_name_or_path)
	logger.info("Training/evaluation parameters %s", args)
	if args.load_model_path is not None:
	logger.info("reload model from {}".format(args.load_model_path + "/pytorch_model.bin"))
	model.load_state_dict(torch.load(args.load_model_path + "/pytorch_model.bin"))
	model.to(args.device)

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

	if args.do_train:
	# Prepare training data loader
	train_examples = read_examples(args.train_filename)
	train_features = convert_examples_to_features(train_examples, tokenizer,args,stage='train')
	all_source_ids = pad_sequence([f.source_ids for f in train_features], batch_first=True, padding_value=tokenizer.pad_token_id)
	all_source_mask = pad_sequence([f.source_mask for f in train_features], batch_first=True, padding_value=0)
	all_target_ids = pad_sequence([f.target_ids for f in train_features], batch_first=True, padding_value=tokenizer.pad_token_id)
	all_target_mask = pad_sequence([f.target_mask for f in train_features], batch_first=True, padding_value=0)
	train_data = TensorDataset(all_source_ids,all_source_mask,all_target_ids,all_target_mask)
	train_sampler = RandomSampler(train_data)
	train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size // args.gradient_accumulation_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=int(len(train_dataloader)args.num_train_epochs0.1),
	num_training_steps=len(train_dataloader)*args.num_train_epochs)

	#Start training
	logger.info("*** Running training ***")
	logger.info(" Num examples = %d", len(train_examples))
	logger.info(" Batch size = %d", args.train_batch_size * args.gradient_accumulation_steps)
	logger.info(" Num epoch = %d", args.num_train_epochs)


	model.train()
	patience, best_score, losses, dev_dataset = 0, 0, [], {}
	for epoch in range(args.num_train_epochs):
	for idx,batch in enumerate(train_dataloader):
	batch = tuple(t.to(device) for t in batch)
	source_ids,source_mask,target_ids,target_mask = batch
	# loss,_,_ = model(source_ids=source_ids,target_ids=target_ids)

	loss = model(input_ids=source_ids, attention_mask=source_mask.gt(0),
	labels=target_ids, decoder_attention_mask=target_mask.gt(0)).loss

	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

	losses.append(loss.item())
	loss.backward()
	if len(losses) % args.gradient_accumulation_steps == 0:
	#Update parameters
	optimizer.step()
	optimizer.zero_grad()
	scheduler.step()
	if len(losses) // args.gradient_accumulation_steps % 100 == 0:
	logger.info("epoch {} step {} loss {}".format(epoch,
	len(losses)//args.gradient_accumulation_steps,
	round(np.mean(losses[-100*args.gradient_accumulation_steps:]),4)))
	if args.do_eval:
	#Eval model with dev dataset
	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')
	all_source_ids = pad_sequence([f.source_ids for f in eval_features], batch_first=True, padding_value=tokenizer.pad_token_id)
	all_source_mask = pad_sequence([f.source_mask for f in eval_features], batch_first=True, padding_value=0)
	all_target_ids = pad_sequence([f.target_ids for f in eval_features], batch_first=True, padding_value=tokenizer.pad_token_id)
	all_target_mask = pad_sequence([f.target_mask for f in eval_features], batch_first=True, padding_value=0)
	eval_data = TensorDataset(all_source_ids,all_source_mask,all_target_ids,all_target_mask)
	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)

	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,target_ids,target_mask = batch
	with torch.no_grad():
	loss = model(input_ids=source_ids, attention_mask=source_mask,
	labels=target_ids, decoder_attention_mask=target_mask).loss

	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
	eval_loss += loss.item()
	tokens_num += 1
	#Pring loss of dev dataset
	model.train()
	eval_loss = eval_loss / tokens_num
	result = {'eval_ppl': round(np.exp(eval_loss),5)}
	for key in sorted(result.keys()):
	logger.info(" %s = %s", key, str(result[key]))
	logger.info(" "+""20)

	#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)
	eval_features = convert_examples_to_features(eval_examples, tokenizer, args,stage='test')
	all_source_ids = pad_sequence([f.source_ids for f in eval_features], batch_first=True, padding_value=tokenizer.pad_token_id)
	all_source_mask = pad_sequence([f.source_mask for f in eval_features], batch_first=True, padding_value=0)
	all_target_ids = pad_sequence([f.target_ids for f in eval_features], batch_first=True, padding_value=tokenizer.pad_token_id)
	all_target_mask = pad_sequence([f.target_mask for f in eval_features], batch_first=True, padding_value=0)
	eval_data = TensorDataset(all_source_ids,all_source_mask,all_target_ids,all_target_mask)
	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)

	model.eval()
	p=[]
	for batch in eval_dataloader:
	batch = tuple(t.to(device) for t in batch)
	source_ids,source_mask,target_ids,target_mask = batch
	with torch.no_grad():
	preds = model.module.generate(source_ids, attention_mask=source_mask, use_cache=True,
	num_beams=args.beam_size, max_new_tokens =args.max_target_length)

	# convert ids to text
	for pred in preds:
	# print(pred)
	text = tokenizer.decode(pred, skip_special_tokens=True, clean_up_tokenization_spaces=False)
	p.append(text)

	model.train()
	predictions = []
	res_list = []
	EM = []
	is_gened = False
	with open(args.output_dir+"/dev.output",'w') as f, open(args.output_dir+"/dev.gold",'w') as f1:
	for ref,gold in zip(p,eval_examples):
	predictions.append(ref)
	if len(ref) > 0:
	is_gened = True
	f.write(ref+'\n')
	f1.write(gold.target+'\n')
	EM.append(ref.split()==gold.target.split())
	res_list.append([ref,gold.target])
	if is_gened:
	dev_bleu = _bleu(args.output_dir+"/dev.gold", args.output_dir+"/dev.output")
	else:
	dev_bleu = 0
	logger.info(" %s = %s "%("Epoch",str(epoch)))
	logger.info(" %s = %s "%("bleu-4",str(dev_bleu)))
	logger.info(" %s = %s "%("EM",str(round(np.mean(EM)*100,2))))
	logger.info(" "+""20)
	dev_score = (dev_bleu+round(np.mean(EM)*100,2))
	if dev_score>best_score:
	best_score=dev_score
	# Save best checkpoint for best bleu
	output_dir = args.output_dir
	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)
	patience = 0
	else:
	patience += 1
	if patience == 3:
	break
	output_dir = args.output_dir
	with open(output_dir + "/last_training_result.jsonl", 'w') as wf:
	for line in res_list:
	dic = {}
	dic["Pred"] = line[0]
	dic["GT"] = line[1]
	wf.write(json.dumps(dic))
	wf.write("\n")

	logger.info(" Best score:%s",best_score)
	logger.info(" "+""20)
	if args.do_test:
	res_list = []

	if args.load_model_path is not None:
	checkpoint_prefix = 'pytorch_model.bin'
	output_dir = os.path.join(args.output_dir, checkpoint_prefix)
	model_to_load = model.module if hasattr(model, 'module') else model
	model_to_load.load_state_dict(torch.load(output_dir))



	eval_examples = read_examples(args.test_filename)
	eval_features = convert_examples_to_features(eval_examples, tokenizer, args,stage='test')
	all_source_ids = pad_sequence([f.source_ids for f in eval_features], batch_first=True, padding_value=tokenizer.pad_token_id)
	all_source_mask = pad_sequence([f.source_mask for f in eval_features], batch_first=True, padding_value=0)
	all_target_ids = pad_sequence([f.target_ids for f in eval_features], batch_first=True, padding_value=tokenizer.pad_token_id)
	all_target_mask = pad_sequence([f.target_mask for f in eval_features], batch_first=True, padding_value=0)
	eval_data = TensorDataset(all_source_ids,all_source_mask,all_target_ids,all_target_mask)

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

	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, _, _ = batch
	with torch.no_grad():
	preds = model.module.generate(source_ids, attention_mask=source_mask, use_cache=True,
	num_beams=args.beam_size, max_new_tokens =args.max_target_length)
	for pred in preds:
	# print(pred)
	text = tokenizer.decode(pred, skip_special_tokens=True, clean_up_tokenization_spaces=False)
	p.append(text)

	predictions=[]
	EM = []
	edit_dis = 0
	cnt = 0
	with open(args.output_dir+"/test.output",'w') as f, open(args.output_dir+"/test.gold",'w') as f1:
	for ref,gold in zip(p,eval_examples):
	res_list.append([ref,gold.target])
	predictions.append(ref)
	f.write(ref+'\n')
	f1.write(gold.target+'\n')
	EM.append(ref.split()==gold.target.split())
	edit_dis += fuzz.ratio(ref, gold.target)
	cnt += 1

	dev_bleu = _bleu(args.output_dir+"/test.gold", args.output_dir+"/test.output")
	logger.info(" %s = %s "%("bleu-4",str(dev_bleu)))
	logger.info(" %s = %s "%("EM",str(round(np.mean(EM)*100,2))))
	logger.info(" %s = %s "%("Edit Distance",str(round(float(edit_dis)/cnt,2))))
	logger.info(" "+""20)


	with open(args.output_dir + "/last_training_result.jsonl", 'w') as wf:
	for line in res_list:
	dic = {}
	dic["Pred"] = line[0]
	dic["GT"] = line[1]
	wf.write(json.dumps(dic))
	wf.write("\n")

	if __name__ == "__main__":
	main()