train_torch.py

import argparse
import logging
import timeit
import gluonnlp as nlp
import numpy as np
import pandas as pd
import torch
import GPUtil
from gluonnlp.data import SentencepieceTokenizer
from kogpt2.pytorch_kogpt2 import get_pytorch_kogpt2_model
from kogpt2.utils import get_tokenizer
from pytorch_lightning import Trainer
from pytorch_lightning.callbacks import ModelCheckpoint
from pytorch_lightning.core.lightning import LightningModule
from torch.utils.data import DataLoader, Dataset
from transformers.optimization import AdamW, get_cosine_schedule_with_warmup
from flask import Flask, request, jsonify
from flask_cors import CORS, cross_origin
#from log.chat.logging_chat import log_chat
import time
import logging
import datetime
from log.train.logging_train import log_train

app = Flask(__name__)
CORS(app, support_credentials=True)



parser = argparse.ArgumentParser(description='Simsimi based on KoGPT-2')

parser.add_argument('--chat',
                    action='store_true',
                    default=False,
                    help='response generation on given user input')

parser.add_argument('--sentiment',
                    type=str,
                    default='0',
                    help='sentiment for system. 0 is neutral, 1 is negative, 2 is positive.')

parser.add_argument('--model_params',
                    type=str,
                    default='model_chp/model_last.ckpt',
                    help='model binary for starting chat')

parser.add_argument('--train',
                    action='store_true',
                    default=False,
                    help='for training')

#logger = logging.getLogger()
#logger.setLevel(logging.INFO)

U_TKN = '<usr>'
S_TKN = '<sys>'
BOS = '<s>'
EOS = '</s>'
MASK = '<unused0>'
SENT = '<unused1>'


for handler in logging.root.handlers[:]:
    logging.root.removeHandler(handler)

def log_chat(gpu_util, question, answer, generation_time):
    GPUs = GPUtil.getGPUs()
    for GPU in GPUs:
        logging.basicConfig(filename='./log/chat/{}_chat.log'.format(datetime.datetime.now().strftime('%Y-%m-%d')), level=logging.DEBUG)
        print('/KoGPT2chatbot/log/chat/{}_chat.log'.format(datetime.datetime.now().strftime('%Y-%m-%d')))
        logging.debug('{}, {}, used_memory={}/{}, gpu_max_utilization={}, q={}, a={}, generation_time={}'.format(datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S'), GPU.name, GPU.memoryUsed, GPU.memoryTotal, gpu_util, question, answer, generation_time))


class CharDataset(Dataset):
    def __init__(self, chats, tok_path, vocab, max_len=32):
        self._data = chats
        self._tok_path = tok_path
        self.tokenizer = None
        self.first = True
        self.q_token = U_TKN
        self.a_token = S_TKN
        self.sent_token = SENT
        self.bos = BOS
        self.eos = EOS
        self.maskt = MASK
        self.vocab = vocab
        self.max_len = max_len
        self.padder = nlp.data.PadSequence(
            max_len, pad_val=self.vocab[self.vocab.padding_token])

    def _activate_sp(self):
        self.tokenizer = nlp.data.SentencepieceTokenizer(self._tok_path, 0, 0)

    def __len__(self):
        return len(self._data)

    def __getitem__(self, idx):
        if self.tokenizer is None:
            self._activate_sp()
        turn = self._data.iloc[idx]
        q = turn['Q']
        a = turn['A']
        sentiment = str(turn['label'])
        q_toked = [
            self.q_token,
        ] + self.tokenizer(q) + [
            self.eos,
        ] + [self.sent_token] + self.tokenizer(sentiment) + [
            self.eos,
        ]
        q_len = len(q_toked)
        a_toked = [
            self.a_token,
        ] + self.tokenizer(a) + [
            self.eos,
        ]
        a_len = len(a_toked)
        if q_len + a_len > self.max_len:
            a_len = self.max_len - q_len
            if a_len <= 0:
                q_toked = q_toked[-(int(self.max_len/2)):]
                q_len = len(q_toked)
                a_len = self.max_len - q_len
                assert a_len > 0
            a_toked = a_toked[:a_len]
            a_len = len(a_toked)
            assert a_len == len(a_toked), f'{a_len} ==? {len(a_toked)}'
        # [mask, mask, ...., mask, ..., <bos>,..A.. <eos>, <pad>....]
        labels = [
            self.maskt,
        ] * q_len + a_toked[1:]
        if self.first:
            #logging.info("contexts : {}".format(q))
            #logging.info("toked ctx: {}".format(q_toked))
            #logging.info("response : {}".format(a))
            #logging.info("toked response : {}".format(a_toked))
            #logging.info('labels {}'.format(labels))
            self.first = False
        mask = [0] * q_len + [1] * a_len + [0] * (self.max_len - q_len - a_len)
        return (self.padder(self.vocab[q_toked + a_toked]), np.array(mask),
                self.padder(self.vocab[labels]))


class KoGPT2Chat(LightningModule):
    def __init__(self, hparams, **kwargs):
        super(KoGPT2Chat, self).__init__()
        self.hparams = hparams
        self.tok_path = get_tokenizer()
        self.neg = -1e18
        self.kogpt2, self.vocab = get_pytorch_kogpt2_model("cuda")
        self.loss_function = torch.nn.CrossEntropyLoss(reduction='none')
        self.max_gpu_load_train= 0
        self.max_memory_used_train=0.0

    @staticmethod
    def add_model_specific_args(parent_parser):
        # add model specific args
        parser = argparse.ArgumentParser(parents=[parent_parser], add_help=False)
        parser.add_argument('--max-len',
                            type=int,
                            default=32,
                            help='max sentence length on input (default: 32)')

        parser.add_argument('--batch-size',
                            type=int,
                            default=16,
                            help='batch size for training (default: 96)')
        parser.add_argument('--lr',
                            type=float,
                            default=5e-5,
                            help='The initial learning rate')
        parser.add_argument('--warmup_ratio',
                            type=float,
                            default=0.1,
                            help='warmup ratio')
        return parser

    def forward(self, inputs):
        # (batch, seq_len, hiddens)
        output, _ = self.kogpt2(inputs)
        return output

    def training_step(self, batch, batch_idx):
        token_ids, mask, label = batch
        
        if GPUtil.getGPUs()[0].load > self.max_gpu_load_train:
            self.max_gpu_load_train = GPUtil.getGPUs()[0].load
        if GPUtil.getGPUs()[0].memoryUsed > self.max_memory_used_train:
            self.max_memory_used_train = GPUtil.getGPUs()[0].memoryUsed

        out = self(token_ids)
        mask_3d = mask.unsqueeze(dim=2).repeat_interleave(repeats=out.shape[2], dim=2)
        mask_out = torch.where(mask_3d == 1, out, self.neg * torch.ones_like(out))
        loss = self.loss_function(mask_out.transpose(2, 1), label)
        loss_avg = loss.sum() / mask.sum()
        
        if GPUtil.getGPUs()[0].memoryUsed > self.max_memory_used_train:
            self.max_memory_used_train = GPUtil.getGPUs()[0].memoryUsed
        if GPUtil.getGPUs()[0].load > self.max_gpu_load_train:
            self.max_gpu_load_train = GPUtil.getGPUs()[0].load
        
        
        tensorboard_logs = {'train_loss': loss_avg}
        return {'loss': loss_avg, 'log': tensorboard_logs}

    def configure_optimizers(self):
        # Prepare optimizer
        param_optimizer = list(self.named_parameters())
        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
        optimizer_grouped_parameters = [
            {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
            {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]
        optimizer = AdamW(optimizer_grouped_parameters,
                          lr=self.hparams.lr, correct_bias=False)
        # warm up lr
        num_train_steps = len(self.train_dataloader()) * self.hparams.max_epochs
        num_warmup_steps = int(num_train_steps * self.hparams.warmup_ratio)
        scheduler = get_cosine_schedule_with_warmup(
            optimizer,
            num_warmup_steps=num_warmup_steps, num_training_steps=num_train_steps)
        lr_scheduler = {'scheduler': scheduler, 'name': 'cosine_schedule_with_warmup',
                        'monitor': 'loss', 'interval': 'step',
                        'frequency': 1}
        return [optimizer], [lr_scheduler]
    
    def _collate_fn(self, batch):
        data = [item[0] for item in batch]
        mask = [item[1] for item in batch]
        label = [item[2] for item in batch]
        return torch.LongTensor(data), torch.LongTensor(mask), torch.LongTensor(label)

    def train_dataloader(self):
        data = pd.read_csv('Chatbot_data/ChatbotData.csv')
        self.train_set = CharDataset(data, self.tok_path, self.vocab, max_len=self.hparams.max_len)
        train_dataloader = DataLoader(
            self.train_set, batch_size=self.hparams.batch_size, num_workers=2,
            shuffle=True, collate_fn=self._collate_fn)
        return train_dataloader

    def chat(self, sent='0', text=''):
        start = timeit.default_timer()
        self.tok_path
        tok = SentencepieceTokenizer(self.tok_path, num_best=0, alpha=0)
        sent_tokens = tok(sent)
        with torch.no_grad():
            max_gpu_load = 0
            q = text
            if q == 'quit':
                return
            q_tok = tok(q)
            a = ''
            a_tok = []
            while 1:
                input_ids = torch.LongTensor([
                    self.vocab[U_TKN]] + self.vocab[q_tok] +
                    self.vocab[EOS, SENT] + self.vocab[sent_tokens] +
                    self.vocab[EOS, S_TKN] +
                    self.vocab[a_tok]).unsqueeze(dim=0).to("cuda")
                pred = self(input_ids)
                
                if GPUtil.getGPUs()[0].load > max_gpu_load:
                    max_gpu_load = GPUtil.getGPUs()[0].load
                
                gen = self.vocab.to_tokens(
                    torch.argmax(
                        pred,
                        dim=-1).cpu().squeeze().numpy().tolist())[-1]
                
                if GPUtil.getGPUs()[0].load > max_gpu_load:
                    max_gpu_load = GPUtil.getGPUs()[0].load

                if gen == EOS:
                    break
                a += gen.replace('▁', ' ')
                a_tok = tok(a)
                app.logger.info(tok(a))
            stop = timeit.default_timer()
            #log_chat(max_gpu_load, sent, a.strip(), stop-start)
            app.logger.info(a.strip())
            return a.strip()

parser = KoGPT2Chat.add_model_specific_args(parser)
parser = Trainer.add_argparse_args(parser)
args = parser.parse_args()
#logging.info(args)

model = KoGPT2Chat.load_from_checkpoint(args.model_params)

@app.route('/test', methods = ['POST'])
@cross_origin(supports_credentials=True)
def testSentence():
  text = request.get_json(force=True)
  app.logger.info(text['message'])
  return jsonify(model.chat(text['message']))


if __name__ == "__main__":
    if args.train:
        checkpoint_callback = ModelCheckpoint(
            filepath='model_chp/{epoch:02d}-{loss:.2f}',
            verbose=True,
            save_last=True,
            monitor='loss',
            mode='min',
            prefix='model_'
        )
        model = KoGPT2Chat(args)
        start = time.time()
        model.train()
        trainer = Trainer.from_argparse_args(
            args,
            checkpoint_callback=checkpoint_callback, gradient_clip_val=1.0)
        trainer.fit(model)
        #log_train(model.max_memory_used_train, model.max_gpu_load_train, start, model.hparams.max_epochs, model.hparams.batch_size)

    if args.chat:
        app.run(host="0.0.0.0",port=5000)