2021-12-3

 2021SC@SDUSC

在readme.md文档中有一个步骤要执行

python divide_inhouse_data.py

但是找遍项目目录结构，没有这个文件，只好略过。

按照readme.md文档，此时应该执行main.py，涉及两个文件main.py和train.py。train.py文件里面创建了一个重要的图对象GraphC_QAModel，对象中提供了几个方法。

下面是做初始化。

class GraphC_QAModel(object):

    def __init__(self, args, local_rank):

        super(GraphC_QAModel, self).__init__()

        self.save_args = args

        args['cnn_filters'] = list(zip(args['cnn_filters'][:-1:2], args['cnn_filters'][1::2]))

        args = collections.namedtuple("HParams", sorted(args.keys()))(**args)

        if not os.path.exists(args.ckpt):

            os.mkdir(args.ckpt)

        self.args = args

        self.local_rank = local_rank

这个函数根据AMR模型来建图，这里首选GPU,GPU可以并行训练多个batch 数据，提高性能。

 def _build_model(self):

        print(self.args, 'final')

        self.device = torch.device("cuda:"+str(self.args.gpus[0]) if torch.cuda.is_available() else "cpu")

        print(self.device, 'here')

        vocabs, lexical_mappings = [], []

        config_class, model_class, tokenizer_class = MODEL_CLASSES[self.args.encoder_type]

        self.bert_config = config_class.from_pretrained(

            self.args.lm_model,

        )

        self.bert_tokenizer = tokenizer_class.from_pretrained(

            self.args.lm_model

        )

        if self.args.bert_pretrained_file == None:

            self.bert_model = model_class.from_pretrained(

                self.args.lm_model,

                config=self.args.lm_model

            ).to(self.device)

        else:

            self.bert_model = model_class.from_pretrained(

                self.args.bert_pretrained_file,

            ).to(self.device)

            print('bert_pretrained')

        # self.device = torch.device('cuda', self.args.gpus[0])

        if self.args.encoder_type in ['ACB_dual']:

            vocabs, lexical_mapping = self._prepare_data()

            self.model = Reasoning_AMR_CN_DUAL(vocabs,

                                               self.args.concept_char_dim, self.args.concept_dim,

                                               self.args.cnn_filters, self.args.char2concept_dim,

                                               self.args.rel_dim, self.args.rnn_hidden_size, self.args.rnn_num_layers,

                                               self.args.embed_dim, self.args.bert_embed_dim, self.args.ff_embed_dim,

                                               self.args.num_heads,

                                               self.args.dropout,

                                               self.args.snt_layer,

                                               self.args.graph_layers,

                                               self.args.pretrained_file, self.device, self.args.batch_size,

                                               self.args.lm_model, self.bert_config, self.bert_model, self.bert_tokenizer, self.args.bert_max_length,

                                               self.args.n_answers,

                                               self.args.encoder_type,

                                               self.args.max_conceptnet_length,

                                               self.args.conceptnet_path,

            )

        else:

            pass

        self.model.to(self.device)

        self.criterion = nn.CrossEntropyLoss()

        return vocabs, lexical_mapping

定义函数average_gradients，计算平均梯度。

def _average_gradients(self, model):

        size = float(dist.get_world_size())

        for param in model.parameters():

            if param.grad is not None:

                dist.all_reduce(param.grad.data, op=dist.ReduceOp.SUM)

                param.grad.data /= size

把图中每一个结点的数据分为5个维度：

def _prepare_data(self):

        vocabs = dict()

        vocabs['concept'] = Vocab(self.args.concept_vocab, 5, [CLS])

        vocabs['token'] = Vocab(self.args.token_vocab, 5, [STR, END])

        vocabs['token_char'] = Vocab(self.args.token_char_vocab, 100, [STR, END])

        vocabs['concept_char'] = Vocab(self.args.concept_char_vocab, 100, [STR, END])

        vocabs['relation'] = Vocab(self.args.relation_vocab, 5, [CLS, rCLS, SEL, TL])

        lexical_mapping = LexicalMap()

定义train函数，训练模块，针对训练、开发、测试三种类型的实验，参数不同，调用同一个数据加载程序。

def train(self):

        from datetime import datetime

        current_time = datetime.now().strftime('%b%d_%H-%M-%S')

        task = self.args.task

        tb_writer = SummaryWriter(log_dir='./runs/'+task+"/"+current_time+self.args.prefix, comment=self.args.prefix)

        vocabs, lexical_mapping = self._build_model()

        train_data = DataLoader(self.args, vocabs, lexical_mapping, self.args.train_data, self.args.batch_size,

                                for_train=True)

        dev_data = DataLoader(self.args, vocabs, lexical_mapping, self.args.dev_data, self.args.batch_size,

                              for_train=False)

        test_data = DataLoader(self.args, vocabs, lexical_mapping, self.args.test_data, self.args.batch_size,

                               for_train='Eval')

下面要去分析数据加载的程序。