models.py 文件源码

def __init__(self, config):
        super(BiLSTM, self).__init__()
        self.drop = nn.Dropout(config['dropout'])
        self.encoder = nn.Embedding(config['ntoken'], config['ninp'])
        self.bilstm = nn.LSTM(config['ninp'], config['nhid'], config['nlayers'], dropout=config['dropout'],
                              bidirectional=True)
        self.nlayers = config['nlayers']
        self.nhid = config['nhid']
        self.pooling = config['pooling']
        self.dictionary = config['dictionary']
#        self.init_weights()
        self.encoder.weight.data[self.dictionary.word2idx['<pad>']] = 0
        if os.path.exists(config['word-vector']):
            print('Loading word vectors from', config['word-vector'])
            vectors = torch.load(config['word-vector'])
            assert vectors[2] >= config['ninp']
            vocab = vectors[0]
            vectors = vectors[1]
            loaded_cnt = 0
            for word in self.dictionary.word2idx:
                if word not in vocab:
                    continue
                real_id = self.dictionary.word2idx[word]
                loaded_id = vocab[word]
                self.encoder.weight.data[real_id] = vectors[loaded_id][:config['ninp']]
                loaded_cnt += 1
            print('%d words from external word vectors loaded.' % loaded_cnt)

    # note: init_range constraints the value of initial weights