def __init__(self, config):
super(BertForSequenceClassification, self).__init__(config)
self.num_labels = config.num_labels
self.bert = BertModel(config)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
self.batch_size = 32
self.hidden_size = config.hidden_size
self.num_layers = 2
n_filters = 200
filter_sizes = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
self.classifier = nn.Linear(
len(filter_sizes) * n_filters * 3 + config.hidden_size * 3 +
config.hidden_size * 3, config.num_labels)
self.convs = TextCNN1D(config.hidden_size, n_filters, filter_sizes)
self.resnet = ResnetCNN1D(config.hidden_size)
# 最后的全连接层
self.fc = nn.Sequential(
nn.Linear(
len(filter_sizes) * n_filters * 3 * 2, config.hidden_size),
nn.ReLU(inplace=True), nn.Dropout(0.5),
nn.Linear(config.hidden_size, self.num_labels))
self.layer_norm = nn.LayerNorm(self.hidden_size)
self.FeedForward2 = nn.Sequential(
# nn.BatchNorm1d(num_features=self.channel_size),
nn.Linear(self.hidden_size, self.hidden_size),
nn.ReLU(inplace=True),
nn.Dropout(0.1))
self.triple_attention = TripleAttentionHighWay(
hidden_size=self.hidden_size)
self.self_attention = ScaledDotProductAttention()
self.apply(self.init_weights)
def __init__(self, config):
super(BertForSequenceClassification, self).__init__(config)
self.num_labels = config.num_labels
self.bert = BertModel(config)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
self.batch_size = 32
self.hidden_size = config.hidden_size
rnn_hidden_size = 768
num_layers = 2
dropout = 0.2
self.rnn = nn.LSTM(config.hidden_size * 3,
rnn_hidden_size,
num_layers,
bidirectional=True,
batch_first=True,
dropout=dropout)
self.W2 = nn.Linear(config.hidden_size + 2 * rnn_hidden_size,
config.hidden_size)
# self.classifier = nn.Linear(rnn_hidden_size * 2, config.num_labels)
n_filters = 200
filter_sizes = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
self.classifier = nn.Linear(
len(filter_sizes) * n_filters * 3 + config.hidden_size * 3 +
config.hidden_size * 3, config.num_labels)
self.convs = TextCNN1D(config.hidden_size * 3, n_filters, filter_sizes)
# 最后的全连接层
self.fc = nn.Sequential(
nn.Linear(
len(filter_sizes) * n_filters * 3 + config.hidden_size * 3 +
config.hidden_size * 3, config.hidden_size),
nn.ReLU(inplace=True), nn.Dropout(0.5),
nn.Linear(config.hidden_size, self.num_labels))
self.layer_norm = nn.LayerNorm(self.hidden_size)
self.FeedForward2 = nn.Sequential(
# nn.BatchNorm1d(num_features=self.channel_size),
nn.Linear(self.hidden_size, self.hidden_size),
nn.ReLU(inplace=True),
nn.Dropout(0.1))
self.triple_attention = TripleAttention(hidden_size=self.hidden_size)
self.kmax_pooling = KMaxPooling1D(k=5)
self.self_attention = ScaledDotProductAttention()
self.apply(self.init_weights)