def __init__(self, vocab_size, config):
super(BiLSTM, self).__init__()
self.model_name = 'BiLSTM'
self.word_dim = config.word_dim
self.pos_size = config.pos_size
self.pos_dim = config.pos_dim
self.hidden_dim = config.hidden_dim
self.lstm_layers = config.lstm_layers
self.last_hn = config.last_hn
self.out_dim = config.relation_type
self.dropout = config.dropout
self.embedding = Embedding(vocab_size, self.word_dim, self.pos_size,
self.pos_dim)
self.input_dim = self.word_dim + self.pos_dim * 2
self.lstm = VarLenLSTM(self.input_dim,
self.hidden_dim,
self.lstm_layers,
dropout=self.dropout,
last_hn=self.last_hn)
if self.last_hn:
linear_input_dim = self.hidden_dim * 2
else:
linear_input_dim = self.hidden_dim * 2 * self.lstm_layers
self.fc1 = nn.Linear(linear_input_dim, self.hidden_dim)
self.fc2 = nn.Linear(self.hidden_dim, self.out_dim)
def __init__(self, vocab_size, config):
super(Capsule, self).__init__()
self.model_name = 'Capsule'
self.vocab_size = vocab_size
self.word_dim = config.model.word_dim
self.pos_size = config.model.pos_size
self.pos_dim = config.model.pos_dim
self.hidden_dim = config.model.hidden_dim
self.num_primary_units = config.capsule.num_primary_units
self.num_output_units = config.capsule.num_output_units
self.primary_channels = config.capsule.primary_channels
self.primary_unit_size = config.capsule.primary_unit_size
self.output_unit_size = config.capsule.output_unit_size
self.num_iterations = config.capsule.num_iterations
self.embedding = Embedding(self.vocab_size, self.word_dim,
self.pos_size, self.pos_dim)
self.input_dim = self.word_dim + self.pos_dim * 2
self.lstm = VarLenLSTM(
self.input_dim,
self.hidden_dim,
)
self.capsule = CapsuleNet(self.num_primary_units,
self.num_output_units, self.primary_channels,
self.primary_unit_size,
self.output_unit_size, self.num_iterations)
def __init__(self, vocab_size, config):
super(GCN, self).__init__()
self.model_name = 'GCN'
self.vocab_size = vocab_size
self.word_dim = config.model.word_dim
self.pos_size = config.model.pos_size
self.pos_dim = config.model.pos_dim
self.hidden_dim = config.model.hidden_dim
self.dropout = config.model.dropout
self.num_layers = config.gcn.num_layers
self.out_dim = config.relation_type
self.embedding = Embedding(self.vocab_size, self.word_dim,
self.pos_size, self.pos_dim)
self.input_dim = self.word_dim + self.pos_dim * 2
self.fc1 = nn.Linear(self.input_dim, self.hidden_dim)
self.fc2 = nn.Linear(self.hidden_dim, self.hidden_dim)
self.fc3 = nn.Linear(self.hidden_dim, self.out_dim)
self.dropout = nn.Dropout(self.dropout)
def __init__(self, vocab_size, config):
super(CNN, self).__init__()
self.model_name = 'CNN'
self.vocab_size = vocab_size
self.word_dim = config.model.word_dim
self.pos_size = config.model.pos_size
self.pos_dim = config.model.pos_dim
self.hidden_dim = config.model.hidden_dim
self.dropout = config.model.dropout
self.use_pcnn = config.cnn.use_pcnn
self.out_channels = config.cnn.out_channels
self.kernel_size = config.cnn.kernel_size
self.out_dim = config.relation_type
if isinstance(self.kernel_size, int):
self.kernel_size = [self.kernel_size]
for k in self.kernel_size:
assert k % 2 == 1, "kernel size has to be odd numbers."
self.embedding = Embedding(self.vocab_size, self.word_dim,
self.pos_size, self.pos_dim)
# PCNN embedding
self.mask_embed = nn.Embedding(4, 3)
masks = torch.tensor([[0, 0, 0], [100, 0, 0], [0, 100, 0], [0, 0,
100]])
self.mask_embed.weight.data.copy_(masks)
self.mask_embed.weight.requires_grad = False
self.input_dim = self.word_dim + self.pos_dim * 2
self.convs = nn.ModuleList([
nn.Conv1d(in_channels=self.input_dim,
out_channels=self.out_channels,
kernel_size=k,
padding=k // 2,
bias=None) for k in self.kernel_size
])
self.conv_dim = len(self.kernel_size) * self.out_channels
if self.use_pcnn:
self.conv_dim *= 3
self.fc1 = nn.Linear(self.conv_dim, self.hidden_dim)
self.fc2 = nn.Linear(self.hidden_dim, self.out_dim)
self.dropout = nn.Dropout(self.dropout)
def __init__(self, vocab_size, config):
super(Transformer, self).__init__()
self.model_name = 'Transformer'
self.word_dim = config.word_dim
self.pos_size = config.pos_size
self.pos_dim = config.pos_dim
self.hidden_dim = config.hidden_dim
self.dropout = config.dropout
self.out_dim = config.relation_type
self.layers = config.transformer_layers
self.embedding = Embedding(vocab_size, self.word_dim, self.pos_size,
self.pos_dim)
self.feature_dim = self.word_dim + self.pos_dim * 2
self.att = MultiHeadAttention(self.feature_dim, num_head=4)
self.norm1 = nn.LayerNorm(self.feature_dim)
self.ffn = nn.Sequential(nn.Linear(self.feature_dim, self.hidden_dim),
nn.ReLU(),
nn.Linear(self.hidden_dim, self.feature_dim),
nn.Dropout(self.dropout))
self.norm2 = nn.LayerNorm(self.feature_dim)
self.fc = nn.Linear(self.feature_dim, self.out_dim)