def __init__(self, embedding_matrix_list, opt):
super(BILSTM_SSA, self).__init__()
self.embed0 = nn.Embedding.from_pretrained(torch.tensor(
embedding_matrix_list[0], dtype=torch.float),
freeze=True)
self.input_drop = nn.Dropout(bilstm_ssa_config['input_drop'])
self.lstm = DynamicLSTM(bilstm_ssa_config['embed_dim'],
bilstm_ssa_config['hidden_dim'],
num_layers=bilstm_ssa_config['num_layers'],
batch_first=True,
bidirectional=True,
dropout=bilstm_ssa_config['lstm_drop'])
self.self_attn = _SSA(input_size=bilstm_ssa_config['hidden_dim'] * 2,
hidden_size=bilstm_ssa_config['hidden_dim'],
n_hop=bilstm_ssa_config['n_hop'])
self.drop = nn.Dropout(0.5)
self.dense = nn.Linear(bilstm_ssa_config['hidden_dim'] * 2,
opt.polarities_dim)
def __init__(self, embedding_matrix, opt):
super(BiLSTM, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.dropout = nn.Dropout(0.5)
self.clause_cnn = ClauseCNN(opt)
self.bilstm = DynamicLSTM(opt.hidden_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
bidirectional=True)
self.emotion_fc = nn.Linear(2 * opt.hidden_dim, 2 * opt.MLP_out_dim)
self.cause_fc = nn.Linear(2 * opt.hidden_dim, 2 * opt.MLP_out_dim)
self.pair_biaffine = Biaffine(opt,
opt.MLP_out_dim,
opt.MLP_out_dim,
1,
bias=(True, False))
self.emotion_fc1 = nn.Linear(opt.MLP_out_dim, opt.polarities_dim)
self.cause_fc1 = nn.Linear(opt.MLP_out_dim, opt.polarities_dim)
def __init__(self, _, hidden_dim, batch_size, max_length, n_class, n_tag, embedding_matrix):
super(Joint_model, self).__init__()
self.hidden_dim = hidden_dim
self.batch_size = batch_size
self.max_length = max_length
self.n_class = n_class
self.n_tag = n_tag
self.LayerNorm = LayerNorm(self.hidden_dim, eps=1e-12)
self.emb_drop = nn.Dropout(config.emb_dorpout)
self.embed = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float), padding_idx=0)
self.embed.weight.requires_grad = True
self.biLSTM = DynamicLSTM(config.emb_dim, config.hidden_dim // 2, bidirectional=True, batch_first=True,
dropout=config.lstm_dropout, num_layers=1)
self.intent_fc = nn.Linear(self.hidden_dim, self.n_class)
self.slot_fc = nn.Linear(self.hidden_dim, self.n_tag)
self.I_S_Emb = Label_Attention(self.intent_fc, self.slot_fc)
self.T_block1 = I_S_Block(self.intent_fc, self.slot_fc, self.hidden_dim)
self.T_block2 = I_S_Block(self.intent_fc, self.slot_fc, self.hidden_dim)
self.T_block3 = I_S_Block(self.intent_fc, self.slot_fc, self.hidden_dim)
self.crflayer = CRF(self.n_tag)
self.criterion = nn.CrossEntropyLoss()
def __init__(self, embedding_matrix, opt):
super(ASBIGCN, self).__init__()
self.opt = opt
# self.mul1=mutualatt(2*opt.hidden_dim)
# self.mul2=mutualatt(2*opt.hidden_dim)
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.text_lstm = DynamicLSTM(768,
opt.hidden_dim,
num_layers=1,
batch_first=True,
bidirectional=True)
# self.text_lstm1 = DynamicLSTM(2*opt.hidden_dim, opt.hidden_dim, num_layers=1, batch_first=True, bidirectional=True)
# self.text_lstm2 = DynamicLSTM(opt.embed_dim, 384, num_layers=1, batch_first=True, bidirectional=True)
# self.gc1 = GraphConvolution(2*opt.hidden_dim, 2*opt.hidden_dim)
# self.gc2 = GraphConvolution(2*opt.hidden_dim, 2*opt.hidden_dim)
self.gc = simpleGraphConvolutionalignment(2 * opt.hidden_dim,
2 * opt.hidden_dim,
opt.edge_size,
bias=True)
self.fc = nn.Linear(8 * opt.hidden_dim, opt.polarities_dim)
# self.fc1 = nn.Linear(768*2,768)
self.text_embed_dropout = nn.Dropout(0.1)
def __init__(self, embedding_matrix_list, opt):
super(BILSTM_ATTENTION, self).__init__()
self.embed = nn.Embedding.from_pretrained(torch.tensor(
embedding_matrix_list[0], dtype=torch.float),
freeze=True)
# self.input_drop = nn.Dropout(bilstm_attention_config['input_drop'])
self.bilstm = DynamicLSTM(
bilstm_attention_config['embed_dim'],
bilstm_attention_config['hidden_dim'],
num_layers=bilstm_attention_config['num_layers'],
batch_first=True,
bidirectional=True)
self.weight1 = nn.Parameter(
torch.Tensor(bilstm_attention_config['hidden_dim'] * 2,
bilstm_attention_config['hidden_dim'] * 2))
self.weight2 = nn.Parameter(
torch.Tensor(bilstm_attention_config['hidden_dim'] * 2, 1))
nn.init.uniform_(self.weight1, -0.1, 0.1)
nn.init.uniform_(self.weight2, -0.1, 0.1)
self.dense = nn.Linear(bilstm_attention_config['hidden_dim'] * 2,
opt.polarities_dim)
def __init__(self,
embedding_matrix,
opt,
dropout=0.6,
alpha=0.2,
nheads=6):
"""Dense version of GAT."""
super(GAT, self).__init__()
self.dropout = dropout
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.text_lstm = DynamicLSTM(opt.embed_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
bidirectional=True)
self.opt = opt
# nfeat, nhid
self.attentions = [
GraphAttentionLayer(2 * opt.hidden_dim,
2 * opt.hidden_dim,
dropout=dropout,
alpha=alpha,
concat=True) for _ in range(nheads)
]
for i, attention in enumerate(self.attentions):
self.add_module('attention_{}'.format(i), attention)
self.out_att = GraphAttentionLayer(2 * opt.hidden_dim,
2 * opt.hidden_dim,
dropout=dropout,
alpha=alpha,
concat=False)
self.text_embed_dropout = nn.Dropout(0.3)
self.fc = nn.Linear(2 * opt.hidden_dim, opt.polarities_dim)
def __init__(self, embedding_matrix, opt):
super(TD_LSTM, self).__init__()
self.embed = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float))
self.lstm_l = DynamicLSTM(opt.embed_dim, opt.hidden_dim, num_layers=1, batch_first=True)
self.lstm_r = DynamicLSTM(opt.embed_dim, opt.hidden_dim, num_layers=1, batch_first=True)
self.dense = nn.Linear(opt.hidden_dim*2, opt.polarities_dim)
def __init__(self, embedding_matrix, opt):
super(LSTM, self).__init__()
self.embed = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float), freeze=False)
self.lstm = DynamicLSTM(opt.embed_dim, opt.hidden_dim, num_layers=1, batch_first=True, bidirectional=True)
self.dense = nn.Linear(opt.hidden_dim, opt.polarities_dim)
def __init__(self, embedding_matrix, opt):
super(MMFUSION, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.lstm_aspect = DynamicLSTM(
opt.embed_dim, opt.hidden_dim, num_layers=1,
batch_first=True) #, dropout = opt.dropout_rate
self.lstm_l = DynamicLSTM(
opt.embed_dim, opt.hidden_dim, num_layers=1,
batch_first=True) #, dropout = opt.dropout_rate
self.lstm_r = DynamicLSTM(
opt.embed_dim, opt.hidden_dim, num_layers=1,
batch_first=True) #, dropout = opt.dropout_rate
self.attention_l = Attention2(opt.hidden_dim,
score_function='bi_linear',
dropout=opt.dropout_rate)
self.attention_r = Attention2(opt.hidden_dim,
score_function='bi_linear',
dropout=opt.dropout_rate)
self.visaspect_att_l = MMAttention(opt.hidden_dim,
score_function='bi_linear',
dropout=opt.dropout_rate)
self.visaspect_att_r = MMAttention(opt.hidden_dim,
score_function='bi_linear',
dropout=opt.dropout_rate)
self.ltext2hidden = nn.Linear(opt.hidden_dim, opt.hidden_dim)
self.laspect2hidden = nn.Linear(opt.hidden_dim, opt.hidden_dim)
self.rtext2hidden = nn.Linear(opt.hidden_dim, opt.hidden_dim)
self.raspect2hidden = nn.Linear(opt.hidden_dim, opt.hidden_dim)
self.dropout = nn.Dropout(self.opt.dropout_rate)
#self.viscontext_att_aspect = MMAttention(opt.hidden_dim, score_function='mlp', dropout=opt.dropout_rate)
#self.visaspect_att_context = MMAttention(opt.hidden_dim, score_function='mlp', dropout=opt.dropout_rate)
self.aspect2text = nn.Linear(opt.hidden_dim, opt.hidden_dim)
self.vismap2text = nn.Linear(2048, opt.hidden_dim)
self.vis2text = nn.Linear(2048, opt.hidden_dim)
self.gate = nn.Linear(2048 + 4 * opt.hidden_dim, opt.hidden_dim)
self.madality_attetion = nn.Linear(opt.hidden_dim, 1)
#blinear interaction between text vectors and image vectors
#self.text2hidden = nn.Linear(opt.hidden_dim*3, opt.hidden_dim)
#self.vis2hidden = nn.Linear(opt.hidden_dim, opt.hidden_dim)
#self.hidden2final = nn.Linear(opt.hidden_dim, opt.hidden_dim)
#self.text2hiddentext = nn.Linear(opt.hidden_dim*4, opt.hidden_dim*4)
#self.vis2hiddentext = nn.Linear(opt.hidden_dim, opt.hidden_dim*4)
self.text2hiddenvis = nn.Linear(opt.hidden_dim * 4, opt.hidden_dim)
self.vis2hiddenvis = nn.Linear(opt.hidden_dim, opt.hidden_dim)
#self.dense_2 = nn.Linear(opt.hidden_dim*2, opt.polarities_dim)
#self.dense_3 = nn.Linear(opt.hidden_dim*3, opt.polarities_dim)
#self.dense_4 = nn.Linear(opt.hidden_dim*4, opt.polarities_dim)
#self.dense_5 = nn.Linear(opt.hidden_dim*5, opt.polarities_dim)
#self.dense_10 = nn.Linear(opt.hidden_dim*10, opt.polarities_dim)
if self.opt.att_mode == 'vis_concat_attimg' or self.opt.att_mode == 'vis_concat':
self.dense_5 = nn.Linear(opt.hidden_dim * 5, opt.polarities_dim)
elif self.opt.att_mode == 'vis_concat_attimg_gate':
if self.opt.tfn:
self.dense_6 = nn.Linear(opt.hidden_dim * opt.hidden_dim,
opt.polarities_dim)
else:
self.dense_6 = nn.Linear(opt.hidden_dim * 6,
opt.polarities_dim)
