def __init__(self, bert, opt):
super(ATAE_LSTM, self).__init__()
self.opt = opt
self.bert = bert
self.squeeze_embedding = SqueezeEmbedding()
self.dropout = nn.Dropout(opt.dropout)
self.squeeze_embedding = SqueezeEmbedding()
self.lstm = DynamicLSTM(opt.bert_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True)
self.attention = NoQueryAttention(opt.hidden_dim + opt.bert_dim,
score_function='bi_linear')
self.dense = nn.Linear(opt.hidden_dim, opt.polarities_dim)
def __init__(self, embedding_matrix, opt):
super(AEN, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding()
self.attn_k = Attention(opt.embed_dim,
out_dim=opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.attn_q = Attention(opt.embed_dim,
out_dim=opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.ffn_c = PositionwiseFeedForward(opt.hidden_dim,
dropout=opt.dropout)
self.ffn_t = PositionwiseFeedForward(opt.hidden_dim,
dropout=opt.dropout)
self.attn_s1 = Attention(opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.dense = nn.Linear(opt.hidden_dim * 3, opt.polarities_dim)
def __init__(self, bert, opt):
super(AEN_BERT, self).__init__()
self.opt = opt
self.bert = bert
self.squeeze_embedding = SqueezeEmbedding()
self.dropout = nn.Dropout(opt.dropout)
self.attn_k = Attention(opt.bert_dim,
out_dim=opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.attn_q = Attention(opt.bert_dim,
out_dim=opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.ffn_c = PositionwiseFeedForward(opt.hidden_dim,
dropout=opt.dropout)
self.ffn_t = PositionwiseFeedForward(opt.hidden_dim,
dropout=opt.dropout)
self.attn_s1 = Attention(opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.dense = nn.Linear(opt.hidden_dim * 3, opt.polarities_dim)
def __init__(self, embedding_matrix, opt):
super(Test, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding()
self.layer_stack = nn.ModuleList([
EncoderLayer(d_model,
d_inner,
n_head,
d_k,
d_v,
dropout=opt.dropout) for _ in range(n_layers)
])
# self.ffn_c = PositionwiseFeedForward(opt.hidden_dim, dropout=opt.dropout)
# self.ffn_t = PositionwiseFeedForward(opt.hidden_dim, dropout=opt.dropout)
self.attn_s1 = Attention(opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.dense = nn.Linear(opt.hidden_dim * 3, opt.polarities_dim)
def __init__(self, bert, opt):
"""
注意力编码器网络, Attentional Encoder Network for Targeted Sentiment Classification
:param bert:
:param opt:
"""
super(AEN_BERT, self).__init__()
self.opt = opt
self.bert = bert
self.squeeze_embedding = SqueezeEmbedding()
self.dropout = nn.Dropout(opt.dropout)
# attn_k和 attn_q的初始化
self.attn_k = Attention(opt.bert_dim,
out_dim=opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.attn_q = Attention(opt.bert_dim,
out_dim=opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
# 初始化ffn_c, PCT层
self.ffn_c = PositionwiseFeedForward(opt.hidden_dim,
dropout=opt.dropout)
self.ffn_t = PositionwiseFeedForward(opt.hidden_dim,
dropout=opt.dropout)
# 目标特定的注意力层初始化
self.attn_s1 = Attention(opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
# 最终输出层定义
self.dense = nn.Linear(opt.hidden_dim * 3, opt.polarities_dim)
def __init__(self, bert, opt):
super(AEN_BERT, self).__init__()
#print(" 1 In AEN_BERT ")
self.opt = opt
self.bert = bert
self.squeeze_embedding = SqueezeEmbedding()
self.dropout = nn.Dropout(opt.dropout)
#print(" 2 In AEN_BERT ")
self.attn_k = Attention(opt.bert_dim,
out_dim=opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.attn_q = Attention(opt.bert_dim,
out_dim=opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.ffn_c = PositionwiseFeedForward(opt.hidden_dim,
dropout=opt.dropout)
self.ffn_t = PositionwiseFeedForward(opt.hidden_dim,
dropout=opt.dropout)
#print(" 3 In AEN_BERT ")
self.attn_s1 = Attention(opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.hat = False
self.last = torch.nn.ModuleList()
for t in range(self.opt.taskcla):
self.last.append(nn.Linear(opt.hidden_dim * 3, opt.polarities_dim))
def __init__(self, embedding_matrix, opt, type='cabasc'):
super(Cabasc, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding(batch_first=True)
self.attention = Attention(opt.embed_dim,
score_function='mlp',
dropout=opt.dropout) # content attention
self.m_linear = nn.Linear(opt.embed_dim, opt.embed_dim, bias=False)
self.mlp = nn.Linear(opt.embed_dim, opt.embed_dim) # W4
self.dense = nn.Linear(opt.embed_dim, opt.polarities_dim) # W5
# context attention layer
self.rnn_l = DynamicLSTM(opt.embed_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
rnn_type='GRU')
self.rnn_r = DynamicLSTM(opt.embed_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
rnn_type='GRU')
self.mlp_l = nn.Linear(opt.hidden_dim, 1)
self.mlp_r = nn.Linear(opt.hidden_dim, 1)
def __init__(self, embedding_matrix, opt):
super(ATAE_LSTM, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding()
self.lstm = DynamicLSTM(opt.embed_dim * 2,
opt.hidden_dim,
num_layers=1,
batch_first=True)
self.attention = NoQueryAttention(opt.hidden_dim + opt.embed_dim,
score_function='bi_linear')
self.dense = nn.Linear(opt.hidden_dim, opt.polarities_dim)
self.cls_lstm = DynamicLSTM(opt.embed_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
bidirectional=True)
self.cls_asp_lstm = DynamicLSTM(opt.embed_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
bidirectional=True)
self.h_h_proj = nn.Linear(4 * opt.hidden_dim, opt.hidden_dim)
self.cls_dense = nn.Linear(opt.hidden_dim, opt.polarities_dim)
self.dense_total = nn.Linear(opt.hidden_dim + opt.hidden_dim,
opt.polarities_dim)
self.conv = nn.Conv1d(300, 2 * opt.hidden_dim, 3)
def __init__(self, embedding_matrix, opt):
super(MemNet, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding(batch_first=True)
self.attention = Attention(opt.embed_dim, score_function='mlp')
self.x_linear = nn.Linear(opt.embed_dim, opt.embed_dim)
self.dense = nn.Linear(opt.embed_dim, opt.polarities_dim)
def __init__(self, config, opt):
super(MemNet, self).__init__()
self.opt = opt
self.bert = BertModel(config)
self.squeeze_embedding = SqueezeEmbedding(batch_first=True)
self.attention = Attention(opt.embed_dim, score_function='mlp')
self.x_linear = nn.Linear(opt.embed_dim, opt.embed_dim)
self.dense = nn.Linear(opt.embed_dim, opt.output_dim)
def __init__(self, embedding_matrix, opt): super(ATAE_LSTM, self).__init__() self.opt = opt #Parser arguments assigned self.embed = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float)) #Initialized the word embedding matrix with the GloVe word embeddings for fine tuning self.squeeze_embedding = SqueezeEmbedding() #Instance of Squeeze Embedding Class from layers.squeeze_embedding created self.lstm = DynamicLSTM(opt.embed_dim*2, opt.hidden_dim, num_layers=1, batch_first=True) #Instance of DynamicLSTM Class from layers.dynamic_rnn created parameters passed, input_size=opt.embed_dim*2, hidden_size=opt.hidden_dim self.attention = NoQueryAttention(opt.hidden_dim+opt.embed_dim, score_function='bi_linear') #Instance of NoQueryAttention Class from layers.attention created parameters passes, embed_dim=opt.hidden_dim+opt.embed_dim and score function is 'bi_linear' self.dense = nn.Linear(opt.hidden_dim, opt.polarities_dim) #A linear layer is created with in_features=opt.hidden_dim, out_features=opt.polarities_dim
def __init__(self, embedding_matrix, opt):
super(Gated_CNN, self).__init__()
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.in_layer = GLU_Block(opt.kernel_size, opt.embed_dim,
opt.in_channels, opt.downbot)
self.squeeze_embedding = SqueezeEmbedding()
self.glu_layers = self.make_glu_layers(opt)
self.attention = NoQueryAttention(opt.in_channels,
score_function='bi_linear')
self.dense = nn.Linear(opt.in_channels, opt.polarities_dim)
def __init__(self, embedding_matrix, opt):
super(AEGCN, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding() #
self.text_lstm = DynamicLSTM(opt.embed_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
bidirectional=True)
self.aspect_lstm = DynamicLSTM(opt.embed_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
bidirectional=True)
self.attn_k = Attention(opt.embed_dim * 2,
out_dim=opt.hidden_dim,
n_head=opt.head,
score_function='mlp',
dropout=opt.dropout) #
self.attn_a = Attention(opt.embed_dim * 2,
out_dim=opt.hidden_dim,
n_head=opt.head,
score_function='mlp',
dropout=opt.dropout) #
# self.attn_s1 = Attention(opt.embed_dim*2, out_dim=opt.hidden_dim, n_head=3, score_function='mlp', dropout=0.5)
self.attn_q = Attention(opt.embed_dim * 2,
out_dim=opt.hidden_dim,
n_head=opt.head,
score_function='mlp',
dropout=opt.dropout) #
self.gc1 = GraphConvolution(2 * opt.hidden_dim, 2 * opt.hidden_dim,
opt)
self.gc2 = GraphConvolution(2 * opt.hidden_dim, 2 * opt.hidden_dim,
opt)
self.attn_k_q = Attention(opt.hidden_dim,
n_head=opt.head,
score_function='mlp',
dropout=opt.dropout) #
self.attn_k_a = Attention(opt.hidden_dim,
n_head=opt.head,
score_function='mlp',
dropout=opt.dropout)
#self.fc = nn.Linear(2*opt.hidden_dim, opt.polarities_dim)
self.text_embed_dropout = nn.Dropout(opt.dropout)
self.aspect_embed_dropout = nn.Dropout(opt.dropout)
self.dense = nn.Linear(opt.hidden_dim * 3, opt.polarities_dim)
def __init__(self, config, opt):
super(ATAE_LSTM, self).__init__()
self.opt = opt
# self.embed = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float))
self.bert = BertModel(config)
self.squeeze_embedding = SqueezeEmbedding()
self.lstm = DynamicLSTM(opt.embed_dim * 2,
opt.hidden_dim,
num_layers=1,
batch_first=True)
self.attention = NoQueryAttention(opt.hidden_dim + opt.embed_dim,
score_function='bi_linear')
self.dense = nn.Linear(opt.hidden_dim, opt.output_dim)
def __init__(self, embedding_matrix, opt):
super(GCAE, self).__init__()
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding()
self.cnn_context_x = nn.ModuleList(
[self.cnn_layer(opt, k) for k in [3, 4, 5]])
self.cnn_context_y = nn.ModuleList(
[self.cnn_layer(opt, k) for k in [3, 4, 5]])
self.cnn_aspect = nn.Sequential(
nn.ConstantPad2d((0, 0, 2, 0), 0),
nn.Conv2d(opt.embed_dim, opt.in_channels, (3, 1)))
self.dropout = nn.Dropout(opt.dropout)
self.dense = nn.Linear(3 * opt.in_channels, opt.polarities_dim)
def __init__(self, embedding_matrix, opt):
super(BaseB, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding(batch_first=True)
self.attention = Attention(opt.embed_dim, score_function='mlp')
self.m_linear = nn.Linear(
opt.embed_dim, opt.embed_dim, bias=False
) # FwNN3: I think this should be moved to attention module
self.x_linear = nn.Linear(opt.embed_dim, opt.embed_dim, bias=False)
self.s_linear = nn.Linear(opt.embed_dim, opt.embed_dim, bias=False)
self.mlp = nn.Linear(opt.embed_dim, opt.embed_dim) # W4
self.dense = nn.Linear(opt.embed_dim, opt.polarities_dim) # W5
def __init__(self, bert, opt):
super(SDGCN_NEW, self).__init__()
self.opt = opt
self.bert = bert
self.squeeze_embedding = SqueezeEmbedding()
self.dropout = nn.Dropout(opt.dropout)
self.lstm_context = sLSTM(opt.bert_dim, 768, window_size=2)
self.lstm_aspect = sLSTM(opt.bert_dim, 768, window_size=2)
# self.lstm_context = DynamicLSTM(opt.bert_dim, opt.hidden_dim, num_layers=1, batch_first=True,bidirectional=True)
# self.lstm_aspect = DynamicLSTM(opt.bert_dim, opt.hidden_dim, num_layers=1, batch_first=True,bidirectional=True)
self.attention_aspect = Attention(768, score_function='bi_linear')
self.attention_context = Attention(768, score_function='bi_linear')
self.dense = nn.Linear(768 * 2, opt.polarities_dim)
def __init__(self, embedding_matrix, opt):
super(MemNet2, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(torch.tensor(
embedding_matrix, dtype=torch.float),
freeze=False)
self.squeeze_embedding = SqueezeEmbedding(batch_first=True)
self.bi_lstm_img = DynamicLSTM(opt.embed_dim_img,
opt.embed_dim,
num_layers=1,
batch_first=True,
bidirectional=False)
self.attention = Attention(opt.embed_dim, score_function='mlp')
self.x_linear = nn.Linear(opt.embed_dim, opt.embed_dim)
self.dense = nn.Linear(opt.embed_dim * 2, opt.polarities_dim)
def __init__(self, embedding_matrix, opt):
super(Cabasc, self).__init__()
self.opt = opt
self.dropout = nn.Dropout(0.5)
self.embed = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float))
if opt.freeze_embeddings == "yes":
self.embed.weight.requieres_grad = False
self.squeeze_embedding = SqueezeEmbedding(batch_first=True)
self.linear1 = nn.Linear(3*opt.embed_dim, opt.embed_dim)
self.linear2 = nn.Linear(opt.embed_dim, 1, bias=False)
self.mlp = nn.Linear(opt.embed_dim, opt.embed_dim)
self.dense = nn.Linear(opt.embed_dim, opt.polarities_dim)
if opt.model_name == 'cabasc':
self.rnn_l = DynamicLSTM(opt.embed_dim, opt.hidden_dim, num_layers=1, batch_first=True, rnn_type = 'GRU')
self.rnn_r = DynamicLSTM(opt.embed_dim, opt.hidden_dim, num_layers=1, batch_first=True, rnn_type = 'GRU')
self.mlp_l = nn.Linear(opt.hidden_dim, 1)
self.mlp_r = nn.Linear(opt.hidden_dim, 1)
def __init__(self, bert, opt):
super(SDGCN, self).__init__()
self.opt = opt
self.bert = bert
self.squeeze_embedding = SqueezeEmbedding()
self.dropout = nn.Dropout(opt.dropout)
self.lstm_context = DynamicLSTM(opt.bert_dim,
opt.hidden_dim,
bidirectional=True,
num_layers=1,
batch_first=True)
# self.lstm_aspect = DynamicLSTM(opt.bert_dim, opt.hidden_dim, num_layers=1, batch_first=True)
# self.attention_aspect = Attention(opt.hidden_dim, score_function='bi_linear')
# self.attention_context = Attention(opt.hidden_dim, score_function='bi_linear')
self.gc1 = GraphConvolution(2 * opt.hidden_dim, 2 * opt.hidden_dim)
self.gc2 = GraphConvolution(2 * opt.hidden_dim, 2 * opt.hidden_dim)
self.fc = nn.Linear(2 * opt.hidden_dim, opt.polarities_dim)
def __init__(self, bert, opt):
super(RAM, self).__init__()
self.opt = opt
self.bert = bert
self.squeeze_embedding = SqueezeEmbedding()
self.dropout = nn.Dropout(opt.dropout)
self.bi_lstm_context = DynamicLSTM(opt.bert_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
bidirectional=True)
self.att_linear = nn.Linear(opt.hidden_dim * 2 + 1 + opt.bert_dim * 2,
1)
self.gru_cell = nn.GRUCell(opt.hidden_dim * 2 + 1, opt.bert_dim)
self.dense = nn.Linear(opt.bert_dim, opt.polarities_dim)
def __init__(self, bert, opt):
super(SDGCN, self).__init__()
self.opt = opt
self.bert = bert
self.squeeze_embedding = SqueezeEmbedding()
self.dropout = nn.Dropout(opt.dropout)
self.lstm_context = DynamicLSTM(opt.bert_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True)
self.lstm_aspect = DynamicLSTM(opt.bert_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True)
self.attention_aspect = Bilinear_Attention(opt.hidden_dim)
self.attention_context = Bilinear_Attention(opt.hidden_dim)
self.dense = nn.Linear(opt.hidden_dim * 2, opt.polarities_dim)
def __init__(self, bert_or_embedding_matrix, opt):
super(BERT_ALBERT_GCN, self).__init__()
self.opt = opt
bert = bert_or_embedding_matrix
self.context_bert = bert
self.squeeze_embedding = SqueezeEmbedding()
if not opt.no_gnn:
self.gcn = GNN_RELU_DIFF(opt.bert_dim,
opt.bert_dim,
step=opt.gnn_step,
drop=0.0)
self.encoder_fn = nn.Linear(opt.bert_dim, opt.bert_dim)
self.norm = LayerNorm(opt.bert_dim)
if not opt.no_sa:
self.bert_self_att = SelfAttention(bert.config, opt)
if self.opt.pool_tp == 'bert_pool':
self.bert_pooler = BertPooler(bert.config)
elif self.opt.pool_tp == 'max_pool':
self.pool_fc = nn.Linear(opt.bert_dim, opt.bert_dim)
self.fc = nn.Linear(opt.bert_dim, 3)
def __init__(self, bert, opt):
super(AEN_BERT_HAT, self).__init__()
self.opt = opt
self.bert = bert
self.squeeze_embedding = SqueezeEmbedding()
self.dropout = nn.Dropout(opt.dropout)
self.attn_k = Attention(opt.bert_dim,
out_dim=opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.attn_q = Attention(opt.bert_dim,
out_dim=opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.ffn_c = PositionwiseFeedForward(opt.hidden_dim,
dropout=opt.dropout)
self.ffn_t = PositionwiseFeedForward(opt.hidden_dim,
dropout=opt.dropout)
self.attn_s1 = Attention(opt.hidden_dim,
n_head=8,
score_function='mlp',
dropout=opt.dropout)
self.gate = torch.nn.Sigmoid()
self.ec0 = torch.nn.Embedding(self.opt.taskcla, opt.hidden_dim)
self.ec1 = torch.nn.Embedding(self.opt.taskcla, opt.hidden_dim)
self.ec2 = torch.nn.Embedding(self.opt.taskcla, opt.hidden_dim)
self.ec3 = torch.nn.Embedding(self.opt.taskcla, opt.hidden_dim)
self.ec4 = torch.nn.Embedding(self.opt.taskcla, opt.hidden_dim)
self.hat = True
self.last = torch.nn.ModuleList()
for t in range(self.opt.taskcla):
self.last.append(nn.Linear(opt.hidden_dim * 3, opt.polarities_dim))
def __init__(self, embedding_matrix, opt):
super(AEN_GloVe, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding()
self.attn_k = Attention(opt.embed_dim, out_dim=opt.hidden_dim, n_head=8, score_function='mlp', dropout=opt.dropout)
self.attn_q = Attention(opt.embed_dim, out_dim=opt.hidden_dim, n_head=8, score_function='mlp', dropout=opt.dropout)
self.ffn_c = PositionwiseFeedForward(opt.hidden_dim, dropout=opt.dropout)
self.ffn_t = PositionwiseFeedForward(opt.hidden_dim, dropout=opt.dropout)
self.attn_s1 = Attention(opt.hidden_dim, n_head=8, score_function='mlp', dropout=opt.dropout)
self.dense = nn.Linear(opt.hidden_dim*3, opt.polarities_dim)
self.cls_lstm = DynamicLSTM(opt.embed_dim, opt.hidden_dim, num_layers=1, batch_first=True, bidirectional=True)
self.cls_asp_lstm = DynamicLSTM(opt.embed_dim, opt.hidden_dim, num_layers=1, batch_first=True, bidirectional=True)
self.dense_total = nn.Linear(3*opt.hidden_dim*2, opt.polarities_dim)
self.h_h_proj = nn.Linear(4*opt.hidden_dim,3*opt.hidden_dim)
self.cls_dense = nn.Linear(3*opt.hidden_dim,3)
self.conv = nn.Conv1d(300,2*opt.hidden_dim,3)
def __init__(self, embedding_matrix, opt):
super(LifelongABSA, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding(batch_first=True)
self.attention = Attention(opt.embed_dim, score_function='mlp')
#self.x_linear = nn.Linear(opt.embed_dim, opt.embed_dim)
# self.dense = nn.Linear(opt.embed_dim, opt.polarities_dim)
self.polarities_dim = opt.polarities_dim
self.last = torch.nn.ModuleList()
self.hat = False
self.word_in_domain = torch.zeros(1, dtype=torch.int64)
self.context_attention = dict()
self.aspect_context_attention = dict()
self.currenttask = -1
self.currentSentence = 0
for t in range(self.opt.taskcla):
self.last.append(nn.Linear(opt.polarities_dim, opt.polarities_dim))
self.context_attention[t] = dict()
self.aspect_context_attention[t] = dict()
#Parameter(torch.Tensor(out_features, in_features))
self.W = torch.nn.Parameter(torch.randn(opt.embed_dim, opt.polarities_dim ))
self.opt.initializer(self.W)
#where the negative, neutral, and positive classes are denoted
# as [1, 0 ,0], [0, 1 ,0] and [0, 0 ,1] respectively
self.An = torch.tensor( np.array([[1,-1,-1]]), dtype=torch.float32, requires_grad=False )
self.Bn = torch.tensor(np.array([[1, 0, 0]]), dtype=torch.float32, requires_grad=False)
self.Ap = torch.tensor(np.array([[-1, -1, 1]]), dtype=torch.float32, requires_grad=False)
self.Bp = torch.tensor(np.array([[0, 0, 1]]), dtype=torch.float32, requires_grad=False)
self.L2MN = False
def __init__(self, embedding_matrix, opt):
super(AEN, self).__init__()
self.opt = opt
self.embed = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding()
self.attn_k = Attention(opt.embed_dim, out_dim=opt.hidden_dim, n_head=8, score_function='mlp', dropout=0.1)
self.attn_q = Attention(opt.embed_dim, out_dim=opt.hidden_dim, n_head=8, score_function='mlp', dropout=0.1)
###
D = opt.hidden_dim
Co = 100
Ks = [3, 4, 5]
C = 3
self.convs1 = nn.ModuleList([nn.Conv1d(D, Co, K) for K in Ks])
self.convs2 = nn.ModuleList([nn.Conv1d(D, Co, K) for K in Ks])
self.convs3 = nn.ModuleList([nn.Conv1d(D, Co, K, padding=K - 2) for K in [3]])
# self.convs3 = nn.Conv1d(D, 300, 3, padding=1), smaller is better
self.dropout = nn.Dropout(opt.dropout)
self.fc1 = nn.Linear(len(Ks) * Co, C)
self.fc_aspect = nn.Linear(100, Co)
def __init__(self, config, opt, _type='c'):
super(Cabasc, self).__init__()
self.opt = opt
self.type = _type
self.bert = BertModel(config)
self.squeeze_embedding = SqueezeEmbedding(batch_first=True)
self.linear1 = nn.Linear(3 * opt.embed_dim, opt.embed_dim)
self.linear2 = nn.Linear(opt.embed_dim, 1, bias=False)
self.mlp = nn.Linear(opt.embed_dim, opt.embed_dim)
self.dense = nn.Linear(opt.embed_dim, opt.output_dim)
# context attention layer
self.rnn_l = DynamicLSTM(opt.embed_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
rnn_type='GRU')
self.rnn_r = DynamicLSTM(opt.embed_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
rnn_type='GRU')
self.mlp_l = nn.Linear(opt.hidden_dim, 1)
self.mlp_r = nn.Linear(opt.hidden_dim, 1)
def __init__(self, bert, opt):
super(AEN_SIMPLE, self).__init__()
self.opt = opt
self.bert = bert
self.squeeze_embedding = SqueezeEmbedding()
self.dropout = nn.Dropout(opt.dropout)
self.attn_q = Attention(opt.bert_dim,
out_dim=opt.hidden_dim,
n_head=opt.mha_heads,
score_function='mlp',
dropout=opt.dropout)
self.ffn_t = PositionwiseFeedForward(opt.hidden_dim,
dropout=opt.dropout)
self.lstm = nn.LSTM(opt.hidden_dim,
opt.lstm_hid,
opt.lstm_layer,
batch_first=True,
bidirectional=opt.lstm_bidir)
self.dense = nn.Linear(opt.lstm_hid * (2 if opt.lstm_bidir else 1),
opt.polarities_dim)
def __init__(self, embedding_matrix, opt, _type='c'):
super(Cabasc, self).__init__()
self.opt = opt
self.type = _type
self.embed = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding(batch_first=True)
self.linear1 = nn.Linear(3 * opt.embed_dim, opt.embed_dim)
self.linear2 = nn.Linear(opt.embed_dim, 1, bias=False)
self.mlp = nn.Linear(opt.embed_dim, opt.embed_dim)
self.dense = nn.Linear(opt.embed_dim, opt.polarities_dim)
# context attention layer
self.rnn_l = DynamicLSTM(opt.embed_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
rnn_type='GRU')
self.rnn_r = DynamicLSTM(opt.embed_dim,
opt.hidden_dim,
num_layers=1,
batch_first=True,
rnn_type='GRU')
self.mlp_l = nn.Linear(opt.hidden_dim, 1)
self.mlp_r = nn.Linear(opt.hidden_dim, 1)