def __init__(self, opts, vocab, char_vocab, label_vocab):
super(Char_CNN, self).__init__()
random.seed(opts.seed)
torch.manual_seed(opts.seed)
torch.cuda.manual_seed(opts.seed)
self.embed_dim = opts.embed_size
self.char_embed_dim = opts.char_embed_size
self.word_num = vocab.m_size
self.char_num = char_vocab.m_size
self.pre_embed_path = opts.pre_embed_path
self.string2id = vocab.string2id
self.char_string2id = char_vocab.string2id
self.embed_uniform_init = opts.embed_uniform_init
self.stride = opts.stride
self.kernel_size = opts.kernel_size
self.kernel_num = opts.kernel_num
self.label_num = label_vocab.m_size
self.embed_dropout = opts.embed_dropout
self.fc_dropout = opts.fc_dropout
self.use_cuda = opts.use_cuda
self.word_embeddings = nn.Embedding(self.word_num, self.embed_dim)
self.char_embeddings = nn.Embedding(self.char_num, self.char_embed_dim)
if opts.pre_embed_path != '':
embedding = Embedding.load_predtrained_emb_zero(
self.pre_embed_path, self.string2id)
self.word_embeddings.weight.data.copy_(embedding)
else:
nn.init.uniform_(self.word_embeddings.weight.data,
-self.embed_uniform_init, self.embed_uniform_init)
nn.init.uniform_(self.char_embeddings.weight.data,
-self.embed_uniform_init, self.embed_uniform_init)
word_char_embed_dim = self.embed_dim + len(
self.kernel_size) * self.kernel_num
self.word_char_convs = nn.ModuleList([
nn.Conv2d(1,
self.kernel_num, (K, word_char_embed_dim),
stride=self.stride,
padding=(K // 2, 0)) for K in self.kernel_size
])
self.char_convs = nn.ModuleList([
nn.Conv2d(1,
self.kernel_num, (K, self.char_embed_dim),
stride=self.stride,
padding=(K // 2, 0)) for K in self.kernel_size
])
infea = len(self.kernel_size) * self.kernel_num
self.linear1 = nn.Linear(infea, infea // 2)
self.linear2 = nn.Linear(infea // 2, self.label_num)
self.embed_dropout = nn.Dropout(self.embed_dropout)
self.fc_dropout = nn.Dropout(self.fc_dropout)
def __init__(self, opts, vocab, label_vocab, rel_vocab):
super(biChildSumTreeLSTM, self).__init__()
random.seed(opts.seed)
torch.manual_seed(opts.seed)
torch.cuda.manual_seed(opts.seed)
self.embed_dim = opts.embed_size
self.word_num = vocab.m_size
self.pre_embed_path = opts.pre_embed_path
self.string2id = vocab.string2id
self.embed_uniform_init = opts.embed_uniform_init
self.label_num = label_vocab.m_size
self.rel_num = rel_vocab.m_size
self.dropout = opts.dropout
self.hidden_size = opts.hidden_size
self.hidden_num = opts.hidden_num
self.bidirectional = opts.bidirectional
self.use_cuda = opts.use_cuda
self.debug = False
self.embeddings = nn.Embedding(self.word_num, self.embed_dim)
self.dropout = nn.Dropout(self.dropout)
if opts.pre_embed_path != '':
embedding = Embedding.load_predtrained_emb_zero(self.pre_embed_path, self.string2id)
self.embeddings.weight.data.copy_(embedding)
self.dt_tree = DTTreeLSTM(self.embed_dim, self.hidden_size, opts.dropout)
self.td_tree = TDTreeLSTM(self.embed_dim, self.hidden_size, opts.dropout)
self.linear = nn.Linear(self.hidden_size * 2, self.label_num)
def __init__(self, opts, vocab, label_vocab):
super(Pooling, self).__init__()
random.seed(opts.seed)
torch.manual_seed(opts.seed)
torch.cuda.manual_seed(opts.seed)
self.embed_dim = opts.embed_size
self.word_num = vocab.m_size
self.pre_embed_path = opts.pre_embed_path
self.string2id = vocab.string2id
self.embed_uniform_init = opts.embed_uniform_init
self.label_num = label_vocab.m_size
self.embed_dropout = opts.embed_dropout
self.fc_dropout = opts.fc_dropout
self.embeddings = nn.Embedding(self.word_num, self.embed_dim)
if opts.pre_embed_path != '':
embedding = Embedding.load_predtrained_emb_zero(self.pre_embed_path, self.string2id)
self.embeddings.weight.data.copy_(embedding)
else:
nn.init.uniform_(self.embeddings.weight.data, -self.embed_uniform_init, self.embed_uniform_init)
self.linear1 = nn.Linear(self.embed_dim, self.embed_dim // 2)
self.linear2 = nn.Linear(self.embed_dim // 2, self.label_num)
self.embed_dropout = nn.Dropout(self.embed_dropout)
self.fc_dropout = nn.Dropout(self.fc_dropout)
def __init__(self, opts, vocab, label_vocab):
super(LSTM_CNN, self).__init__()
random.seed(opts.seed)
torch.manual_seed(opts.seed)
torch.cuda.manual_seed(opts.seed)
self.embed_dim = opts.embed_size
self.word_num = vocab.m_size
self.pre_embed_path = opts.pre_embed_path
self.string2id = vocab.string2id
self.embed_uniform_init = opts.embed_uniform_init
self.label_num = label_vocab.m_size
self.embed_dropout = opts.embed_dropout
self.fc_dropout = opts.fc_dropout
# CNN
self.stride = opts.stride
self.kernel_size = opts.kernel_size
self.kernel_num = opts.kernel_num
# RNN
self.hidden_num = opts.hidden_num
self.hidden_size = opts.hidden_size
self.hidden_dropout = opts.hidden_dropout
self.bidirectional = opts.bidirectional
self.flag = 2 if self.bidirectional else 1
self.embeddings = nn.Embedding(self.word_num, self.embed_dim)
if opts.pre_embed_path != '':
embedding = Embedding.load_predtrained_emb_zero(
self.pre_embed_path, self.string2id)
self.embeddings.weight.data.copy_(embedding)
else:
nn.init.uniform_(self.embeddings.weight.data,
-self.embed_uniform_init, self.embed_uniform_init)
self.convs = nn.ModuleList([
nn.Conv2d(1,
self.kernel_num, (K, self.hidden_size * self.flag),
stride=self.stride,
padding=(K // 2, 0)) for K in self.kernel_size
])
self.lstm = nn.LSTM(self.embed_dim,
self.hidden_size,
dropout=self.hidden_dropout,
num_layers=self.hidden_num,
batch_first=True,
bidirectional=self.bidirectional)
in_fea = len(self.kernel_size) * self.kernel_num
self.linear1 = nn.Linear(in_fea, in_fea // 2)
self.linear2 = nn.Linear(in_fea // 2, self.label_num)
self.embed_dropout = nn.Dropout(self.embed_dropout)
self.fc_dropout = nn.Dropout(self.fc_dropout)
def __init__(self, opts, vocab, label_vocab):
super(LSTM_TreeLSTM, self).__init__()
random.seed(opts.seed)
torch.manual_seed(opts.seed)
torch.cuda.manual_seed(opts.seed)
self.embed_dim = opts.embed_size
self.word_num = vocab.m_size
self.pre_embed_path = opts.pre_embed_path
self.string2id = vocab.string2id
self.embed_uniform_init = opts.embed_uniform_init
self.label_num = label_vocab.m_size
self.embed_dropout = opts.embed_dropout
self.fc_dropout = opts.fc_dropout
self.hidden_size = opts.hidden_size
self.hidden_num = opts.hidden_num
self.hidden_dropout = opts.hidden_dropout
self.bidirectional = opts.bidirectional
self.flag = 2 if self.bidirectional else 1
self.use_cuda = opts.use_cuda
self.debug = False
self.embeddings = nn.Embedding(self.word_num, self.embed_dim)
if opts.pre_embed_path != '':
embedding = Embedding.load_predtrained_emb_zero(
self.pre_embed_path, self.string2id)
self.embeddings.weight.data.copy_(embedding)
self.lstm = nn.LSTM(self.embed_dim,
self.hidden_size,
num_layers=self.hidden_num,
dropout=self.hidden_dropout,
batch_first=True,
bidirectional=self.bidirectional)
# build lstm
self.ix = nn.Linear(self.embed_dim, self.hidden_size)
self.ih = nn.Linear(self.hidden_size, self.hidden_size)
self.fx = nn.Linear(self.embed_dim, self.hidden_size)
self.fh = nn.Linear(self.hidden_size, self.hidden_size)
self.ox = nn.Linear(self.embed_dim, self.hidden_size)
self.oh = nn.Linear(self.hidden_size, self.hidden_size)
self.ux = nn.Linear(self.embed_dim, self.hidden_size)
self.uh = nn.Linear(self.hidden_size, self.hidden_size)
self.linear1 = nn.Linear(self.hidden_size * (self.flag + 1),
self.hidden_size // 2)
self.linear2 = nn.Linear(self.hidden_size // 2, self.label_num)
self.hidden_dropout = nn.Dropout(self.hidden_dropout)
self.embed_dropout = nn.Dropout(self.embed_dropout)
self.fc_dropout = nn.Dropout(self.fc_dropout)
def __init__(self, opts, vocab, label_vocab):
super(ChildSumTreeLSTM, self).__init__()
random.seed(opts.seed)
torch.manual_seed(opts.seed)
torch.cuda.manual_seed(opts.seed)
# embedding parameters
self.embed_dim = opts.embed_size
self.word_num = vocab.m_size
self.pre_embed_path = opts.pre_embed_path
self.str2idx = vocab.str2idx
self.embed_uniform_init = opts.embed_uniform_init
self.label_num = label_vocab.m_size
self.embed_dropout = opts.embed_dropout
# network parameters
self.fc_dropout = opts.fc_dropout
self.hidden_size = opts.hidden_size
self.use_cuda = opts.use_cuda
self.embeddings = nn.Embedding(self.word_num, self.embed_dim)
if opts.pre_embed_path != '':
embedding = Embedding.load_predtrained_emb_zero(self.pre_embed_path, self.str2idx)
self.embeddings.weight.data.copy_(embedding)
# build lstm: following the notations as the paper
# input unit
self.ix = nn.Linear(self.embed_dim, self.hidden_size)
self.ih = nn.Linear(self.hidden_size, self.hidden_size)
# forget unit
self.fx = nn.Linear(self.embed_dim, self.hidden_size)
self.fh = nn.Linear(self.hidden_size, self.hidden_size)
# output unit
self.ox = nn.Linear(self.embed_dim, self.hidden_size)
self.oh = nn.Linear(self.hidden_size, self.hidden_size)
# utility
self.ux = nn.Linear(self.embed_dim, self.hidden_size)
self.uh = nn.Linear(self.hidden_size, self.hidden_size)
self.out = nn.Linear(self.hidden_size, self.label_num)
self.embed_dropout = nn.Dropout(self.embed_dropout)
self.fc_dropout = nn.Dropout(self.fc_dropout)
if self.use_cuda:
self.loss = self.loss.cuda()
def __init__(self, opts, vocab, label_vocab):
super(CNN, self).__init__()
random.seed(opts.seed)
torch.cuda.manual_seed(opts.gpu_seed)
# embedding parameters
self.embed_dim = opts.embed_size
self.vocab_size = vocab.m_size
self.pre_embed_path = opts.pre_embed_path
self.str2idx = vocab.str2idx
self.embed_uniform_init = opts.embed_uniform_init
# network parameters
self.stride = opts.stride
self.kernel_size = opts.kernel_size
self.kernel_num = opts.kernel_num
self.label_num = label_vocab.m_size
self.embed_dropout = opts.embed_dropout
self.fc_dropout = opts.fc_dropout
# embeddings
self.embeddings = nn.Embedding(self.vocab_size, self.embed_dim)
if opts.pre_embed_path != '':
embedding = Embedding.load_predtrained_emb_zero(
self.pre_embed_path, self.str2idx)
self.embeddings.weight.data.copy_(embedding)
else:
nn.init.uniform_(self.embeddings.weight.data,
-self.embed_uniform_init, self.embed_uniform_init)
self.convs = nn.ModuleList([
nn.Conv2d(1,
self.kernel_num, (K, self.embed_dim),
stride=self.stride,
padding=(K // 2, 0)) for K in self.kernel_size
])
# torch.nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True)
in_fea = len(self.kernel_size) * self.kernel_num
# dense layer
# torch.nn.Linear(in_features, out_features, bias=True)
self.linear1 = nn.Linear(in_fea, in_fea // 2)
self.linear2 = nn.Linear(in_fea // 2, self.label_num)
self.embed_dropout = nn.Dropout(self.embed_dropout)
self.fc_dropout = nn.Dropout(self.fc_dropout)
def __init__(self, opts, vocab, label_vocab):
super(ChildSumTreeLSTM, self).__init__()
random.seed(opts.seed)
torch.manual_seed(opts.seed)
torch.cuda.manual_seed(opts.seed)
self.embed_dim = opts.embed_size
self.word_num = vocab.m_size
self.pre_embed_path = opts.pre_embed_path
self.string2id = vocab.string2id
self.embed_uniform_init = opts.embed_uniform_init
self.label_num = label_vocab.m_size
self.embed_dropout = opts.embed_dropout
self.fc_dropout = opts.fc_dropout
self.hidden_size = opts.hidden_size
self.use_cuda = opts.use_cuda
self.embeddings = nn.Embedding(self.word_num, self.embed_dim)
if opts.pre_embed_path != '':
embedding = Embedding.load_predtrained_emb_zero(
self.pre_embed_path, self.string2id)
self.embeddings.weight.data.copy_(embedding)
# build lstm
self.ix = nn.Linear(self.embed_dim, self.hidden_size)
self.ih = nn.Linear(self.hidden_size, self.hidden_size)
self.fx = nn.Linear(self.embed_dim, self.hidden_size)
self.fh = nn.Linear(self.hidden_size, self.hidden_size)
self.ox = nn.Linear(self.embed_dim, self.hidden_size)
self.oh = nn.Linear(self.hidden_size, self.hidden_size)
self.ux = nn.Linear(self.embed_dim, self.hidden_size)
self.uh = nn.Linear(self.hidden_size, self.hidden_size)
self.out = nn.Linear(self.hidden_size, self.label_num)
self.embed_dropout = nn.Dropout(self.embed_dropout)
self.fc_dropout = nn.Dropout(self.fc_dropout)
if self.use_cuda:
self.loss = self.loss.cuda()
def __init__(self, opts, vocab, label_vocab):
super(Child_Sum_Tree_LSTM, self).__init__()
random.seed(opts.seed)
torch.manual_seed(opts.seed)
torch.cuda.manual_seed(opts.seed)
self.embed_dim = opts.embed_size
self.word_num = vocab.m_size
self.pre_embed_path = opts.pre_embed_path
self.string2id = vocab.string2id
self.embed_uniform_init = opts.embed_uniform_init
self.label_num = label_vocab.m_size
self.embed_dropout = opts.embed_dropout
self.fc_dropout = opts.fc_dropout
self.embeddings = nn.Embedding(self.word_num, self.embed_dim)
if opts.pre_embed_path != '':
embedding = Embedding.load_predtrained_emb_zero(self.pre_embed_path, self.string2id)
self.embeddings.weight.data.copy_(embedding)
def __init__(self, opts, vocab, label_vocab):
super(GRU, self).__init__()
random.seed(opts.seed)
torch.manual_seed(opts.seed)
torch.cuda.manual_seed(opts.seed)
self.embed_dim = opts.embed_size
self.word_num = vocab.m_size
self.pre_embed_path = opts.pre_embed_path
self.string2id = vocab.string2id
self.embed_uniform_init = opts.embed_uniform_init
self.label_num = label_vocab.m_size
self.embed_dropout = opts.embed_dropout
self.fc_dropout = opts.fc_dropout
self.hidden_num = opts.hidden_num
self.hidden_size = opts.hidden_size
self.hidden_dropout = opts.hidden_dropout
self.bidirectional = opts.bidirectional
self.embeddings = nn.Embedding(self.word_num, self.embed_dim)
if opts.pre_embed_path != '':
embedding = Embedding.load_predtrained_emb_zero(self.pre_embed_path, self.string2id)
self.embeddings.weight.data.copy_(embedding)
else:
nn.init.uniform_(self.embeddings.weight.data, -self.embed_uniform_init, self.embed_uniform_init)
self.gru = nn.GRU(
self.embed_dim,
self.hidden_size,
dropout=self.hidden_dropout,
num_layers=self.hidden_num,
batch_first=True,
bidirectional=self.bidirectional
)
self.embed_dropout = nn.Dropout(self.embed_dropout)
self.fc_dropout = nn.Dropout(self.fc_dropout)
self.linear1 = nn.Linear(self.hidden_size * 2, self.hidden_size // 2)
self.linear2 = nn.Linear(self.hidden_size // 2, self.label_num)
def __init__(self, opts, vocab, label_vocab):
super(CNN_TreeLSTM, self).__init__()
random.seed(opts.seed)
torch.cuda.manual_seed(opts.gpu_seed)
self.embed_dim = opts.embed_size
self.word_num = vocab.m_size
self.pre_embed_path = opts.pre_embed_path
self.string2id = vocab.string2id
self.embed_uniform_init = opts.embed_uniform_init
self.stride = opts.stride
self.kernel_size = opts.kernel_size
self.kernel_num = opts.kernel_num
self.hidden_size = opts.hidden_size
self.use_cuda = opts.use_cuda
self.label_num = label_vocab.m_size
self.embed_dropout = opts.embed_dropout
self.fc_dropout = opts.fc_dropout
self.hidden_dropout = opts.hidden_dropout
self.debug = True
self.embeddings = nn.Embedding(self.word_num, self.embed_dim)
if opts.pre_embed_path != '':
embedding = Embedding.load_predtrained_emb_zero(
self.pre_embed_path, self.string2id)
self.embeddings.weight.data.copy_(embedding)
else:
nn.init.uniform_(self.embeddings.weight.data,
-self.embed_uniform_init, self.embed_uniform_init)
self.convs = nn.ModuleList([
nn.Conv2d(1,
self.kernel_num, (K, self.embed_dim),
stride=self.stride,
padding=(K // 2, 0)) for K in self.kernel_size
])
in_fea = len(self.kernel_size) * self.kernel_num + self.hidden_size
self.linear1 = nn.Linear(in_fea, in_fea // 4)
self.linear2 = nn.Linear(in_fea // 4, self.label_num)
# build lstm
self.ix = nn.Linear(self.embed_dim, self.hidden_size)
self.ih = nn.Linear(self.hidden_size, self.hidden_size)
self.fx = nn.Linear(self.embed_dim, self.hidden_size)
self.fh = nn.Linear(self.hidden_size, self.hidden_size)
self.ox = nn.Linear(self.embed_dim, self.hidden_size)
self.oh = nn.Linear(self.hidden_size, self.hidden_size)
self.ux = nn.Linear(self.embed_dim, self.hidden_size)
self.uh = nn.Linear(self.hidden_size, self.hidden_size)
# self.out = nn.Linear(self.hidden_size, self.label_num)
#dropout
self.hidden_dropout = nn.Dropout(self.hidden_dropout)
self.embed_dropout = nn.Dropout(self.embed_dropout)
self.fc_dropout = nn.Dropout(self.fc_dropout)