def __init__(self,
vocab_size,
word_dim,
embed_size,
use_abs=False,
att_units=300,
hops=30,
gru_units=None,
num_layers=None,
norm_words=False):
super(AttentiveTextEncoder, self).__init__()
self.use_abs = use_abs
self.embed_size = embed_size
self.hops = hops
self.att_units = att_units
# word embedding
self.embed = nn.Embedding(vocab_size, word_dim)
self.attention = SelfAttentiveEncoder(nb_features=word_dim,
att_units=att_units,
hops=hops)
att_out_size = word_dim * hops
self.fc = False
if att_out_size != embed_size:
self.fc = nn.Linear(att_out_size, embed_size)
global_initializer(self)
def __init__(self,
vocab_size,
word_dim,
embed_size,
use_abs=False,
att_units=200,
hops=15,
gru_units=1024,
num_layers=1,
norm_words=None):
super(GRUAttentiveTextEncoder, self).__init__()
self.use_abs = use_abs
self.embed_size = embed_size
self.hops = hops
self.att_units = att_units
# word embedding
self.embed = nn.Embedding(vocab_size, word_dim)
self.norm_words = norm_words
# caption embedding
self.rnn = nn.GRU(word_dim, gru_units, num_layers, batch_first=True)
self.attention = SelfAttentiveEncoder(nb_features=gru_units,
att_units=att_units,
hops=hops)
self.fc = nn.Linear(gru_units * hops, embed_size)
global_initializer(self)
def __init__(self,
vocab_size,
word_dim,
embed_size,
use_abs=False,
att_units=300,
hops=30,
gru_units=None,
num_layers=1,
norm_words=None):
super(ConvAttentiveTextEncoder, self).__init__()
self.use_abs = use_abs
self.embed_size = embed_size
self.hops = hops
self.att_units = att_units
conv_filters = 100
# word embedding
self.embed = nn.Embedding(vocab_size, word_dim)
self.conv1 = ConvBlock(
in_channels=word_dim,
out_channels=conv_filters,
kernel_size=2,
padding=1,
activation='ReLU',
batchnorm=True,
)
self.conv2 = ConvBlock(
in_channels=word_dim,
out_channels=conv_filters,
kernel_size=3,
padding=1,
activation='ReLU',
batchnorm=True,
)
self.att_emb = SelfAttentiveEncoder(nb_features=word_dim,
att_units=att_units,
hops=hops)
self.att_conv1 = SelfAttentiveEncoder(nb_features=conv_filters,
att_units=att_units,
hops=hops)
self.att_conv2 = SelfAttentiveEncoder(nb_features=conv_filters,
att_units=att_units,
hops=hops)
self.fc = nn.Linear((conv_filters * 2 + word_dim) * hops, embed_size)
global_initializer(self)
