def __init__(self, config, pretrained_word_embedding):
super(NewsEncoder, self).__init__()
self.config = config
word_embedding = nn.Embedding.from_pretrained(
pretrained_word_embedding,
freeze=config.freeze_word_embeddings,
padding_idx=0)
category_embedding = nn.Embedding(config.num_categories,
config.category_embedding_dim,
padding_idx=0)
self.title_encoder = TimeDistributed(TextEncoder(
word_embedding, config.word_embedding_dim, config.num_filters,
config.window_size, config.query_vector_dim,
config.dropout_probability),
batch_first=True)
self.abstract_encoder = TimeDistributed(TextEncoder(
word_embedding, config.word_embedding_dim, config.num_filters,
config.window_size, config.query_vector_dim,
config.dropout_probability),
batch_first=True)
self.category_encoder = CategoryEncoder(category_embedding,
config.category_embedding_dim,
config.num_filters)
self.sub_category_encoder = CategoryEncoder(
category_embedding, config.category_embedding_dim,
config.num_filters)
self.final_attention = AdditiveAttention(config.query_vector_dim,
config.num_filters)
def __init__(self, config):
super(UserEncoder, self).__init__()
self.config = config
self.mh_selfattention = nn.MultiheadAttention(
config.word_embedding_dim, config.num_attention_heads)
self.additive_attention = AdditiveAttention(config.query_vector_dim,
config.word_embedding_dim)
def __init__(self, word_embedding, word_embedding_dim, num_filters,
window_size, query_vector_dim, dropout_probability):
super(TextEncoder, self).__init__()
self.dropout_probability = dropout_probability
self.word_embedding = word_embedding
assert window_size >= 1 and window_size % 2 == 1
self.CNN = nn.Conv1d(1,
num_filters, (window_size, word_embedding_dim),
padding=(int((window_size - 1) / 2), 0))
self.additive_attention = AdditiveAttention(query_vector_dim,
num_filters)
def __init__(self, config, pretrained_word_embedding):
super(NewsEncoder, self).__init__()
self.config = config
self.word_embedding = nn.Embedding.from_pretrained(
pretrained_word_embedding,
freeze=config.freeze_word_embeddings,
padding_idx=0)
self.mh_selfattention = nn.MultiheadAttention(
config.word_embedding_dim, config.num_attention_heads)
self.additive_attention = AdditiveAttention(config.query_vector_dim,
config.word_embedding_dim)
def __init__(self, config, pretrained_word_embedding):
super(NewsEncoder, self).__init__()
self.config = config
self.word_embedding = nn.Embedding.from_pretrained(
pretrained_word_embedding,
freeze=config.freeze_word_embeddings,
padding_idx=0)
assert config.window_size >= 1 and config.window_size % 2 == 1
self.title_CNN = nn.Conv1d(
1,
config.num_filters,
(config.window_size, config.word_embedding_dim),
padding=(int((config.window_size - 1) / 2), 0))
self.title_attention = AdditiveAttention(config.query_vector_dim,
config.num_filters)
def __init__(self, config):
super(UserEncoder, self).__init__()
self.additive_attention = AdditiveAttention(config.query_vector_dim,
config.num_filters)