def __init__(self, fields_tuples, options):
super().__init__(fields_tuples)
self.words_vocab_size = len(self.fields_dict['words'].vocab)
self.explainer_attn_top_k = options.explainer_attn_top_k
self.message_type = options.message_type
self.emb_size = options.word_embeddings_size
self.explainer_cheat_ratio = options.explainer_cheat_ratio
self.explainer_idf = options.explainer_idf
self.explainer_ignore_top_words = options.explainer_ignore_top_words
self.valid_top_word_ids = None
# create an embedding layer to reencode the words
embeddings_weight = None
if self.fields_dict['words'].vocab.vectors is not None:
embeddings_weight = self.fields_dict['words'].vocab.vectors
options.word_embeddings_size = embeddings_weight.size(1)
self.word_emb_explainer = nn.Embedding(
num_embeddings=len(self.fields_dict['words'].vocab),
embedding_dim=options.word_embeddings_size,
padding_idx=constants.PAD_ID,
_weight=embeddings_weight,
)
if options.freeze_embeddings:
self.word_emb_explainer.weight.requires_grad = False
hidden_size = options.hidden_size[0]
self.lstm_explainer = nn.LSTM(options.word_embeddings_size,
hidden_size,
batch_first=True,
bidirectional=True)
n = 2 if options.bidirectional else 1
concat_features_size = n * options.hidden_size[0]
concat_features_size += self.nb_classes
self.scorer = OperationScorer(concat_features_size,
n * options.hidden_size[0],
concat_features_size // 4,
op='add',
scaled=True)
self.attn = Attention(self.scorer,
dropout=options.attn_dropout,
max_activation=options.attn_max_activation)
# save for later
self.lstm_hidden = None
self.lstm_out = None
init_xavier(self.lstm_explainer, dist='uniform')
init_xavier(self.attn, dist='uniform')
self.rec_layer_ae = nn.Linear(self.words_vocab_size, 100)
self.rec_layer_ae_2 = nn.Linear(100, self.words_vocab_size)
self.rec_layer_pred = nn.Linear(n * hidden_size, n * hidden_size)
self.nb_documents = 0
self.idf = None
if self.explainer_idf in ['embs', 'scores']:
self.calc_idf(options)
def init_weights(self):
init_xavier(self.linear_mid, dist='uniform')
init_xavier(self.linear_message, dist='uniform')
init_xavier(self.rnn_hyp, dist='uniform')
def init_weights(self):
init_xavier(self.linear_input, dist='uniform')
init_xavier(self.rnn_hyp, dist='uniform')
init_xavier(self.linear_hyp, dist='uniform')
init_xavier(self.linear_merge, dist='uniform')