def _build_conversation_layer(self):
self.register_buffer(
'START', torch.tensor([self.start_token_idx], dtype=torch.long))
self.conv_encoder = TransformerEncoder(
n_heads=self.n_heads,
n_layers=self.n_layers,
embedding_size=self.token_emb_dim,
ffn_size=self.ffn_size,
vocabulary_size=self.vocab_size,
embedding=self.token_embedding,
dropout=self.dropout,
attention_dropout=self.attention_dropout,
relu_dropout=self.relu_dropout,
padding_idx=self.pad_token_idx,
learn_positional_embeddings=self.learn_positional_embeddings,
embeddings_scale=self.embeddings_scale,
reduction=self.reduction,
n_positions=self.n_positions,
)
self.conv_entity_norm = nn.Linear(self.kg_emb_dim, self.ffn_size)
self.conv_entity_attn_norm = nn.Linear(self.kg_emb_dim, self.ffn_size)
self.conv_word_norm = nn.Linear(self.kg_emb_dim, self.ffn_size)
self.conv_word_attn_norm = nn.Linear(self.kg_emb_dim, self.ffn_size)
self.copy_norm = nn.Linear(self.ffn_size * 3, self.token_emb_dim)
self.copy_output = nn.Linear(self.token_emb_dim, self.vocab_size)
copy_mask = np.load(os.path.join(self.dpath,
"copy_mask.npy")).astype(bool)
if self.replace_token:
if self.replace_token_idx < len(copy_mask):
copy_mask[self.replace_token_idx] = False
else:
copy_mask = np.insert(copy_mask, len(copy_mask), False)
self.copy_mask = torch.as_tensor(copy_mask).to(self.device)
self.conv_decoder = TransformerDecoderKG(
self.n_heads,
self.n_layers,
self.token_emb_dim,
self.ffn_size,
self.vocab_size,
embedding=self.token_embedding,
dropout=self.dropout,
attention_dropout=self.attention_dropout,
relu_dropout=self.relu_dropout,
embeddings_scale=self.embeddings_scale,
learn_positional_embeddings=self.learn_positional_embeddings,
padding_idx=self.pad_token_idx,
n_positions=self.n_positions)
self.conv_loss = nn.CrossEntropyLoss(ignore_index=self.pad_token_idx)
logger.debug('[Finish build conv layer]')
def _build_conversation_layer(self):
self.register_buffer(
'START', torch.tensor([self.start_token_idx], dtype=torch.long))
self.dialog_encoder = TransformerEncoder(
self.n_heads, self.n_layers, self.token_emb_dim, self.ffn_size,
self.vocab_size, self.token_embedding, self.dropout,
self.attention_dropout, self.relu_dropout, self.pad_token_idx,
self.learn_positional_embeddings, self.embeddings_scale,
self.reduction, self.n_positions)
self.decoder = TransformerDecoder(
self.n_heads, self.n_layers, self.token_emb_dim, self.ffn_size,
self.vocab_size, self.token_embedding, self.dropout,
self.attention_dropout, self.relu_dropout, self.embeddings_scale,
self.learn_positional_embeddings, self.pad_token_idx,
self.n_positions)
self.user_proj_1 = nn.Linear(self.user_emb_dim, self.user_proj_dim)
self.user_proj_2 = nn.Linear(self.user_proj_dim, self.vocab_size)
self.conv_loss = nn.CrossEntropyLoss(ignore_index=self.pad_token_idx)
logger.debug('[Build conversation layer]')
def _build_conversation_layer(self):
self.register_buffer(
'START', torch.tensor([self.start_token_idx], dtype=torch.long))
self.conv_encoder = TransformerEncoder(
n_heads=self.n_heads,
n_layers=self.n_layers,
embedding_size=self.token_emb_dim,
ffn_size=self.ffn_size,
vocabulary_size=self.vocab_size,
embedding=self.token_embedding,
dropout=self.dropout,
attention_dropout=self.attention_dropout,
relu_dropout=self.relu_dropout,
padding_idx=self.pad_token_idx,
learn_positional_embeddings=self.learn_positional_embeddings,
embeddings_scale=self.embeddings_scale,
reduction=self.reduction,
n_positions=self.n_positions,
)
self.conv_decoder = TransformerDecoder(
self.n_heads,
self.n_layers,
self.token_emb_dim,
self.ffn_size,
self.vocab_size,
embedding=self.token_embedding,
dropout=self.dropout,
attention_dropout=self.attention_dropout,
relu_dropout=self.relu_dropout,
embeddings_scale=self.embeddings_scale,
learn_positional_embeddings=self.learn_positional_embeddings,
padding_idx=self.pad_token_idx,
n_positions=self.n_positions)
self.conv_loss = nn.CrossEntropyLoss(ignore_index=self.pad_token_idx)
logger.debug('[Finish build conv layer]')