def __init__(self, num_classes, hparams):
super(G2PTransformer, self).__init__()
self.bert = BertModel.from_pretrained('./bert/bert-base-chinese')
self.poly_phoneme_classifier = Poly_Phoneme_Classifier(hparams)
self.linear = nn.Linear(1024, num_classes)
self.bert_embedding_features_dim = 768
self.transformer_embedding_features_dim = 1324
self.embedding_features_dim = 1024
self.select_model_hidden_dim = 512
self.linear_pre = nn.Sequential(
nn.Linear(self.bert_embedding_features_dim, self.select_model_hidden_dim),
parse_nk.LayerNormalization(self.select_model_hidden_dim),
nn.ReLU(),
nn.Linear(self.select_model_hidden_dim, self.transformer_embedding_features_dim),
)
self.linear_aft = nn.Sequential(
nn.Linear(self.embedding_features_dim, self.select_model_hidden_dim),
parse_nk.LayerNormalization(self.select_model_hidden_dim),
nn.ReLU(),
nn.Linear(self.select_model_hidden_dim, num_classes),
)
class G2PTransformer(nn.Module):
def __init__(self, num_classes):
super(G2PTransformer, self).__init__()
self.bert = BertModel.from_pretrained('./bert/bert-base-chinese')
self.poly_phoneme_classifier = Poly_Phoneme_Classifier(hparams)
self.linear = nn.Linear(1024, num_classes)
self.bert_embedding_features_dim = 768
self.transformer_embedding_features_dim = 1324
self.embedding_features_dim = 1024
self.select_model_hidden_dim = 512
self.linear_pre = nn.Sequential(
nn.Linear(self.bert_embedding_features_dim,
self.select_model_hidden_dim),
parse_nk.LayerNormalization(self.select_model_hidden_dim),
nn.ReLU(),
nn.Linear(self.select_model_hidden_dim,
self.transformer_embedding_features_dim),
)
self.linear_aft = nn.Sequential(
nn.Linear(self.embedding_features_dim,
self.select_model_hidden_dim),
parse_nk.LayerNormalization(self.select_model_hidden_dim),
nn.ReLU(),
nn.Linear(self.select_model_hidden_dim, num_classes),
)
def forward(self, input_ids, attention_mask, poly_ids):
batch_size = input_ids.size(0)
inputs = {"input_ids": input_ids, "attention_mask": attention_mask}
outputs = self.bert(**inputs)
hidden = outputs[0][0]
hidden = self.linear_pre(hidden)
transformer_output = self.poly_phoneme_classifier.forward_train_polyphonic(
hidden)
poly_hidden = transformer_output[torch.arange(batch_size), poly_ids]
logits = self.linear_aft(poly_hidden)
return logits