cpf-nlpr / incrementaled Goto Github PK

Thank you for the great job!
During studying the source code, I found something difficult for me to understand.

First

The loss function in 'train.py' is nn.CrossEntropyLoss().

logits, _ = self.model(train_X, train_mask)
p = logits[:, :, :self.seen_cls+1]
loss = criterion(p.view(-1, self.seen_cls+1), train_Y.view(-1))

The variable 'p' is the first self.seen_cls+1 colums of each matrix in logits tensor, but we usually utilize the index of max value in logits as the prediction result. Could you please tell me why do you choose the self.seen_cls+1 rows as input? And the shape of input is [batch_size * max_length, self.seen_cls+1], which is different with the shape of target 'train_Y.view(-1)'.

Second

outputs = self.bert(data_x, attention_mask = mask_x)
bert_enc = outputs[0]
logits = self.fc(bert_enc)
return logits, outputs[0]

The shape of the variable 'bert_enc' is [batch_size, max_length, hidden_size], and the shape of 'logits' is [batch_size, max_length, y_num]. Do you want to get the pooled result of bert encoder and the corresponding predictions?

Thanks for your great work! However, I wonder whether you could give some explanations about the data format in your codes? To exemplify this, what does the 7 mean in the following sample?
The same official said the IMF also insists that the good sections of the two banks be merged . 7 contact