Classifier selection problem

[nlp-wh]

Hello, when I choose LSTM as a classifier, I always give an error: AssertionError, assert(False) #Wrong classifier, how should I choose LSTM as a classifier?

[nreimers]

Hi. You can only choose between softmax and crf. The classifier is only the last layer of the network. An LSTM network is always applied before the classifier

[nlp-wh]

Ok, I understand, thank you, I saw a part of the code that can choose LSTM as a classifier, I thought it could be similar to seq2seq.

elif isinstance(classifier, (list, tuple)) and classifier[0] == 'LSTM':

                size = classifier[1]
                if isinstance(self.params['dropout'], (list, tuple)): 
                    output = Bidirectional(LSTM(size, return_sequences=True, dropout=self.params['dropout'][0], recurrent_dropout=self.params['dropout'][1]), name=modelName+'_varLSTM_'+str(cnt))(output)
                else:
                    """ Naive dropout """ 
                    output = Bidirectional(LSTM(size, return_sequences=True), name=modelName+'_LSTM_'+str(cnt))(output) 
                    if self.params['dropout'] > 0.0:
                        output = TimeDistributed(Dropout(self.params['dropout']), name=modelName+'_dropout_'+str(self.params['dropout'])+"_"+str(cnt))(output)                    
            else:
                assert(False) #Wrong classifier

[nreimers]

For multi task learning you can have task specific LSTM layers by passing a list as argument to the classifier option. But there the last layer must be softmax or CRF

[nlp-wh]

Hello, I have already understood the part of multi-tasking learning in your code. Is this part of the code suitable for joint learning? For example, named entity recognition+relationship classification, because it involves different input formats, how to modify the code of input formats?

Best regards,

[nreimers]

Hi I think changing the code to join learning would require substantial work. It was not really designed for that, and multi-task learning and joint learning of two tasks on the same data are quite different (=> requires a different setup)

[nlp-wh]

Thank you for your answer. Recently I plan to do joint learning, but I don’t have any ideas. Smh...