In fastrnn’s custom LSTM, we could conveniently define JIT LayerNormLSTM. I am wondering that how could we also make it support PackedSequence as input data for varied lengths of tensors ?
This should be supported:
In fastrnn’s custom LSTM, we could conveniently define JIT LayerNormLSTM. I am wondering that how could we also make it support PackedSequence as input data for varied lengths of tensors ?
This should be supported: