How to write a dynamically unrolled RNN with GLUON


#1

Gluon’s LSTM layer should be to unroll all the samples in a batch to the T dimension of the matrix of the batch, but the problem is that many samples in a batch may be Pad to the T dimension, so how to get each sample to unroll to itself The PAD position is no longer calculated? Just like tf.nn.dynamic_rnn() in Tensorflow can be controlled by passing in the length of each sample. What should I do in Gluon? Please help me.


#2

As per my knowledge dynamic unrolling is not supported yet. You can however, set fixed sequence lengths (buckets) and fully unroll for that T. Here is an example https://github.com/apache/incubator-mxnet/blob/master/example/rnn/bucketing/lstm_bucketing.py
https://mxnet.incubator.apache.org/faq/bucketing.html