pytorch implementation of the IndRNN Paper (https://arxiv.org/pdf/1803.04831.pdf)
The test functions are adapted from the tensorflow implementation (https://github.com/batzner/indrnn) and the theano implementation (https://github.com/Sunnydreamrain/IndRNN_Theano_Lasagne). Tested with Python3.7 and pytorch 1.0.0
IndRNNv2 version should be faster with GPUs, especially for bidirectional networks. Seconds per 100 iterations with GPU-P100 on the addition test and batch size of 50:
| IndRNN | IndRNNv2 |
|---|---|
| 6.7 | 3.65 |
Seconds per epoch with GPU-P100 on SeqMNIST and batch size of 256:
| IndRNN | IndRNNv2 |
|---|---|
| 394 | 114 |
TODOs: -get parameters for MNIST experiments -add permutation MNIST test