https://github.com/stefoe/indrnn-pytorch
pytorch implementation of Independently Recurrent Neural Networks https://arxiv.org/abs/1803.04831
https://github.com/stefoe/indrnn-pytorch
indrnn paper-implementations pytorch rnn
Last synced: 9 months ago
JSON representation
pytorch implementation of Independently Recurrent Neural Networks https://arxiv.org/abs/1803.04831
- Host: GitHub
- URL: https://github.com/stefoe/indrnn-pytorch
- Owner: StefOe
- Created: 2018-03-15T17:15:31.000Z (about 8 years ago)
- Default Branch: master
- Last Pushed: 2018-12-30T01:02:59.000Z (over 7 years ago)
- Last Synced: 2025-07-30T16:24:21.281Z (10 months ago)
- Topics: indrnn, paper-implementations, pytorch, rnn
- Language: Python
- Homepage:
- Size: 39.1 KB
- Stars: 121
- Watchers: 8
- Forks: 33
- Open Issues: 8
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# indrnn-pytorch
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