https://github.com/hughperkins/char-lstm
char-rnn, tweaked to use Element-Research rnn modules
https://github.com/hughperkins/char-lstm
Last synced: 3 months ago
JSON representation
char-rnn, tweaked to use Element-Research rnn modules
- Host: GitHub
- URL: https://github.com/hughperkins/char-lstm
- Owner: hughperkins
- License: bsd-2-clause
- Created: 2015-12-29T12:49:47.000Z (almost 10 years ago)
- Default Branch: master
- Last Pushed: 2016-01-22T16:05:10.000Z (over 9 years ago)
- Last Synced: 2025-05-13T02:09:14.666Z (5 months ago)
- Language: Lua
- Size: 472 KB
- Stars: 9
- Watchers: 5
- Forks: 2
- Open Issues: 2
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# char-lstm
char-rnn, tweaked to use Element-Research rnn modules
What this does, and the way it works, is closely based on how Karpathy's https://github.com/karpathy/char-rnn works, but tweaked to use Element Research's https://github.com/element-research/rnn modules instead.
## Status
Draft, not yet fully working
Update:
- both training and sampling are implemented now, but seems to be some critical bug in training for some reason. I'm working on this :-)
## Differences from original char-rnn
* uses Element Research's rnn modules
* weights are stored as a FloatTensor, rather than CudaTensor etc
* can train using any of cuda/cl/cpu, and sample using the same, or different, up to you
* the sequences used to train each epoch are offset by 1 character from the previous epoch, which hopefully will improve generalization
* each thread is exposed to the entire training set, rather than a 1/batchSize portion of it, which hopefully means can use really large batch sizes, for speed of execution
## Does it support cuda and OpenCL?
* of course it supports OpenCL :-)
* and it supports CUDA :-)
## To do
* currently, calls 'forget' before each sequence. It should not
* implement sampling
* add command-line options
## How to use
### Training
```
th train.lua
```
### Sampling
eg:
```
th sample.lua out/weights_tinyshakespeare_1_501.t7
```
## Naming
If you can think of a better name, please raise an issue to suggest :-)
## License
Original char-rnn code is MIT. New code in this repo are BSDv2