Hello,
I'm happy to announce a new major version of the CSLM toolkit.
It is now available on github, just type:
git clone [email protected]:hschwenk/cslm-toolkit.git
The major changes are:
- bug fixes:
- deterministic sorting of the word-list for short lists
- corrected race condition in MachTab on GPU leading to concurrent
updates
now the results are identical to CPU version
- neural network architectures and training:
- added classes at the output layer
- introduced learning rate schemes
- layer-specific learning rates
- support for auxiliary data at the network input
- flexible sharing of parameters
- simplified network configuration
- refactorisation of GPU code
- use of CUDA streams
- better GPU kernels for activation functions
- more options to select CUDA devices, automatic selection with "-N"
- several missing functions are now available on GPU
- data handling:
- sentence scores
- cross-sentence n-grams
- arbitrary target position (still experimental)
- fast loading of phrases
- new tools
- dump embeddings
- added more documentation
- improved tutorial
Within the next days, we will provide a new tutorial with highlights
some of the new features.
feed-back, bug reports and contributions are welcome !
--
Holger Schwenk
membre IUF sénior
professeur en Informatique
LIUM - Université du Maine
email : [email protected]
_______________________________________________
Mt-list site list
[email protected]
http://lists.eamt.org/mailman/listinfo/mt-list
