English
资源说明：NN-LM
Hello... THis is the NNLM tool readme file.
With this README, you will also find in this ( or in /home/ankurgan/tools/NNLM) folder :
nnLM_conf.ini
scripts/
## Dependencies
This is a theano-based NNLM training tool. Hence, you must have
1. A gpu on the machine
2. cuda libraries installed
3. numpy, scipy
4. Theano toolkit installed
5. Theano dependencies : g++, python-dev, BLAS
## Training
All Training resources and information are provided via a configuration file, such as nnLM_defaults.ini/ nnlm_conf.ini
[inputs] : Provide location of your training,dev and test files in text format. These should be preprocessed(tokenization,cleaning,etc)
vocab_file is the vocabulary to be used by the LM. (If you wish to provide the vocab ids as well, the format is " "
vocab_freq_file is optional file to give freq information to the NNLM. Needed only for more complex models.
[outputs] : All output description files and matrices will be written to this directory
[training_params] : The network structure and training parameters are defined here. Most are self explanatory.
add_singleton_as_unk : If no OOVs are present in data, NNLM doesnt train them well. Hence you have an option of allowing the model to learn OOV probabilities by treating sigleton's as OOVs. A vocb frequeny file is required for this
use_singleton_as_unk : This will collapse all singletons into a single output node.
write_ngram_files : Mainly for debugging. If you want to write the ngram files being used by the traning , set is to True. The files have word ids and NOT words.
gpu_copy_size : For large training data size, gpu memory is insufficient. Hence, we need to split the data and copy in parts. A good estimate is : for 100k tokens, a copy size of 75000.
To run training on device gpu0, execute :
sh scripts/run.sh nnlm_conf.ini gpu0
## Testing
To test the file, you need to have a trained model directory somewhere.
To run test on some file on gpu0 , execture
sh scripts/run_test.sh None gpu0
It outputs 2 perplexities: 1. Accounting the probabilities of OOVs 2. Skipping the probabilities of OOVs.
本源码包内暂不包含可直接显示的源代码文件，请下载源码包。
