Please find the documentation page here
Nabu is an ASR framework for end-to-end networks built on top of TensorFlow. Nabu's design focusses on adaptibility, making it easy for the designer to adjust everything from the model structure to the way it is trained.
Last built with TensorFlow version: 1.8.0
Using Nabu
Nabu works in several stages: data prepation, training and finally testing and decoding. Each of these stages uses a recipe for a specific model and database. The recipe contains configuration files for the all components and defines all the necesary parameters for the database and the model. You can find more information on the components in a recipe here.
Data preperation
In the data preperation stage all the data is prepared (feature computation, target normalization etc.) for training and testing. Before running the data preperation you should create a database.conf file in the recipe directory based on the database.cfg that should already be there, and fill in all the paths. Should you want to modify parameters in the processors, you can modify the config files that are pointed to in the database config. You can find more information about processors here.
You can run the data prepation with:
- recipe: points to the directory containing the recipe you want to prepare the data for.
- expdir: the absolute path to a directory where you can write to. In this directory all files will be stored, like the configurations and logs
- computing [default: standard]: the distributed computing software you want to use. One of standard or condor. standard means that no distributed computing software is used and the job will run on the machine where nabu is called from. the condor option uses HTCondor. More information can be found here.
Training
In the training stage the model will be trained to minimize a loss function. During training the model can be evaluated to adjust the learning rate if necessary. Multiple configuration files in the recipe are used during training:
- model.cfg: model parameters
- trainer.cfg: training parameters
- validation_evaluator.cfg: validation parameters
You can find more information about models here, about trainers here and about evaluators here.
You can run the training with:
The parameters are the same as the data preperation script (see above) with one extra parameter; mode (default: non_distributed). Mode is the distribution mode. This should be one of non_distributed, single_machine or multi_machine. You can find more information about this here
Testing
In the testing stage the performance of the model is evaluated on a testing set. To modify the way the model in is evaluated you can modify the test_evaluator.cfg file in the recipe dir. You can find more information on evaluators here.
You can run testing with
The parameters for this script are similar to the training script (see above). You should use the same expdir that you used for training the model.
Decoding
In the decoding stage the model is used to decode the test set and the resulting nbest lists are written do disk in the expdir. To modify the way the model is used for decoding look into the recognizer.cfg file. You can find more information about decoders here.
You can run decoding with
The parameters for this script are similar to the training script (see above). You should use the same expdir that you used for training the model.
Parameter search
You can automatically do a parameter search using Nabu. To do this you should create a sweep file. A sweep file contain blocks of parameters, each block will change the parameters in the recipe and run a script. A sweep file looks like this:
For example, if you want to try several number of layers and number of units:
The parameter sweep can then be executed as follows:
where command can be any of the commands discussed above.
Kaldi and Nabu
There are some scripts avaiable to use a Nabu Neural Network in the Kaldi framework. Kaldi is an ASR toolkit. You can find more information here.
Using Kaldi with nabu happens in several steps: 1) Data preperation 2) GMM-HMM training 3) Aligning the data 4) computing the prior 5) Training the Neural Network 6) Decoding and scoring
Data preperation
The data preperation is database dependent. Kaldi has many scripts for data preperation and you should use them.
GMM-HMM training
You can train the GMM-HMM model as folows:
With the folowing arguments:
- datadir: the directory containing the training data (created in data prep)
- langdir: the directory containing the language model for training (created in data prep)
- langdir-test: the directory containg the language model that should be used for decoding (created in data prep)
- traindir: The directory where the training files (logs, models, ...) will be written
- kaldi: the location of your kaldi installation
The script will compute the features, train the GMM-HMM models and align the training data, so you do not have to do this anymore in the coming step. The alignments for the training set can be found in <traindir>/pdfs.
Aligning the data
The training data has already been aligned in the previous step, but if you want to align e.g. the validation set you can do that as follows:
the datadir should point to the data you want to align, the traindir should be the traindir you used in the previous step and the targetdir is the directory where the alignments will be written. The alignments can be found in <targetdir>/pdfs
Computing the prior
The prior is needed to convert the pdf posteriors to pdf pseudo-likelihoods. The prior can be computed with:
traindir should be the same as the traindir in the previous step. the prior can then be found in numpy format in <traindir>/prior.npy
Training the neural net
Training the neural network happens using the Nabu framework. In order to do this you should create a recipe for doing so (see the section on training). You can find an example recipe for this in config/recipes/DNN/WSJ. You can use this recipe, but you should still create the database.conf file based on the database.cfg file. In your database configuration you should create sections for the features which is the same as you would do for a normal Nabu neural network and sections for the alignments. The alignment sections should get the special type "alignments". A section should look something like this:
dir is just the directory where the processed alignments will be written.
The rest of the training procedure is the same as the normal procedure, so folow the instructions in the sections above.
Decoding and scoring
To decode the using the trained system you should first compute the pseudo-likelihoods as folows:
The pseudo likelihoods can the be found in <expdir>/decode/decoded/alignments.
You can then do the Kaldi decoding and scoring with:
The arguments are similar as the arguments in the script above. The outputs will be written to the <outputs> folder.
Designing in Nabu
As mentioned in the beginning Nabu focusses on adaptibility. Everything in the recipe can be modified (more information about recipes here). Most classes used in Nabu have a general class that defines an interface and common functionality for all children and a factory that is used to create the necessary class. Look into the respective README files to see how to implement a new class.
In general, if you want to add your own type of class (like a new model) you should follow these steps:
- Create a file in the class directory
- Create your child class, let it inherit from the general class and overwrite the abstract methods
- Add your class to the factory method. In the factory method you give your class a name, this does not have to be the name of the class. You will use this name in the configuration file for your model so Nabu knows which class to use.
- Add your file to the package in __init__.py
- create a configuration file for your class and put it in templates. You should then add this configuration file in whichever recipe you want to use it for or create your own recipe using your new class.
Example experiment
As an example you can find the results for running the LAS/TIMIT recipe. The training loss in average cross entropy is plotted in the following image:
The performance on the validation set measured in character error rate is plotted in the folowing image (I don't know why the extra lines are there)
The error rate on the test set is 21.6% and training took just under 4 hours on a Nvidia GeForce GTX 1080 Ti.
