CNTK Library API

Python API

CNTK Python API consists of abstractions for model definition and compute, learning algorithms, data reading and distributed training.

• Flexibility and compactness: These abstractions orthogonally compose offering both flexibility and conciseness in definition and training of arbitrary neural networks.
• Efficient data interfaces: Simple yet light-weight data interfaces allow users to efficiently feed data in the form of native numpy arrays to the compute engine.
• Built-in data readers: CNTKs built-in optimized and scalable data readers for image, text format and speech HTK data formats are also available from the python APIs for ease of directly training with existing data without users having to author any data reading code.
• Highly scalable learning: The API exposes CNTK's highly scalable distributed training capabilities (parallelization algorithms like 1-Bit SGD) in a very easy to use form. The distributed training example illustrates the training parallelization API.
• Concise network definition: The API includes a high level layers library that enables concise advanced neural networks definition including recurrences similar to CNTK V1. The toolkit supports representation of recurrent models in symbolic form as cycles in the neural network instead of requiring static unrolling of the recurrence steps. This results in much more general, concise and efficient representation and execution of recurrent neural networks.

All the core computation, learning, and data reading API abstractions in the CNTK python API are very easily extensible from both Python and C++ allowing users to easily implement new operators, learners and data readers which freely compose with the built-in facilities of the library. The extensibility features are not available in the initial Beta release and will become available in subsequent updates in a few weeks.

The API introduces new Protocol Buffers based model serialization format which supports backwards and upwards compatibility for saved models.

C++ API

The V2 C++ library included in this release exposes CNTK's core computational, neural network composition & training, efficient data reading capabilities, and scalable model training facilities for developers. The C++ APIs are fully featured for both model training as well as evaluation, allowing both training and model serving to be driven from native code. This enables your native model serving code to simultaneously refine the online model tuning using new data seen as part of serving requests (i.e. online learning).

The initial beta release includes examples illustrating the use of the C++ APIs for evaluating previously trained CNTK models. We are also actively working on creating C++ examples illustrating how model training can be performed using the native C++ API and these will become available in a few weeks.

Currently the best source of API documentation is inline in the API header file (CNTKLibrary.h) that contains the full C++ API definition.

Examples

The best way to learn about the APIs currently is to look at the following examples in the [CNTK clone root]/bindings/python/examples directory:

• MNIST: A fully connected feed-forward model for classification of MNIST images. (follow the instructions in Examples/Image/DataSets/MNIST/README.md)

• CifarRest: An image classification ResNet model for training on the CIFAR image dataset. (follow the instructions in Examples/Image/DataSets/CIFAR-10/README.md to get the CIFAR dataset and convert it to the CNTK supported format)

• SequenceClassification: An LSTM sequence classification model for text data.

• Sequence2Sequence: A sequence to sequence grapheme to phoneme translation model that trains on the CMUDict corpus.

• NumpyInterop – numpy interop example showing how to train a simple feed-forward network with training data fed using numpy arrays.

• SLUHandson – Language Understanding.