CNTK Library Evaluation on Windows

The CNTK library on Windows is currently available in C++, C#/.NET and Python.

Using the CNTK Library Managed API

The CNTK Library Managed Eval API provides a managed (.NET) library for evaluation of CNTK models on CPU and GPU using C# and other .NET languages. The Nuget package CNTK.CPUOnly is for evaluation on CPU, and the CNTK.GPU is for evaluation on devices with NVIDIA GPU. These Nuget packages are available at Nuget.org (search for CNTK). Please refer to the NuGet Package page for details on how to get started with CNTK Library NuGet Packages.

The CNTK Library Managed Eval API allows you to load pre-trained models onto the specified device (CPU or GPU), retrieve model inputs and outputs information, prepare your input data for evaluation. It supports both dense, one-hot vector and sparse CSC format data. The input data can be a single sample, batch of multiple samples, a sequence of samples, or a batch of sequences with variable lengths. A detailed description about batch and sequence format can be found at the page CNTK Concepts. After evaluation, the API provides convenient methods to get results.

The usual steps for model evaluation using CNTK Library Managed API include:

1. Install the appropriate Nuget package (CPU or GPU depending on usage) to your. Details are described in the CNTK NuGet Package.
2. Load the model by calling Function.LoadModel(). The returned value is a Function instance referring to the loaded model.
3. If needed, get inputs and outputs of the loaded model by calling the properties Arguments, Output or Outputs of the Function instance returned by LoadModel(). The inputs and outputs are represented as Variable. Further information about each individual inputs or outputs, e.g. Name or Shape can be found by calling corresponding properties of Variable.
4. For each input, prepare input data for evaluation. Depending on usage, input data might consist of a single sample, batch of multiple samples, a sequence of samples, or a batch of sequences with variable length. Regarding storage, input data can be in dense format, one-hot vector, or sparse CSC format. The Value object is used to represent the input data with different contents and formats. There are different methods provided to easily create the Value object from plain user data.
   • CreateBatch() creates a Value object containing a batch of single or multiple samples.
   • CreateSequence() creates a Value object containing a sequence with multiple samples. It is possible to specify whether this sequence is a new sequence or continuation of a previous sequence at the same index in the sequences list from a previous call to this method.
   • CreateBatchOfSequences() creates a Value object containing a batch of sequences. Each sequence consists of variable length samples. All the methods above support both dense, one-hot vector or sparse CSC format as input.
5. Call the Evaluate() method of the Function instance returned by LoadModel() to start model evaluation. Besides the input Variables and their associated Values, Evaluate() also requires to specify the output Variables and possible Values. The caller may explicitly pass the Value object for a output variable or pass null in which case CNTK Library allocates the actual storage for the output value. On return the Value objects associated with the output variables contain the results of evaluation.
6. Call the GetDenseData()/GetOneHotData() method of the output Value to get the results in plain data format. It copies the data stored in the Value object into the buffer as a list of sequences with variable length samples, either in dense format or one-hot vector format (support for the sparse format will be added soon).

In order to evaluate multiple evaluation requests in parallel, clone() should be called to create multiple Function instances before Evaluate(). The EvaluateMultipleImagesInParallel() demonstrates how to evaluate concurrent requests using CNTK C#/.NET Managed API.

The CNTK Library Managed API is described in the CNTK Library Managed API page.

The CNTKLibraryEvalExamples shows how to evaluate a model in C# using CNTK Library NuGet packages. Please see the Eval Examples page for building and running examples.

If you do not want to use NuGet Package, you can add Cntk.Core.Managed-<VERSION>.dll as reference to your project. The Cntk.Core.Managed dll and all dependent dlls can be found in the CNTK binary release package on the CNTK Releases page. (Please refer here for CNTK binary naming scheme). Please make sure in this case that the path to Cntk.Core.Managed dll and its dependencies are included in the search path of dlls for your application.

Using the C++ CNTK Library API for evaluation

The CNTK Nuget packages CNTK.CPUOnly and CNTK.GPU also contain C++ CNTK Library. Please refer to the NuGet Package page for details on how to get CNTK NuGet Packages. Please note that C++ CNTK Library requires Visual Studio 2015 Update 3 and the target platform in the project property should be X64.

To use C++ CNTK Library in your project, you just need to install the appropriate Nuget package to your project. No additional configuration of include and library directories is needed, as it is automatically added to your project property by the Nuget package. The Nuget package supports both Debug and Release build.

The following steps describe how to use the C++ CNTK Library for model evaluation.

1. Link the Cntk.Core-<VERSION>.lib import library into the application. This step can be omitted if you use the Nuget package.
2. Include the evaluation header file "CNTKLibrary.h".
3. Load the model by CNTK::Function::LoadModel(). The returned Function object represents the computation graph of the model.
4. Prepare data for each input variable. You can use CNTK::Value::CreateBatch(), CNTK::Value::CreateSequence(), or CNTK::Value::CreateBatchOfSequences() to create a Value object from your input data representing a batch of samples, a sequence of samples, and a batch of sequences of samples respectively.
5. Call CNTK::Function::Evaluate() for evaluation. The Evaluate() expects as input an unordered_map for input variables and input values, and an unordered_map for output variables and output values. The output value object could be null which means that CNTK Library allocates the actual storage for this output value. On return the Value objects associated with the output variables contain the results of evaluation.
6. Get output data from evaluation results. Use CNTK::Value::CopyVariableTo() to copy the data stored in the Value object into the buffer as a list of sequences with variable length samples.

For concurrent evaluation of multiple requests, CNTK::Function::Clone() should be called before Evaluate(). The MultiThreadsEvaluationWithLoadModel() demonstrates how to evaluate multiple requests in parallel using CNTK C++ Library API.

The C++ examples CNTKLibraryCPPEvalCPUOnlyExamples and CNTKLibraryCPPEvalGPUExamples illustrate the usage pattern above. The Eval Examples page provides detailed steps about building and running examples.

For details on C++ CNTK Library API for evaluation, please refer to the CNTK Library Evaluation Interface page in this wiki.

Alternatively, you can use C++ CNTK Library without the NuGet Package. You can either get Cntk.Core-<VERSION>.lib and all dependent dlls from CNTK release packages on the CNTK Releases page or build them from CNTK source code. (Please refer here for CNTK binary naming scheme.) In this case, you need to configure include and library path to point to the correct directory, and make sure that the build configuration (Debug or Release) of the CNTK Library is the same as that of your project setting (Otherwise you will get unexpected exceptions). The CNTK release package contains only the release build of binaries. Please also note that CNTK Library requires 64bit platform, and the path to Cntk.Core and its dependencies should be in the search path of dlls for your application.

Shipping CNTK Library with your Windows application

If you own application uses CNTK Library C++ API you need to distribute these DLLs with your application. (Please refer here for CNTK binary naming scheme):

• Cntk.Core-<VERSION>.dll
• Cntk.Math-<VERSION>.dll
• libiomp5md.dll
• mkl_cntk_p.dll

If you use the Managed Library, beside the dlls mentioned above, you need to additionally include the following DLLs:

• Cntk.Core.Managed-<VERSION>.dll
• Cntk.Core.CSBinding-<VERSION>.dll

For using GPU, you need in addition to include the following NVIDIA CUDA related DLLs:

• cublas64_80.dll
• cudart64_80.dll
• cudnn64_5.dll
• curand64_80.dll
• cusparse64_80.dll

All these dlls can be found in the CNTK binary release version, see the CNTK Releases page.

Using Python

You can use Python to evaluate a pre-trained model. Examples can be found here.