Deep Learning on FPGA Workflow

This figure illustrates deep learning on FPGA workflow.

To use the workflow:

  1. Load deep learning neural network

    You can load the various deep learning neural networks such as Alexnet, VGG and GoogleNet onto the MATLAB® framework. When you compile the network, the network parameters are saved into a structure that consists of NetConfigs and layerConfigs. NetConfigs consists of the weights and biases of the trained network. layerConfig consists of various configuration values of the trained network.

  2. Modify pretrained neural network on MATLAB using transfer learning

    The internal network developed on the MATLAB framework is trained and modified according to the parameters of the external neural network. See also Get Started with Transfer Learning.

  3. Compile user network

    Compilation of the user network usually begins with validating the architecture, types of layers present , data type of input and output parameters, and maximum number of activations. This FPGA solution supports series network architecture with data types of single and int8. For more details, see "Product Description". If the user network features are different, the compiler produces an error and stops. The compiler also performs sanity check by using weight compression and weight quantization.

  4. Deploy on target FPGA board

    By using specific APIs and the NetConfigs and layerConfigs, deploying the compiled network converts the user-trained network into a fixed bitstream and then programs the bitstream on the target FPGA.

  5. Predict outcome

    To classify objects in the input image, use the deployed framework on the FPGA board.

See Also