DSP++ Autodiff Iris Classification Example

Description:

Trains a small neural network to classify Iris flowers using CMSIS-DSP reverse-mode automatic differentiation. All model parameters, intermediate buffers, gradients, and optimizer state use single-precision float (f32). The dataset is included in iris_data.hpp. The CMSIS-DSP pack supplies the DSP++ autodiff headers and CMSIS-DSP C functions.

Network and Training:

Four inputs describe sepal length, sepal width, petal length, and petal width. A fully connected layer with eight outputs is followed by ReLU, a fully connected layer with three outputs, and softmax. The outputs represent Iris-setosa, Iris-versicolor, and Iris-virginica.

Cross-entropy measures the classification loss. Adam updates the network's 67 parameters with a learning rate of 0.01 over 120 epochs. The 120 training samples are shuffled before each epoch using a deterministic random sequence. Normalization uses fixed population statistics over all 150 samples.

Automatic Differentiation:

The tape registers the fully connected, ReLU, softmax, and cross-entropy operators. Model weights and biases are registered as parameters, while features and one-hot labels are inputs. For each training sample, the graph is rewound, the forward expressions are evaluated, gradients are cleared, and the loss is differentiated before the optimizer updates the parameters. During testing, a recording scope disables graph recording for inference.

Memory:

The training state is allocated on the heap to keep the model, sample buffers, autodiff arena, and optimizer state off the stack. The arena has a capacity of 2048 bytes. The Cortex-M55 configuration provides a 16 KiB stack and a 64 KiB heap.

Results:

Every fifth sample of each class is held out from training, giving 30 test samples. The program prints the mean loss every 20 epochs, the expected and detected class for each test sample, and the final test accuracy.

Build and Run:

Open autodiff.csolution.yml with the CMSIS Solution VS Code extension and install the required tools and packs. The default configuration is Iris.Release+ARMCM55 with the CLANG compiler. AC6 and GCC are also selectable. Run the executable on the Corstone-300 FVP with the supplied fvp_config.txt. Console output and program termination use semihosting.

Refer autodiff_example/Iris/main.cpp