pulsatrix::MnistConvNet Class Reference

Conv2D(1,8,5,5) -> ReLU -> Flatten -> Linear(4608,10), trained via CrossEntropyLoss + Adam, one real MNIST image at a time (this library has no batch dimension anywhere, same constraint XorNetwork already works under). More...

#include <mnist_classifier_example.hpp>

Public Member Functions
	MnistConvNet (DeviceBackend *backend, unsigned seed=42)
	Constructs the network with randomly initialized weights.
Tensor	forward (const Tensor &image)
	Runs the network forward.
int64_t	predict (const Tensor &image)
	forward() plus argmax – the predicted class index.
float	train_step (const Tensor &image, int64_t target_class, AdamOptimizer &optimizer, MetricsSink &sink, int step)
	Runs one training step: forward, cross-entropy loss, backward through every layer, one Adam update per layer's parameters, and logs the loss.
const Tensor &	classifier_weight () const
	Test/inspection accessor.
std::vector< Module * >	modules ()
	The network's layers in forward order (Conv2D, ReLU, Flatten, Linear), for building an ExplainerContext over the trained model – so every explainer (Saliency, IntegratedGradients, GradCAM, LRP, ...) and build_circuit_graph() can run on exactly the weights train_step() learned. The pointers alias this object's members: they stay valid only while this MnistConvNet is alive.

Detailed Description

Conv2D(1,8,5,5) -> ReLU -> Flatten -> Linear(4608,10), trained via CrossEntropyLoss + Adam, one real MNIST image at a time (this library has no batch dimension anywhere, same constraint XorNetwork already works under).

Note

Reuses grad_cam_mnist_demo.cpp's exact network shape (kernel size 5, 8 output channels) deliberately – once this network is trained on real data, a Grad-CAM call against it produces a meaningful heatmap for the first time in this project, not just pipeline mechanics on random weights.

Random (not hand-picked) initial weights – unlike XorNetwork's 12-weight network, hand-picking ~4600+ asymmetric values isn't practical. A fixed RNG seed keeps this reproducible; see mission Recon for why zero-init specifically doesn't work (identical rationale to XorNetwork's own note).

Constructor & Destructor Documentation

MnistConvNet()

pulsatrix::MnistConvNet::MnistConvNet ( DeviceBackend *  backend, unsigned  seed = 42  )

explicit

Constructs the network with randomly initialized weights.

Parameters

backend	Backend to compute through. Not owned; must outlive this network.
seed	RNG seed for weight initialization.

Member Function Documentation

classifier_weight()

const Tensor & pulsatrix::MnistConvNet::classifier_weight ( ) const

inline

Test/inspection accessor.

forward()

Tensor pulsatrix::MnistConvNet::forward

(

const Tensor &

image

)

Runs the network forward.

Parameters

image

Shape (1, 28, 28), pixel values normalized to [0,1].

Returns

Shape (10,) raw logits (pre-softmax).

modules()

std::vector< Module * > pulsatrix::MnistConvNet::modules ( )

inline

The network's layers in forward order (Conv2D, ReLU, Flatten, Linear), for building an ExplainerContext over the trained model – so every explainer (Saliency, IntegratedGradients, GradCAM, LRP, ...) and build_circuit_graph() can run on exactly the weights train_step() learned. The pointers alias this object's members: they stay valid only while this MnistConvNet is alive.

predict()

int64_t pulsatrix::MnistConvNet::predict

(

const Tensor &

image

)

forward() plus argmax – the predicted class index.

Parameters

image

Shape (1, 28, 28).

Returns

Predicted class, 0-9.

train_step()

float pulsatrix::MnistConvNet::train_step	(	const Tensor &	image,
		int64_t	target_class,
		AdamOptimizer &	optimizer,
		MetricsSink &	sink,
		int	step
	)

Runs one training step: forward, cross-entropy loss, backward through every layer, one Adam update per layer's parameters, and logs the loss.

Parameters

image	Shape (1, 28, 28).
target_class	Ground-truth class, 0-9.
optimizer	Optimizer to update this network's parameters with.
sink	Where the loss value is logged (tag "loss").
step	Training step number, passed through to sink.

Returns

The loss value for this example, before the update.

---

The documentation for this class was generated from the following file:

• include/pulsatrix/mnist_classifier_example.hpp