CNN implementation: initialisation, forward pass, backward pass, SGD-with-momentum update. More...

#include "cnn.h"
#include <math.h>
#include <stdlib.h>
#include <string.h>
#include <stdio.h>

Include dependency graph for cnn.c:

Functions
static float	randn (void)
	Box-Muller transform: sample one value from N(0, 1).
static void	relu_inplace (float *x, size_t length)
	Apply ReLU element-wise: f(x) = max(0, x).
static void	softmax_inplace (float *x, size_t length)
	Apply softmax in-place over an array of length n.
CNN *	cnn_create (void)
	Allocate and initialise a CNN with He-initialised weights.
void	cnn_free (CNN *net)
	Free all memory associated with a CNN.
static void	forward_conv (CNN *net)
	Convolution stage: compute net->act.conv_out.
static void	forward_pool (CNN *net)
	Max-pooling stage: compute net->act.pool_out.
static void	forward_flatten (CNN *net)
	Flatten stage: copy pool_out into act->flat, row-major.
static void	forward_dense1 (CNN *net)
	Dense layer 1 (2704 → 128) followed by ReLU.
static void	forward_dense2 (CNN *net)
	Dense layer 2 (128 → 26) followed by Softmax.
void	cnn_forward (CNN net, const float image)
	Run a full forward pass and populate net->act.
float	cnn_loss (const CNN *net, int label)
	Compute cross-entropy loss for the current forward-pass output.
void	cnn_backward (CNN *net, int label)
	Compute gradients via backpropagation.
void	cnn_zero_grads (CNN *net)
	Zero all gradient accumulators in net->grads.
static void	sgd_update (float w, float v, const float *g, size_t n, float inv_batch)
	Apply SGD+momentum to one weight array.
void	cnn_update (CNN *net)
	Apply one SGD-with-momentum update step to the weights.
int	cnn_predict (CNN net, const float image)
	Predict the most likely class for a single image.

Detailed Description

CNN implementation: initialisation, forward pass, backward pass, SGD-with-momentum update.

Function Documentation

◆ cnn_backward()

void cnn_backward	(	CNN *	net,
		int	label )

Compute gradients via backpropagation.

Must be called after cnn_forward(). Accumulates gradients into net->grads.

Parameters

net	CNN after a forward pass.
label	True class index in [0, CNN_N_CLASSES).

Note: This is a stub: the full implementation is deferred.

Here is the caller graph for this function:

◆ cnn_create()

CNN * cnn_create ( void )

Allocate and initialise a CNN with He-initialised weights.

All conv kernels and dense weights are drawn from a zero-mean normal distribution with variance 2/fan_in (He, 2015). Biases are zeroed.

Returns: Pointer to a heap-allocated CNN, or NULL on allocation failure.

Note: The caller must free the returned pointer with cnn_free().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ cnn_forward()

void cnn_forward	(	CNN *	net,
		const float *	image )

Run a full forward pass and populate net->act.

Stages: Conv2D → ReLU → MaxPool → Flatten → Dense → ReLU → Dense → Softmax.

Parameters

net	Initialised CNN.
image	Flat array of CNN_IMG_H * CNN_IMG_W normalised float pixels, row-major order.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ cnn_free()

void cnn_free ( CNN * net )

Free all memory associated with a CNN.

Parameters

net	Pointer returned by cnn_create(). No-op if NULL.

Here is the caller graph for this function:

◆ cnn_loss()

float cnn_loss	(	const CNN *	net,
		int	label )

Compute cross-entropy loss for the current forward-pass output.

loss = -log(output[label])

Parameters

net	CNN after cnn_forward().
label	True class index.

Returns: Scalar cross-entropy loss (≥ 0).

Here is the caller graph for this function:

◆ cnn_predict()

int cnn_predict	(	CNN *	net,
		const float *	image )

Predict the most likely class for a single image.

Runs cnn_forward() internally.

Parameters

net	Initialised CNN with trained weights.
image	Flat float array, same layout as cnn_forward().

Returns: Class index in [0, CNN_N_CLASSES), i.e. 0 = 'A', 25 = 'Z'.

Here is the call graph for this function:

◆ cnn_update()

void cnn_update ( CNN * net )

Apply one SGD-with-momentum update step to the weights.

w_new = w - lr * (β * v + g) where v is the momentum velocity and g is the gradient.

Parameters

net	CNN with filled net->grads from cnn_backward().

Note: Resets net->grads to zero after the update.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ cnn_zero_grads()

void cnn_zero_grads ( CNN * net )

Zero all gradient accumulators in net->grads.

Parameters

net	CNN whose gradients should be cleared.

Here is the caller graph for this function:

◆ forward_conv()

void forward_conv ( CNN * net )

static

Convolution stage: compute net->act.conv_out.

For each filter f and output position (r, c): conv_out[f][r][c] = bias[f] + sum_{dr,dc} kernel[f][dr][dc] * input[r+dr][c+dc]

Parameters

net	CNN with loaded weights and net->act.input filled.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ forward_dense1()

void forward_dense1 ( CNN * net )

static

Dense layer 1 (2704 → 128) followed by ReLU.

z1[h] = b1[h] + sum_f W1[h][f] * flat[f] h1[h] = ReLU(z1[h])

Parameters

net	CNN after forward_flatten().

Here is the caller graph for this function:

◆ forward_dense2()

void forward_dense2 ( CNN * net )

static

Dense layer 2 (128 → 26) followed by Softmax.

z2[k] = b2[k] + sum_h W2[k][h] * h1[h] output = softmax(z2)

Parameters

net	CNN after forward_dense1().

Here is the call graph for this function:

Here is the caller graph for this function:

◆ forward_flatten()

void forward_flatten ( CNN * net )

static

Flatten stage: copy pool_out into act->flat, row-major.

Index formula: flat[f * POOL_OUT_H * POOL_OUT_W + r * POOL_OUT_W + c]

Parameters

net	CNN after forward_pool().

Here is the caller graph for this function:

◆ forward_pool()

void forward_pool ( CNN * net )

static

Max-pooling stage: compute net->act.pool_out.

2×2 window, stride 2. Stores the winning row/col within each window in pool_max_r / pool_max_c for use during backprop.

Parameters

net	CNN after forward_conv().

Here is the caller graph for this function:

◆ randn()

float randn ( void )

static

Box-Muller transform: sample one value from N(0, 1).

Uses the standard two-uniform-samples formula. Not thread-safe due to static state; call from a single thread during initialisation.

Returns: One sample from the standard normal distribution.

Here is the caller graph for this function:

◆ relu_inplace()

void relu_inplace	(	float *	x,
		size_t	length )

static

Apply ReLU element-wise: f(x) = max(0, x).

Parameters

x	Input/output array (modified in-place).
length	Number of elements.

Here is the caller graph for this function:

◆ sgd_update()

void sgd_update	(	float *	w,
		float *	v,
		const float *	g,
		size_t	n,
		float	inv_batch )

static

Apply SGD+momentum to one weight array.

For each weight w, velocity v, gradient g: v = β * v + g w -= lr * v

Parameters

w	Weight array (modified in-place).
v	Velocity array (modified in-place).
g	Gradient array (read-only).
n	Number of elements.

Here is the caller graph for this function:

◆ softmax_inplace()

void softmax_inplace	(	float *	x,
		size_t	length )

static

Apply softmax in-place over an array of length n.

Subtracts the maximum value before exponentiation for numerical stability.

Parameters

x	Input/output array (modified in-place).
length	Number of elements.

Here is the caller graph for this function:

Functions

Detailed Description

Function Documentation

◆ cnn_backward()

◆ cnn_create()

◆ cnn_forward()

◆ cnn_free()

◆ cnn_loss()

◆ cnn_predict()

◆ cnn_update()

◆ cnn_zero_grads()

◆ forward_conv()

◆ forward_dense1()

◆ forward_dense2()

◆ forward_flatten()

◆ forward_pool()

◆ randn()

◆ relu_inplace()

◆ sgd_update()

◆ softmax_inplace()