initial commit

.gitignore

@@ -1 +1,3 @@
 build/
+data/
+

.vscode/settings.json

@@ -6,6 +6,8 @@
 		"cltest.c": "c",
 		"*.tcc": "c",
 		"string": "c",
-		"string_view": "c"
+		"string_view": "c",
+		"stdint.h": "c",
+		"inttypes.h": "c"
 	}
 }

Makefile

@@ -5,7 +5,7 @@ CFLAGS=-Wall -g
 .PHONY: all
 all:
 	mkdir -p $(BUILD)
-	gcc -lOpenCL $(CFLAGS) -o $(BUILD)/cltest $(SRC)/cltest.c $(SRC)/clm.c
+	gcc -lOpenCL -lm $(CFLAGS) -o $(BUILD)/cltest $(SRC)/cltest.c $(SRC)/clm.c
 
 .PHONY: run
 run: all

src/clm.c

@@ -5,36 +5,8 @@
 
 #include "clm.h"
 
-clm_Matrix INVALID_MATRIX = { .rows = 0, .cols = 0, .values = NULL };
-
-clm_Vector clm_createVector(int length) {
-	clm_Vector vector;
-	vector.values = calloc(length, sizeof(float));
-	vector.length = length;
-	return vector;
-}
-
-void clm_freeVector(clm_Vector vector) {
-	free(vector.values);
-}
-
-clm_Layer clm_createLayer(unsigned int prevNumNeurons, unsigned int numLayers) {
-	clm_Layer layer;
-	for(unsigned int i = 0; i < numLayers; i++) {
-
-	}
-	return layer;
-}
-
-clm_NN clm_createNN(unsigned int numLayers, unsigned int *layerSizes) {
-	clm_NN nn;
-	nn.numLayers = numLayers;
-	for(unsigned int i = 0; i < numLayers; i++) {
-		clm_Layer layer;
-
-	}
-	return nn;
-}
+const clm_Matrix INVALID_MATRIX = { .rows = 0, .cols = 0, .values = NULL };
+const clm_Vector INVALID_VECTOR = { .length = 0, .values = NULL };
 
 clm_Matrix clm_createMatrix(unsigned int rows, unsigned int cols) {
 	clm_Matrix mat;
@@ -60,7 +32,7 @@ void clm_freeMatrix(clm_Matrix mat) {
 	free(mat.values);
 }
 
-clm_Matrix clm_copyMatrix(clm_Matrix mat) {
+clm_Matrix clm_matrixCopy(clm_Matrix mat) {
 	clm_Matrix copy = clm_createMatrix(mat.rows, mat.cols);
 
 	memcpy(copy.values, mat.values, mat.rows * mat.cols * sizeof(float));
@@ -86,7 +58,7 @@ clm_Matrix clm_matrixAddMatrix(clm_Matrix mat, clm_Matrix other) {
 		mat.values[i] += other.values[i];
 	}
 
-	clm_freeMatrix(other);
+	//clm_freeMatrix(other);
 
 	return mat;
 }
@@ -101,7 +73,7 @@ clm_Matrix clm_matrixSubtractMatrix(clm_Matrix mat, clm_Matrix other) {
 		mat.values[i] -= other.values[i];
 	}
 
-	clm_freeMatrix(other);
+	//clm_freeMatrix(other);
 
 	return mat;
 }
@@ -114,7 +86,7 @@ clm_Matrix clm_matrixTranspose(clm_Matrix mat) {
 		}
 	}
 
-	clm_freeMatrix(mat);
+	//clm_freeMatrix(mat);
 
 	return tr;
 }
@@ -136,12 +108,25 @@ clm_Matrix clm_matrixMultiplyMatrix(clm_Matrix a, clm_Matrix b) {
 		}
 	}
 
-	clm_freeMatrix(a);
-	clm_freeMatrix(b);
+	//clm_freeMatrix(a);
+	//clm_freeMatrix(b);
 
 	return r;
 }
 
+clm_Matrix clm_matrixMultiplyMatrixElements(clm_Matrix mat, clm_Matrix other) {
+	if(mat.rows != other.rows || mat.cols != other.cols) {
+		printf("Cannot multiply matrices element-wise\n");
+		return INVALID_MATRIX;
+	}
+
+	for(unsigned int i = 0; i < mat.cols * mat.rows; i++) {
+		mat.values[i] *= other.values[i];
+	}
+
+	return mat;
+}
+
 clm_Matrix clm_matrixMultiplyScalar(clm_Matrix mat, float scalar) {
 	for(unsigned int i = 0; i < mat.cols * mat.rows; i++) {
 		mat.values[i] *= scalar;
@@ -153,7 +138,7 @@ clm_Matrix clm_matrixMultiplyScalar(clm_Matrix mat, float scalar) {
 clm_Matrix clm_matrixSigmoid(clm_Matrix mat) {
 	for(unsigned int i = 0; i < mat.rows; i++) {
 		for(unsigned int j = 0; j < mat.cols; j++) {
-			mat.data[i * mat.cols + j] = 1 / (1 + exp(-mat.values[i * mat.cols + j]));
+			mat.values[i * mat.cols + j] = 1 / (1 + exp(-mat.values[i * mat.cols + j]));
 		}
 	}
 
@@ -163,14 +148,51 @@ clm_Matrix clm_matrixSigmoid(clm_Matrix mat) {
 clm_Matrix clm_matrixDSigmoid(clm_Matrix mat) {
 	for(unsigned int i = 0; i < mat.rows; i++) {
 		for(unsigned int j = 0; j < mat.cols; j++) {
-			float v = mat.data[i * mat.cols + j];
-			mat.data[i * mat.cols + j] = v * (1 - v);
+			float v = mat.values[i * mat.cols + j];
+			mat.values[i * mat.cols + j] = v * (1 - v);
 		}
 	}
 
 	return mat;
 }
 
+clm_Matrix clm_matrixFromArray(float *array, unsigned int length) {
+	clm_Matrix matrix = clm_createMatrix(length, 1);
+	memcpy(matrix.values, array, length * sizeof(float));
+	return matrix;
+}
+
+bool clm_matrixIsInvalid(clm_Matrix mat) {
+	return mat.values == NULL;
+}
+
+clm_Vector clm_vectorCreate(unsigned int length) {
+	clm_Vector vector;
+	vector.length = length;
+	vector.values = calloc(length, sizeof(float));
+	return vector;
+}
+
+bool clm_vectorIsInvalid(clm_Vector vec) {
+	return vec.values != NULL;
+}
+
+clm_Linear clm_linearCreateRandom(unsigned int inputs, unsigned int outputs) {
+	clm_Linear linear;
+	linear.weights = clm_createMatrixRandom(outputs, inputs);
+	linear.bias = clm_createMatrixRandom(outputs, 1);
+	return linear;
+}
+
+void clm_freeVector(clm_Vector vector) {
+	free(vector.values);
+}
+
+void clm_freeLinear(clm_Linear linear) {
+	clm_freeMatrix(linear.weights);
+	clm_freeMatrix(linear.bias);
+}
+
 void clm_matrixPrint(clm_Matrix mat) {
 	printf("[\n");
 	for(unsigned int i = 0; i < mat.rows; i++) {
@@ -181,3 +203,7 @@ void clm_matrixPrint(clm_Matrix mat) {
 	}
 	printf("]\n");
 }
+
+void clm_matrixPrintShape(clm_Matrix mat) {
+	printf("(%dx%d)\n", mat.rows, mat.cols);
+}

src/clm.h

@@ -1,6 +1,8 @@
 #ifndef _CLM_H_
 #define _CLM_H_
 
+#include <stdbool.h>
+
 typedef struct {
 	float *values;
 	unsigned int rows;
@@ -9,22 +11,22 @@ typedef struct {
 
 typedef struct {
 	float *values;
-	int length;
+	unsigned int length;
 } clm_Vector;
 
-clm_Vector clm_createVector(int length);
-void clm_freeVector(clm_Vector vector);
+typedef struct {
+	clm_Matrix weights;
+	clm_Matrix bias;
+} clm_Linear;
 
 typedef struct {
-	clm_Matrix weights, bias;
-} clm_Layer;
-
-typedef struct {
-	clm_Layer *layers;
+	clm_Linear *layers;
 	unsigned int numLayers;
+	float learnRate;
 } clm_NN;
 
-clm_NN clm_createNN(unsigned int numLayers, unsigned int *layerSizes);
+extern const clm_Matrix INVALID_MATRIX;
+extern const clm_Vector INVALID_VECTOR;
 
 clm_Matrix clm_createMatrixRandom(unsigned int rows, unsigned int cols);
 clm_Matrix clm_matrixAddScalar(clm_Matrix mat, float scalar);
@@ -32,8 +34,26 @@ clm_Matrix clm_matrixAddMatrix(clm_Matrix mat, clm_Matrix other);
 clm_Matrix clm_matrixSubtractMatrix(clm_Matrix mat, clm_Matrix other);
 clm_Matrix clm_matrixTranspose(clm_Matrix mat);
 clm_Matrix clm_matrixMultiplyMatrix(clm_Matrix a, clm_Matrix b);
+clm_Matrix clm_matrixMultiplyMatrixElements(clm_Matrix mat, clm_Matrix other);
 clm_Matrix clm_matrixMultiplyScalar(clm_Matrix mat, float scalar);
+clm_Matrix clm_matrixSigmoid(clm_Matrix mat);
+clm_Matrix clm_matrixDSigmoid(clm_Matrix mat);
+
+clm_Matrix clm_matrixFromArray(float *array, unsigned int length);
+clm_Matrix clm_matrixCopy(clm_Matrix matrix);
+
+bool clm_matrixIsInvalid(clm_Matrix mat);
+
+clm_Vector clm_vectorCreate(unsigned int length);
+
+bool clm_vectorIsInvalid(clm_Vector vec);
+
+clm_Linear clm_linearCreateRandom(unsigned int inputs, unsigned int outputs);
 
 void clm_matrixPrint(clm_Matrix mat);
+void clm_matrixPrintShape(clm_Matrix mat);
+void clm_freeMatrix(clm_Matrix mat);
+void clm_freeVector(clm_Vector vector);
+void clm_freeLinear(clm_Linear linear);
 
 #endif

src/cltest.c

@@ -1,23 +1,283 @@
 #include <stdio.h>
+#include <stdlib.h>
+#include <inttypes.h>
 
 #include "clm.h"
 
-int main() {
-	clm_Matrix mat = clm_createMatrixRandom(2, 1);
-	clm_matrixPrint(mat);
+float train_data_x[4][2] = {
+	{0, 0},
+	{0, 1},
+	{1, 0},
+	{1, 1}
+};
 
-	mat = clm_matrixTranspose(mat);
-	clm_matrixPrint(mat);
+float train_data_y[4][1] = {
+	{0},
+	{1},
+	{1},
+	{0}
+};
 
-	mat = clm_matrixAddScalar(mat, 1);
-	clm_matrixPrint(mat);
+float *predict(clm_NN nn, float *x, unsigned int length) {
+	clm_Matrix xM = clm_matrixFromArray(x, length);
 
-	clm_Matrix b = clm_createMatrixRandom(2, 2);
-	clm_matrixPrint(b);
+	for(unsigned int i = 0; i < nn.numLayers; i++) {
+		clm_Linear layer = nn.layers[i];
+		clm_Matrix newX = clm_matrixMultiplyMatrix(layer.weights, xM);
 
-	clm_Matrix mul = clm_matrixMultiplyMatrix(mat, b);
-	clm_matrixPrint(mul);
-
-	mul = clm_matrixMultiplyScalar(mul, 0.5);
-	clm_matrixPrint(mul);
+		if(clm_matrixIsInvalid(newX)) {
+			printf("Failed to predict\n");
+			return NULL;
+		}
+
+		clm_matrixAddMatrix(newX, layer.bias);
+		clm_matrixSigmoid(newX);
+		clm_freeMatrix(xM);
+		xM = newX;
+	}
+
+	return xM.values;
+}
+
+void train(clm_NN nn, float *x, unsigned int xL, float *y, unsigned int yL) {
+	clm_Matrix xM = clm_matrixFromArray(x, xL);
+	clm_Matrix yM = clm_matrixFromArray(y, yL);
+
+	// TODO: potential compute/memory tradeoff? (recalculate matrices every time <-> keep everything cached)
+
+	// Forward pass
+	clm_Matrix *outputs = calloc(nn.numLayers + 1 /* 1 for input */, sizeof(clm_Matrix));
+	outputs[0] = xM;
+	for(unsigned int i = 0; i < nn.numLayers; i++) {
+		clm_Linear layer = nn.layers[i];
+		clm_Matrix newX = clm_matrixMultiplyMatrix(layer.weights, xM);
+		if(clm_matrixIsInvalid(newX)) {
+			printf("Forward pass failed\n");
+			return;
+		}
+
+		clm_matrixAddMatrix(newX, layer.bias);
+		clm_matrixSigmoid(newX);
+		xM = newX;
+		outputs[i + 1] = xM;
+	}
+
+	clm_Matrix dError = clm_matrixSubtractMatrix(yM, outputs[nn.numLayers]); // yhat - y
+
+	clm_Matrix lastGradient = clm_matrixDSigmoid(clm_matrixCopy(outputs[nn.numLayers])); // dsig(yhat)
+	clm_matrixMultiplyMatrixElements(lastGradient, dError); // (yhat - y) . dsig(yhat)
+	clm_matrixMultiplyScalar(lastGradient, nn.learnRate);
+
+	clm_Matrix lastInputT = clm_matrixTranspose(outputs[nn.numLayers - 1]);
+	clm_Matrix lastDeltaW = clm_matrixMultiplyMatrix(lastGradient, lastInputT);
+	clm_freeMatrix(lastInputT);
+
+	clm_matrixAddMatrix(nn.layers[nn.numLayers - 1].weights, lastDeltaW);
+	clm_matrixAddMatrix(nn.layers[nn.numLayers - 1].bias, lastGradient);
+
+	clm_freeMatrix(lastDeltaW);
+	clm_freeMatrix(lastGradient);
+
+	for(int i = nn.numLayers - 2; i >= 0; i--) {
+		clm_Linear layer = nn.layers[i];
+		clm_Matrix inputToThisLayer = outputs[i];
+		clm_Matrix outputOfThisLayer = outputs[i + 1];
+
+		clm_Matrix weightsT = clm_matrixTranspose(nn.layers[i + 1].weights);
+		clm_Matrix newDError = clm_matrixMultiplyMatrix(weightsT, dError);
+		clm_freeMatrix(weightsT);
+		clm_freeMatrix(dError);
+		dError = newDError;
+
+		clm_Matrix gradient = clm_matrixDSigmoid(clm_matrixCopy(outputOfThisLayer));
+		clm_matrixMultiplyMatrixElements(gradient, dError);
+		clm_matrixMultiplyScalar(gradient, nn.learnRate);
+
+		clm_Matrix inputT = clm_matrixTranspose(inputToThisLayer);
+		clm_Matrix deltaW = clm_matrixMultiplyMatrix(gradient, inputT);
+		clm_freeMatrix(inputT);
+
+		clm_matrixAddMatrix(layer.weights, deltaW);
+		clm_matrixAddMatrix(layer.bias, gradient);
+
+		clm_freeMatrix(deltaW);
+		clm_freeMatrix(gradient);
+	}
+
+	clm_freeMatrix(dError);
+
+	for(unsigned int i = 0; i <= nn.numLayers; i++) {
+		clm_freeMatrix(outputs[i]);
+	}
+
+	free(outputs);
+}
+
+void loadLabels(clm_Vector **labelsOut, unsigned int *labelsCountOut) {
+	FILE *file = fopen("data/train-labels.idx1-ubyte", "r");
+	if(!file) {
+		perror("Failed to open labels\n");
+		return;
+	}
+
+	unsigned char magicBytes[4];
+	fread(magicBytes, sizeof(magicBytes), 1, file);
+
+	printf("%d\n", (magicBytes[0] << 24) | (magicBytes[1] << 16) | (magicBytes[2] << 8) | magicBytes[3]);
+
+	unsigned char lengthBytes[4];
+	fread(lengthBytes, sizeof(lengthBytes), 1, file);
+
+	uint32_t length = (lengthBytes[0] << 24) | (lengthBytes[1] << 16) | (lengthBytes[2] << 8) | lengthBytes[3];
+	printf("%" PRId32 "\n", length);
+
+	clm_Vector *vectors = calloc(length, sizeof(clm_Vector));
+
+	for(unsigned int i = 0; i < length; i++) {
+		unsigned char label;
+		fread(&label, sizeof(unsigned char), 1, file);
+
+		clm_Vector vector = clm_vectorCreate(10);
+		for(unsigned int j = 0; j < 10; j++) {
+			vector.values[j] = label == j ? 1 : 0;
+		}
+		vectors[i] = vector;
+	}
+
+	*labelsOut = vectors;
+	*labelsCountOut = length;
+}
+
+void loadImages(clm_Vector **imagesOut, unsigned int *imageCountOut) {
+	FILE *file = fopen("data/train-images.idx3-ubyte", "r");
+	if(!file) {
+		perror("Failed to open images\n");
+		return;
+	}
+
+	unsigned char magicBytes[4];
+	fread(magicBytes, sizeof(magicBytes), 1, file);
+
+	printf("%d\n", (magicBytes[0] << 24) | (magicBytes[1] << 16) | (magicBytes[2] << 8) | magicBytes[3]);
+
+	unsigned char lengthBytes[4];
+	fread(lengthBytes, sizeof(lengthBytes), 1, file);
+	uint32_t length = (lengthBytes[0] << 24) | (lengthBytes[1] << 16) | (lengthBytes[2] << 8) | lengthBytes[3];
+	printf("%" PRId32 "\n", length);
+
+	unsigned char rowsBytes[4];
+	fread(rowsBytes, sizeof(rowsBytes), 1, file);
+	uint32_t rows = (rowsBytes[0] << 24) | (rowsBytes[1] << 16) | (rowsBytes[2] << 8) | rowsBytes[3];
+	printf("%" PRId32 "\n", rows);
+
+	unsigned char colsBytes[4];
+	fread(colsBytes, sizeof(colsBytes), 1, file);
+	uint32_t cols = (colsBytes[0] << 24) | (colsBytes[1] << 16) | (colsBytes[2] << 8) | colsBytes[3];
+	printf("%" PRId32 "\n", cols);
+
+	clm_Vector *images = calloc(length, sizeof(clm_Vector));
+	for(unsigned int i = 0; i < length; i++) {
+		clm_Vector vec = clm_vectorCreate(cols * rows);
+		unsigned char image[cols * rows];
+		fread(image, sizeof(image), 1, file);
+		for(unsigned int j = 0; j < cols * rows; j++) {
+			vec.values[j] = (float) image[j];
+		}
+		images[i] = vec;
+	}
+
+	*imagesOut = images;
+	*imageCountOut = length;
+}
+
+int main() {
+	clm_Vector *labels = NULL;
+	unsigned int labelCount;
+	loadLabels(&labels, &labelCount);
+	printf("LENGTH: %u\n", labelCount);
+
+	clm_Vector *images = NULL;
+	unsigned int imageCount;
+	loadImages(&images, &imageCount);
+
+	imageCount = 60000;
+
+	printf("%f\n", images[0].values[0]);
+
+	srand(1);
+
+	unsigned int
+		i = 784,
+		h = 30,
+		o = 10;
+
+	clm_Linear layer1;
+	layer1.weights = clm_createMatrixRandom(h, i);
+	layer1.bias = clm_createMatrixRandom(h, 1);
+
+	clm_Linear layer2;
+	layer2.weights = clm_createMatrixRandom(o, h);
+	layer2.bias = clm_createMatrixRandom(o, 1);
+
+	clm_Linear layers[] = {layer1, layer2};
+	clm_NN nn = { layers, sizeof(layers) / sizeof(clm_Linear), 0.01 };
+
+	for(unsigned int epoch = 0; epoch < 10;  epoch++) {
+		printf("Epoch %u\n", epoch);
+
+		for(unsigned int idx = 0; idx < imageCount; idx++) { // Each train sample
+			if(idx % 1000 == 0) {
+				printf("%.2f%%\n", idx / (float) imageCount * 100);
+			}
+			//printf("%u\n", idx);
+			//train(nn, train_data_x[idx], 2, train_data_y[idx], 1);
+			/*for(unsigned int f = 0; f < images[idx].length; f++) {
+				printf("%.2f ", images[idx].values[f]);
+			}
+			printf("\n");
+			for(unsigned int f = 0; f < labels[idx].length; f++) {
+				printf("%.2f ", labels[idx].values[f]);
+			}
+			printf("\n");*/
+			//printf("%.2f\n", labels.values[idx]);
+
+			train(nn, images[idx].values, images[idx].length, labels[idx].values, labels[idx].length);
+			//train(nn, test, 784, target, 10);
+			//predict(nn, test, 784);
+		}
+	}
+
+	unsigned int correct = 0;
+	for(unsigned int idx = 0; idx < imageCount; idx++) { // Each train sample
+		//printf("pred(%.2f, %.2f) = %.2f\n", train_data_x[idx][0], train_data_x[idx][1], predict(nn, train_data_x[idx], 2)[0]);
+		float *pred = predict(nn, images[idx].values, images[idx].length);
+		unsigned int predDigit = 0;
+		float max = -1;
+		for(unsigned int j = 0; j < 10; j++) {
+			//printf("%.2f ", pred[j]);
+			if(pred[j] > max || max < 0) {
+				max = pred[j];
+				predDigit = j;
+			}
+		}
+		if(idx < 100) printf("%u (confidence: %.2f)\n", predDigit, max);
+
+		unsigned int actDigit = 0;
+		float maxA = -1;
+		for(unsigned int j = 0; j < 10; j++) {
+			//printf("%.2f ", pred[j]);
+			if(labels[idx].values[j] > maxA || maxA < 0) {
+				maxA = labels[idx].values[j];
+				actDigit = j;
+			}
+		}
+		if(idx < 100) printf("Actual: %u\n", actDigit);
+		//printf("\n");
+
+		if(predDigit == actDigit) correct++;
+	}
+
+	printf("Correct: %u -> %.2f", correct, (float) correct / imageCount * 100);
+
+	printf("\n");
 }