fix float opts, add readme

Author: atzuur

.gitignore

@@ -50,5 +50,4 @@ modules.order
 Module.symvers
 Mkfile.old
 dkms.conf
-*.png
 .vscode/

Makefile

@@ -1,11 +1,11 @@
-CC=clang
+CC=gcc
 FILES=src/*.c
 
-WARNS=-std=c2x -Wall -Wextra -Wpedantic -Wshadow -Wformat=2 -Werror=return-type
+WARNS=-std=c23 -Wall -Wextra -Wpedantic -Wshadow -Wformat=2 -Werror=return-type -Wno-unknown-pragmas
 ifdef debug
 CFLAGS=-lm -g3 $(sanitize:%=-fsanitize=%)
 else
-CFLAGS=-lm -DNDEBUG -O3 -march=native -flto -ffast-math -fno-reciprocal-math
+CFLAGS=-lm -DNDEBUG -O3 -march=native -flto -ffast-math
 endif
 
 default:

README.md

@@ -1 +1,30 @@
-# mnist-c
+# mnist-c
+C implementation of a neural network for classification.
+![model classifying "5" with 99.8% confidence](img/demo.png)
+
+## Performance
+Using default settings training takes about 1.5s per epoch on my Ryzen 5 2600.
+- **[MNIST Handwritten Digits dataset](http://yann.lecun.com/exdb/mnist/)**: ~97% accuracy
+- **[Zalando Fashion-MNIST dataset](https://github.com/zalandoresearch/fashion-mnist)**: ~85% accuracy
+
+## Architecture
+Three fully connected layers (64-32-10 by default) with sigmoid activations in between, though this is trivially modifiable. 
+
+## Usage
+### Training
+```sh
+git clone https://github.com/atzuur/mnist-c
+cd mnist-c
+make # optionally add `CC=clang EXTRA="-fuse-ld=lld"` to use clang
+./a.exe train
+```
+### Prediction
+1. Draw a 28x28 MNIST-like image (see [demo](img/demo.png))
+2. Convert to raw floating point using e.g. [FFmpeg](https://ffmpeg.org):
+```sh
+ffmpeg -v error -i YOUR_IMAGE.png -pix_fmt grayf32 -f rawvideo YOUR_IMAGE
+```
+3. Run prediction using the model:
+```sh
+./a.exe YOUR_IMAGE
+```

img/demo.png

@@ -58,9 +58,11 @@ static int test_mnist(network* nn, const char* image) {
     }
 
     float* image_data = malloc(nn->layer_sizes[0] * sizeof *image_data);
-    if (fread(image_data, sizeof *image_data, MNIST_SAMPLE_LEN, image_file) !=
-        (size_t)MNIST_SAMPLE_LEN) {
+    int64_t n_read = fread(image_data, sizeof *image_data, MNIST_SAMPLE_LEN, image_file);
+    if (n_read != MNIST_SAMPLE_LEN) {
         free(image_data);
+        printf("expected %" PRId64 " bytes in image, got %" PRId64 "\n",
+               MNIST_SAMPLE_LEN * sizeof *image_data, n_read * sizeof *image_data);
         perror("fread");
         return 1;
     }
@@ -91,12 +93,12 @@ int main(int argc, const char** argv) {
     }
 
     network nn;
-    int64_t layer_sizes[] = {MNIST_SAMPLE_LEN, 50, 40, 10};
+    int64_t layer_sizes[] = {MNIST_SAMPLE_LEN, 64, 32, 10};
     nn_init(&nn, 4, layer_sizes);
 
     int ret;
     if (strcmp(argv[1], "train") == 0) {
-        ret = train_mnist(&nn, 30, 10, 3.0f);
+        ret = train_mnist(&nn, 30, 8, 3.0f);
     } else {
         ret = test_mnist(&nn, argv[1]);
     }

src/main.c

@@ -10,9 +10,15 @@
 #include <string.h>
 #include <time.h>
 
-static float sigmoid(float x) {
+#pragma float_control(precise, on, push)
+#if defined(__GNUC__) && !defined(__clang__)
+__attribute__((optimize("no-finite-math-only")))
+#endif
+static float
+sigmoid(float x) {
     return 1.0f / (expf(-x) + 1);
 }
+#pragma float_control(pop)
 
 static float d_sigmoid(float x) {
     return sigmoid(x) * (1.0f - sigmoid(x));