Update k-means example to show graphical result #2521
Conversation
929e63d to
c10cdbf
Compare
c10cdbf to
9c9fae5
Compare
|
Thanks for reviving this good example 👍 Perhaps, we should use a smaller image size so that the example runs relatively fast on lower level NVIDIA/AMD GPUs. I have tried few images from our examples-images-assets, diff --git a/examples/machine_learning/kmeans.cpp b/examples/machine_learning/kmeans.cpp
index 81f1503a..e3a80018 100644
--- a/examples/machine_learning/kmeans.cpp
+++ b/examples/machine_learning/kmeans.cpp
@@ -113,7 +113,7 @@ int kmeans_demo(int k, bool console) {
printf("** ArrayFire K-Means Demo (k = %d) **\n\n", k);
array img =
- loadImage(ASSETS_DIR "/examples/images/vegetable-woman.jpg", true) /
+ loadImage(ASSETS_DIR "/examples/images/spider.jpg", false) /
255; // [0-255]
int w = img.dims(0), h = img.dims(1), c = img.dims(2);
@@ -133,16 +133,16 @@ int kmeans_demo(int k, bool console) {
array out_half = moddims(means_half(span, clusters_half, span), img.dims());
array out_dbl = moddims(means_dbl (span, clusters_dbl , span), img.dims());
- af::Window wnd("ArrayFire K-Means Demo");
- wnd.grid(1, 4);
+ af::Window wnd(1024, 1024, "ArrayFire K-Means Demo");
+ wnd.grid(2, 2);
std::string out_full_caption = "k = " + std::to_string(k);
std::string out_half_caption = "k = " + std::to_string(k / 2);
std::string out_dbl_caption = "k = " + std::to_string(k * 2);
while (!wnd.close()) {
wnd(0, 0).image(img, "Input Image");
wnd(0, 1).image(out_full, out_full_caption.c_str());
- wnd(0, 2).image(out_half, out_half_caption.c_str());
- wnd(0, 3).image(out_dbl, out_dbl_caption.c_str());
+ wnd(1, 0).image(out_half, out_half_caption.c_str());
+ wnd(1, 1).image(out_dbl, out_dbl_caption.c_str());
wnd.show();
}
`` |
|
@9prady9 Thanks. I tried the spider example, and it looks like it's being processed faster indeed on my laptop, I assume largely because it is monochrome. I also think changing the grid layout to 2x2 is good, but I think it's better to have the window
|
ca48a0f to
cfb7c55
Compare
|
Yes, it is three times less data in a monochrome image. Sure, you can reduce it to 800x800 and cluster parameter to 8. The reason I suggested smaller image is since CUDA/OpenCL takes so long, on CPU backend the example will seem like it is running forever. |
The k-means example has a non-working code stub that looks like it was supposed to be updated once arrayfire had the functionality to show windows. This change simply converts that code stub into a working implementation.
Here is a screenshot of the result (for k = 16, which is the example's default setting):
