def gpu_fft_run_times_with_img_copy():
image_pth = Path(os.path.dirname(
os.path.realpath(__file__))) / Path("../simulation/screenshot.png")
screenshot = SimulPLIF(img_path=image_pth, num_repeats=1)
img_height, img_width = screenshot[0].squeeze(0).numpy().shape
train_dataloader = torch.utils.data.DataLoader(screenshot,
batch_size=1,
pin_memory=PIN_MEMORY)
with open("gpu_fftconv_with_img_copy_run_times.csv", "w",
newline="") as csvfile:
writer = csv.writer(csvfile)
writer.writerow(["n_bin", "max_sigma", "time"])
for i in range(REPEATS):
for max_sigma in range(min_sigma + 1, mx_sigma + 1):
for n_bin in range(min_bin, max_bin + 1):
with torch.no_grad():
model = DifferenceOfGaussiansFFT(
img_height=img_height,
img_width=img_width,
min_sigma=min_sigma,
max_sigma=max_sigma,
overlap=0.9,
threshold=0.012,
prune=False,
sigma_bins=n_bin,
).to(DEVICE, non_blocking=PIN_MEMORY)
for p in model.parameters():
p.requires_grad = False
model.eval()
torch.cuda.synchronize(DEVICE)
START = time.monotonic()
img_tensor = next(iter(train_dataloader))
img_tensor = img_tensor.to(DEVICE,
non_blocking=PIN_MEMORY)
torch.cuda.synchronize(DEVICE)
mask, local_maxima = model(img_tensor)
m, l = next(zip(mask, local_maxima))
blobs = model.make_blobs(m, l)
torch.cuda.synchronize(DEVICE)
END = time.monotonic()
res = [n_bin, max_sigma, END - START]
writer.writerow(res)
print(res)
def gpu_accuracy():
test_img_dir = Path(os.path.dirname(
os.path.realpath(__file__))) / Path("../simulation/test_data/")
screenshot = PLIF(plif_dir=test_img_dir, ext="png", include_paths=True)
img_height, img_width = screenshot[0][0].squeeze(0).numpy().shape
train_dataloader = DataLoader(screenshot,
batch_size=1,
pin_memory=PIN_MEMORY,
shuffle=False)
with torch.no_grad():
model = DifferenceOfGaussiansFFT(
img_height=img_height,
img_width=img_width,
min_sigma=min_sigma,
max_sigma=mx_sigma,
overlap=0.5,
threshold=0.1,
prune=False,
sigma_bins=max_bin,
).to(DEVICE, non_blocking=PIN_MEMORY)
for p in model.parameters():
p.requires_grad = False
model.eval()
for img_tensor, image_pth in train_dataloader:
img_tensor = img_tensor.to(DEVICE, non_blocking=PIN_MEMORY)
mask, local_maxima = model(img_tensor)
m, l = next(zip(mask, local_maxima))
blobs = model.make_blobs(m, l)
# make_circles_fig(screenshot[0][0].numpy(), blobs).show()
# break
image_pth = list(image_pth)[0]
fn = str(
Path(os.path.dirname(os.path.realpath(__file__))) /
f"accuracy_results/gpu/{os.path.basename(image_pth)}.res")
np.savetxt(fn, blobs)
print(fn)