def test_plot_gaussian_function():
mu = 0.5
sigma = 0.4
kp = torch.randn(1, 1, 2, requires_grad=True) * sigma + mu
z = MF.gaussian_like_function(kp, 14, 14,
sigma=0.1).squeeze().detach().numpy()
#plot_heightmap3d(z)
plot_heatmap2d(z)
def test_gaussian_like_batch():
ss = knn.SpatialSoftmax()
k = ss(heatmap_batch)
hm = MF.gaussian_like_function(k, 5, 5)
print('')
print(hm)
def test_bottlneck_grads():
heatmap = torch.rand(1, 1, 5, 5, requires_grad=True)
h = heatmap.neg()
ss = knn.SpatialSoftmax()
kp = ss(h)
ss = MF.gaussian_like_function(kp, 5, 5)
loss = torch.sum(ss)
loss.backward()
print(heatmap.grad)
print(kp[0].grad, kp[1].grad)
def test_gaussian_function_grads():
x, y = torch.rand(1, 5, requires_grad=True), torch.rand(1,
5,
requires_grad=True)
kp = x.neg(), y.neg()
ss = MF.gaussian_like_function(kp, 5, 5)
loss = torch.sum(ss)
loss.backward()
print(ss)
print(x, y)
print(x.grad, y.grad)
def test_co_ords():
height, width = 16, 16
hm = torch.zeros(1, 1, height, width)
hm[0, 0, 0, 15] = 20.0
k, p = MF.spacial_softmax(hm, probs=True)
g = MF.gaussian_like_function(k, height, width)
#plot_heightmap3d(hm[0, 0].detach().numpy())
#plot_heightmap3d(g[0, 0].detach().numpy(), k[0, 0])
#plot_single_channel(hm[0, 0])
#plot_single_channel(g[0, 0])
d = UniImageViewer()
def test_align_kp_with_gaussian():
hm = heatmap()
# image = TVF.to_pil_image(heatmap[0])
ss = knn.SpatialSoftmax()
kp = ss(hm)
z = MF.gaussian_like_function(kp, 5, 5,
sigma=0.1).squeeze().detach().numpy()
img = plot_keypoints_on_image(kp[0], hm[0])
plt.imshow(img)
plot_heatmap2d(z)
def test_gaussian_like():
heatmap = torch.tensor([
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 5, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
]).expand(1, 1, 5, 5).float()
ss = knn.SpatialSoftmax()
k = ss(heatmap)
hm = MF.gaussian_like_function(k, 5, 5)
print('')
print(hm)
def __getitem__(self, item):
kp = torch.rand(1, self.keypoints, 2)
pointmap = MF.point_map(kp, self.height, self.width)
mask = MF.gaussian_like_function(kp, self.height, self.width)
mask, _ = torch.max(mask, dim=1, keepdim=True)
return pointmap.squeeze(0), mask.squeeze(0)