def test_PoolingWindows_project(self):
im = torch.rand((1, 1, 256, 256), dtype=torch.float32)
pw = pooling.PoolingWindows(.5, im.shape[2:])
pooled = pw(im)
pw.project(pooled)
pw = pooling.PoolingWindows(.5, im.shape[2:], num_scales=3)
pooled = pw(im)
pw.project(pooled)
def test_PoolingWindows_caching(self, tmp_path):
im = torch.rand((1, 1, 256, 256), dtype=torch.float32)
# first time we save, second we load
pw = pooling.PoolingWindows(.8,
im.shape[-2:],
num_scales=2,
cache_dir=tmp_path)
pw = pooling.PoolingWindows(.8,
im.shape[-2:],
num_scales=2,
cache_dir=tmp_path)
def test_PoolingWindows_nonsquare(self):
# test PoolingWindows with weirdly-shaped iamges
im = torch.rand((1, 1, 256, 256), dtype=torch.float32)
for sh in [(256, 128), (256, 127), (256, 125), (125, 125), (127, 125)]:
tmp = im[..., :sh[0], :sh[1]]
pw = pooling.PoolingWindows(.9, tmp.shape[-2:])
pw(tmp)
def test_PoolingWindows_cosine(self, num_scales, transition_region_width):
im = torch.rand((1, 1, 256, 256), dtype=torch.float32)
pw = pooling.PoolingWindows(
.5,
im.shape[2:],
num_scales=num_scales,
transition_region_width=transition_region_width,
window_type='cosine',
)
pw(im)
def test_PoolingWindows_plotting(self):
im = torch.rand((1, 1, 256, 256), dtype=torch.float32)
pw = pooling.PoolingWindows(.8, im.shape[-2:], num_scales=2)
pw.plot_window_areas()
pw.plot_window_widths()
for i in range(2):
pw.plot_window_areas('pixels', i)
pw.plot_window_widths('pixels', i)
fig = po.imshow(im)
pw.plot_windows(fig.axes[0])
plt.close('all')
def test_PoolingWindows(self, num_scales):
im = torch.rand((1, 1, 256, 256), dtype=torch.float32)
pw = pooling.PoolingWindows(.5,
im.shape[2:],
num_scales=num_scales,
window_type='gaussian',
std_dev=1)
pw(im)
# we only support std_dev=1
with pytest.raises(Exception):
pooling.PoolingWindows(.5,
im.shape[2:],
num_scales=num_scales,
window_type='gaussian',
std_dev=2)
with pytest.raises(Exception):
pooling.PoolingWindows(.5,
im.shape[2:],
num_scales=num_scales,
window_type='gaussian',
std_dev=.5)
def test_reweighting(self, num_scales, input_fmt):
pw = pooling.PoolingWindows(.5, (256, 256), num_scales=num_scales)
im = {(i, ): torch.rand((1, 1, 256 // 2**i, 256 // 2**i),
dtype=torch.float32)
for i in range(num_scales)}
if input_fmt == 'dict':
pw(im, weights=torch.ones(num_scales, 1, 1, 1, 1))
elif input_fmt == 'tensor':
for i in range(num_scales):
pw(im[(i, )],
idx=i,
weights=torch.ones(num_scales, 1, 1, 1, 1))
def test_PoolingWindows_sep(self):
# test the window and pool function separate of the forward function
im = torch.rand((1, 1, 256, 256), dtype=torch.float32)
pw = pooling.PoolingWindows(.5, im.shape[2:])
pw.pool(pw.window(im))