def test_tile_resizer(self):
from flowdec import data as fd_data
image = fd_data.bars_25pct().data.astype(np.uint16)
tile = image[np.newaxis, :, np.newaxis, ...]
self.assertTrue(tile.ndim == 5,
'Expecting 5D shape but got {}'.format(tile.shape))
config = cytokit_simulation.get_example_config(example_name='ex1')
# Equate resolution to do fair comparison between anisotropic skimage resize
# and tensorflow isotropic resize
config._conf['acquisition']['axial_resolution'] = 1.
config._conf['acquisition']['lateral_resolution'] = 1.
# Set resizing factors and implementation to ensure the operation will run
config._conf['processor']['tile_resize']['factors'] = [.75, .75, .75]
config._conf['processor']['tile_resize']['implementation'] = 'skimage'
resizer = tile_resize.CytokitTileResize(config).initialize()
sk_res = resizer.run(tile)[0, :, 0]
self.assertTrue(sk_res.ndim == 3,
'Expecting 3D shape but got {}'.format(sk_res.shape))
# Run again using tensorflow implementation
config._conf['processor']['tile_resize'][
'implementation'] = 'tensorflow'
resizer = tile_resize.CytokitTileResize(config).initialize()
tf_res = resizer.run(tile)[0, :, 0]
self.assertTrue(tf_res.ndim == 3,
'Expecting 3D shape but got {}'.format(tf_res.shape))
# Ensure that differences are both minor and present to such a degree that
# the same implementation was not used for comparison
ssim = compare_ssim(sk_res, tf_res)
self.assertTrue(.99 <= ssim < 1,
'Expecting SSIM in [.99, 1) but got {}'.format(ssim))
def test_observer(self):
acq = fd_data.bars_25pct()
imgs = []
def observer(img, *_):
imgs.append(img)
self.assertEqual(acq.data.shape, img.shape, msg='Observer image and original shapes not equal')
algo = fd_restoration.RichardsonLucyDeconvolver(n_dims=3, observer_fn=observer).initialize()
algo.run(acq, niter=5)
self.assertEqual(len(imgs), 5)
def run_deconvolution(device):
import os
import tensorflow as tf
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = device
#acq = fd_data.load_celegans_channel('CY3')
for i in range(10):
acq = fd_data.bars_25pct()
algo = fd_restoration.RichardsonLucyDeconvolver(3).initialize()
res = algo.run(acq, niter=50, session_config=config).data
time.sleep(5)
return res
def test_bars(self):
"""Test reconstruction of blurred "Hollow Bars" volume"""
n_iter = 10
thresholds = {
# Results should always improve on the original
'original': 1.,
# Results should also be better than DL2 and scikit-image most of the time
'sk': .75,
'dl2': .75 if dl2_enabled() else None
}
acq = fd_data.bars_25pct()
# Initialize list of acquisitions to deconvolve
acqs = [acq]
# Add translations to images and kernels to ensure that there aren't
# any issues supporting non-symmetric inputs
acqs += [
tfv.reblur(tfv.shift(acq, data_shift=(0, 10, 10))),
tfv.reblur(tfv.shift(acq, data_shift=(0, -10, -10))),
tfv.reblur(tfv.shift(acq, kern_shift=(0, 10, 10))),
tfv.reblur(tfv.shift(acq, kern_shift=(0, -10, -10))),
tfv.reblur(
tfv.shift(acq, data_shift=(-3, 5, -5), kern_shift=(-3, 5, -5)))
]
# Subset image and kernel to make sure that padding of inputs for fft is added
# and then cropped out correctly (tests asymmetries and non-power-of-2 dimensions)
acqs += [
tfv.reblur(
tfv.subset(acq,
kern_slice=[
slice(None, 24),
slice(None, 48),
slice(None, 48)
])),
]
# Validate that downsampling the original volume also causes no issues
acqs += [
tfv.reblur(tfv.downsample(acq, data_factor=.8, kern_factor=.4)),
tfv.reblur(tfv.downsample(acq, data_factor=.5, kern_factor=.5))
]
self._validate_restoration(acqs, n_iter, thresholds=thresholds)
def test_bars(self):
"""Test reconstruction of blurred "Hollow Bars" volume"""
n_iter = 10
thresholds = {
# Results should always improve on the original
'original': 1.0,
'dl2': .75 if dl2_enabled() else None
}
acq = fd_data.bars_25pct()
self.assertTrue(acq.data.shape == acq.kernel.shape == (32, 64, 64))
# shape: {'actual': (32, 64, 64), 'kernel': (32, 64, 64), 'data': (32, 64, 64)}
# Initialize list of acquisitions to deconvolve
acqs = [acq]
# Slice to give shape with odd dimensions as this is a better test of padding features
acq_slice = [slice(2, 29), slice(None, 57), slice(None, 57)]
acq = tfv.subset(acq, data_slice=acq_slice, kern_slice=acq_slice)
# shape: {'data': (27, 57, 57), 'actual': (27, 57, 57), 'kernel': (27, 57, 57)}
acqs += [acq]
# Add translations to images and kernels to ensure that there aren't
# any issues supporting non-symmetric inputs
acqs += [
tfv.reblur(tfv.shift(acq, data_shift=(0, 4, 4))),
tfv.reblur(tfv.shift(acq, data_shift=(0, -4, -4))),
tfv.reblur(tfv.shift(acq, kern_shift=(0, 4, 4))),
tfv.reblur(tfv.shift(acq, kern_shift=(0, -4, -4))),
tfv.reblur(
tfv.shift(acq, data_shift=(-3, 5, -5), kern_shift=(-3, 5, -5)))
]
# Validate that downsampling the original volume also causes no issues
acqs += [
tfv.reblur(tfv.downsample(acq, data_factor=.8, kern_factor=.4)),
tfv.reblur(tfv.downsample(acq, data_factor=.5, kern_factor=.5))
]
self._validate_restoration(acqs,
n_iter,
thresholds=thresholds,
dtype=np.uint16)
def load_simulated_bars_experiment(blur=False):
ref_img = fd_data.bars_25pct().data if blur else fd_data.bars_25pct().actual
# Subset image to not be equal in x and y
ref_img = ref_img[:, :48, :]
return experiment_from_img(ref_img)
