def test_normalize_minmse():
rng = np.random.RandomState(42)
for _ in range(50):
target = rng.uniform(-100,100,size=(32,32,32))
x = rng.uniform(-500,500)*target + rng.uniform(-500,500)
assert np.allclose(normalize_minmse(x,target),target)
x, target = x.astype(np.float32), target.astype(np.float32)
assert np.max(np.abs(normalize_minmse(x,target)-target)) < 1e-3
def _run_multi(y, x, models, d, ix, args):
""" Trying an alternative normalization to see if that's causing the
incong. with the paper results
"""
# Prep data
axes = 'ZYX'
def get_output(name, y, x):
# Define comparisons
return [name, np.sqrt(mse(x, y)), ssim(x, y)]
# Normalize GT dynamic range to enable comparisons w/ numbers
yn = normalize(y, pmin=0.1, pmax=99.9)
# Get the comparison for normalized input im
if args.rescale_x:
xn = normalize_minmse(x, yn)
else:
xn = normalize(x)
d['output'].append(get_output('input', yn, xn))
for m in models:
model = CARE(config=None, name=m, basedir='models')
pred = model.predict(xn, axes, n_tiles=(1, 4, 4), normalizer=None)
pred = normalize_minmse(pred, yn)
d['output'].append(get_output(m, yn, pred))
# Save the first test volume for each model
if args.save_predictions:
if ix == 0:
save_prediction(pred, m, d['condition'])
# Save the first test volume input and GT
if args.save_predictions:
if ix == 0:
save_prediction(xn, 'input', d['condition'])
save_prediction(yn, None, d['condition'])
return d
def _run_input(y, x, models, d, ix, args):
""" Trying an alternative normalization to see if that's causing the
incong. with the paper results
"""
# Prep data
axes = 'ZYX'
def get_output(name, y, x):
# Define comparisons
return [name, np.sqrt(mse(x, y)), ssim(x, y)]
# Normalize GT dynamic range to enable comparisons w/ numbers
yn = normalize(y, pmin=0.1, pmax=99.9)
# Get the comparison for normalized input im
if args.rescale_x:
xn = normalize_minmse(x, yn)
else:
xn = normalize(x)
d['output'].append(get_output('input', yn, xn))
return d
def _run_multi(y, x, models, d):
# Prep data
axes = 'ZYX'
# Define comparisons
def get_output(name, y, x):
return [name, np.sqrt(mse(x, y)), ssim(x, y)]
# Normalize GT dynamic range to enable comparisons w/ numbers
yn = normalize(y, pmin=0.1, pmax=99.9)
# Get the comparison for normalized input im
xn = normalize(x)
d['output'].append(get_output('input', yn, xn))
for m in models:
model = CARE(config=None, name=m, basedir='models')
# None normalizer, already normalizing x, and this way can report the
# exact pmin/max params used
pred = model.predict(xn, n_tiles=(1, 4, 4), axes=axes, normalizer=None)
pred = normalize_minmse(pred, yn)
d['output'].append(get_output(m, yn, pred))
return d