pytest general test speed-up (#211)

* initial commit, restoration re-denoising test removed

* removed the commented out lines

* regression subpackage tests are improved

* using smaller input data where possible

* black fixes

* to keep changes

* aydin.it.test_fgr.py addressed

* black fix

aydin/it/classic_denoisers/test/test_bilateral.py

@@ -7,7 +7,7 @@
 
 def est_bilateral():
     # TODO: bilateral filter in scikit image seems broken
-    assert demo_bilateral(cropped_newyork(), display=False) >= 0.40 - 0.1
+    assert demo_bilateral(cropped_newyork(crop_amount=384), display=False) >= 0.40 - 0.1
 
 
 def test_bilateral_nd():

aydin/it/classic_denoisers/test/test_butterworth.py

@@ -8,7 +8,9 @@
 
 
 def test_butterworth():
-    ssim_denoised, parameters = demo_butterworth(cropped_newyork(), display=False)
+    ssim_denoised, parameters = demo_butterworth(
+        cropped_newyork(crop_amount=384), display=False
+    )
     assert ssim_denoised >= 0.608 - 0.035
 
 

aydin/it/classic_denoisers/test/test_gaussian.py

@@ -6,7 +6,9 @@
 
 
 def test_gaussian():
-    assert demo_gaussian(cropped_newyork(), display=False) >= 0.600 - 0.02
+    assert (
+        demo_gaussian(cropped_newyork(crop_amount=384), display=False) >= 0.600 - 0.02
+    )
 
 
 def test_gaussian_nd():

aydin/it/classic_denoisers/test/test_gm.py

@@ -6,7 +6,7 @@
 
 
 def test_gm():
-    assert demo_gm(cropped_newyork(), display=False) >= 0.610 - 0.03
+    assert demo_gm(cropped_newyork(crop_amount=384), display=False) >= 0.610 - 0.03
 
 
 def test_gm_nd():

aydin/it/classic_denoisers/test/test_lipschitz.py

@@ -7,7 +7,9 @@
 
 
 def test_lipschitz():
-    assert demo_lipschitz(cropped_newyork(), display=False) >= 0.50 - 0.02
+    assert (
+        demo_lipschitz(cropped_newyork(crop_amount=384), display=False) >= 0.50 - 0.02
+    )
 
 
 def test_lipschitz_nd():

aydin/it/test/test_fgr.py

@@ -20,54 +20,61 @@
 @pytest.mark.heavy
 def test_it_fgr_linear():
     regressor = LinearRegressor()
-    do_it_fgr_with_self_supervised(regressor, min_psnr=18, min_ssim=0.65)
+    do_it_fgr(regressor, min_psnr=18, min_ssim=0.65)
 
 
 def test_it_fgr_rf():
     regressor = RandomForestRegressor(max_num_estimators=128)
-    do_it_fgr_with_self_supervised(regressor, min_ssim=0.75)
+    do_it_fgr(regressor, min_ssim=0.75)
 
 
 @pytest.mark.heavy
 def test_it_fgr_svr():
     regressor = SupportVectorRegressor()
-    do_it_fgr_with_self_supervised(regressor, min_psnr=22, min_ssim=0.71)
+    do_it_fgr(regressor, min_psnr=22, min_ssim=0.71)
 
 
 def test_it_fgr_lgbm():
     regressor = LGBMRegressor(max_num_estimators=256)
-    do_it_fgr_with_self_supervised(regressor, min_ssim=0.77)
+    do_it_fgr(regressor, min_ssim=0.77)
 
 
 @pytest.mark.heavy
 def test_it_fgr_nn():
     regressor = PerceptronRegressor(max_epochs=64)
-    do_it_fgr_with_self_supervised(regressor, min_ssim=0.70)
+    do_it_fgr(regressor, min_ssim=0.70)
 
 
+@pytest.mark.heavy  # we skip this as we run same for saveload
 def test_it_fgr_cb():
     regressor = CBRegressor(max_num_estimators=256, min_num_estimators=64)
-    do_it_fgr_with_self_supervised(regressor, min_ssim=0.779)
+    do_it_fgr(regressor, min_ssim=0.779)
 
 
 def test_it_fgr_cb_supervised():
     regressor = CBRegressor(max_num_estimators=256, min_num_estimators=64)
-    do_it_fgr_with_supervised(regressor, min_ssim=0.77)
+    do_it_fgr(regressor, min_ssim=0.77, supervised=True)
 
 
-def do_it_fgr_with_self_supervised(regressor, min_psnr=22, min_ssim=0.75):
+def do_it_fgr(regressor, min_psnr=22, min_ssim=0.75, supervised=False):
     """
     Test for self-supervised denoising using camera image with synthetic noise
     """
 
     image = normalise(camera().astype(numpy.float32))
     noisy = add_noise(image)
 
-    generator = StandardFeatureGenerator(max_level=5, include_spatial_features=True)
+    if supervised:
+        generator = StandardFeatureGenerator(include_spatial_features=False)
+    else:
+        generator = StandardFeatureGenerator(max_level=5, include_spatial_features=True)
     it = ImageTranslatorFGR(feature_generator=generator, regressor=regressor)
 
     start = time.time()
-    it.train(noisy, noisy)
+    if supervised:
+        it.train(noisy, image)
+    else:
+        it.train(noisy, noisy)
     stop = time.time()
     print(f"####### Training: elapsed time:  {stop - start} sec")
 
@@ -94,45 +101,3 @@ def do_it_fgr_with_self_supervised(regressor, min_psnr=22, min_ssim=0.75):
     # do not change the number below, but instead, fix the problem -- most likely a parameter.
 
     assert psnr_denoised > min_psnr and ssim_denoised > min_ssim
-
-
-def do_it_fgr_with_supervised(regressor, min_psnr=23, min_ssim=0.74):
-    """
-    Test for self-supervised denoising using camera image with synthetic noise
-    """
-
-    image = normalise(camera().astype(numpy.float32))
-    noisy = add_noise(image)
-
-    generator = StandardFeatureGenerator(include_spatial_features=False)
-    it = ImageTranslatorFGR(feature_generator=generator, regressor=regressor)
-
-    start = time.time()
-    it.train(noisy, image)
-    stop = time.time()
-    print(f"####### Training: elapsed time:  {stop - start} sec")
-
-    start = time.time()
-    denoised = it.translate(noisy)
-    stop = time.time()
-    print(f"####### Inference: elapsed time:  {stop - start} sec")
-
-    image = numpy.clip(image, 0, 1)
-    noisy = numpy.clip(noisy, 0, 1)
-    denoised = numpy.clip(denoised, 0, 1)
-
-    psnr_noisy = psnr(noisy, image)
-    ssim_noisy = ssim(noisy, image)
-    print("noisy", psnr_noisy, ssim_noisy)
-
-    psnr_denoised = psnr(denoised, image)
-    ssim_denoised = ssim(denoised, image)
-    print("denoised", psnr_denoised, ssim_denoised)
-
-    assert psnr_denoised > psnr_noisy and ssim_denoised > ssim_noisy
-    assert psnr_denoised > psnr_noisy and ssim_denoised > ssim_noisy
-
-    # if the line below fails, then the parameters of the image the regressor have been broken.
-    # do not change the number below, but instead, fix the problem -- most likely a parameter.
-
-    assert psnr_denoised > min_psnr and ssim_denoised > min_ssim

aydin/regression/test/test_regressors.py

@@ -1,6 +1,7 @@
 import time
 from os.path import join
 import numpy
+import pytest
 from skimage.data import camera
 from skimage.metrics import peak_signal_noise_ratio as psnr
 from skimage.metrics import structural_similarity as ssim
@@ -17,51 +18,51 @@
 from aydin.regression.support_vector import SupportVectorRegressor
 
 
-def test_linear_regressor():
-    regressor = LinearRegressor()
-    with_regressor(regressor, min_ssim=0.6)
+@pytest.fixture(scope="session")
+def data():
+    image = camera()[:256, :256].astype(numpy.float32)
+    image = normalise(image)
+    noisy = add_noise(image)
 
+    # feature generator requires images in 'standard' form: BCTZYX
+    noisy = noisy[numpy.newaxis, numpy.newaxis, ...]
 
-def test_rf_regressor():
-    regressor = RandomForestRegressor()
-    with_regressor(regressor, min_ssim=0.6)
+    generator = StandardFeatureGenerator(max_level=6)
 
+    features = generator.compute(noisy, exclude_center_value=True)
 
-def test_svr_regressor():
-    regressor = SupportVectorRegressor()
-    with_regressor(regressor, min_ssim=0.65)
+    x = features.reshape(-1, features.shape[-1])
+    y = noisy.reshape(-1)
 
+    return image, noisy, x, y
 
-def test_lgbm_regressor():
-    regressor = LGBMRegressor(max_num_estimators=600)
-    with_regressor(regressor, min_ssim=0.75)
 
+def test_linear_regressor(data):
+    with_regressor(data, LinearRegressor(), min_ssim=0.6)
 
-def test_cb_regressor():
-    regressor = CBRegressor(max_num_estimators=600)
-    with_regressor(regressor, min_ssim=0.75)
 
+def test_rf_regressor(data):
+    with_regressor(data, RandomForestRegressor(), min_ssim=0.6)
 
-def test_nn_regressor():
-    regressor = PerceptronRegressor(max_epochs=6, depth=6)
-    with_regressor(regressor, min_ssim=0.64)
 
+def test_svr_regressor(data):
+    with_regressor(data, SupportVectorRegressor(), min_ssim=0.65)
 
-def with_regressor(regressor, min_ssim=0.8):
 
-    image = camera().astype(numpy.float32)
-    image = normalise(image)
-    noisy = add_noise(image)
+def test_lgbm_regressor(data):
+    with_regressor(data, LGBMRegressor(max_num_estimators=600), min_ssim=0.75)
 
-    # feature generator requires images in 'standard' form: BCTZYX
-    noisy = noisy[numpy.newaxis, numpy.newaxis, ...]
 
-    generator = StandardFeatureGenerator(max_level=6)
+def test_cb_regressor(data):
+    with_regressor(data, CBRegressor(max_num_estimators=600), min_ssim=0.75)
 
-    features = generator.compute(noisy, exclude_center_value=True)
 
-    x = features.reshape(-1, features.shape[-1])
-    y = noisy.reshape(-1)
+def test_nn_regressor(data):
+    with_regressor(data, PerceptronRegressor(max_epochs=6, depth=6), min_ssim=0.64)
+
+
+def with_regressor(data, regressor, min_ssim=0.8):
+    image, noisy, x, y = data
 
     regressor.fit(x, y)
 

aydin/restoration/denoise/test/test_n2s_fgr.py

@@ -1,14 +1,5 @@
 from pprint import pprint
 
-import numpy as np
-from skimage.data import camera
-from skimage.exposure import rescale_intensity
-from skimage.metrics import peak_signal_noise_ratio as psnr
-from skimage.metrics import structural_similarity as ssim
-
-from aydin.io.datasets import add_noise
-from aydin.it.transforms.padding import PaddingTransform
-from aydin.it.transforms.range import RangeTransform
 from aydin.restoration.denoise.noise2selffgr import Noise2SelfFGR
 
 
@@ -22,23 +13,3 @@ def test_configure():
 
     implementations_description = Noise2SelfFGR().implementations_description
     pprint(implementations_description)
-
-
-def test_run_n2s_fgr():
-    # Prepare the noisy classic_denoisers camera image
-    image = camera().astype(np.float32)
-    image = rescale_intensity(image, in_range='image', out_range=(0, 1))
-    noisy_image = add_noise(image)
-
-    # Call the Noise2Self restoration
-    transforms = [
-        {"class": RangeTransform, "kwargs": {}},
-        {"class": PaddingTransform, "kwargs": {}},
-    ]
-    n2s = Noise2SelfFGR(variant="cb", it_transforms=transforms)
-    n2s.train(noisy_image)
-    denoised_image = n2s.denoise(noisy_image).clip(0, 1)
-
-    # Check if denoised image satisfies some checks
-    assert psnr(denoised_image, image) >= 20.0
-    assert ssim(denoised_image, image) >= 0.7

aydin/restoration/denoise/test/test_saveload.py

@@ -9,15 +9,31 @@
 from aydin import Classic
 from aydin.io.datasets import normalise, add_noise
 from aydin.io.folders import get_temp_folder
+from aydin.it.transforms.padding import PaddingTransform
+from aydin.it.transforms.range import RangeTransform
 from aydin.restoration.denoise.noise2selffgr import Noise2SelfFGR
 
 
+transforms = [
+    {"class": RangeTransform, "kwargs": {}},
+    {"class": PaddingTransform, "kwargs": {}},
+]
+
+
 def test_saveload_classic_gaussian():
-    saveload(Classic(variant="gaussian"), min_psnr=19, min_ssim=0.61)
+    saveload(
+        Classic(variant="gaussian", it_transforms=transforms),
+        min_psnr=19,
+        min_ssim=0.61,
+    )
 
 
 def test_saveload_noise2selffgr():
-    saveload(Noise2SelfFGR(), min_psnr=20, min_ssim=0.78)
+    saveload(
+        Noise2SelfFGR(variant="cb", it_transforms=transforms),
+        min_psnr=20,
+        min_ssim=0.78,
+    )
 
 
 def saveload(denoiser, min_psnr=22, min_ssim=0.75):
@@ -48,6 +64,10 @@ def saveload(denoiser, min_psnr=22, min_ssim=0.75):
     ssim_denoised = ssim(denoised, image)
     print("denoised", psnr_denoised, ssim_denoised)
 
+    # Check if denoised image satisfies some checks
+    assert psnr_denoised >= 20.0
+    assert ssim_denoised >= 0.7
+
     assert psnr_denoised > psnr_noisy and ssim_denoised > ssim_noisy
     assert psnr_denoised > psnr_noisy and ssim_denoised > ssim_noisy