FIX: Resolved characteristic wavelength overestimation via re-scaling

examples/example_1/config

@@ -5,5 +5,5 @@ method = response_distance
 [response_distance]
 shapelet_order = default
 num_clusters = 20
-ux = [50, 80]
-uy = [150, 180]
+ux = [109, 158]
+uy = [283, 322]

examples/example_1/example_1.py

@@ -29,8 +29,8 @@
 image_name = "lamSIM1.png"
 shapelet_order = 'default' # can also be integer value to set upper bound
 num_clusters = 20 # default is 20, can be any other positive integer 
-ux = [50, 80]
-uy = [150, 180]
+ux = [109, 158]
+uy = [283, 322]
 
 ## Section 3: code
 

examples/example_3/example_3.py

@@ -33,7 +33,7 @@
 
 # 3.1: image and output directory handling
 image_path = os.path.join(Path(__file__).parents[0], 'images')
-image = read_image(image_name = image_name, image_path = image_path)
+image = read_image(image_name = image_name, image_path = image_path, do_rescale=False)
 save_path = os.path.join(Path(__file__).parents[0], 'output')
 if not os.path.exists(save_path): 
     os.mkdir(save_path)

shapelets/self_assembly/misc.py

@@ -32,7 +32,7 @@
     'trim_image'
 ]
 
-def read_image(image_name: str, image_path: str, verbose: bool = True) -> np.ndarray:
+def read_image(image_name: str, image_path: str, do_rescale: bool = True, verbose: bool = True) -> np.ndarray:
     r""" 
     Read an image using OpenCV, with some extra handling. By default, re-scales images as greyscale on [-1, 1].
 
@@ -42,6 +42,8 @@ def read_image(image_name: str, image_path: str, verbose: bool = True) -> np.nda
         * The filename of the image (including extension)
     * image_path: str
         * The path holding the image
+    * do_rescale : bool, optional
+        * Automatically re-scale in accordance with requirements for wavelength algorithm. Default is True
     * verbose: bool, optional
         * True (default) to print image-related information
 
@@ -55,33 +57,34 @@ def read_image(image_name: str, image_path: str, verbose: bool = True) -> np.nda
     Re-scaling of image to greyscale on [-1, 1] is intentional to align with the minimum and maximum of shapelet function values.
 
     """
-    # Ensure image_path provided exists
     if not os.path.exists(image_path):
         raise RuntimeError(f'Image path: {image_path} does not exist.')
 
-    # read image then ensure image does exist (by evaluating result of cv2.imread)
     f = cv2.imread(os.path.join(image_path, image_name))
     if not isinstance(f, np.ndarray):
         raise RuntimeError(f"Could not read image: {image_name}. Ensure it is located in {image_path} and is of image format.")
 
-    # convert to grayscale
     f = cv2.cvtColor(f, cv2.COLOR_BGR2GRAY)
 
-    # check if re-scaling is needed b/c k-means clustering cannot handle very large images
-    init_shape = f.shape
-    resized = False 
-    while any(dim >= 1000 for dim in f.shape):
-        dim0, dim1 = round(f.shape[0]*0.8), round(f.shape[1]*0.8)
-        f = cv2.resize(f, dsize=(dim0, dim1))
-        resized = True
+    # NOTE: image may require rescaling based on two observed issues:
+    #   1. scipy.cluster.vq.kmeans yields abort: coredump if one dim of image large (~2000 pixels)
+    #   2. wavelength algorithm fails if image is too small... need to upsample
 
-    # rescale to [-1, 1] greyscale 
+    if do_rescale:
+        while any(dim >= 2000 for dim in f.shape):
+            height, width = round(f.shape[0]*0.8), round(f.shape[1]*0.8)
+            f = cv2.resize(f, dsize=(width, height)) # dsize is opposite numpy convention
+
+        while any(dim <= 500 for dim in f.shape):
+            height, width = round(f.shape[0]*1.2), round(f.shape[1]*1.2)
+            f = cv2.resize(f, dsize=(width, height)) # dsize is opposite numpy convention
+
+    # rescale to [-1, 1] greyscale to match shapelet kernel min/max values
     f = ( ((f-f.min()) / (f.max()-f.min())) * 2 ) - 1
 
     if verbose: 
         print(f"Successfully loaded image: {image_name}")
-        if resized: print(f"New image dimensions are: {f.shape} as initial size {init_shape} too large")
-        else: print(f"Image dimensions are: {f.shape}")
+        print(f"Image loaded with dimensions: {f.shape}")
         print(f"Image {image_name} normalized to greyscale on [-1, 1] to align with shapelet kernels")
 
     return f

shapelets/self_assembly/wavelength.py

@@ -25,7 +25,7 @@
 ]
 
 
-def get_wavelength(image: np.ndarray, rng: list = [0, 70], verbose: bool = True) -> float:
+def get_wavelength(image: np.ndarray, rng: list = [0, 100], verbose: bool = True) -> float:
     r""" 
     Find characteristic wavelength of an image. Computed using method described in ref. [1].
 

shapelets/tests/test_self_assembly_basic.py

@@ -49,7 +49,7 @@ def test_a_read_image(self) -> None:
         self.assertTrue(isinstance(self.lamSIM, np.ndarray))
         self.assertEqual(self.lamSIM.min(), -1)
         self.assertEqual(self.lamSIM.max(), 1)
-        self.assertEqual(self.lamSIM.shape, (296, 296))
+        self.assertEqual(self.lamSIM.shape, (511, 511))
 
         self.assertTrue(isinstance(self.hexSIM, np.ndarray))
         self.assertEqual(self.hexSIM.min(), -1)
@@ -63,27 +63,27 @@ def test_scaling(self) -> None:
 
         lamSIM_wvl = get_wavelength(image = self.lamSIM, verbose = False)
         self.assertTrue(isinstance(lamSIM_wvl, numbers.Real))
-        self.assertAlmostEqual(lamSIM_wvl, 10.231, places = 3)
+        self.assertAlmostEqual(lamSIM_wvl, 17.60, places = 2)
 
         lamSIM_beta_n0 = lambda_to_beta_n0(3, lamSIM_wvl)
         self.assertTrue(isinstance(lamSIM_beta_n0, numbers.Real))
-        self.assertAlmostEqual(lamSIM_beta_n0, 3.41, places = 2)
+        self.assertAlmostEqual(lamSIM_beta_n0, 5.87, places = 2)
 
         lamSIM_beta_n1 = lambda_to_beta_n1(3, lamSIM_wvl)
         self.assertTrue(isinstance(lamSIM_beta_n1, numbers.Real))
-        self.assertAlmostEqual(lamSIM_beta_n1, 7.3, places = 5)
+        self.assertAlmostEqual(lamSIM_beta_n1, 12.4, places = 1)
 
         hexSIM_wvl = get_wavelength(image = self.hexSIM, verbose = False)
         self.assertTrue(isinstance(hexSIM_wvl, numbers.Real))
-        self.assertAlmostEqual(hexSIM_wvl, 16.882, places = 3)
+        self.assertAlmostEqual(hexSIM_wvl, 16.88, places = 2)
 
         hexSIM_beta_n0 = lambda_to_beta_n0(6, hexSIM_wvl)
         self.assertTrue(isinstance(hexSIM_beta_n0, numbers.Real))
         self.assertAlmostEqual(hexSIM_beta_n0, 6.89, places = 2)
 
         hexSIM_beta_n1 = lambda_to_beta_n1(6, hexSIM_wvl)
         self.assertTrue(isinstance(hexSIM_beta_n1, numbers.Real))
-        self.assertAlmostEqual(hexSIM_beta_n1, 9.1, places = 5)
+        self.assertAlmostEqual(hexSIM_beta_n1, 9.1, places = 1)
 
 if __name__ == "__main__":
     unittest.main()

shapelets/tests/test_self_assembly_methods.py

@@ -41,9 +41,10 @@ def setUpClass(cls) -> None:
         cls.dir = __file__.replace(os.path.basename(__file__), 'images/')
 
         # Ensure core functions have appropriate outputs before proceeding.
-        # Note that read_image and get_wavelength are not tested here, so inline asserts
-        #   are used to ensure correct output.
+        # NOTE: read_image & get_wavelength are not tested here, so inline asserts are used to ensure correct output.
+
         cls.image = read_image(image_name="hexSIM1.png", image_path=cls.dir, verbose=False)
+
         assert isinstance(cls.image, np.ndarray)
 
         cls.omega, cls.phi = convresponse_n0(cls.image, shapelet_order='default', verbose=False)