Merge pull request #14 from dimitri-yatsenko/main

refactored for better package design

src/poisson_numcodecs/__init__.py

@@ -1,4 +1,3 @@
-from .poisson import Poisson
-
-__all__ = ["Poisson"]
+from .codec import PoissonCodec
 
+__all__ = ["PoissonCodec"]

src/poisson_numcodecs/codec.py

@@ -0,0 +1,98 @@
+"""
+Numcodecs Codec implementation for Poisson noise calibration
+"""
+import numpy as np
+import numcodecs
+from numcodecs.abc import Codec
+
+def make_anscombe_lookup(
+          sensitivity: float, 
+          input_max: int=0x7fff,
+          zero_level: int=0,
+          beta: float=0.5,
+          output_type='uint8'):
+    """
+    Compute the Anscombe lookup table.
+    The lookup converts a linear grayscale image into a uniform variance image. 
+    :param sensitivity: the size of one photon in the linear input image.
+    :param input_max: the maximum value in the input
+    :param beta: the grayscale quantization step expressed in units of noise std dev
+    """
+    xx = (np.r_[:input_max + 1] - zero_level) / sensitivity  # input expressed in photon rates
+    zero_slope = 1 / beta / np.sqrt(3/8)  # slope for negative values
+    offset = zero_level * zero_slope / sensitivity
+    lookup_table = np.round(offset +
+          (xx < 0) * (xx * zero_slope) +
+          (xx >= 0) * (2.0 / beta * (np.sqrt(np.maximum(0, xx) + 3/8) - np.sqrt(3/8))))
+    lookup = lookup_table.astype(output_type)
+    assert np.diff(lookup_table).min() >= 0, "non-monotonic lookup generated"
+    return lookup
+
+
+def make_inverse_lookup(lookup_table: np.ndarray, output_type='int16') -> np.ndarray:
+    """Compute the inverse lookup table for a monotonic forward lookup table."""
+    _, inv1 = np.unique(lookup_table, return_index=True)  # first entry
+    _, inv2 = np.unique(lookup_table[::-1], return_index=True)  # last entry
+    inverse = (inv1 + lookup_table.size - 1 - inv2)/2
+    return inverse.astype(output_type)
+
+
+def lookup(movie: np.ndarray, lookup_table: np.ndarray) -> np.ndarray:
+    """Apply lookup table to movie"""
+    return lookup_table[np.maximum(0, np.minimum(movie, lookup_table.size-1))]
+
+
+class PoissonCodec(Codec):
+    """Codec for 3-dimensional Filter. The codec assumes that input data are of shape:
+    (time, x, y).
+
+    Parameters
+    ----------
+    zero_level : float
+        Signal level when no photons are recorded. 
+        This should pre-computed or measured directly on the instrument.
+    photon_sensitivity : float
+        Conversion scalor to convert the measure signal into absolute photon numbers.
+        This should pre-computed or measured directly on the instrument.
+    """
+    codec_id = "poisson"
+
+    def __init__(self, 
+                 zero_level, 
+                 photon_sensitivity, 
+                 encoded_dtype='int8', 
+                 decoded_dtype='int16',
+                 ): 
+        self.zero_level = zero_level
+        self.photon_sensitivity = photon_sensitivity
+        self.encoded_dtype = encoded_dtype
+        self.decoded_dtype = decoded_dtype
+
+    def encode(self, buf: np.ndarray) -> np.ndarray:
+        lookup_table = make_anscombe_lookup(
+            self.photon_sensitivity, 
+            output_type=self.encoded_dtype,
+            zero_level=self.zero_level,
+        )
+        encoded = lookup(buf, lookup_table)
+        shape = [encoded.ndim] + list(encoded.shape)
+        shape = np.array(shape, dtype='uint8')
+        return shape.tobytes() + encoded.astype(self.encoded_dtype).tobytes()
+
+    def decode(self, buf: bytes, out=None) -> np.ndarray:
+        lookup_table = make_anscombe_lookup(
+            self.photon_sensitivity, 
+            output_type=self.encoded_dtype,
+            zero_level=self.zero_level,
+        )
+        inverse_table = make_inverse_lookup(
+            lookup_table,
+            output_type=self.decoded_dtype
+        )
+        ndims = int(buf[0])
+        shape = [int(_) for _ in buf[1:ndims+1]]
+        decoded = np.frombuffer(buf[ndims+1:], dtype=self.encoded_dtype).reshape(shape)
+        return lookup(decoded, inverse_table).astype(self.decoded_dtype)
+
+
+numcodecs.register_codec(PoissonCodec)

src/poisson_numcodecs/estimate.py

@@ -1,6 +1,5 @@
 import numpy as np
 from sklearn.linear_model import HuberRegressor as Regressor
-import imageio as io
 
 
 def _longest_run(bool_array: np.ndarray) -> slice:
@@ -22,7 +21,7 @@ def compute_sensitivity(movie: np.array, count_weight_gamma: float=0.2) -> dict:
     """Calculate photon sensitivity
 
     Args:
-        movie (np.array):  A movie in the format (height, width, time).
+        movie (np.array):  A movie in the format (time, height, width).
         count_weight_gamma: 0.00001=weigh each intensity level equally, 
             1.0=weigh each intensity in proportion to pixel counts.
 
@@ -40,10 +39,10 @@ def compute_sensitivity(movie: np.array, count_weight_gamma: float=0.2) -> dict:
     ), f"A three dimensional (Height, Width, Time) grayscale movie is expected, got {movie.ndim}"
 
     movie = np.maximum(0, movie.astype(np.int32, copy=False))
-    intensity = (movie[:, :, :-1] + movie[:, :, 1:] + 1) // 2
-    difference = movie[:, :, :-1].astype(np.float32) - movie[:, :, 1:]
+    intensity = (movie[:-1, :, :] + movie[1:, :, :] + 1) // 2
+    difference = movie[:-1, :, :].astype(np.float32) - movie[1:, :, :]
 
-    select = intensity > 0
+    select = intensity > 0   # discard non-positive values
     intensity = intensity[select]
     difference = difference[select]
 
@@ -77,39 +76,3 @@ def compute_sensitivity(movie: np.array, count_weight_gamma: float=0.2) -> dict:
         sensitivity=sensitivity,
         zero_level=zero_level,
     )
-
-def make_anscombe_lookup(sensitivity: float, input_max: int=0x7fff, beta: float=0.5):
-	"""
-	Compute the Anscombe lookup table.
-	The lookup converts a linear grayscale image into a uniform variance image. 
-	:param sensitivity: the size of one photon in the linear input image.
-    :param input_max: the maximum value in the input
-	:param beta: the grayscale quantization step expressed in units of noise std dev
-	"""
-	# produce anscombe lookup_table
-	xx = np.r_[:input_max + 1] / sensitivity
-	lookup = np.uint8(
-        2.0 / beta * (np.sqrt(np.maximum(0, xx) + 3/8) - np.sqrt(3/8)))
-	return lookup
-
-
-def make_inverse_lookup(lookup):
-    _, inverse = np.unique(lookup, return_index=True)
-    inverse += (np.r_[:inverse.size] / inverse.size * (inverse[-1] - inverse[-2])/2).astype(np.int16)
-    return inverse 
-
-
-def lookup(movie, LUT):
-    """
-    Apply lookup table LUT to input movie
-    """
-    return LUT[np.maximum(0, np.minimum(movie, LUT.size-1))]
-
-
-def save_movie(movie, path, scale=1, format='gif'):
-    if format == "gif":
-        with io.get_writer(path, mode='I', duration=.01, loop=False) as f:
-            for frame in movie:
-                f.append_data(scale * frame)
-    else:
-        raise NotImplementedError(f"Format {format} is not implemented")

tests/test_poisson_codec.py

@@ -1,4 +1,4 @@
-from poisson_numcodecs import Poisson
+from poisson_numcodecs import PoissonCodec
 import numpy as np
 import pytest
 
@@ -11,28 +11,30 @@ def make_poisson_ramp_signals(shape=(10, 1, 1), min_rate=1, max_rate=5, dtype="i
     output_array = np.zeros(shape, dtype=dtype)
     for x_ind in range(x):
         for y_ind in range(y):
-            output_array[x_ind, y_ind, :] = np.random.poisson(
+            output_array[x_ind, y_ind, :] = sensitivity * np.random.poisson(
                 np.arange(min_rate, max_rate, (max_rate-min_rate)/times))
-    return output_array
+    return output_array.astype('int16')
+
+sensitivity = 100.0
 
 @pytest.fixture
 def test_data(dtype="int16"):
-    test2d = make_poisson_ramp_signals(shape=(50, 1, 1), min_rate=1, max_rate=5, dtype=dtype)
+    test2d =  make_poisson_ramp_signals(shape=(50, 1, 1), min_rate=1, max_rate=5, dtype=dtype)
     test2d_long = make_poisson_ramp_signals(shape=(1, 50, 1), min_rate=1, max_rate=5, dtype=dtype)
     return [test2d, test2d_long]
 
 def test_poisson_encode_decode(test_data):
-    poisson_codec = Poisson(
+    poisson_codec = PoissonCodec(
         zero_level=0,
-        photon_sensitivity=1.0,
+        photon_sensitivity=sensitivity,
         encoded_dtype='uint8', 
         decoded_dtype='int16'
     )
 
     for example_data in test_data:
         encoded = poisson_codec.encode(example_data)
         decoded = poisson_codec.decode(encoded)
-        np.testing.assert_allclose(decoded, example_data, atol=1e-3)
+        assert np.abs(decoded - example_data).max() < sensitivity / 2
 
 if __name__ == '__main__':
     list_data = test_data("int16")