Yh cyber windowed blending sampler

.gitignore

@@ -12,6 +12,7 @@ dist/
 # tool caches
 .pytest_cache/
 .ruff_cache/
+.coverage
 
 node_modules/
 

pyproject.toml

@@ -22,6 +22,7 @@ terrain-diffusion = "terrain_diffusion.cli:main"
 dev = [
     "pytest>=8.3",
     "ruff>=0.9",
+    "pytest-mock>=3.15.1",
 ]
 
 
@@ -81,4 +82,4 @@ markers = [
 [tool.coverage.run]
 omit = [
     # add omissions here
-]
+]

src/terrain_diffusion/sampler.py

@@ -10,3 +10,112 @@
 - Asks the Model Pipeline to clean patches.
 - Writes results into the Terrain Store and reads them back.
 """
+
+# What value does the window sampler take from generation orchestration?
+# Seed, Region Size, and Region coordinates
+
+# What does it expect from the model inference?
+# It gives a noisy patch to the model and gets a processed full resolution one back
+
+# What does the sampler generate?
+# The sampler generates noise, overlapping windows over region, processes each one through model pipleine, then stores in terrain cache. Terrain Store makes and and returns heightgrid (2D numpy array).
+
+# What do we store in the terrain cache?
+# It holds weight grids for tiles being generated and finished (Sum grid: running total of value × weight and Weight grid: running total of weight)
+
+import numpy as np
+
+
+def window_positions(
+    region_height: int, region_width: int, window_size: int, step: int
+) -> list[tuple[int, int]]:
+    """Takes a region height and width, a window size, and a step size, and returns the list of top left positions to place windows at.
+    The step is smaller than the window, which is what makes them overlap.
+    If window_size does not divide evenly, throws an assertion error."""
+
+    assert (region_height - window_size) % step == 0
+    assert (region_width - window_size) % step == 0
+    assert step <= window_size
+
+    # Row, Column Position
+    rows = np.arange(0, region_height - window_size + 1, step)
+    columns = np.arange(0, region_width - window_size + 1, step)
+
+    row_grid, column_grid = np.meshgrid(
+        rows, columns, indexing="ij"
+    )  # gives 2-D matrix array: [0, 2, 4] -> [[0, 0, 0], [2, 2, 2]...]; (every row/column repeated across column/row)
+    positions = np.stack(
+        (row_grid.ravel(), column_grid.ravel()), axis=1
+    )  # Ravel flattens array; Stacks in (Row, Column)
+
+    return positions
+
+
+def weight_grid(edge_len: int) -> np.ndarray:
+    """Create a function that returns a grid of weights the size of a patch, since the weights get applied to what is written into the store.
+    Weights should be largest in the middle and get smaller toward the edges. Every weight must be greater than zero.
+    A weight of exactly zero means a cell in the corner of a region, covered by only one window, can never be filled in.
+    The same grid is used for every window so it only needs to be worked out once."""
+
+    # NOTES:
+    # Distance-Based Weighting For Vignettes or Radial Masks - linear distance decay function: each (row, column) = 1 - distance to center/maximum patch radius
+    # numpy array: [[row 1 contents], [row 2 contents]]
+    # indexing in 2D Array: array[row, column]
+
+    assert edge_len > 1
+
+    # create 1D arrays
+    positions = np.arange(edge_len)
+
+    # find center (-1 because we start from 0)
+    center = (edge_len - 1) / 2
+
+    # distance from center
+    distance = np.abs(positions - center)
+
+    # weight: apply formula. multiplied 0.9 so values stay above 0
+    weight = 1 - 0.9 * distance / center
+    # combine
+    weights = np.outer(weight, weight)
+
+    return weights
+
+
+def generate_noise_from_seed(seed: int, height: int, width: int) -> np.ndarray:
+    "Takes a seed and a canvas size and returns a grid of random numbers that size"
+    generator = np.random.default_rng(seed)
+    return generator.random((height, width))
+
+
+def produce_region(
+    seed: int,
+    height: int,
+    width: int,
+    window_size: int,
+    step: int,
+    pipeline,
+) -> tuple[np.ndarray]:
+    """Make noise canvas of given dimensions. Make noise and weight grid.
+    For each window position, cut the window out of the noise canvas, send it to pipeline, add the processed output and its weight to Terrain Store (at that position).
+    Read the finished height grid from store and return it."""
+
+    noise = generate_noise_from_seed(seed, height, width)
+
+    positions = window_positions(height, width, window_size, step)
+    weights = weight_grid(window_size)  # weight grid made on window_size
+
+    weighted_sum = np.zeros((height, width))
+    weight_sum = np.zeros((height, width))
+
+    for row, column in positions:
+        window = noise[row : row + window_size, column : column + window_size]
+
+        processed_patch = pipeline.generate(window)
+
+        # From FakeStore()
+        weighted_sum[row : row + window_size, column : column + window_size] += (
+            processed_patch * weights
+        )
+        weight_sum[row : row + window_size, column : column + window_size] += weights
+
+    return weighted_sum, weight_sum

tests/test_sampler.py

@@ -0,0 +1,220 @@
+"""
+Testing for window blending sampler.
+"""
+
+import numpy as np
+import pytest
+
+from terrain_diffusion.sampler import (
+    generate_noise_from_seed,
+    produce_region,
+    weight_grid,
+    window_positions,
+)
+
+
+class TestWindowPositions:
+    def test_all_covered(self):
+        "Assert every cell in the region is covered by at least one window"
+        height = 4
+        width = 4
+        window = 2
+        step = 2
+        positions = window_positions(height, width, window, step)
+        assert all(
+            any(
+                window_r <= row < window_r + window and window_c <= column < window_c + window
+                for window_r, window_c in positions
+            )
+            for row in range(height)
+            for column in range(width)
+        )
+
+    def test_exceed_region(self):
+        "Assert no window exceeds past the region"
+        height = 4
+        width = 4
+        size = 2
+        step = 1
+        positions = window_positions(height, width, size, step)
+        assert all(x[0] + size <= height and x[1] + size <= width for x in positions)
+
+    def test_one_window(self):
+        "Assert a region exactly one window in size returns one position"
+        WindowRegionSize = 4
+        positions = window_positions(
+            WindowRegionSize, WindowRegionSize, WindowRegionSize, WindowRegionSize
+        )
+        assert len(positions) == 1
+
+    # Had to modify test because added assertion to original function
+    def test_region_not_divisible(self):
+        with pytest.raises(AssertionError):
+            window_positions(10, 8, 4, 3)
+
+
+class TestWeights:
+    def test_grid_equal_patch(self):
+        "Assert the grid is the size of a patch"
+        edge = 10
+        weights = weight_grid(edge)
+        assert weights.shape == (edge, edge)
+
+    def test_palidrome(self):
+        "Assert it reads the same forwards and backwards in both directions"
+        edge = 10
+        weights = weight_grid(edge)
+        assert np.array_equal(weights, weights[::-1, :])  # vertical
+        assert np.array_equal(weights, weights[:, ::-1])  # horizontal
+
+    def test_large_middle(self):
+        "Assert the largest value is in the middle"
+        edge = 11
+        weights = weight_grid(edge)
+
+        # middle
+        center = edge // 2
+
+        assert weights[center, center] == weights.max()
+
+    def test_edges_smaller(self):
+        "Assert values at the edges are smaller than values in the middle"
+        edge = 11
+        weights = weight_grid(edge)
+
+        # middle
+        center = edge // 2
+        middle_val = weights[center, center]
+
+        # R/L edges
+        assert all(weights[x, 0] < middle_val for x in range(edge))
+        assert all(weights[x, edge - 1] < middle_val for x in range(edge))
+
+        # T/B edges
+        assert all(weights[0, y] < middle_val for y in range(edge))
+        assert all(weights[edge - 1, y] < middle_val for y in range(edge))
+
+    def test_greater_zero(self):
+        "Assert every value is greater than zero"
+        weights = weight_grid(10)
+        assert np.all(weights > 0)
+
+
+# the requested test for determinitic grid has been added at line 205
+class TestSeed:
+    def test_same_seed(self):
+        """Assert the same seed twice gives identical grids."""
+        seed = 123
+        height = 10
+        width = 20
+        noise1 = generate_noise_from_seed(seed, height, width)
+        noise2 = generate_noise_from_seed(seed, height, width)
+        assert np.array_equal(noise1, noise2)
+
+    def test_diff_seed(self):
+        """Assert two different seeds give different grids."""
+        seed1 = 123
+        seed2 = 456
+        height = 10
+        width = 20
+        noise1 = generate_noise_from_seed(seed1, height, width)
+        noise2 = generate_noise_from_seed(seed2, height, width)
+        assert not np.array_equal(noise1, noise2)
+
+    def test_right_size(self):
+        """Assert the grid is the size asked for"""
+        seed = 123
+        height = 10
+        width = 20
+        noise = generate_noise_from_seed(seed, height, width)
+        assert noise.shape == (height, width)
+
+
+class TestRegionProduction:
+    @pytest.fixture
+    def pipeline(self, mocker):
+        pipeline = mocker.Mock()  # make it a Mock object, this way can count calls.
+        pipeline.generate.side_effect = lambda patch: np.full(
+            patch.shape, 5
+        )  # added side_effect to keep it a Mock object
+        return pipeline
+
+    def test_all_fives(self, pipeline):
+        """Assert the finished grid is all fives everywhere, including the overlaps and the corners.
+        If the overlaps read higher then the weights are not being divided out"""
+
+        seed = 123
+        height = 8
+        width = 8
+        window_size = 4
+        step = 2
+
+        weighted_sum, weight_sum = produce_region(seed, height, width, window_size, step, pipeline)
+        result = weighted_sum / weight_sum  # doing job of store
+        assert np.allclose(
+            result, 5
+        )  # All close because was getting float error as some are 4.9999 due to the store
+
+    def test_full_size(self, pipeline):
+        """Assert the finished grid is the region's full resolution size"""
+
+        seed = 123
+        height = 8
+        width = 8
+        window_size = 4
+        step = 2
+
+        weighted_sum, weight_sum = produce_region(seed, height, width, window_size, step, pipeline)
+        result = weighted_sum / weight_sum  # doing job of store
+
+        assert result.shape == (height, width)
+
+    def test_once_per_window(self, pipeline):
+        """Assert the fake pipeline was called once per window position and no more"""
+
+        seed = 123
+        height = 8
+        width = 8
+        window_size = 4
+        step = 2
+
+        positions = window_positions(height, width, window_size, step)
+
+        produce_region(seed, height, width, window_size, step, pipeline)
+
+        assert pipeline.generate.call_count == len(positions)
+
+    def test_same_seed_grid(self, pipeline):
+        """Assert the same seed and region run twice give identical grids"""
+        seed = 123
+        height = 8
+        width = 8
+        window_size = 4
+        step = 2
+
+        weighted_sum1, weight_sum1 = produce_region(
+            seed, height, width, window_size, step, pipeline
+        )
+        weighted_sum2, weight_sum2 = produce_region(
+            seed, height, width, window_size, step, pipeline
+        )
+        result1 = weighted_sum1 / weight_sum1
+        result2 = weighted_sum2 / weight_sum2
+
+        assert np.array_equal(result1, result2)
+
+    def test_deterministic_region(self, pipeline):
+        """Assert the output matches the expected grid."""
+        seed = 123
+        height = 9
+        width = 9
+        window_size = 3
+        step = 2
+        pipeline.generate.side_effect = lambda patch: np.ones(patch.shape)
+
+        weighted_sum, weight_sum = produce_region(seed, height, width, window_size, step, pipeline)
+        result = weighted_sum / weight_sum
+
+        expected = np.ones((9, 9))
+
+        assert np.allclose(result, expected)