Tested code for for python > 3.9 and numpy >= 2.0

.github/workflows/test_pull_requests.yml

@@ -21,24 +21,30 @@ jobs:
         os: [ubuntu-latest, macos-latest, windows-latest]
         # see supported versions at
         # https://raw.githubusercontent.com/actions/python-versions/master/versions-manifest.json
-        python-version: [2.7, 3.5, 3.6, 3.7, 3.8]
-        # test only 2.7 and the latest 3.x on mac
+        python-version: ['3.7', '3.8', '3.9', '3.10', '3.11', '3.12']
+        # test only the latest 3.x on mac
         # test only the latest 3.x on windows
         exclude:
           - os: macos-latest
-            python-version: 3.5
+            python-version: '3.7'
           - os: macos-latest
-            python-version: 3.6
+            python-version: '3.8'
           - os: macos-latest
-            python-version: 3.7
+            python-version: '3.9'
+          - os: macos-latest
+            python-version: '3.10'
+          - os: macos-latest
+            python-version: '3.11'
+          - os: windows-latest
+            python-version: '3.7'
           - os: windows-latest
-            python-version: 2.7  # causes a Shapely install error
+            python-version: '3.8'
           - os: windows-latest
-            python-version: 3.5
+            python-version: '3.9'
           - os: windows-latest
-            python-version: 3.6
+            python-version: '3.10'
           - os: windows-latest
-            python-version: 3.7
+            python-version: '3.11'
     env:
       OS: ${{ matrix.os }}
       PYTHON: ${{ matrix.python-version }}
@@ -103,27 +109,12 @@ jobs:
         restore-keys: |
           ${{ runner.os }}-pip-
 
-    # ----------------
-    # Install dependencies
-    # ----------------
-    - name: Install dependencies
-      run: |
-        pip install -r requirements.txt
-
-    - name: Install test dependencies
-      run: |
-        pip install --upgrade -r test/requirements.txt
-
-    - name: Install further test tools
-      run: |
-        pip install coverage pytest-cov flake8
-
     # ----------------
     # Install library
     # ----------------
     - name: Install library
       run: |
-        pip install .
+        pip install .[dev]
 
     # ----------------
     # Run checks and tests

README.md

@@ -84,7 +84,8 @@ It converts a set of input images into a new, much larger set of slightly altere
 4. [Recent Changes](#recent_changes)
 5. [Example Images](#example_images)
 6. [Code Examples](#code_examples)
-7. [Citation](#citation)
+7. [Development](#development)
+8. [Citation](#citation)
 
 
 <a name="features"/>
@@ -121,7 +122,7 @@ It converts a set of input images into a new, much larger set of slightly altere
 
 ## Installation
 
-The library supports python 2.7 and 3.4+.
+The library supports python 3.7+.
 
 ### Installation: Anaconda
 
@@ -1179,6 +1180,20 @@ aug = iaa.WithChannels(
 images_aug = aug(images=images)
 ```
 
+<a name="development"/>
+
+## Development
+
+To install this packages with all development requirments, clone this repository, and run
+```
+pip install -e .[dev]
+```
+
+You can run the tests locally via
+```
+./test/run_tests.sh
+```
+
 
 <a name="citation"/>
 

imgaug/augmentables/polys.py

@@ -950,7 +950,7 @@ def extract_from_image(self, image):
         rr_face, cc_face = skimage.draw.polygon(
             yy_mask, xx_mask, shape=(height_mask, width_mask))
 
-        mask = np.zeros((height_mask, width_mask), dtype=np.bool)
+        mask = np.zeros((height_mask, width_mask), dtype=bool)
         mask[rr_face, cc_face] = True
 
         if image.ndim == 3:

imgaug/augmenters/arithmetic.py

@@ -175,7 +175,7 @@ def _add_scalar_to_uint8_(image, value):
         return cv2.add(image, value, dst=image, dtype=cv2.CV_8U)
 
     input_shape = image.shape
-    image = image.ravel()
+    image = image.reshape(-1, 1)
     values = np.array(value)
     if not is_channelwise:
         values = np.broadcast_to(values, image.shape)
@@ -1442,14 +1442,16 @@ def _invert_uint8_lut_pregenerated_(arr, min_value, max_value, threshold,
 # Added in 0.5.0.
 def _invert_uint8_subtract_(arr, max_value):
     # seems to work with arr.base.shape[0] > 1
+    if arr.size == 0:
+        return arr
     if arr.base is not None and arr.base.shape[0] == 1:
         arr = np.copy(arr)
     if not arr.flags["C_CONTIGUOUS"]:
         arr = np.ascontiguousarray(arr)
 
     input_shape = arr.shape
     if len(input_shape) > 2 and input_shape[-1] > 1:
-        arr = arr.ravel()
+        arr = arr.reshape(-1, 1)
     # This also supports a mask, which would help for thresholded invert, but
     # it seems that all non-masked components are set to zero in the output
     # array. Tackling this issue seems to rather require more time than just

imgaug/augmenters/geometric.py

@@ -237,6 +237,9 @@ def _warp_affine_arr_skimage(arr, matrix, cval, mode, order, output_shape):
     if input_dtype == iadt._BOOL_DTYPE and order != 0:
         arr = arr.astype(np.float32)
 
+    if input_dtype == np.float16 and order == 0:
+        arr = arr.astype(np.float32)
+
     image_warped = tf.warp(
         arr,
         np.linalg.inv(matrix),
@@ -3204,19 +3207,26 @@ def _augment_images_by_samples(self, images, samples):
                 if image.dtype.kind == "b":
                     image = image.astype(np.float64)
 
+                if image.dtype == np.float16 and samples.order[i] == 0:
+                    image = image.astype(np.float32)
+
+                cval = samples.get_clipped_cval(i, image.dtype)
                 image_warped = tf.warp(
                     image,
                     transformer,
                     order=samples.order[i],
                     mode=samples.mode[i],
-                    cval=samples.get_clipped_cval(i, image.dtype),
+                    cval=cval,
                     preserve_range=True,
                     output_shape=images[i].shape
                 )
 
                 if input_dtype.kind == "b":
                     image_warped = image_warped > 0.5
                 else:
+                    # fill skimage.transform.warp output nan values with cval
+                    image_warped = np.where(np.isnan(image_warped), cval.astype(input_dtype), image_warped)
+
                     # warp seems to change everything to float64, including
                     # uint8, making this necessary
                     image_warped = iadt.restore_dtypes_(
@@ -4789,7 +4799,7 @@ def _map_coordinates(self, image, dx, dy, order=1, cval=0, mode="constant"):
             result = np.empty_like(image)
 
             for c in sm.xrange(image.shape[2]):
-                remapped_flat = ndimage.interpolation.map_coordinates(
+                remapped_flat = ndimage.map_coordinates(
                     image[..., c],
                     (y_shifted.flatten(), x_shifted.flatten()),
                     order=order,
@@ -4878,7 +4888,6 @@ def generate(self, shapes, alphas, sigmas, random_state):
         # somewhere by 2. It is fastes to multiply the (fewer) alphas, which
         # we will have to multiply the shift maps with anyways.
         alphas *= 2
-
         # Configuration for each chunk.
         # switch dx / dy, flip dx lr, flip dx ud, flip dy lr, flip dy ud
         switch = [False, True]
@@ -4948,9 +4957,8 @@ def _flip(cls, dx, dy, flips):
     # Added in 0.5.0.
     @classmethod
     def _mul_alpha(cls, dx, dy, alpha):
-        # performance drops for cv2.multiply here
-        dx = dx * alpha
-        dy = dy * alpha
+        dx = (dx * alpha).astype(dx.dtype)
+        dy = (dy * alpha).astype(dy.dtype)
         return dx, dy
 
     # Added in 0.5.0.
@@ -4961,6 +4969,8 @@ def _smoothen_(cls, dx, dy, sigma):
             dy = blur_lib.blur_gaussian_(dy, sigma)
         else:
             ksize = int(round(2*sigma))
+            dx = np.ascontiguousarray(dx)
+            dy = np.ascontiguousarray(dy)
             dx = cv2.blur(dx, (ksize, ksize), dst=dx)
             dy = cv2.blur(dy, (ksize, ksize), dst=dy)
         return dx, dy

imgaug/augmenters/imgcorruptlike.py

@@ -69,11 +69,13 @@
 """
 from __future__ import print_function, division, absolute_import
 
+import functools
 import warnings
 
 import six.moves as sm
 import numpy as np
 import skimage.filters
+import skimage
 
 import imgaug as ia
 from ..imgaug import _numbajit
@@ -93,6 +95,8 @@
     "of the package."
 )
 
+SK_VERSION = {k:int(v) for k,v in zip(['major', 'minor'], skimage.__version__.split('.')[:2])}
+
 
 # Added in 0.4.0.
 def _clipped_zoom_no_scipy_warning(img, zoom_factor):
@@ -353,7 +357,12 @@ def apply_impulse_noise(x, severity=1, seed=None):
         Corrupted image.
 
     """
-    return _call_imgcorrupt_func("impulse_noise", seed, False, x, severity)
+    return functools.partial(
+        _call_imgcorrupt_func,
+        "impulse_noise",
+        seed,
+        False,
+     )(x, severity, seed)
 
 
 def apply_speckle_noise(x, severity=1, seed=None):
@@ -473,9 +482,14 @@ def _apply_glass_blur_imgaug(x, severity=1):
 
     sigma, max_delta, iterations = c
 
+    if SK_VERSION['major'] >= 0 and SK_VERSION['minor'] >= 19:
+        kwargs = {'channel_axis': -1}
+    else: # pre scikit-image 0.19
+        kwargs = {'multichannel': True}
+
     x = (
         skimage.filters.gaussian(
-            np.array(x) / 255., sigma=sigma, multichannel=True
+            np.array(x) / 255., sigma=sigma, **kwargs
         ) * 255
     ).astype(np.uint)
     x_shape = x.shape
@@ -496,7 +510,7 @@ def _apply_glass_blur_imgaug(x, severity=1):
     )
 
     return np.clip(
-        skimage.filters.gaussian(x / 255., sigma=sigma, multichannel=True),
+        skimage.filters.gaussian(x / 255., sigma=sigma, **kwargs),
         0, 1
     ) * 255
 

imgaug/augmenters/meta.py

@@ -3384,7 +3384,7 @@ def _get_augmenter_active(self, nb_rows, random_state):
         # pylint: disable=invalid-name
         nn = self._get_n(nb_rows, random_state)
         nn = [min(n, len(self)) for n in nn]
-        augmenter_active = np.zeros((nb_rows, len(self)), dtype=np.bool)
+        augmenter_active = np.zeros((nb_rows, len(self)), dtype=bool)
         for row_idx, n_true in enumerate(nn):
             if n_true > 0:
                 augmenter_active[row_idx, 0:n_true] = 1

imgaug/imgaug.py

@@ -36,9 +36,9 @@
 # to check if a dtype instance is among these dtypes, use e.g.
 # `dtype.type in  NP_FLOAT_TYPES` do not just use `dtype in NP_FLOAT_TYPES` as
 # that would fail
-NP_FLOAT_TYPES = set(np.sctypes["float"])
-NP_INT_TYPES = set(np.sctypes["int"])
-NP_UINT_TYPES = set(np.sctypes["uint"])
+NP_FLOAT_TYPES = set([np.float16, np.float32, np.float64])
+NP_INT_TYPES = set([np.int8, np.int16, np.int32, np.int64])
+NP_UINT_TYPES = set([np.uint8, np.uint16, np.uint32, np.uint64])
 
 IMSHOW_BACKEND_DEFAULT = "matplotlib"
 
@@ -697,13 +697,13 @@ def angle_between_vectors(v1, v2):
 
     Examples
     --------
-    >>> angle_between_vectors(np.float32([1, 0, 0]), np.float32([0, 1, 0]))
+    >>> angle_between_vectors(np.float32([1, 0, 0]), np.float32([0, 1, 0])).item()
     1.570796...
 
-    >>> angle_between_vectors(np.float32([1, 0, 0]), np.float32([1, 0, 0]))
+    >>> angle_between_vectors(np.float32([1, 0, 0]), np.float32([1, 0, 0])).item()
     0.0
 
-    >>> angle_between_vectors(np.float32([1, 0, 0]), np.float32([-1, 0, 0]))
+    >>> angle_between_vectors(np.float32([1, 0, 0]), np.float32([-1, 0, 0])).item()
     3.141592...
 
     """
@@ -1297,7 +1297,7 @@ def pool(arr, block_size, func, pad_mode="constant", pad_cval=0,
     input_dtype = arr.dtype
 
     arr_reduced = skimage.measure.block_reduce(arr, tuple(block_size), func,
-                                               cval=cval)
+                                               cval=pad_cval)
     if preserve_dtype and arr_reduced.dtype.name != input_dtype.name:
         arr_reduced = arr_reduced.astype(input_dtype)
     return arr_reduced

imgaug/multicore.py

@@ -612,7 +612,7 @@ def __init__(self, load_batch_func, queue_size=50, nb_workers=1,
                 worker = multiprocessing.Process(
                     target=self._load_batches,
                     args=(load_batch_func, self._queue_internal,
-                          self.join_signal, seeds[i])
+                          self.join_signal, int(seeds[i]))
                 )
             worker.daemon = True
             worker.start()
@@ -828,7 +828,7 @@ def __init__(self, batch_loader, augseq, queue_size=50, nb_workers="auto"):
         for i in range(nb_workers):
             worker = multiprocessing.Process(
                 target=self._augment_images_worker,
-                args=(augseq, self.queue_source, self.queue_result, seeds[i])
+                args=(augseq, self.queue_source, self.queue_result, int(seeds[i]))
             )
             worker.daemon = True
             worker.start()

imgaug/parameters.py

@@ -968,7 +968,7 @@ class Deterministic(StochasticParameter):
     --------
     >>> import imgaug.parameters as iap
     >>> param = iap.Deterministic(10)
-    >>> param.draw_sample()
+    >>> param.draw_sample().item()
     10
 
     Will always sample the value 10.
@@ -2588,7 +2588,7 @@ def _draw_samples(self, size, random_state):
         # result = np.float_power(samples, exponents)
         # TODO why was float32 type here replaced with complex number
         #      formulation?
-        result = np.power(samples.astype(np.complex), exponents).real
+        result = np.power(samples.astype(np.complex128), exponents).real
         if result.dtype != samples_dtype:
             result = result.astype(samples_dtype)
 
@@ -3558,7 +3558,7 @@ def _draw_samples_hw(self, height, width, random_state):
             wn = wn.astype(np.float32)
 
         # equivalent but slightly faster then:
-        #   wn_freqs_mul = np.zeros(treal.shape, dtype=np.complex)
+        #   wn_freqs_mul = np.zeros(treal.shape, dtype=np.complex128)
         #   wn_freqs_mul.real = wn[0]
         #   wn_freqs_mul.imag = wn[1]
         #   wn_inv = np.fft.ifft2(wn_freqs_mul).real