Created
February 5, 2018 03:28
-
-
Save Erotemic/75aa64817a4f657915e2356f9c23300b to your computer and use it in GitHub Desktop.
benchmark code
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| import ubelt as ub | |
| import numpy as np | |
| from PIL import Image | |
| import six | |
| import cv2 | |
| from clab.augment import augment_common | |
| from clab.util import imutil | |
| from clab import util | |
| try: | |
| import skimage | |
| import skimage.transform | |
| except ImportError: | |
| pass | |
| class NaNInputer(object): | |
| """ | |
| Methods for removing nan or invalid values | |
| Example: | |
| >>> data = np.array([ | |
| >>> [-1, 20, 30, 40], | |
| >>> [10, -1, 30, -1], | |
| >>> [30, -1, 50, 90], | |
| >>> [30, 11, 13, 90], | |
| >>> ]) | |
| >>> self = NaNInputer(fill='median', nan_value=-1) | |
| >>> self(data) | |
| array([[30, 20, 30, 40], | |
| [10, 30, 30, 30], | |
| [30, 30, 50, 90], | |
| [30, 11, 13, 90]]) | |
| Example: | |
| >>> channel = np.array([[-1, 20], [10, -1]]) | |
| >>> data = np.dstack([channel, channel]) | |
| >>> self = NaNInputer(fill='median', nan_value=-1) | |
| >>> self(data).tolist() | |
| [[[15, 15], [20, 20]], [[10, 10], [15, 15]]] | |
| """ | |
| def __init__(self, fill='median', nan_value=np.nan): | |
| self.nan_value = nan_value | |
| if np.isnan(nan_value): | |
| self.isnan = np.isnan | |
| else: | |
| def isnan(data): | |
| return data == nan_value | |
| self.isnan = isnan | |
| if isinstance(fill, six.string_types): | |
| # Define a local method for inputing fill data | |
| if fill == 'median': | |
| self.getfill = np.median | |
| elif fill == 'mean': | |
| self.getfill = np.mean | |
| else: | |
| raise KeyError(fill) | |
| else: | |
| self.getfill = None | |
| self.fill_value = fill | |
| def __call__(self, data): | |
| mask = self.isnan(data) | |
| if self.getfill is None: | |
| fill_value = self.fill_value | |
| elif np.all(mask): | |
| fill_value = 0 | |
| else: | |
| fill_value = self.getfill(data[~mask].ravel()) | |
| data[mask] = fill_value | |
| return data | |
| class LocalCenterGlobalScale(object): | |
| def __init__(self, std): | |
| self.std = std | |
| def __call__(self, data): | |
| center = np.median(data) | |
| data = (data - center) / self.std | |
| return data | |
| class DTMCenterScale(object): | |
| """ | |
| Overly specific class for urban3d mapper. | |
| Combines NaNInputer and LocalCenterGlobalScale because both use median for a | |
| tiny speed increase. | |
| """ | |
| def __init__(self, std, fill='median', nan_value=-32767.0): | |
| # aux_std = 5.3757350869126723 | |
| self.std = std | |
| self.nan_value = nan_value | |
| self.fill = fill | |
| def __call__(self, data): | |
| # zero the median on a per-chip basis, but use | |
| # the global internal_std to normalize extent | |
| mask = (data == self.nan_value) | |
| if np.all(mask): | |
| center = 0 | |
| else: | |
| center = self._get_center(data[~mask].ravel()) | |
| data[mask] = center | |
| data = (data - center) / self.std | |
| return data | |
| @property | |
| def fill(self): | |
| return self._fill | |
| @fill.setter | |
| def fill(self, fill): | |
| valid_center_funcs = { | |
| 'median': np.median | |
| } | |
| self._fill = fill | |
| self._get_center = valid_center_funcs[fill] | |
| def __getstate__(self): | |
| return {'std': self.std, 'nan_value': self.nan_value, 'fill': self.fill} | |
| def __setstate__(self, state): | |
| for key, value in state.items(): | |
| setattr(self, key, value) | |
| class ImageCenterScale(ub.NiceRepr): | |
| def __init__(self, im_mean, im_scale): | |
| self.im_mean = im_mean | |
| self.im_scale = im_scale | |
| def __nice__(self): | |
| try: | |
| return '{:g}, {:g}'.format(self.im_mean, self.im_scale) | |
| except: | |
| return '{}, {}'.format(self.im_mean, self.im_scale) | |
| def __call__(self, data): | |
| return (data - self.im_mean) / self.im_scale | |
| def invert(self, data): | |
| return (data * self.im_scale) + self.im_mean | |
| def __getstate__(self): | |
| return self.__dict__ | |
| def __setstate__(self, state): | |
| for key, value in state.items(): | |
| setattr(self, key, value) | |
| class ZipTransforms(): | |
| def __init__(self, transforms): | |
| self.transforms = transforms | |
| def __call__(self, inputs): | |
| return [ | |
| transform(input) | |
| for transform, input in zip(self.transforms, inputs) | |
| ] | |
| class AffineWarp(object): | |
| """ | |
| Wrapper around three different methods for affine warping. The fastest is | |
| cv2, followed by pil, followed by skimage. The most flexible is skimage, | |
| and the least flexible is pil, but cv2 has lanczos resampling while the | |
| others do not. However, cv2.warpAffine has known bugs that sometimes occur | |
| with multiprocessing, so its not always 100% safe to use. | |
| Benchmark: | |
| >>> from clab.transforms import * | |
| >>> import itertools as it | |
| >>> fpath = ub.grabdata('http://i.imgur.com/JGrqMnV.png', fname='lena.png') | |
| >>> img_orig = util.imread(fpath) | |
| >>> shape = img_orig.shape | |
| >>> aff = AffineWarp() | |
| >>> x, y = shape[1] // 2, shape[0] // 2 | |
| >>> matrix = aff.around_mat3x3(x, y, sx=.8, theta=.3, shear=.001, tx=2.5) | |
| >>> N = 100 | |
| >>> def _compare_results(results, scale, thresh): | |
| >>> diffs = [] | |
| >>> for key1, key2 in it.combinations(results.keys(), 2): | |
| >>> ave_diff = np.abs(results[key1].astype(np.float64) - results[key2].astype(np.float64)).mean() / scale | |
| >>> #print('{} {} ave_diff = {!r}'.format(key1, key2, ave_diff)) | |
| >>> diffs.append(ave_diff) | |
| >>> assert np.all(np.array(diffs) < thresh) | |
| >>> def _assert_range(results, min, max): | |
| >>> for key, result in results.items(): | |
| >>> assert result.max() <= max, '{} {}'.format(key, result.max()) | |
| >>> assert result.min() >= min, '{} {}'.format(key, result.min()) | |
| >>> # Results for rgb float01 | |
| >>> img = img_orig.astype(np.float32) / 255 | |
| >>> results = ub.odict() | |
| >>> for back in aff.backends: | |
| >>> _warper = aff.make_warper(backend=back, shape=img.shape, matrix=matrix, mode='float01') | |
| >>> results[back] = _warper(img) | |
| >>> ub.Timerit(N, verbose=1, label=back).call(_warper, img) | |
| >>> _assert_range(results, 0, 1) | |
| >>> _compare_results(results, scale=1, thresh=.02) | |
| Timed ski for: 10 loops, best of 3 | |
| time per loop : best=67.75 ms, mean=68.14 ± 0.35 ms | |
| Timed pil for: 10 loops, best of 3 | |
| time per loop : best=23.44 ms, mean=23.7 ± 0.22 ms | |
| Timed cv2 for: 10 loops, best of 3 | |
| time per loop : best=3.403 ms, mean=3.713 ± 0.2 ms | |
| >>> # Results for grayscale float01 | |
| >>> img = (img_orig.astype(np.float32) / 255).mean(axis=2) | |
| >>> results = ub.odict() | |
| >>> for back in aff.backends: | |
| >>> _warper = aff.make_warper(backend=back, shape=img.shape, matrix=matrix, mode='float01') | |
| >>> results[back] = _warper(img) | |
| >>> ub.Timerit(N, verbose=1, label=back).call(_warper, img) | |
| >>> _compare_results(results, scale=1, thresh=.02) | |
| >>> _assert_range(results, 0, 1) | |
| Timed ski for: 10 loops, best of 3 | |
| time per loop : best=21.31 ms, mean=22.5 ± 1.9 ms | |
| Timed pil for: 10 loops, best of 3 | |
| time per loop : best=9.877 ms, mean=10.24 ± 0.37 ms | |
| Timed cv2 for: 10 loops, best of 3 | |
| time per loop : best=1.753 ms, mean=1.926 ± 0.15 ms | |
| >>> # Results for grayscale uint8 | |
| >>> img = ((img_orig.astype(np.float32) / 255).mean(axis=2) * 255).astype(np.uint8) | |
| >>> results = ub.odict() | |
| >>> for back in aff.backends: | |
| >>> _warper = aff.make_warper(backend=back, shape=img.shape, matrix=matrix, mode='uint8') | |
| >>> results[back] = _warper(img) | |
| >>> ub.Timerit(N, verbose=1, label=back).call(_warper, img) | |
| >>> _compare_results(results, scale=255, thresh=.02) | |
| Timed ski for: 10 loops, best of 3 | |
| time per loop : best=21.46 ms, mean=21.56 ± 0.14 ms | |
| Timed pil for: 10 loops, best of 3 | |
| time per loop : best=11.22 ms, mean=11.9 ± 0.76 ms | |
| Timed cv2 for: 10 loops, best of 3 | |
| time per loop : best=1.518 ms, mean=1.596 ± 0.053 ms | |
| >>> # Results for rgb uint8 | |
| >>> img = img_orig.copy() | |
| >>> results = ub.odict() | |
| >>> for back in aff.backends: | |
| >>> _warper = aff.make_warper(shape=img.shape, matrix=matrix, backend=back, mode='uint8') | |
| >>> results[back] = _warper(img) | |
| >>> ub.Timerit(N, verbose=1, label=back).call(_warper, img) | |
| >>> _compare_results(results, scale=255, thresh=.02) | |
| Timed ski for: 10 loops, best of 3 | |
| time per loop : best=67.65 ms, mean=67.9 ± 0.25 ms | |
| Timed pil for: 10 loops, best of 3 | |
| time per loop : best=21.38 ms, mean=21.65 ± 0.39 ms | |
| Timed cv2 for: 10 loops, best of 3 | |
| time per loop : best=2.927 ms, mean=3.318 ± 0.3 ms | |
| Ignore: | |
| >>> # +SKIP | |
| >>> import plottool as pt | |
| >>> pt.qtensure() | |
| >>> pnum_ = pt.make_pnum_nextgen(nCols=len(results) + 1) | |
| >>> pt.imshow(img, pnum=pnum_(), fnum=1) | |
| >>> for key, result in results.items(): | |
| >>> pt.imshow(result, pnum=pnum_(), fnum=1, title=key) | |
| Ignore: | |
| >>> # +SKIP | |
| >>> %timeit aff.make_warper(shape=img.shape, matrix=matrix, backend='ski')(img) | |
| >>> %timeit aff.make_warper(shape=img.shape, matrix=matrix, backend='pil')(img) | |
| >>> %timeit aff.make_warper(shape=img.shape, matrix=matrix, backend='cv2')(img) | |
| >>> %timeit aff.make_warper(shape=img.shape, matrix=matrix, backend='cv2_inv')(img) | |
| 10 loops, best of 3: 22.6 ms per loop | |
| 100 loops, best of 3: 7.97 ms per loop | |
| 1000 loops, best of 3: 1.41 ms per loop | |
| """ | |
| backends = ['cv2', 'pil', 'ski'] | |
| skimage_interp_lookup = { | |
| 'nearest' : 0, | |
| 'linear' : 1, | |
| 'quadradic' : 2, | |
| 'cubic' : 3, | |
| 'lanczos': NotImplemented, | |
| } | |
| pil_interp_lookup = { | |
| 'nearest' : Image.NEAREST, | |
| 'linear' : Image.BILINEAR, | |
| 'quadradic' : NotImplemented, | |
| 'cubic' : Image.BICUBIC, | |
| 'lanczos': NotImplemented, | |
| } | |
| cv2_interp_lookup = { | |
| 'nearest' : cv2.INTER_NEAREST, | |
| 'linear' : cv2.INTER_LINEAR, | |
| 'quadradic' : NotImplemented, | |
| 'cubic' : cv2.INTER_CUBIC, | |
| 'lanczos': cv2.INTER_LANCZOS4, | |
| } | |
| skimage_border_lookup = { | |
| 'constant': 'constant', | |
| 'reflect': 'reflect', | |
| } | |
| cv2_border_lookup = { | |
| 'constant': cv2.BORDER_CONSTANT, | |
| 'reflect': cv2.BORDER_REFLECT, | |
| } | |
| pil_border_lookup = { | |
| 'constant': 'constant', | |
| 'reflect': NotImplemented, | |
| } | |
| def __init__(self): | |
| self.defaults = { | |
| 'backend': 'cv2', | |
| 'interp': 'cubic', | |
| 'border_mode': 'constant', | |
| 'mode': 'float01', | |
| 'clip': 'auto', | |
| } | |
| def warp(self, img, **kwargs): | |
| _warper = self.make_warper(**kwargs) | |
| return _warper(img) | |
| def around_mat3x3(self, x, y, sx=1.0, sy=1.0, theta=0.0, shear=0.0, tx=0.0, | |
| ty=0.0, flip_ud=False, flip_lr=False): | |
| mat_2x3 = augment_common.affine_around_mat2x3(x, y, sx, sy, theta, | |
| shear, tx, ty, | |
| flip_ud=flip_ud, | |
| flip_lr=flip_lr) | |
| mat_3x3 = np.array(mat_2x3 + [[0, 0, 1]]) | |
| return mat_3x3 | |
| def mat3x3(self, sx=1.0, sy=1.0, theta=0.0, shear=0.0, tx=0.0, | |
| ty=0.0, flip_ud=False, flip_lr=False): | |
| mat_2x3 = augment_common.affine_mat2x3(sx, sy, theta, shear, tx, ty, | |
| flip_ud=flip_ud, | |
| flip_lr=flip_lr) | |
| mat_3x3 = np.array(mat_2x3 + [[0, 0, 1]]) | |
| return mat_3x3 | |
| def make_warper(self, shape, matrix, interp=None, border_mode=None, | |
| clip=None, mode=None, backend=None): | |
| if backend is None: | |
| backend = self.defaults['backend'] | |
| if interp is None: | |
| interp = self.defaults['interp'] | |
| if border_mode is None: | |
| border_mode = self.defaults['border_mode'] | |
| if mode is None: | |
| mode = self.defaults['mode'] | |
| if clip is None: | |
| clip = self.defaults['clip'] | |
| if backend == 'skimage': | |
| backend = 'ski' | |
| if clip == 'auto': | |
| # Find a "good" value for clip | |
| if mode == 'float01': | |
| # We only need to clip for higher order interpolation methods | |
| if interp not in {'nearest', 'linear'}: | |
| clip = True | |
| else: | |
| clip = False | |
| elif mode == 'uint8': | |
| # only clip uint8 for skimage | |
| if backend == 'ski': | |
| clip = True | |
| else: | |
| clip = False | |
| return { | |
| 'cv2': self.make_cv2_warper, | |
| 'pil': self.make_pil_warper, | |
| 'ski': self.make_skimage_warper, | |
| }[backend](shape, matrix, interp, border_mode, clip, mode) | |
| def make_skimage_warper(self, shape, matrix, interp, border_mode, clip, mode): | |
| inv_matrix = np.linalg.inv(matrix) | |
| order = self.skimage_interp_lookup[interp] | |
| skaff = skimage.transform.AffineTransform(matrix=inv_matrix) | |
| def _sk_warper(img): | |
| imaug = skimage.transform.warp( | |
| img, skaff, output_shape=img.shape, order=order, | |
| mode=border_mode, clip=clip, preserve_range=True | |
| ) | |
| imaug = imaug.astype(img.dtype) | |
| return imaug | |
| return _sk_warper | |
| def make_pil_warper(self, shape, matrix, interp, border_mode, clip, mode): | |
| inv_matrix = np.linalg.inv(matrix) | |
| pil_aff_params = list(inv_matrix.ravel()[0:6]) | |
| h, w = shape[0:2] | |
| resample = self.pil_interp_lookup[interp] | |
| assert border_mode == 'constant', 'pil can only do constant right now' | |
| need_convert_uint8 = False | |
| if len(shape) > 2 and shape[2] >= 3: | |
| # TODO: fixme when floating point values are not in the 0-1 range | |
| if mode == 'float01': | |
| # PIL cannot handle multi-channel float images. | |
| # Need to convert to uint8 | |
| need_convert_uint8 = True | |
| clip = False | |
| def _pil_warper(img): | |
| if need_convert_uint8: | |
| orig_dtype = img.dtype | |
| img = (img * 255.0).astype(np.uint8) | |
| imaug = np.array(Image.fromarray(img).transform((w, h), | |
| Image.AFFINE, | |
| pil_aff_params, | |
| resample=resample)) | |
| if clip: | |
| np.clip(imaug, 0, 1, out=imaug) | |
| if need_convert_uint8: | |
| imaug = imaug.astype(orig_dtype) / 255.0 | |
| return imaug | |
| return _pil_warper | |
| def make_cv2_warper(self, shape, matrix, interp, border_mode, clip, mode): | |
| h, w = shape[0:2] | |
| borderMode = self.cv2_border_lookup[border_mode] | |
| cv2_interp = self.cv2_interp_lookup[interp] | |
| # It is slightly faster (5-10%) to pass an inverted matrix to cv2 | |
| inv_mat_2x3 = np.linalg.inv(matrix)[0:2] | |
| inv_flags = cv2_interp | cv2.WARP_INVERSE_MAP | |
| # TODO: only clip if we are in float01 space | |
| if clip: | |
| def _cv2_warper(img): | |
| imaug = cv2.warpAffine(img, inv_mat_2x3, dsize=(w, h), | |
| flags=inv_flags, borderMode=borderMode) | |
| np.clip(imaug, 0, 1, out=imaug) | |
| return imaug | |
| else: | |
| def _cv2_warper(img): | |
| imaug = cv2.warpAffine(img, inv_mat_2x3, dsize=(w, h), | |
| flags=inv_flags, borderMode=borderMode) | |
| return imaug | |
| return _cv2_warper | |
| class RandomWarpAffine(object): | |
| def __init__(self, rng=None, backend='skimage', **kw): | |
| self.rng = util.ensure_rng(rng) | |
| self.augkw = augment_common.PERTERB_AUG_KW.copy() | |
| self.augkw.update(kw) | |
| # self.interp = 'nearest' | |
| # self.border_mode = 'reflect' | |
| # self.backend = 'skimage' | |
| # self.backend = 'cv2' | |
| # self.backend = 'pil' | |
| self.border_mode = 'constant' | |
| self.interp = 'nearest' | |
| self.backend = backend | |
| # @ub.memoize | |
| # def _memo_empty_uint8(shape): | |
| # data = np.empty(shape, dtype=np.uint8) | |
| # return data | |
| def __call__(self, data): | |
| # raise NotImplementedError( | |
| # 'Use explicit calls to tie params between instances') | |
| params = self.random_params() | |
| return self.warp(data, params, interp=self.interp, | |
| border_mode=self.border_mode, backend=self.backend) | |
| def random_params(self): | |
| affine_args = augment_common.random_affine_args(rng=self.rng, | |
| **self.augkw) | |
| return affine_args | |
| def warp(self, img, params, interp='nearest', border_mode='constant', backend=None): | |
| """ | |
| >>> from clab.transforms import * | |
| >>> import plottool as pt | |
| >>> pt.qtensure() | |
| >>> img_orig = util.imread(ut.grab_test_imgpath('lena.png')) | |
| >>> def _test_backend(img, backend, fnum): | |
| >>> try: | |
| >>> pnum_ = pt.make_pnum_nextgen(nSubplots=20) | |
| >>> self = RandomWarpAffine(0, backend=backend) | |
| >>> for _ in range(20): | |
| >>> imgaug = self.warp(img, self.random_params()) | |
| >>> #pt.imshow(imgaug, fnum=fnum, pnum=pnum_()) | |
| >>> except Exception as ex: | |
| >>> print('{} failed {}'.format(backend, repr(ex))) | |
| >>> pass | |
| >>> def _test_all_backends(img): | |
| >>> _test_backend(img, backend='skimage', fnum=1) | |
| >>> _test_backend(img, backend='pil', fnum=2) | |
| >>> _test_backend(img, backend='cv2', fnum=3) | |
| >>> # --- RGB dtype uint8 --- | |
| >>> img = img_orig.copy() | |
| >>> _test_all_backends(img) | |
| >>> # --- RGB dtype float32 --- (PIL FAILS) | |
| >>> img = img_orig.copy().astype(np.float32) / 255 | |
| >>> _test_all_backends(img) | |
| >>> # --- GRAY dtype float32, 0-1 --- | |
| >>> img = (img_orig.astype(np.float32) / 255).mean(axis=2) | |
| >>> _test_all_backends(img) | |
| >>> # --- RGB dtype uint16 --- | |
| >>> img = ((img_orig.astype(np.float32) / 255) * (2 ** 16) - 1).astype(np.uint16) | |
| >>> _test_all_backends(img) | |
| >>> # --- GRAY dtype float32, -1000-1000 --- | |
| >>> img = (img_orig.astype(np.float32) / 255).mean(axis=2) * 2000 - 1000 | |
| >>> print(ut.get_stats(RandomWarpAffine(0, backend='pil')(img))) | |
| >>> print(ut.get_stats(RandomWarpAffine(0, backend='skimage')(img))) | |
| >>> print(ut.get_stats(RandomWarpAffine(0, backend='cv2')(img))) | |
| >>> _test_all_backends(img) | |
| SpeedPlots: | |
| >>> from clab.transforms import * | |
| >>> import plottool as pt | |
| >>> pt.qtensure() | |
| >>> img_orig = util.imread(ut.grab_test_imgpath('lena.png')) | |
| >>> img = (img_orig.astype(np.float32) / 255).mean(axis=2) * 2000 - 1000 | |
| >>> xdata = [2, 4, 8, 16, 32, 64, 80, 128, 192, 256, 384, 512, 768, 1024, 2048] | |
| >>> # | |
| >>> ub.Timerit.DEFAULT_VERBOSE = False | |
| >>> ydatas = ub.ddict(list) | |
| >>> N = 30 | |
| >>> for size in xdata: | |
| >>> img = cv2.resize(img_orig, (size, size)) | |
| >>> ydatas['ski'] += [ub.Timerit(N, bestof=3).call(RandomWarpAffine(0, backend='skimage'), img).min()] | |
| >>> ydatas['pil'] += [ub.Timerit(N, bestof=3).call(RandomWarpAffine(0, backend='pil'), img).min()] | |
| >>> ydatas['cv2'] += [ub.Timerit(N, bestof=3).call(RandomWarpAffine(0, backend='cv2'), img).min()] | |
| >>> pt.multi_plot(xdata, ydatas, title='affine warp speed (RGB uint8)', xlabel='image size', ylabel='seconds', fnum=99, ymax=.1) | |
| >>> pt.gca().set_yscale('log') | |
| >>> pt.gca().set_ylim(5e-5, 1e-1) | |
| >>> pt.gcf().savefig('rgb_uint8.png') | |
| >>> # | |
| >>> ydatas_f32 = ub.ddict(list) | |
| >>> for size in xdata: | |
| >>> img = cv2.resize(img_orig, (size, size)).astype(np.float32).mean(axis=2) | |
| >>> ydatas_f32['ski'] += [ub.Timerit(N, bestof=3).call(RandomWarpAffine(0, backend='skimage'), img).min()] | |
| >>> ydatas_f32['pil'] += [ub.Timerit(N, bestof=3).call(RandomWarpAffine(0, backend='pil'), img).min()] | |
| >>> ydatas_f32['cv2'] += [ub.Timerit(N, bestof=3).call(RandomWarpAffine(0, backend='cv2'), img).min()] | |
| >>> pt.multi_plot(xdata, ydatas_f32, title='affine warp speed (GRAY float32)', xlabel='image size', ylabel='seconds', fnum=100, ymax=.1) | |
| >>> pt.gca().set_yscale('log') | |
| >>> pt.gca().set_ylim(5e-5, 1e-1) | |
| >>> pt.gcf().savefig('gray_float32.png') | |
| >>> # | |
| >>> ydatas_f64 = ub.ddict(list) | |
| >>> for size in xdata: | |
| >>> img = cv2.resize(img_orig, (size, size)).astype(np.float64).mean(axis=2) | |
| >>> ydatas_f64['ski'] += [ub.Timerit(N, bestof=3).call(RandomWarpAffine(0, backend='skimage'), img).min()] | |
| >>> ydatas_f64['pil'] += [ub.Timerit(N, bestof=3).call(RandomWarpAffine(0, backend='pil'), img).min()] | |
| >>> ydatas_f64['cv2'] += [ub.Timerit(N, bestof=3).call(RandomWarpAffine(0, backend='cv2'), img).min()] | |
| >>> pt.multi_plot(xdata, ydatas_f64, title='affine warp speed (GRAY float64)', xlabel='image size', ylabel='seconds', fnum=101, ymax=.1) | |
| >>> pt.gca().set_yscale('log') | |
| >>> pt.gca().set_ylim(5e-5, 1e-1) | |
| >>> pt.gcf().savefig('gray_float64.png') | |
| >>> # | |
| >>> ydatas_f64 = ub.ddict(list) | |
| >>> for size in xdata: | |
| >>> img = (cv2.resize(img_orig, (size, size)).astype(np.float64).mean(axis=2) * 255).astype(np.uint8) | |
| >>> ydatas_f64['ski'] += [ub.Timerit(N, bestof=3).call(RandomWarpAffine(0, backend='skimage'), img).min()] | |
| >>> ydatas_f64['pil'] += [ub.Timerit(N, bestof=3).call(RandomWarpAffine(0, backend='pil'), img).min()] | |
| >>> ydatas_f64['cv2'] += [ub.Timerit(N, bestof=3).call(RandomWarpAffine(0, backend='cv2'), img).min()] | |
| >>> pt.multi_plot(xdata, ydatas_f64, title='affine warp speed (GRAY uint8)', xlabel='image size', ylabel='seconds', fnum=102, ymax=.1) | |
| >>> pt.gca().set_yscale('log') | |
| >>> pt.gca().set_ylim(5e-5, 1e-1) | |
| >>> pt.gcf().savefig('gray_uint8.png') | |
| Ignore: | |
| >>> %timeit RandomWarpAffine(0, backend='skimage')(img) | |
| 4.76 ms ± 86.8 µs per loop (mean ± std. dev. of 7 runs, 100 loops each) | |
| >>> %timeit RandomWarpAffine(0, backend='pil')(img) | |
| 1.11 ms ± 9.85 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) | |
| >>> %timeit RandomWarpAffine(0, backend='cv2')(img) | |
| 25.6 ms ± 498 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) | |
| >>> img = (img_orig.astype(np.float32) / 255).mean(axis=2) | |
| >>> ub.Timerit(50, bestof=5).call(RandomWarpAffine(0, backend='skimage'), img) | |
| time per loop : 5.0 ms ± 0.26 | |
| >>> ub.Timerit(50, bestof=5).call(RandomWarpAffine(0, backend='pil'), img) | |
| time per loop : 1.334 ms ± 0.19 | |
| >>> ub.Timerit(50, bestof=5).call(RandomWarpAffine(0, backend='cv2'), img) | |
| time per loop : 20.24 ms ± 5.4 | |
| >>> img = (img_orig.copy())[0:8, 0:8] | |
| >>> ub.Timerit(100, bestof=10).call(RandomWarpAffine(0, backend='skimage'), img) | |
| >>> ub.Timerit(100, bestof=10).call(RandomWarpAffine(0, backend='pil'), img) | |
| >>> ub.Timerit(100, bestof=10).call(RandomWarpAffine(0, backend='cv2'), img) | |
| """ | |
| x, y = img.shape[1] / 2, img.shape[0] / 2 | |
| if backend is None: | |
| backend = self.backend | |
| affwarp = AffineWarp() | |
| matrix = affwarp.around_mat3x3(x, y, *params) | |
| _warp = affwarp.make_warper(shape=img.shape, matrix=matrix, | |
| interp=interp, backend=backend, | |
| border_mode=border_mode) | |
| imaug = _warp(img) | |
| return imaug | |
| def sseg_warp(self, im, aux_channels, gt, border_mode='constant'): | |
| """ | |
| Specialized warping for semantic segmentation problems | |
| """ | |
| params = self.random_params() | |
| im_aug = self.warp(im, params, interp='cubic', border_mode=border_mode) | |
| aux_channels_aug = [ | |
| self.warp(aux, params, interp='nearest', border_mode=border_mode) | |
| for aux in aux_channels | |
| ] | |
| gt_aug = self.warp(gt, params, interp='nearest', border_mode=border_mode) | |
| return im_aug, aux_channels_aug, gt_aug | |
| # class RandomIntensity(object): | |
| # def __init__(self, rng): | |
| # self.rng = rng | |
| # def __call__(self, data): | |
| # if self.rng.rand() > .5: | |
| # gamma = self.rng.rand() * 2 + .5 | |
| # img = imutil.adjust_gamma(data, gamma=gamma) | |
| # if self.rng.rand() > .5: | |
| # k = self.rng.randint(1, 2) | |
| # img = cv2.blur(data, (k, k)) | |
| # return img | |
| class RandomBlur(object): | |
| """ | |
| TODO: use imgaug or something similar. | |
| k_pdf should be a class that samples from a distribution | |
| """ | |
| def __init__(self, k_pdf=(2, 4), freq=.5, rng=None): | |
| self.rng = rng | |
| self.freq = freq | |
| self.k_pdf = k_pdf | |
| def __call__(self, data): | |
| if self.rng.rand() < self.freq: | |
| k = self.rng.randint(*self.k_pdf) | |
| data = cv2.blur(data, (k, k)) | |
| return data | |
| class RandomGamma(object): | |
| def __init__(self, input_colorspace='RGB', gamma_pdf=(.5, 2.5), freq=.5, | |
| rng=None): | |
| self.rng = rng | |
| self.gamma_pdf = gamma_pdf | |
| self.input_colorspace = input_colorspace | |
| self.freq = freq | |
| def __call__(self, data): | |
| if self.rng.rand() < self.freq: | |
| mn, mx = self.gamma_pdf | |
| gamma = self.rng.rand() * (mx - mn) + mn | |
| if self.input_colorspace != 'RGB': | |
| if self.input_colorspace == 'LAB': | |
| data = cv2.cvtColor(data, cv2.COLOR_LAB2RGB) | |
| else: | |
| raise KeyError(self.input_colorspace) | |
| data = imutil.adjust_gamma(data, gamma=gamma) | |
| if self.input_colorspace != 'RGB': | |
| if self.input_colorspace == 'LAB': | |
| data = cv2.cvtColor(data, cv2.COLOR_RGB2LAB) | |
| return data |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment