"""Packed random-convolution parity with the previous release backend.""" from unittest import TestCase, skipIf from unittest.mock import patch import pickle import warnings import numpy as np from wisp_fast import features as fx @skipIf(fx._cis_conv is None or fx._kernels is None, 'native convolution unavailable') class TestPackedConvolution(TestCase): def test_exact_tails_and_pairwise_reductions(self): rng = np.random.default_rng(707) cases = [15, 16, 17, 18, 19, 23, 32, 35, 37, 63, 96, 126, 127, 128, 129, 130, 131, 159, 160, 161, 259, 2049] for index, length in enumerate(cases): for c in [1, 2, 3, 5]: with self.subTest(length=length, channels=c): x = rng.normal(size=(3, length, c)) if index % 4 == 0: x = np.rint(x) if index % 3 == 0: x[0, 3, 0] = np.nan x[2, :, -1] = 0. for turbo in [False, True]: model = fx.RandomConvSketch(n_kernels=16, kernel_lengths=(1, 5, 7, 9, 15), fast_profile=turbo, random_state=312) model.fit(x) with patch.object(fx, '_cis_conv', None): old = model.transform(x) with warnings.catch_warnings(): warnings.simplefilter('ignore', RuntimeWarning) out = model.transform(x) np.testing.assert_array_equal(out, old) def test_validation_and_fallback(self): x = np.random.default_rng(874).normal(size=(3, 32, 3)) m = fx.RandomConvSketch(n_kernels=8).fit(x) expected = m.transform(x) with patch.object(fx, '_verified_conv_blas', return_value=False): np.testing.assert_array_equal(m.transform(x), expected) kernel = m.kernels_[0] kernel['weights'] *= 0. kernel['bias'] = 0. m.invalidate_plan() np.testing.assert_array_equal(m.transform(x)[:, :4], np.zeros((3,4))) p = pickle.loads(pickle.dumps(m)) self.assertNotIn('_conv_plan', vars(p)) np.testing.assert_array_equal(m.transform(x), p.transform(x)) layout, weights, bias, _, _ = m._conv_plan[1] for wrong in ('channel', 'length', 'stride', 'output', 'dtype'): l = layout.copy() if wrong == 'channel': l[0,3] = 4 if wrong == 'length': l[1,0] = 0 out = np.empty((3,8,5 if wrong == 'output' else 4)) with self.assertRaises((ValueError, BufferError)): fx._cis_conv.transform(np.ascontiguousarray(x.transpose(0,2,1)),l, weights.astype('f4') if wrong=='dtype' else weights, bias,40 if wrong=='stride' else 1,False,False,out)