def test_max_neighborhood_sizei_with_arr_out():
arr_in = np.arange(4*4*3).reshape((4, 4, 3)).astype(DTYPE)
neighborhood = (4, 3)
arr_out = np.empty((1, 2, 3)).astype(DTYPE)
arr_out = lcdnorm3(arr_in, neighborhood, arr_out=arr_out)
def test_stride():
arr_in = np.arange(4 * 4 * 3).reshape((4, 4, 3)).astype(DTYPE)
neighborhood = (4, 3)
arr_out = lcdnorm3(arr_in, neighborhood, stride=2)
assert arr_out.shape == (1, 1, 3)
def test_stride():
arr_in = np.arange(4*4*3).reshape((4, 4, 3)).astype(DTYPE)
neighborhood = (4, 3)
arr_out = lcdnorm3(arr_in, neighborhood, stride=2)
assert arr_out.shape == (1, 1, 3)
def test_max_neighborhood_size():
arr_in = np.arange(4 * 4 * 3).reshape((4, 4, 3)).astype(DTYPE)
neighborhood = (4, 3)
arr_out = lcdnorm3(arr_in, neighborhood)
assert arr_out.shape == (1, 2, 3)
def test_max_neighborhood_sizei_with_arr_out():
arr_in = np.arange(4 * 4 * 3).reshape((4, 4, 3)).astype(DTYPE)
neighborhood = (4, 3)
arr_out = np.empty((1, 2, 3)).astype(DTYPE)
arr_out = lcdnorm3(arr_in, neighborhood, arr_out=arr_out)
def test_max_neighborhood_size():
arr_in = np.arange(4*4*3).reshape((4, 4, 3)).astype(DTYPE)
neighborhood = (4, 3)
arr_out = lcdnorm3(arr_in, neighborhood)
assert arr_out.shape == (1, 2, 3)
def test_input_d_1_default():
arr_in = np.zeros((20, 30, 1), dtype=DTYPE)
neighborhood = 5, 5
arr_out = np.zeros((16, 26, 1), dtype=DTYPE)
np.random.seed(42)
arr_in[:] = np.random.randn(*arr_in.shape)
idx = [[4, 3], [20, 12]]
gt = np.array([[0.23064923], [0.20238316]], dtype=DTYPE)
lcdnorm3(arr_in, neighborhood, arr_out=arr_out)
gv = arr_out[idx]
assert_allclose(gv, gt, rtol=RTOL, atol=ATOL)
arr_out = lcdnorm3(arr_in, neighborhood)
gv = arr_out[idx]
assert_allclose(gv, gt, rtol=RTOL, atol=ATOL)
def test_input_d_1_default():
arr_in = np.zeros((20, 30, 1), dtype=DTYPE)
neighborhood = 5, 5
arr_out = np.zeros((16, 26, 1), dtype=DTYPE)
np.random.seed(42)
arr_in[:] = np.random.randn(*arr_in.shape)
idx = [[4, 3], [20, 12]]
gt = np.array([[0.23064923],
[0.20238316]], dtype=DTYPE)
lcdnorm3(arr_in, neighborhood, arr_out=arr_out)
gv = arr_out[idx]
assert_allclose(gv, gt, rtol=RTOL, atol=ATOL)
arr_out = lcdnorm3(arr_in, neighborhood)
gv = arr_out[idx]
assert_allclose(gv, gt, rtol=RTOL, atol=ATOL)
def test_lena_npy_array():
arr_in = lena()[::32, ::32].astype(DTYPE)
arr_in.shape = arr_in.shape[:2] + (1, )
neighborhood = 3, 3
idx = [[4, 2], [4, 2]]
gt = np.array([[-0.53213698], [-0.0816204]], dtype=DTYPE)
arr_out = lcdnorm3(arr_in, neighborhood)
gv = arr_out[idx]
assert_allclose(gv, gt, rtol=RTOL, atol=ATOL)
def test_input_d_4_default():
arr_in = np.zeros((20, 30, 4), dtype=DTYPE)
neighborhood = 5, 5
arr_out = np.zeros((16, 26, 4), dtype=DTYPE)
np.random.seed(42)
arr_in[:] = np.random.randn(np.prod(arr_in.shape)).reshape(arr_in.shape)
idx = [[4, 3], [20, 12]]
gt = np.array(
[[+1.27813682e-01, -9.97862518e-02, -2.48777084e-02, -5.16409911e-02],
[-2.00690944e-02, -2.42322776e-02, 7.76741435e-05, -3.73861268e-02]],
dtype=DTYPE)
lcdnorm3(arr_in, neighborhood, arr_out=arr_out)
gv = arr_out[idx]
assert_allclose(gv, gt, rtol=RTOL, atol=ATOL)
arr_out = lcdnorm3(arr_in, neighborhood)
gv = arr_out[idx]
assert_allclose(gv, gt, rtol=RTOL, atol=ATOL)
def test_lena_npy_array():
arr_in = lena()[::32, ::32].astype(DTYPE)
arr_in.shape = arr_in.shape[:2] + (1,)
neighborhood = 3, 3
idx = [[4, 2], [4, 2]]
gt = np.array([[-0.53213698], [-0.0816204]],
dtype=DTYPE)
arr_out = lcdnorm3(arr_in, neighborhood)
gv = arr_out[idx]
assert_allclose(gv, gt, rtol=RTOL, atol=ATOL)
def test_input_d_4_default():
arr_in = np.zeros((20, 30, 4), dtype=DTYPE)
neighborhood = 5, 5
arr_out = np.zeros((16, 26, 4), dtype=DTYPE)
np.random.seed(42)
arr_in[:] = np.random.randn(np.prod(arr_in.shape)).reshape(arr_in.shape)
idx = [[4, 3], [20, 12]]
gt = np.array([[+1.27813682e-01, -9.97862518e-02,
-2.48777084e-02, -5.16409911e-02],
[-2.00690944e-02, -2.42322776e-02,
7.76741435e-05, -3.73861268e-02]],
dtype=DTYPE)
lcdnorm3(arr_in, neighborhood, arr_out=arr_out)
gv = arr_out[idx]
assert_allclose(gv, gt, rtol=RTOL, atol=ATOL)
arr_out = lcdnorm3(arr_in, neighborhood)
gv = arr_out[idx]
assert_allclose(gv, gt, rtol=RTOL, atol=ATOL)
def test_smoke_max_neighborhood_size():
np.random.seed(42)
arr_in = np.random.randn(4, 4, 3).astype(DTYPE)
neighborhood = (4, 3)
arr_out = lcdnorm3(arr_in, neighborhood)
idx = (0, 1, 2)
ref = 0.0123399
res = arr_out[idx]
assert (ref - res) < ATOL
def test_smoke_max_neighborhood_size():
np.random.seed(42)
arr_in = np.random.randn(4, 4, 3).astype(DTYPE)
neighborhood = (4, 3)
arr_out = lcdnorm3(arr_in, neighborhood)
idx = (0, 1, 2)
ref = 0.0123399
res = arr_out[idx]
assert (ref - res) < ATOL
def slm(arr_in):
assert arr_in.ndim == 3
inh, inw, ind = arr_in.shape
fbs = []
fbd = arr_in.shape[-1]
for nf in nfs:
fbshape = nb + (fbd, nf)
print 'generating', fbshape, 'filterbank'
fb = np.random.randn(*fbshape).astype('f')
fbs += [fb]
fbd = nf
#for nf in nfs:
for fb in fbs:
n1 = lcdnorm3(arr_in, nb, threshold=1.0)
f1 = fbcorr(n1, fb)
p1 = lpool3(f1, nb, order=2, stride=2)
arr_in = p1
return p1
def slm(arr_in):
assert arr_in.ndim == 3
inh, inw, ind = arr_in.shape
fbs = []
fbd = arr_in.shape[-1]
for nf in nfs:
fbshape = nb + (fbd, nf)
print 'generating', fbshape, 'filterbank'
fb = np.random.randn(*fbshape).astype('f')
fbs += [fb]
fbd = nf
#for nf in nfs:
for fb in fbs:
n1 = lcdnorm3(arr_in, nb, threshold=1.0)
f1 = fbcorr(n1, fb)
p1 = lpool3(f1, nb, order=2, stride=2)
arr_in = p1
return p1
def _get_feature_map(self, tmp_in,
layer_idx, op_idx, kwargs,
op_params, op_name):
if op_name == 'lnorm':
inker_shape = kwargs['inker_shape']
outker_shape = kwargs['outker_shape']
remove_mean = kwargs['remove_mean']
stretch = kwargs['stretch']
threshold = kwargs['threshold']
# SLM PLoS09 / FG11 constraints:
assert inker_shape == outker_shape
tmp_out = lcdnorm3(tmp_in, inker_shape,
contrast=remove_mean,
stretch=stretch,
threshold=threshold)
elif op_name == 'fbcorr':
max_out = kwargs['max_out']
min_out = kwargs['min_out']
fbkey = layer_idx, op_idx
if fbkey not in self.filterbanks:
initialize = op_params['initialize']
if isinstance(initialize, np.ndarray):
fb = initialize
if len(fb.shape) == 3:
fb = fb[..., np.newaxis]
else:
filter_shape = list(initialize['filter_shape'])
generate = initialize['generate']
n_filters = initialize['n_filters']
fb_shape = [n_filters] + filter_shape + [tmp_in.shape[-1]]
# generate filterbank data
method_name, method_kwargs = generate
assert method_name == 'random:uniform'
rseed = method_kwargs.get('rseed', None)
rng = np.random.RandomState(rseed)
fb = rng.uniform(size=fb_shape)
for fidx in xrange(n_filters):
filt = fb[fidx]
# zero-mean, unit-l2norm
filt -= filt.mean()
filt_norm = np.linalg.norm(filt)
assert filt_norm != 0
filt /= filt_norm
fb[fidx] = filt
fb = np.ascontiguousarray(np.rollaxis(fb, 0, 4)).astype(DTYPE)
self.filterbanks[fbkey] = fb
print fb.shape
fb = self.filterbanks[fbkey]
# -- filter
assert tmp_in.dtype == np.float32
tmp_out = fbcorr3(tmp_in, fb)
# -- activation
min_out = -np.inf if min_out is None else min_out
max_out = +np.inf if max_out is None else max_out
# insure that the type is right before calling numexpr
min_out = np.array([min_out], dtype=tmp_in.dtype)
max_out = np.array([max_out], dtype=tmp_in.dtype)
# call numexpr
tmp_out = ne.evaluate('where(tmp_out < min_out, min_out, tmp_out)')
tmp_out = ne.evaluate('where(tmp_out > max_out, max_out, tmp_out)')
assert tmp_out.dtype == tmp_in.dtype
elif op_name == 'lpool':
ker_shape = kwargs['ker_shape']
order = kwargs['order']
stride = kwargs['stride']
tmp_out = lpool3(tmp_in, ker_shape, order=order, stride=stride)
else:
raise ValueError("operation '%s' not understood" % op_name)
assert tmp_out.dtype == tmp_in.dtype
assert tmp_out.dtype == np.float32
return tmp_out