def main(model_dir, train_dir, dev_dir,
is_runtime=False,
nr_hidden=64, max_length=100, # Shape
dropout=0.5, learn_rate=0.001, # General NN config
nb_epoch=5, batch_size=32, nr_examples=-1): # Training params
model_dir = pathlib.Path(model_dir)
train_dir = pathlib.Path(train_dir)
dev_dir = pathlib.Path(dev_dir)
if is_runtime:
dev_texts, dev_labels = read_data(dev_dir)
acc = evaluate(model_dir, dev_texts, dev_labels, max_length=max_length)
print(acc)
else:
print("Read data")
train_texts, train_labels = read_data(train_dir, limit=nr_examples)
dev_texts, dev_labels = read_data(dev_dir, limit=nr_examples)
print("Using GPU 0")
#chainer.cuda.get_device(0).use()
train_labels = xp.asarray(train_labels, dtype='i')
dev_labels = xp.asarray(dev_labels, dtype='i')
lstm = train(train_texts, train_labels, dev_texts, dev_labels,
{'nr_hidden': nr_hidden, 'max_length': max_length, 'nr_class': 2,
'nr_vector': 2000, 'nr_dim': 32},
{'dropout': 0.5, 'lr': learn_rate},
{},
nb_epoch=nb_epoch, batch_size=batch_size)
def _convert_array(xs, array_module):
if array_module == 'all_numpy':
return xs
elif array_module == 'all_cupy':
return cupy.asarray(xs)
else:
return [cupy.asarray(x) if numpy.random.random_integers(0, 1)
else x for x in xs]
def _array_to_gpu(array, device, stream):
if array is None:
return None
if isinstance(array, chainerx.ndarray):
# TODO(niboshi): Update this logic once both CuPy and ChainerX support
# the array interface.
if array.device.backend.name == 'cuda':
# Convert to cupy.ndarray on the same device as source array
array = cupy.ndarray(
array.shape,
array.dtype,
cupy.cuda.MemoryPointer(
cupy.cuda.UnownedMemory(
array.data_ptr + array.offset,
array.data_size,
array,
array.device.index),
0),
strides=array.strides)
else:
array = chainerx.to_numpy(array)
elif isinstance(array, (numpy.number, numpy.bool_)):
array = numpy.asarray(array)
elif isinstance(array, intel64.mdarray):
array = numpy.asarray(array)
if isinstance(array, ndarray):
if array.device == device:
return array
is_numpy = False
elif isinstance(array, numpy.ndarray):
is_numpy = True
else:
raise TypeError(
'The array sent to gpu must be an array or a NumPy scalar.'
'\nActual type: {0}.'.format(type(array)))
if stream is not None:
with device:
with stream:
if is_numpy:
return cupy.asarray(array)
# Need to make a copy when an array is copied to another device
return cupy.array(array, copy=True)
with device:
if is_numpy:
return cupy.asarray(array)
# Need to make a copy when an array is copied to another device
return cupy.array(array, copy=True)
def _fftconv(a, b, axes=(0, 1)):
"""Patched version of :func:`sporco.linalg.fftconv`."""
if cp.isrealobj(a) and cp.isrealobj(b):
fft = cp.fft.rfftn
ifft = cp.fft.irfftn
else:
fft = cp.fft.fftn
ifft = cp.fft.ifftn
dims = cp.maximum(cp.asarray([a.shape[i] for i in axes]),
cp.asarray([b.shape[i] for i in axes]))
dims = [int(d) for d in dims]
af = fft(a, dims, axes)
bf = fft(b, dims, axes)
return ifft(af * bf, dims, axes)
def to_gpu(array, device=None, stream=None):
"""Copies the given CPU array to specified device.
Args:
array: Array to be sent to GPU.
device: Device specifier.
stream (cupy.cuda.Stream): CUDA stream.
Returns:
cupy.ndarray: Array on GPU.
If ``array`` is already on GPU, then this function just returns
``array`` without performing any copy. Note that this function does not
copy :class:`cupy.ndarray` into specified device.
"""
check_cuda_available()
assert stream is None # TODO(beam2d): FIX IT
with get_device(device):
dev_id = int(get_device(array))
if dev_id != -1 and dev_id != cupy.cuda.device.get_device_id():
# Need to make a copy when an array is copied to another device
return cupy.array(array, copy=True)
else:
return cupy.asarray(array)
def setup_method(self, method):
np.random.seed(12345)
N = 32
self.U = cp.ones((N, N, N))
self.U[:, 0:(old_div(N, 2)), :] = -1
self.V = 1e-1 * cp.asarray(np.random.randn(N, N, N))
self.D = self.U + self.V
def _list2array(lst):
"""Convert a list to a numpy array."""
if lst and isinstance(lst[0], cp.ndarray):
return cp.hstack(lst)
else:
return cp.asarray(lst)
def check(self, func, n, gen, *args):
@cupy.fuse(input_num=n)
def f(*x):
return func(*x)
if type(gen) == tuple:
ndata = [g(*a) for i, g, a in zip(range(n), list(gen), args)]
else:
ndata = [gen(*args) for i in range(n)]
nret = func(*ndata)
fnret = f(*ndata)
nret = list(nret) if type(nret) == tuple else [nret]
fnret = list(fnret) if type(fnret) == tuple else [fnret]
for n, fn in zip(nret, fnret):
numpy.testing.assert_array_almost_equal(n, fn)
cdata = [cupy.asarray(_) for _ in ndata]
cret = func(*cdata)
fcret = f(*cdata)
cret = list(cret) if type(cret) == tuple else [cret]
fcret = list(fcret) if type(fcret) == tuple else [fcret]
for n, c, fc in zip(nret, cret, fcret):
numpy.testing.assert_array_almost_equal(n, c.get())
numpy.testing.assert_array_almost_equal(n, fc.get())
def to_cupy(array): # pragma: no cover
import cupy
if isinstance(array, np.ndarray):
return cupy.asarray(array)
return array
def _get_labelled_sentences(self, docs, doc_labels):
labels = []
sentences = []
for doc, y in izip(docs, doc_labels):
for sent in doc.sents:
sentences.append(sent)
labels.append(y)
return sentences, xp.asarray(labels, dtype='i')
def to_gpu(array, device=None, stream=None):
"""Copies the given CPU array to specified device.
Args:
array: Array to be sent to GPU.
device: Device specifier.
stream (cupy.cuda.Stream): CUDA stream. If not ``None``, the copy runs
asynchronously.
Returns:
cupy.ndarray: Array on GPU.
If ``array`` is already on GPU, then this function just returns
``array`` without performing any copy. Note that this function does not
copy :class:`cupy.ndarray` into specified device.
"""
check_cuda_available()
with _get_device(device):
array_dev = get_device_from_array(array)
if array_dev.id == cupy.cuda.device.get_device_id():
return array
if stream is not None:
warnings.warn(
'The stream option is deprecated in chainer.cuda.to_gpu. '
'Please remove it.', DeprecationWarning)
if stream.ptr != 0:
ret = cupy.empty_like(array)
if array_dev.id == -1:
# cpu to gpu
mem = cupy.cuda.alloc_pinned_memory(array.nbytes)
src = numpy.frombuffer(
mem, array.dtype, array.size).reshape(array.shape)
src[...] = array
ret.set(src, stream)
cupy.cuda.pinned_memory._add_to_watch_list(
stream.record(), mem)
else:
# gpu to gpu
with array_dev:
src = array.copy()
event = Stream.null.record()
stream.wait_event(event)
ret.data.copy_from_device_async(
src.data, src.nbytes, stream)
# to hold a reference until the end of the asynchronous
# memcpy
stream.add_callback(lambda *x: None, (src, ret))
return ret
if array_dev.id == -1:
return cupy.asarray(array)
# Need to make a copy when an array is copied to another device
return cupy.array(array, copy=True)
def asanyarray(a, dtype=None):
"""Converts an object to array.
This is equivalent to cupy.asarray.
.. seealso:: :func:`cupy.asarray`, :func:`numpy.asanyarray`
"""
return cupy.asarray(a, dtype)
def check_usv(self, array, dtype):
a_cpu = numpy.asarray(array, dtype=dtype)
a_gpu = cupy.asarray(array, dtype=dtype)
result_cpu = numpy.linalg.svd(a_cpu, full_matrices=self.full_matrices)
result_gpu = cupy.linalg.svd(a_gpu, full_matrices=self.full_matrices)
self.assertEqual(len(result_cpu), len(result_gpu))
for b_cpu, b_gpu in zip(result_cpu, result_gpu):
# Use abs to support an inverse vector
cupy.testing.assert_allclose(
numpy.abs(b_cpu), cupy.abs(b_gpu), atol=1e-4)
def check_reduce(self, func, n, reduce_f, gen, *args):
@cupy.fuse(input_num=n, reduce=reduce_f)
def f(*x):
return func(*x)
ndata = [gen(*args) for i in range(n)]
fnret = f(*ndata)
cdata = [cupy.asarray(_) for _ in ndata]
fcret = f(*cdata)
numpy.testing.assert_array_almost_equal(fnret, fcret.get())
def check_mode(self, array, mode, dtype):
a_cpu = numpy.asarray(array, dtype=dtype)
a_gpu = cupy.asarray(array, dtype=dtype)
result_cpu = numpy.linalg.qr(a_cpu, mode=mode)
result_gpu = cupy.linalg.qr(a_gpu, mode=mode)
if isinstance(result_cpu, tuple):
for b_cpu, b_gpu in six.moves.zip(result_cpu, result_gpu):
self.assertEqual(b_cpu.dtype, b_gpu.dtype)
cupy.testing.assert_allclose(b_cpu, b_gpu, atol=1e-4)
else:
self.assertEqual(result_cpu.dtype, result_gpu.dtype)
cupy.testing.assert_allclose(result_cpu, result_gpu, atol=1e-4)
def asanyarray(a, dtype=None):
"""Converts an object to array.
This is currently equivalent to :func:`~cupy.asarray`, since there is no
subclass of ndarray in CuPy. Note that the original
:func:`numpy.asanyarray` returns the input array as is if it is an instance
of a subtype of numpy.ndarray.
.. seealso:: :func:`cupy.asarray`, :func:`numpy.asanyarray`
"""
return cupy.asarray(a, dtype)
def _array_to_gpu(array, device, stream):
assert device is DummyDevice or isinstance(device, Device)
if array is None:
return None
if isinstance(array, (numpy.number, numpy.bool_)):
array = numpy.asarray(array)
elif isinstance(array, intel64.mdarray):
array = numpy.asarray(array)
if not isinstance(array, (cupy.ndarray, numpy.ndarray)):
raise TypeError(
'The array sent to gpu must be an array or a NumPy scalar.'
'\nActual type: {0}.'.format(type(array)))
array_dev = get_device_from_array(array)
if array_dev.id == cupy.cuda.device.get_device_id():
return array
if stream is not None and stream.ptr != 0:
ret = cupy.empty_like(array)
if array_dev.id == -1:
# cpu to gpu
mem = cupy.cuda.alloc_pinned_memory(array.nbytes)
src = numpy.frombuffer(
mem, array.dtype, array.size).reshape(array.shape)
src[...] = array
ret.set(src, stream)
cupy.cuda.pinned_memory._add_to_watch_list(
stream.record(), mem)
else:
# gpu to gpu
with array_dev:
src = array.copy()
event = Stream.null.record()
stream.wait_event(event)
ret.data.copy_from_device_async(
src.data, src.nbytes, stream)
# to hold a reference until the end of the asynchronous
# memcpy
stream.add_callback(lambda *x: None, (src, ret))
return ret
if array_dev.id == -1:
return cupy.asarray(array)
# Need to make a copy when an array is copied to another device
return cupy.array(array, copy=True)
def ix_(*args):
"""Construct an open mesh from multiple sequences.
This function takes N 1-D sequences and returns N outputs with N
dimensions each, such that the shape is 1 in all but one dimension
and the dimension with the non-unit shape value cycles through all
N dimensions.
Using `ix_` one can quickly construct index arrays that will index
the cross product. ``a[cupy.ix_([1,3],[2,5])]`` returns the array
``[[a[1,2] a[1,5]], [a[3,2] a[3,5]]]``.
Args:
*args: 1-D sequences
Returns:
tuple of ndarrays:
N arrays with N dimensions each, with N the number of input sequences.
Together these arrays form an open mesh.
Examples
--------
>>> a = cupy.arange(10).reshape(2, 5)
>>> a
array([[0, 1, 2, 3, 4],
[5, 6, 7, 8, 9]])
>>> ixgrid = cupy.ix_([0,1], [2,4])
>>> ixgrid
(array([[0],
[1]]), array([[2, 4]]))
.. seealso:: :func:`numpy.ix_`
"""
out = []
nd = len(args)
for k, new in enumerate(args):
new = cupy.asarray(new)
if new.ndim != 1:
raise ValueError("Cross index must be 1 dimensional")
if new.size == 0:
# Explicitly type empty arrays to avoid float default
new = new.astype(numpy.intp)
if cupy.issubdtype(new.dtype, cupy.bool_):
new, = new.nonzero()
new = new.reshape((1,) * k + (new.size,) + (1,) * (nd - k - 1))
out.append(new)
return tuple(out)
def to_gpu(array, device=None, stream=None):
"""Copies the given CPU array to specified device.
Args:
array: Array to be sent to GPU.
device: Device specifier.
stream (cupy.cuda.Stream): CUDA stream. If not ``None``, the copy runs
asynchronously.
Returns:
cupy.ndarray: Array on GPU.
If ``array`` is already on GPU, then this function just returns
``array`` without performing any copy. Note that this function does not
copy :class:`cupy.ndarray` into specified device.
"""
check_cuda_available()
with get_device(device):
array_dev = get_device(array)
if array_dev.id == cupy.cuda.device.get_device_id():
return array
if stream is not None:
ret = cupy.empty_like(array)
if array_dev.id == -1:
# cpu to gpu
src = array.copy(order='C')
ret.set(src, stream)
else:
# gpu to gpu
with array_dev:
src = array.copy()
ret.data.copy_from_device_async(src.data, src.nbytes, stream)
# to hold a reference until the end of the asynchronous memcpy
stream.add_callback(lambda *x: None, (src, ret))
return ret
if array_dev.id == -1:
return cupy.asarray(array)
# Need to make a copy when an array is copied to another device
return cupy.array(array, copy=True)
def to_gpu(array, device=None, stream=None):
"""Copies the given CPU array to specified device.
Args:
array: Array to be sent to GPU.
device: Device specifier.
stream (cupy.cuda.Stream): CUDA stream.
Returns:
cupy.ndarray: Array on GPU.
If ``array`` is already on GPU, then this function just returns
``array`` without performing any copy. Note that this function does not
copy cupy.ndarray into specified device.
"""
_check_cuda_available()
assert stream is None # TODO(beam2d): FIX IT
with get_device(device):
return cupy.asarray(array)
def affine(volume: np.ndarray,
transform_m: np.ndarray,
interpolation: str = 'linear',
reshape: bool = False,
profile: bool = False,
output=None,
device: str = 'cpu'):
if device not in AVAILABLE_DEVICES:
raise ValueError(
f'Unknown device ({device}), must be one of {AVAILABLE_DEVICES}')
if device == 'cpu':
if profile:
t_start = time.time()
# set parameters for scipy affine transform
if interpolation == 'linear':
order = 1
else:
order = 3
if not interpolation.startswith('filt_bspline'):
prefilter = False
else:
prefilter = True
if reshape:
pad_before, pad_after, output_shape = utils.compute_post_transform_dimensions(
volume.shape, transform_m)
# scipy will take care of padding in this case
# but we need to apply pad_before offset to transform_m get full volume
transform_m = np.dot(
transform_m, translation_matrix(pad_before, transform_m.dtype))
else:
output_shape = volume.shape
# run affine transformation
output_vol = affine_transform(volume,
transform_m,
output_shape=output_shape,
output=output,
order=order,
prefilter=prefilter)
if profile:
t_end = time.time()
time_took = (t_end - t_start) * 1000
print(f'transform finished in {time_took:.3f}ms')
if output is not None:
return output
else:
return output_vol
elif device.startswith('gpu'):
utils.switch_to_device(device)
if profile:
stream = cp.cuda.Stream.null
t_start = stream.record()
if reshape:
pad_before, pad_after, output_shape = utils.compute_post_transform_dimensions(
volume.shape, transform_m)
# manually pad volume
volume = np.pad(volume,
list(zip(pad_before, pad_after)),
mode='constant')
# include pad_before offset: first apply offset, then apply negative offset
transform_m = translation_matrix(
-1 * pad_before) @ transform_m @ translation_matrix(pad_before)
volume = cp.asarray(volume)
volume_shape = volume.shape
# texture setup
ch = cp.cuda.texture.ChannelFormatDescriptor(
32, 0, 0, 0, cp.cuda.runtime.cudaChannelFormatKindFloat)
arr = cp.cuda.texture.CUDAarray(
ch, *volume_shape[::-1]
) # CUDAArray: last dimension=fastest changing dimension
res = cp.cuda.texture.ResourceDescriptor(
cp.cuda.runtime.cudaResourceTypeArray, cuArr=arr)
tex = cp.cuda.texture.TextureDescriptor(
(cp.cuda.runtime.cudaAddressModeBorder,
cp.cuda.runtime.cudaAddressModeBorder,
cp.cuda.runtime.cudaAddressModeBorder),
cp.cuda.runtime.cudaFilterModeLinear,
cp.cuda.runtime.cudaReadModeElementType)
texobj = cp.cuda.texture.TextureObject(res, tex)
# prefilter if required and upload to texture
if interpolation.startswith('filt_bspline'):
volume = _bspline_prefilter(volume)
arr.copy_from(volume)
else:
arr.copy_from(volume)
# kernel setup
kernel = _get_transform_kernel(interpolation)
dims = cp.asarray(volume_shape, dtype=cp.uint32)
xform = cp.asarray(transform_m)
dim_grid, dim_blocks = utils.compute_elementwise_launch_dims(
volume_shape)
if output is None:
volume.fill(0.0) # reuse input array
else:
volume = output
kernel(dim_grid, dim_blocks, (volume, texobj, xform, dims))
if profile:
t_end = stream.record()
t_end.synchronize()
time_took = cp.cuda.get_elapsed_time(t_start, t_end)
print(f'transform finished in {time_took:.3f}ms')
if output is None:
del texobj, xform, dims
return volume.get()
else:
del texobj, xform, dims
return None
else:
raise ValueError(f'No instructions for {device}.')
def test_compare_xp_gpu(self):
noisyimg_gpu = cp.asarray(self.noisyimg)
imgivar_gpu = cp.asarray(self.imgivar)
A4_gpu = cp.asarray(self.A4)
# Compare the "signal" decorrelation method
flux0, ivar0, R0 = ex2d_patch(self.noisyimg,
self.imgivar,
self.A4,
decorrelate='signal')
flux1_gpu, ivar1_gpu, R1_gpu = xp_ex2d_patch(noisyimg_gpu,
imgivar_gpu,
A4_gpu,
decorrelate='signal')
flux1 = cp.asnumpy(flux1_gpu)
ivar1 = cp.asnumpy(ivar1_gpu)
R1 = cp.asnumpy(R1_gpu)
eps_double = np.finfo(np.float64).eps
where = np.where(
~np.isclose(flux0, flux1, rtol=1e5 * eps_double, atol=0))
np.testing.assert_allclose(flux0,
flux1,
rtol=1e5 * eps_double,
atol=0,
err_msg=f"where: {where}")
self.assertTrue(
np.allclose(ivar0, ivar1, rtol=1e3 * eps_double, atol=0))
self.assertTrue(
np.allclose(np.diag(R0),
np.diag(R1),
rtol=1e2 * eps_double,
atol=1e3 * eps_double))
self.assertTrue(
np.allclose(
np.abs(flux0 - flux1) / np.sqrt(1. / ivar0 + 1. / ivar1),
np.zeros_like(flux0)))
# Compare the "noise" decorrelation method
flux0, ivar0, R0 = ex2d_patch(self.noisyimg,
self.imgivar,
self.A4,
decorrelate='noise')
flux1_gpu, ivar1_gpu, R1_gpu = xp_ex2d_patch(noisyimg_gpu,
imgivar_gpu,
A4_gpu,
decorrelate='noise')
flux1 = cp.asnumpy(flux1_gpu)
ivar1 = cp.asnumpy(ivar1_gpu)
R1 = cp.asnumpy(R1_gpu)
self.assertTrue(
np.allclose(flux0, flux1, rtol=1e5 * eps_double, atol=0))
self.assertTrue(
np.allclose(ivar0, ivar1, rtol=1e3 * eps_double, atol=0))
self.assertTrue(
np.allclose(np.diag(R0),
np.diag(R1),
rtol=1e2 * eps_double,
atol=0))
self.assertTrue(
np.allclose(
np.abs(flux0 - flux1) / np.sqrt(1. / ivar0 + 1. / ivar1),
np.zeros_like(flux0)))
def evaluate(self):
self.load_from_file()
# Perform MC reweighting
if _GPU_ENABLED:
import cupy as xp
else:
import numpy as xp
m1_src = self.fiducial_binaries["mass_1_source"]
m2_src = self.fiducial_binaries["mass_2_source"]
spin_1x, spin_1y, spin_1z = [
self.fiducial_binaries[k]
for k in ["spin_1x", "spin_1y", "spin_1z"]
]
spin_2x, spin_2y, spin_2z = [
self.fiducial_binaries[k]
for k in ["spin_2x", "spin_2y", "spin_2z"]
]
pdf_mass_fiducial = self.pdf_mass_fiducial
pdf_spin_fiducial = self.pdf_spin_fiducial
# Move data to GPU if needed
m1_src = xp.asarray(m1_src)
m2_src = xp.asarray(m2_src)
spin_1x = xp.asarray(spin_1x)
spin_1y = xp.asarray(spin_1y)
spin_1z = xp.asarray(spin_1z)
spin_2x = xp.asarray(spin_2x)
spin_2y = xp.asarray(spin_2y)
spin_2z = xp.asarray(spin_2z)
pdf_mass_fiducial = xp.asarray(pdf_mass_fiducial)
pdf_spin_fiducial = xp.asarray(pdf_spin_fiducial)
pdf_mass_pop = self.mass_src_pop_model.prob(
{
"mass_1_source": m1_src,
"mass_2_source": m2_src
}, axis=0)
weights_mass = pdf_mass_pop / pdf_mass_fiducial
pdf_spin_pop = self.spin_src_pop_model.prob({
"spin_1x": spin_1x,
"spin_1y": spin_1y,
"spin_1z": spin_1z,
"spin_2x": spin_2x,
"spin_2y": spin_2y,
"spin_2z": spin_2z,
})
weights_spin = pdf_spin_pop / pdf_spin_fiducial
weights_source = weights_mass * weights_spin
z = self.fiducial_z
pz = NotLensedSourceRedshiftProbDist(
merger_rate_density=self.merger_rate_density_src_pop_model,
optical_depth=self.optical_depth)
pdf_z_fiducial = self.pdf_z_fiducial
pdf_z_pop = pz.prob(z) # NOTE p_z still uses CPU-only code
weights_z = xp.asarray(pdf_z_pop / pdf_z_fiducial)
predictions = xp.asarray(self.predictions)
alpha = xp.sum(predictions * weights_source *
weights_z).astype(float) / (float(self.N_inj))
self.f.close()
# NOTE If using numpy, alpha is a scalar but if using cupy, alpha is a 0-d array
return float(alpha)
def svd_wrapper(matrix, mode, ncomp, verbose, full_output=False,
random_state=None, to_numpy=True):
""" Wrapper for different SVD libraries (CPU and GPU).
Parameters
----------
matrix : numpy ndarray, 2d
2d input matrix.
mode : {'lapack', 'arpack', 'eigen', 'randsvd', 'cupy', 'eigencupy',
'randcupy', 'pytorch', 'eigenpytorch', 'randpytorch'}, str optional
Switch for the SVD method/library to be used.
``lapack``: uses the LAPACK linear algebra library through Numpy
and it is the most conventional way of computing the SVD
(deterministic result computed on CPU).
``arpack``: uses the ARPACK Fortran libraries accessible through
Scipy (computation on CPU).
``eigen``: computes the singular vectors through the
eigendecomposition of the covariance M.M' (computation on CPU).
``randsvd``: uses the randomized_svd algorithm implemented in
Sklearn (computation on CPU).
``cupy``: uses the Cupy library for GPU computation of the SVD as in
the LAPACK version. `
`eigencupy``: offers the same method as with the ``eigen`` option
but on GPU (through Cupy).
``randcupy``: is an adaptation of the randomized_svd algorithm,
where all the computations are done on a GPU (through Cupy). `
`pytorch``: uses the Pytorch library for GPU computation of the SVD.
``eigenpytorch``: offers the same method as with the ``eigen``
option but on GPU (through Pytorch).
``randpytorch``: is an adaptation of the randomized_svd algorithm,
where all the linear algebra computations are done on a GPU
(through Pytorch).
ncomp : int
Number of singular vectors to be obtained. In the cases when the full
SVD is computed (LAPACK, ARPACK, EIGEN, CUPY), the matrix of singular
vectors is truncated.
verbose: bool
If True intermediate information is printed out.
full_output : bool optional
If True the 3 terms of the SVD factorization are returned. If ``mode``
is eigen then only S and V are returned.
random_state : int, RandomState instance or None, optional
If int, random_state is the seed used by the random number generator.
If RandomState instance, random_state is the random number generator.
If None, the random number generator is the RandomState instance used
by np.random. Used for ``randsvd`` mode.
to_numpy : bool, optional
If True (by default) the arrays computed in GPU are transferred from
VRAM and converted to numpy ndarrays.
Returns
-------
V : numpy ndarray
The right singular vectors of the input matrix. If ``full_output`` is
True it returns the left and right singular vectors and the singular
values of the input matrix. If ``mode`` is set to eigen then only S and
V are returned.
References
----------
* For ``lapack`` SVD mode see:
https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.linalg.svd.html
http://www.netlib.org/lapack/
* For ``eigen`` mode see:
https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.linalg.eigh.html
* For ``arpack`` SVD mode see:
https://docs.scipy.org/doc/scipy-0.19.1/reference/generated/scipy.sparse.linalg.svds.html
http://www.caam.rice.edu/software/ARPACK/
* For ``randsvd`` SVD mode see:
https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/utils/extmath.py
Finding structure with randomness: Stochastic algorithms for constructing
approximate matrix decompositions
Halko, et al., 2009 http://arxiv.org/abs/arXiv:0909.4061
* For ``cupy`` SVD mode see:
https://docs-cupy.chainer.org/en/stable/reference/generated/cupy.linalg.svd.html
* For ``eigencupy`` mode see:
https://docs-cupy.chainer.org/en/master/reference/generated/cupy.linalg.eigh.html
* For ``pytorch`` SVD mode see:
http://pytorch.org/docs/master/torch.html#torch.svd
* For ``eigenpytorch`` mode see:
http://pytorch.org/docs/master/torch.html#torch.eig
"""
if matrix.ndim != 2:
raise TypeError('Input matrix is not a 2d array')
if ncomp > min(matrix.shape[0], matrix.shape[1]):
msg = '{} PCs cannot be obtained from a matrix with size [{},{}].'
msg += ' Increase the size of the patches or request less PCs'
raise RuntimeError(msg.format(ncomp, matrix.shape[0], matrix.shape[1]))
if mode == 'eigen':
# building C as np.dot(matrix.T,matrix) is slower and takes more memory
C = np.dot(matrix, matrix.T) # covariance matrix
e, EV = linalg.eigh(C) # EVals and EVs
pc = np.dot(EV.T, matrix) # PCs using a compact trick when cov is MM'
V = pc[::-1] # reverse since we need the last EVs
S = np.sqrt(np.abs(e)) # SVals = sqrt(EVals)
S = S[::-1] # reverse since EVals go in increasing order
for i in range(V.shape[1]):
V[:, i] /= S # scaling EVs by the square root of EVals
V = V[:ncomp]
if verbose:
print('Done PCA with numpy linalg eigh functions')
elif mode == 'lapack':
# n_frames is usually smaller than n_pixels. In this setting taking
# the SVD of M' and keeping the left (transposed) SVs is faster than
# taking the SVD of M (right SVs)
U, S, V = linalg.svd(matrix.T, full_matrices=False)
V = V[:ncomp] # we cut projection matrix according to the # of PCs
U = U[:, :ncomp]
S = S[:ncomp]
if verbose:
print('Done SVD/PCA with numpy SVD (LAPACK)')
elif mode == 'arpack':
U, S, V = svds(matrix, k=ncomp)
if verbose:
print('Done SVD/PCA with scipy sparse SVD (ARPACK)')
elif mode == 'randsvd':
U, S, V = randomized_svd(matrix, n_components=ncomp, n_iter=2,
transpose='auto', random_state=random_state)
if verbose:
print('Done SVD/PCA with randomized SVD')
elif mode == 'cupy':
if no_cupy:
raise RuntimeError('Cupy is not installed')
a_gpu = cupy.array(matrix)
a_gpu = cupy.asarray(a_gpu) # move the data to the current device
u_gpu, s_gpu, vh_gpu = cupy.linalg.svd(a_gpu, full_matrices=True,
compute_uv=True)
V = vh_gpu[:ncomp]
if to_numpy:
V = cupy.asnumpy(V)
if full_output:
S = s_gpu[:ncomp]
if to_numpy:
S = cupy.asnumpy(S)
U = u_gpu[:, :ncomp]
if to_numpy:
U = cupy.asnumpy(U)
if verbose:
print('Done SVD/PCA with cupy (GPU)')
elif mode == 'randcupy':
if no_cupy:
raise RuntimeError('Cupy is not installed')
U, S, V = randomized_svd_gpu(matrix, ncomp, n_iter=2, lib='cupy')
if to_numpy:
V = cupy.asnumpy(V)
S = cupy.asnumpy(S)
U = cupy.asnumpy(U)
if verbose:
print('Done randomized SVD/PCA with cupy (GPU)')
elif mode == 'eigencupy':
if no_cupy:
raise RuntimeError('Cupy is not installed')
a_gpu = cupy.array(matrix)
a_gpu = cupy.asarray(a_gpu) # move the data to the current device
C = cupy.dot(a_gpu, a_gpu.T) # covariance matrix
e, EV = cupy.linalg.eigh(C) # eigenvalues and eigenvectors
pc = cupy.dot(EV.T, a_gpu) # using a compact trick when cov is MM'
V = pc[::-1] # reverse to get last eigenvectors
S = cupy.sqrt(e)[::-1] # reverse since EVals go in increasing order
for i in range(V.shape[1]):
V[:, i] /= S # scaling by the square root of eigvals
V = V[:ncomp]
if to_numpy:
V = cupy.asnumpy(V)
if verbose:
print('Done PCA with cupy eigh function (GPU)')
elif mode == 'pytorch':
if no_torch:
raise RuntimeError('Pytorch is not installed')
a_gpu = torch.Tensor.cuda(torch.from_numpy(matrix.astype('float32').T))
u_gpu, s_gpu, vh_gpu = torch.svd(a_gpu)
V = vh_gpu[:ncomp]
S = s_gpu[:ncomp]
U = torch.transpose(u_gpu, 0, 1)[:ncomp]
if to_numpy:
V = np.array(V)
S = np.array(S)
U = np.array(U)
if verbose:
print('Done SVD/PCA with pytorch (GPU)')
elif mode == 'eigenpytorch':
if no_torch:
raise RuntimeError('Pytorch is not installed')
a_gpu = torch.Tensor.cuda(torch.from_numpy(matrix.astype('float32')))
C = torch.mm(a_gpu, torch.transpose(a_gpu, 0, 1))
e, EV = torch.eig(C, eigenvectors=True)
V = torch.mm(torch.transpose(EV, 0, 1), a_gpu)
S = torch.sqrt(e[:, 0])
for i in range(V.shape[1]):
V[:, i] /= S
V = V[:ncomp]
if to_numpy:
V = np.array(V)
if verbose:
print('Done PCA with pytorch eig function')
elif mode == 'randpytorch':
if no_torch:
raise RuntimeError('Pytorch is not installed')
U, S, V = randomized_svd_gpu(matrix, ncomp, n_iter=2, lib='pytorch')
if to_numpy:
V = np.array(V)
S = np.array(S)
U = np.array(U)
if verbose:
print('Done randomized SVD/PCA with randomized pytorch (GPU)')
else:
raise ValueError('The SVD `mode` is not recognized')
if full_output:
if mode == 'lapack':
return V.T, S, U.T
elif mode == 'pytorch':
if to_numpy:
return V.T, S, U.T
else:
return torch.transpose(V, 0, 1), S, torch.transpose(U, 0, 1)
elif mode in ('eigen', 'eigencupy', 'eigenpytorch'):
return S, V
else:
return U, S, V
else:
if mode == 'lapack':
return U.T
elif mode == 'pytorch':
return U
else:
return V
def preprocess(ctx):
# function rez = preprocessDataSub(ops)
# this function takes an ops struct, which contains all the Kilosort2 settings and file paths
# and creates a new binary file of preprocessed data, logging new variables into rez.
# The following steps are applied:
# 1) conversion to float32
# 2) common median subtraction
# 3) bandpass filtering
# 4) channel whitening
# 5) scaling to int16 values
params = ctx.params
probe = ctx.probe
raw_data = ctx.raw_data
ir = ctx.intermediate
fs = params.fs
fshigh = params.fshigh
fslow = params.fslow
Nbatch = ir.Nbatch
NT = params.NT
NTbuff = params.NTbuff
Wrot = cp.asarray(ir.Wrot)
logger.info("Loading raw data and applying filters.")
with open(ir.proc_path, 'wb') as fw: # open for writing processed data
for ibatch in tqdm(range(Nbatch), desc="Preprocessing"):
# we'll create a binary file of batches of NT samples, which overlap consecutively
# on params.ntbuff samples
# in addition to that, we'll read another params.ntbuff samples from before and after,
# to have as buffers for filtering
# number of samples to start reading at.
i = max(0, (NT - params.ntbuff) * ibatch - 2 * params.ntbuff)
if ibatch == 0:
# The very first batch has no pre-buffer, and has to be treated separately
ioffset = 0
else:
ioffset = params.ntbuff
buff = raw_data[:, i:i + NTbuff]
if buff.size == 0:
logger.error("Loaded buffer has an empty size!")
break # this shouldn't really happen, unless we counted data batches wrong
nsampcurr = buff.shape[
1] # how many time samples the current batch has
if nsampcurr < NTbuff:
buff = np.concatenate(
(buff,
np.tile(buff[:, nsampcurr - 1][:, np.newaxis],
(1, NTbuff))),
axis=1)
# apply filters and median subtraction
buff = cp.asarray(buff, dtype=np.float32)
datr = gpufilter(buff,
chanMap=probe.chanMap,
fs=fs,
fshigh=fshigh,
fslow=fslow)
datr = datr[ioffset:ioffset +
NT, :] # remove timepoints used as buffers
datr = cp.dot(
datr, Wrot) # whiten the data and scale by 200 for int16 range
# convert to int16, and gather on the CPU side
datcpu = cp.asnumpy(datr).astype(np.int16)
# write this batch to binary file
fw.write(datcpu.tobytes('F'))
def to_gpu(x):
import cupy
if type(x) == cupy.ndarray:
return x
return cupy.asarray(x)
y = np.arange(0, int(col), 1)
z = np.arange(0, int(sta), 1)
X, Y, Z = np.meshgrid(x, y, z)
OSS_alpha = float(row)
OSS_alpha_step = (float(row) - 1.0 / float(row)) / (
float(iteration) / float(OSS_interval) - 1.0)
#重心計算関係
cx = cp.arange(0, int(row), 1)
cy = cp.arange(0, int(col), 1)
cz = cp.arange(0, int(sta), 1)
#numpy配列 ⇒ cupy配列に変換
cp_diff_amp = cp.asarray(np_diff_amp, dtype="float32")
cp_sup = cp.asarray(np_sup, dtype="float32")
cp_initial_dens = cp.asarray(np_initial_dens, dtype="float32")
cp_dens = cp.asarray(np_dens)
print("iteration scale_factor Rfactor OS_ratio gamma")
with open(log_path, mode='a') as log:
log.write("iteration scale_factor Rfactor OS_ratio gamma")
for i in range(int(iteration) + int(additional_iteration)):
cp_structure_factor = cp.fft.fftn(cp_dens, axes=(0, 1, 2),
norm="ortho") #【フーリエ変換】
cp_structure_factor = cp.fft.fftshift(
cp_structure_factor) #fftshiftを使ってシフト
cp_amp = cp.absolute(cp_structure_factor) #絶対値をとる
def svd_wrapper(matrix, mode, ncomp, debug, verbose, usv=False,
random_state=None, to_numpy=True):
""" Wrapper for different SVD libraries (CPU and GPU).
Parameters
----------
matrix : array_like, 2d
2d input matrix.
mode : {'lapack', 'arpack', 'eigen', 'randsvd', 'cupy', 'eigencupy',
'randcupy', 'pytorch', 'eigenpytorch', 'randpytorch'}, str optional
Switch for the SVD method/library to be used. ``lapack`` uses the LAPACK
linear algebra library through Numpy and it is the most conventional way
of computing the SVD (deterministic result computed on CPU). ``arpack``
uses the ARPACK Fortran libraries accessible through Scipy (computation
on CPU). ``eigen`` computes the singular vectors through the
eigendecomposition of the covariance M.M' (computation on CPU).
``randsvd`` uses the randomized_svd algorithm implemented in Sklearn
(computation on CPU). ``cupy`` uses the Cupy library for GPU computation
of the SVD as in the LAPACK version. ``eigencupy`` offers the same
method as with the ``eigen`` option but on GPU (through Cupy).
``randcupy`` is an adaptation f the randomized_svd algorithm, where all
the computations are done on a GPU (through Cupy). ``pytorch`` uses the
Pytorch library for GPU computation of the SVD. ``eigenpytorch`` offers
the same method as with the ``eigen`` option but on GPU (through
Pytorch). ``randpytorch`` is an adaptation of the randomized_svd
algorithm, where all the linear algebra computations are done on a GPU
(through Pytorch).
ncomp : int
Number of singular vectors to be obtained. In the cases when the full
SVD is computed (LAPACK, ARPACK, EIGEN, CUPY), the matrix of singular
vectors is truncated.
debug : bool
If True the explained variance ratio is computed and displayed.
verbose: bool
If True intermediate information is printed out.
usv : bool optional
If True the 3 terms of the SVD factorization are returned.
random_state : int, RandomState instance or None, optional
If int, random_state is the seed used by the random number generator.
If RandomState instance, random_state is the random number generator.
If None, the random number generator is the RandomState instance used
by np.random. Used for ``randsvd`` mode.
to_numpy : bool, optional
If True (by default) the arrays computed in GPU are transferred from
VRAM and converted to numpy ndarrays.
Returns
-------
V : array_like
The right singular vectors of the input matrix. If ``usv`` is True it
returns the left and right singular vectors and the singular values of
the input matrix.
References
----------
* For ``lapack`` SVD mode see:
https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.linalg.svd.html
http://www.netlib.org/lapack/
* For ``eigen`` mode see:
https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.linalg.eigh.html
* For ``arpack`` SVD mode see:
https://docs.scipy.org/doc/scipy-0.19.1/reference/generated/scipy.sparse.linalg.svds.html
http://www.caam.rice.edu/software/ARPACK/
* For ``randsvd`` SVD mode see:
https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/utils/extmath.py
Finding structure with randomness: Stochastic algorithms for constructing
approximate matrix decompositions
Halko, et al., 2009 http://arxiv.org/abs/arXiv:0909.4061
* For ``cupy`` SVD mode see:
https://docs-cupy.chainer.org/en/stable/reference/generated/cupy.linalg.svd.html
* For ``eigencupy`` mode see:
https://docs-cupy.chainer.org/en/master/reference/generated/cupy.linalg.eigh.html
* For ``pytorch`` SVD mode see:
http://pytorch.org/docs/master/torch.html#torch.svd
* For ``eigenpytorch`` mode see:
http://pytorch.org/docs/master/torch.html#torch.eig
"""
def reconstruction(ncomp, U, S, V, var=1):
if mode == 'lapack':
rec_matrix = np.dot(U[:, :ncomp],
np.dot(np.diag(S[:ncomp]), V[:ncomp]))
rec_matrix = rec_matrix.T
print(' Matrix reconstruction with {} PCs:'.format(ncomp))
print(' Mean Absolute Error =', MAE(matrix, rec_matrix))
print(' Mean Squared Error =', MSE(matrix, rec_matrix))
# see https://github.com/scikit-learn/scikit-learn/blob/c3980bcbabd9d2527548820581725df2904e4a0d/sklearn/decomposition/pca.py
exp_var = (S ** 2) / (S.shape[0] - 1)
full_var = np.sum(exp_var)
explained_variance_ratio = exp_var / full_var # % of variance explained by each PC
ratio_cumsum = np.cumsum(explained_variance_ratio)
elif mode == 'eigen':
exp_var = (S ** 2) / (S.shape[0] - 1)
full_var = np.sum(exp_var)
explained_variance_ratio = exp_var / full_var # % of variance explained by each PC
ratio_cumsum = np.cumsum(explained_variance_ratio)
else:
rec_matrix = np.dot(U, np.dot(np.diag(S), V))
print(' Matrix reconstruction MAE =', MAE(matrix, rec_matrix))
exp_var = (S ** 2) / (S.shape[0] - 1)
full_var = np.var(matrix, axis=0).sum()
explained_variance_ratio = exp_var / full_var # % of variance explained by each PC
if var == 1:
pass
else:
explained_variance_ratio = explained_variance_ratio[::-1]
ratio_cumsum = np.cumsum(explained_variance_ratio)
msg = ' This info makes sense when the matrix is mean centered '
msg += '(temp-mean scaling)'
print(msg)
lw = 2; alpha = 0.4
fig = plt.figure(figsize=vip_figsize)
fig.subplots_adjust(wspace=0.4)
ax1 = plt.subplot2grid((1, 3), (0, 0), colspan=2)
ax1.step(range(explained_variance_ratio.shape[0]),
explained_variance_ratio, alpha=alpha, where='mid',
label='Individual EVR', lw=lw)
ax1.plot(ratio_cumsum, '.-', alpha=alpha,
label='Cumulative EVR', lw=lw)
ax1.legend(loc='best', frameon=False, fontsize='medium')
ax1.set_ylabel('Explained variance ratio (EVR)')
ax1.set_xlabel('Principal components')
ax1.grid(linestyle='solid', alpha=0.2)
ax1.set_xlim(-10, explained_variance_ratio.shape[0] + 10)
ax1.set_ylim(0, 1)
trunc = 20
ax2 = plt.subplot2grid((1, 3), (0, 2), colspan=1)
# plt.setp(ax2.get_yticklabels(), visible=False)
ax2.step(range(trunc), explained_variance_ratio[:trunc], alpha=alpha,
where='mid', lw=lw)
ax2.plot(ratio_cumsum[:trunc], '.-', alpha=alpha, lw=lw)
ax2.set_xlabel('Principal components')
ax2.grid(linestyle='solid', alpha=0.2)
ax2.set_xlim(-2, trunc + 2)
ax2.set_ylim(0, 1)
msg = ' Cumulative explained variance ratio for {} PCs = {:.5f}'
# plt.savefig('figure.pdf', dpi=300, bbox_inches='tight')
print(msg.format(ncomp, ratio_cumsum[ncomp - 1]))
# --------------------------------------------------------------------------
if matrix.ndim != 2:
raise TypeError('Input matrix is not a 2d array')
if usv:
if mode not in ('lapack', 'arpack', 'randsvd', 'cupy', 'randcupy',
'pytorch', 'randpytorch'):
msg = "Returning USV is supported with modes lapack, arpack, "
msg += "randsvd, cupy, randcupy, pytorch or randpytorch"
raise ValueError(msg)
if ncomp > min(matrix.shape[0], matrix.shape[1]):
msg = '{} PCs cannot be obtained from a matrix with size [{},{}].'
msg += ' Increase the size of the patches or request less PCs'
raise RuntimeError(msg.format(ncomp, matrix.shape[0], matrix.shape[1]))
if mode == 'eigen':
# building C as np.dot(matrix.T,matrix) is slower and takes more memory
C = np.dot(matrix, matrix.T) # covariance matrix
e, EV = linalg.eigh(C) # EVals and EVs
pc = np.dot(EV.T, matrix) # PCs using a compact trick when cov is MM'
V = pc[::-1] # reverse since we need the last EVs
S = np.sqrt(np.abs(e)) # SVals = sqrt(EVals)
S = S[::-1] # reverse since EVals go in increasing order
if debug:
reconstruction(ncomp, None, S, None)
for i in range(V.shape[1]):
V[:, i] /= S # scaling EVs by the square root of EVals
V = V[:ncomp]
if verbose:
print('Done PCA with numpy linalg eigh functions')
elif mode == 'lapack':
# n_frames is usually smaller than n_pixels. In this setting taking the SVD of M'
# and keeping the left (transposed) SVs is faster than taking the SVD of M (right SVs)
U, S, V = linalg.svd(matrix.T, full_matrices=False)
if debug:
reconstruction(ncomp, U, S, V)
V = V[:ncomp] # we cut projection matrix according to the # of PCs
U = U[:, :ncomp]
S = S[:ncomp]
if verbose:
print('Done SVD/PCA with numpy SVD (LAPACK)')
elif mode == 'arpack':
U, S, V = svds(matrix, k=ncomp)
if debug:
reconstruction(ncomp, U, S, V, -1)
if verbose:
print('Done SVD/PCA with scipy sparse SVD (ARPACK)')
elif mode == 'randsvd':
U, S, V = randomized_svd(matrix, n_components=ncomp, n_iter=2,
transpose='auto', random_state=random_state)
if debug:
reconstruction(ncomp, U, S, V)
if verbose:
print('Done SVD/PCA with randomized SVD')
elif mode == 'cupy':
if no_cupy:
raise RuntimeError('Cupy is not installed')
a_gpu = cupy.array(matrix)
a_gpu = cupy.asarray(a_gpu) # move the data to the current device
u_gpu, s_gpu, vh_gpu = cupy.linalg.svd(a_gpu, full_matrices=True,
compute_uv=True)
V = vh_gpu[:ncomp]
if to_numpy:
V = cupy.asnumpy(V)
if usv:
S = s_gpu[:ncomp]
if to_numpy:
S = cupy.asnumpy(S)
U = u_gpu[:, :ncomp]
if to_numpy:
U = cupy.asnumpy(U)
if verbose:
print('Done SVD/PCA with cupy (GPU)')
elif mode == 'randcupy':
if no_cupy:
raise RuntimeError('Cupy is not installed')
U, S, V = randomized_svd_gpu(matrix, ncomp, n_iter=2, lib='cupy')
if to_numpy:
V = cupy.asnumpy(V)
S = cupy.asnumpy(S)
U = cupy.asnumpy(U)
if debug:
reconstruction(ncomp, U, S, V)
if verbose:
print('Done randomized SVD/PCA with cupy (GPU)')
elif mode == 'eigencupy':
if no_cupy:
raise RuntimeError('Cupy is not installed')
a_gpu = cupy.array(matrix)
a_gpu = cupy.asarray(a_gpu) # move the data to the current device
C = cupy.dot(a_gpu, a_gpu.T) # covariance matrix
e, EV = cupy.linalg.eigh(C) # eigenvalues and eigenvectors
pc = cupy.dot(EV.T, a_gpu) # PCs using a compact trick when cov is MM'
V = pc[::-1] # reverse since last eigenvectors are the ones we want
S = cupy.sqrt(e)[::-1] # reverse since eigenvalues are in increasing order
if debug:
reconstruction(ncomp, None, S, None)
for i in range(V.shape[1]):
V[:, i] /= S # scaling by the square root of eigenvalues
V = V[:ncomp]
if to_numpy:
V = cupy.asnumpy(V)
if verbose:
print('Done PCA with cupy eigh function (GPU)')
elif mode == 'pytorch':
if no_torch:
raise RuntimeError('Pytorch is not installed')
a_gpu = torch.Tensor.cuda(torch.from_numpy(matrix.astype('float32').T))
u_gpu, s_gpu, vh_gpu = torch.svd(a_gpu)
V = vh_gpu[:ncomp]
S = s_gpu[:ncomp]
U = torch.transpose(u_gpu, 0, 1)[:ncomp]
if to_numpy:
V = np.array(V)
S = np.array(S)
U = np.array(U)
if verbose:
print('Done SVD/PCA with pytorch (GPU)')
elif mode == 'eigenpytorch':
if no_torch:
raise RuntimeError('Pytorch is not installed')
a_gpu = torch.Tensor.cuda(torch.from_numpy(matrix.astype('float32')))
C = torch.mm(a_gpu, torch.transpose(a_gpu, 0, 1))
e, EV = torch.eig(C, eigenvectors=True)
V = torch.mm(torch.transpose(EV, 0, 1), a_gpu)
S = torch.sqrt(e[:, 0])
if debug:
reconstruction(ncomp, None, S, None)
for i in range(V.shape[1]):
V[:, i] /= S
V = V[:ncomp]
if to_numpy:
V = np.array(V)
if verbose:
print('Done PCA with pytorch eig function')
elif mode == 'randpytorch':
if no_torch:
raise RuntimeError('Pytorch is not installed')
U, S, V = randomized_svd_gpu(matrix, ncomp, n_iter=2, lib='pytorch')
if to_numpy:
V = np.array(V)
S = np.array(S)
U = np.array(U)
if debug:
reconstruction(ncomp, U, S, V)
if verbose:
print('Done randomized SVD/PCA with randomized pytorch (GPU)')
else:
raise ValueError('The SVD mode is not available')
if usv:
if mode == 'lapack':
return V.T, S, U.T
elif mode == 'pytorch':
if to_numpy:
return V.T, S, U.T
else:
return torch.transpose(V, 0, 1), S, torch.transpose(U, 0, 1)
else:
return U, S, V
else:
if mode == 'lapack':
return U.T
elif mode == 'pytorch':
return U
else:
return V
def unique(ar, return_index=False, return_inverse=False,
return_counts=False, axis=None, *, equal_nan=True):
"""Find the unique elements of an array.
Returns the sorted unique elements of an array. There are three optional
outputs in addition to the unique elements:
* the indices of the input array that give the unique values
* the indices of the unique array that reconstruct the input array
* the number of times each unique value comes up in the input array
Args:
ar(array_like): Input array. This will be flattened if it is not
already 1-D.
return_index(bool, optional): If True, also return the indices of `ar`
(along the specified axis, if provided, or in the flattened array)
that result in the unique array.
return_inverse(bool, optional): If True, also return the indices of the
unique array (for the specified axis, if provided) that can be used
to reconstruct `ar`.
return_counts(bool, optional): If True, also return the number of times
each unique item appears in `ar`.
axis(int or None, optional): Not supported yet.
equal_nan(bool, optional): If True, collapse multiple NaN values in the
return array into one.
Returns:
cupy.ndarray or tuple:
If there are no optional outputs, it returns the
:class:`cupy.ndarray` of the sorted unique values. Otherwise, it
returns the tuple which contains the sorted unique values and
followings.
* The indices of the first occurrences of the unique values in the
original array. Only provided if `return_index` is True.
* The indices to reconstruct the original array from the
unique array. Only provided if `return_inverse` is True.
* The number of times each of the unique values comes up in the
original array. Only provided if `return_counts` is True.
.. warning::
This function may synchronize the device.
.. seealso:: :func:`numpy.unique`
"""
if axis is not None:
raise NotImplementedError('axis option is not supported yet.')
ar = cupy.asarray(ar).flatten()
if return_index or return_inverse:
perm = ar.argsort()
aux = ar[perm]
else:
ar.sort()
aux = ar
mask = cupy.empty(aux.shape, dtype=cupy.bool_)
mask[:1] = True
mask[1:] = aux[1:] != aux[:-1]
if equal_nan:
_unique_update_mask_equal_nan(mask[1:], aux[:-1])
ret = aux[mask]
if not return_index and not return_inverse and not return_counts:
return ret
ret = ret,
if return_index:
ret += perm[mask],
if return_inverse:
imask = cupy.cumsum(mask) - 1
inv_idx = cupy.empty(mask.shape, dtype=cupy.intp)
inv_idx[perm] = imask
ret += inv_idx,
if return_counts:
nonzero = cupy.nonzero(mask)[0] # may synchronize
idx = cupy.empty((nonzero.size + 1,), nonzero.dtype)
idx[:-1] = nonzero
idx[-1] = mask.size
ret += idx[1:] - idx[:-1],
return ret
out_h, out_w = h - kh + 1 + ph * 2, w - kw + 1 + pw * 2 # TODO
elif mode == 'valid':
ph, pw = 0, 0
out_h, out_w = h - kh + 1, w - kw + 1 # TODO
else:
raise NotImplementedError
y = cp.empty((n, out_c, out_h, out_w), dtype=in1.dtype)
col = im2col_gpu(in1, kh, kw, 1, 1, ph, pw)
y = cp.tensordot(col, in2, ((1, 2, 3), (1, 2, 3))).astype(in1.dtype,
copy=False)
y = cp.rollaxis(y, 3, 1)
return y.transpose(2, 3, 0, 1)
if __name__ == '__main__':
import cupy as cp
import numpy as np
from scipy.signal import convolve
a = np.random.randn(5, 5, 5, 1) + 1j * np.random.randn(5, 5, 5, 1)
b = np.random.randn(3, 3, 1, 1) + 1j * np.random.randn(3, 3, 1, 1)
y_cpu = convolve(a, b, 'valid')
x = cp.asarray(a)
w = cp.asarray(b)
y_gpu = convolve2d(x, w, 'valid')
np.allclose(y_gpu.get().squeeze(), y_cpu.squeeze(), atol=1e-6)
volume = np.repeat(volume, args.slice, axis=2)
w, h, z = volume.shape
# convert to a tensor
b = c = 1
x = volume.reshape(b, c, w, h, z).astype(np.float32)
# reshape the tensor: [b(=1),c(=1),w,h,z] -> [b*z(=z),c(=1),w,h]
x = x.transpose(0, 4, 1, 2, 3)
x = x.reshape(b * z, c, w, h)
# to gpu
if args.gpu >= 0:
import cupy as xp
x = xp.asarray(x)
x = chainer.Variable(x)
print(x.shape)
# do
radon = Radon(theta=np.linspace(0, 180, args.angle))
if args.gpu >= 0:
chainer.backends.cuda.get_device_from_id(args.gpu).use()
radon.to_gpu()
import tqdm
for _ in tqdm.tqdm(range(args.trial)):
ret = radon(x)
print(ret.shape)
def histogramdd(sample, bins=10, range=None, weights=None, density=False):
"""
Compute the multidimensional histogram of some data.
Parameters
----------
sample : (N, D) array, or (D, N) array_like
The data to be histogrammed.
Note the unusual interpretation of sample when an array_like:
* When an array, each row is a coordinate in a D-dimensional space -
such as ``histogramdd(cupy.array([p1, p2, p3]))``.
* When an array_like, each element is the list of values for single
coordinate - such as ``histogramdd((X, Y, Z))``.
The first form should be preferred.
bins : sequence or int, optional
The bin specification:
* A sequence of arrays describing the monotonically increasing bin
edges along each dimension.
* The number of bins for each dimension (nx, ny, ... =bins)
* The number of bins for all dimensions (nx=ny=...=bins).
range : sequence, optional
A sequence of length D, each an optional (lower, upper) tuple giving
the outer bin edges to be used if the edges are not given explicitly in
`bins`.
An entry of None in the sequence results in the minimum and maximum
values being used for the corresponding dimension.
The default, None, is equivalent to passing a tuple of D None values.
density : bool, optional
If False, the default, returns the number of samples in each bin.
If True, returns the probability *density* function at the bin,
``bin_count / sample_count / bin_volume``.
weights : (N,) array_like, optional
An array of values `w_i` weighing each sample `(x_i, y_i, z_i, ...)`.
The values of the returned histogram are equal to the sum of the
weights belonging to the samples falling into each bin.
Returns
-------
H : ndarray
The multidimensional histogram of sample x. See normed and weights
for the different possible semantics.
edges : list
A list of D arrays describing the bin edges for each dimension.
See Also
--------
histogram: 1-D histogram
histogram2d: 2-D histogram
Examples
--------
>>> r = cupy.random.randn(100,3)
>>> H, edges = cupy.histogramdd(r, bins = (5, 8, 4))
>>> H.shape, edges[0].size, edges[1].size, edges[2].size
((5, 8, 4), 6, 9, 5)
"""
if isinstance(sample, cupy.ndarray):
# Sample is an ND-array.
if sample.ndim == 1:
sample = sample[:, cupy.newaxis]
nsamples, ndim = sample.shape
else:
sample = cupy.stack(sample, axis=-1)
nsamples, ndim = sample.shape
nbin = numpy.empty(ndim, int)
edges = ndim * [None]
dedges = ndim * [None]
if weights is not None:
weights = cupy.asarray(weights)
try:
nbins = len(bins)
if nbins != ndim:
raise ValueError(
"The dimension of bins must be equal to the dimension of the "
" sample x."
)
except TypeError:
# bins is an integer
bins = ndim * [bins]
# normalize the range argument
if range is None:
range = (None,) * ndim
elif len(range) != ndim:
raise ValueError("range argument must have one entry per dimension")
# Create edge arrays
for i in _range(ndim):
if cnp.ndim(bins[i]) == 0:
if bins[i] < 1:
raise ValueError(
"`bins[{}]` must be positive, when an integer".format(i)
)
smin, smax = _get_outer_edges(sample[:, i], range[i])
num = int(bins[i] + 1) # synchronize!
edges[i] = cupy.linspace(smin, smax, num)
elif cnp.ndim(bins[i]) == 1:
edges[i] = cupy.asarray(bins[i])
if (edges[i][:-1] > edges[i][1:]).any():
raise ValueError(
"`bins[{}]` must be monotonically increasing, when an array".format(
i
)
)
else:
raise ValueError(
"`bins[{}]` must be a scalar or 1d array".format(i)
)
nbin[i] = len(edges[i]) + 1 # includes an outlier on each end
dedges[i] = cupy.diff(edges[i])
# Compute the bin number each sample falls into.
ncount = tuple(
# avoid cupy.digitize to work around gh-11022
cupy.searchsorted(edges[i], sample[:, i], side="right")
for i in _range(ndim)
)
# Using digitize, values that fall on an edge are put in the right bin.
# For the rightmost bin, we want values equal to the right edge to be
# counted in the last bin, and not as an outlier.
for i in _range(ndim):
# Find which points are on the rightmost edge.
on_edge = sample[:, i] == edges[i][-1]
# Shift these points one bin to the left.
ncount[i][on_edge] -= 1
# Compute the sample indices in the flattened histogram matrix.
# This raises an error if the array is too large.
xy = cnp.ravel_multi_index(ncount, nbin)
# Compute the number of repetitions in xy and assign it to the
# flattened histmat.
hist = cupy.bincount(xy, weights, minlength=numpy.prod(nbin))
# Shape into a proper matrix
hist = hist.reshape(nbin)
# This preserves the (bad) behavior observed in gh-7845, for now.
hist = hist.astype(float) # Note: NumPy uses casting='safe' here too
# Remove outliers (indices 0 and -1 for each dimension).
core = ndim * (slice(1, -1),)
hist = hist[core]
if density:
# calculate the probability density function
s = hist.sum()
for i in _range(ndim):
shape = [1] * ndim
shape[i] = nbin[i] - 2
hist = hist / dedges[i].reshape(shape)
hist /= s
if any(hist.shape != numpy.asarray(nbin) - 2):
raise RuntimeError("Internal Shape Error")
return hist, edges
def _get_bin_edges(a, bins, range):
"""
Computes the bins used internally by `histogram`.
Args:
a (ndarray): Ravelled data array
bins (int or ndarray): Forwarded argument from `histogram`.
range (None or tuple): Forwarded argument from `histogram`.
Returns:
bin_edges (ndarray): Array of bin edges
uniform_bins (Number, Number, int): The upper bound, lowerbound, and
number of bins, used in the implementation of `histogram` that works on
uniform bins.
"""
# parse the overloaded bins argument
n_equal_bins = None
bin_edges = None
# if isinstance(bins, cupy.ndarray) and bins.ndim == 0:
# # allow uint8 array, etc
# if bins.dtype not in 'bui':
# raise TypeError(
# "`bins` must be an integer, a string, or an array")
# bins = int(bins) # synchronize
if isinstance(bins, int): # will not allow 0-dimensional cupy array
# if cupy.ndim(bins) == 0:
try:
n_equal_bins = operator.index(bins)
except TypeError:
raise TypeError("`bins` must be an integer, a string, or an array")
if n_equal_bins < 1:
raise ValueError("`bins` must be positive, when an integer")
first_edge, last_edge = _get_outer_edges(a, range)
elif isinstance(bins, cupy.ndarray):
if bins.ndim == 1: # cupy.ndim(bins) == 0:
bin_edges = cupy.asarray(bins)
if (bin_edges[:-1] > bin_edges[1:]).any(): # synchronize!
raise ValueError(
"`bins` must increase monotonically, when an array"
)
elif isinstance(bins, str):
raise NotImplementedError("only integer and array bins are implemented")
if n_equal_bins is not None:
# numpy's gh-10322 means that type resolution rules are dependent on
# array shapes. To avoid this causing problems, we pick a type now and
# stick with it throughout.
bin_type = cupy.result_type(first_edge, last_edge, a)
if cupy.issubdtype(bin_type, cupy.integer):
bin_type = cupy.result_type(bin_type, float)
# bin edges must be computed
bin_edges = cupy.linspace(
first_edge,
last_edge,
n_equal_bins + 1,
endpoint=True,
dtype=bin_type,
)
return bin_edges, (first_edge, last_edge, n_equal_bins)
else:
return bin_edges, None
def evolve(model):
# Plot init
plt.style.use('ggplot')
train_loss = []
train_acc = []
test_loss = []
test_acc = []
n_train_batches = N_train / batchsize
# early stopping
patience = 5000
patience_increase = 2
improvement_threshold = 0.995
validation_frequency = min(n_train_batches, patience / 2)
best_validation_loss = np.inf
test_score = 0
done_looping = False
# Learning loop
epoch = 0
while (epoch < n_epoch) and (not done_looping):
epoch = epoch + 1
print 'epoch {}'.format(epoch)
# training
perm = np.random.permutation(N_train)
sum_train_accuracy = 0
sum_train_loss = 0
for i in xrange(0, N_train, batchsize):
x = chainer.Variable(cp.asarray(x_train[perm[i:i + batchsize]]))
t = chainer.Variable(cp.asarray(y_train[perm[i:i + batchsize]]))
# Pass the loss function (Classifier defines it) and its arguments
optimizer.update(model, x, t)
sum_train_loss += float(model.loss.data) * len(t.data)
sum_train_accuracy += float(model.accuracy.data) * len(t.data)
# generate network graph
# if epoch == 1 and i == 0:
# with open('netgraph.dot', 'w') as o:
# g = computational_graph.build_computational_graph(
# (model.loss, ), remove_split=True)
# o.write(g.dump())
# print 'net graph generated'
# validation
batch_index = (i / batchsize)
iter = (epoch - 1) * n_train_batches + batch_index
if (iter + 1) % validation_frequency == 0:
sum_validate_accuracy = 0
sum_validate_loss = 0
for i in xrange(0, N_validate, batchsize):
x = chainer.Variable(cp.asarray(x_test[i:i + batchsize]),
volatile='on')
t = chainer.Variable(cp.asarray(y_test[i:i + batchsize]),
volatile='on')
loss = model(x, t)
sum_validate_loss += float(loss.data) * len(t.data)
sum_validate_accuracy += float(model.accuracy.data) * len(t.data)
this_validate_loss = sum_validate_loss / N_validate
this_validate_accuracy = sum_validate_accuracy / N_validate
print 'validation epoch{}, minibatch{}/{}'.format(epoch, batch_index + 1, n_train_batches)
print ' mean loss={}, accuracy={}'.format(
this_validate_loss, sum_validate_accuracy / N_validate)
if this_validate_loss < best_validation_loss:
if this_validate_loss < best_validation_loss * improvement_threshold:
patience = max(patience, iter * patience_increase)
print " iter {} / patience {}".format(iter+1, patience)
best_validation_loss = this_validate_loss
if patience <= iter:
done_looping = True
break
train_loss.append(sum_train_loss / N_train)
train_acc.append(sum_train_accuracy / N_train)
print 'train mean loss={}, accuracy={}'.format(
sum_train_loss / N_train, sum_train_accuracy / N_train)
# evaluation
sum_test_accuracy = 0
sum_test_loss = 0
for i in xrange(0, N_test, batchsize):
x = chainer.Variable(cp.asarray(x_test[i:i + batchsize]),
volatile='on')
t = chainer.Variable(cp.asarray(y_test[i:i + batchsize]),
volatile='on')
loss = model(x, t)
sum_test_loss += float(loss.data) * len(t.data)
sum_test_accuracy += float(model.accuracy.data) * len(t.data)
test_loss.append(sum_test_loss / N_test)
test_acc.append(sum_test_accuracy / N_test)
print 'test mean loss={}, accuracy={}'.format(
sum_test_loss / N_test, sum_test_accuracy / N_test)
print 'train finish'
print 'draw graph'
# draw graph
# このversionでの推移
plt.figure(figsize=(8,6))
# plt.xlim([0, epoch])
# plt.ylim([0.95, 1.0])
plt.plot(xrange(1,len(train_acc)+1), train_acc)
plt.plot(xrange(1,len(test_acc)+1), test_acc)
plt.legend(["train_acc","test_acc"],loc=4)
plt.title("Accuracy of digit recognition.")
plt.plot()
plt.savefig("graph_v%5d.png" % (version))
plt.close()
# このversionでの推移 範囲[0.95, 1.0]
plt.figure(figsize=(8,6))
plt.xlim([0, epoch])
plt.ylim([0.95, 1.0])
plt.plot(xrange(1,len(train_acc)+1), train_acc)
plt.plot(xrange(1,len(test_acc)+1), test_acc)
plt.legend(["train_acc","test_acc"],loc=4)
plt.title("Accuracy of digit recognition. range [0.95, 1.0]")
plt.plot()
plt.savefig("graph_095_v%5d.png" % (version))
plt.close()
# 各versionにおける精度の変化
version_loss.append(best_validation_loss)
version_acc.append(test_acc[-1])
plt.figure(figsize=(8,6))
# plt.ylim([0.50, 1.0])
plt.plot(xrange(1,len(version_acc)+1), version_acc)
plt.legend(["version_acc"],loc=4)
plt.title("Accuracy of digit recognition. (version)")
plt.plot()
plt.savefig("graph_version_v%5d.png" % (version))
plt.close()
# 今まで全ての推移(x軸epoch)
all_train_acc.extend(train_acc)
all_test_acc.extend(test_acc)
plt.figure(figsize=(8,6))
# plt.ylim([0.95, 1.0])
plt.plot(xrange(1,len(all_train_acc)+1), all_train_acc)
plt.plot(xrange(1,len(all_test_acc)+1), all_test_acc)
plt.legend(["all_train_acc","all_test_acc"],loc=4)
plt.title("Accuracy of digit recognition.")
plt.plot()
plt.savefig("graph_allepoch_v%5d.png" % (version))
plt.close()
plt.figure(figsize=(8,6))
plt.ylim([0.95, 1.0])
plt.plot(xrange(1,len(all_train_acc)+1), all_train_acc)
plt.plot(xrange(1,len(all_test_acc)+1), all_test_acc)
plt.legend(["all_train_acc","all_test_acc"],loc=4)
plt.title("Accuracy of digit recognition. range [0.95, 1.0]")
plt.plot()
plt.savefig("graph_allepoch095_v%5d.png" % (version))
plt.close()
plt.close('all')
# Save the model and the optimizer
print 'save the model'
model.to_cpu()
serializers.save_hdf5("v%5d.model" % (version), model)
print 'save the optimizer'
serializers.save_hdf5('v%5d.state' % (version), optimizer)
finishtime = time.time()
print 'execute time = {}'.format(finishtime - starttime)
# plt.show()
return sum_test_accuracy / N_test
def Allreduce_mean(self, x, **kwargs):
"""Multi-process multi-GPU based mean."""
src = self.pool.reduce_mean(x, **kwargs)
mean = self.mpi.Allreduce(cp.asnumpy(src)) / self.mpi.size
return cp.asarray(mean)
def walsh_transform(self,keys=None):
if keys is None:
keys = ['kernel'] + list(self.constraints.keys())
else:
keys = keys
is_stored = dict()
for key in keys:
is_stored[key] = False
if os.path.exists(self.fname):
with h5py.File(self.fname,mode='r') as f:
for key in keys:
try:
if '3' in f[key].keys():
is_stored[key] = True
if key == 'depth':
res = f['depth']['constraint'][:] - self.constraints['depth']
res = np.linalg.norm(res)/np.linalg.norm(self.constraints['depth'])
if res > 1.0e-3:
is_stored[key] = False
except KeyError:
continue
self._gen_walsh_matrix()
logn = int(np.ceil(np.log2(self._nx*self._ny*self._nz)))
norm_walsh = 1./(np.sqrt(2)**logn)
blocks = ['0','1','2','3']
matvec_op = {'kernel':self.kernel_op.gtoep.matvec,
'dx': lambda x: self._dxyzvec(x,key='dx'),
'dy': lambda x: self._dxyzvec(x,key='dy'),
'dz': lambda x: self._dxyzvec(x,key='dz'),
'refer': lambda x: self._diagvec(x,diag=self.constraints['refer']),
'depth': lambda x: self._diagvec(x,diag=np.sqrt(self.constraints['depth']))
}
is_stored['refer'] = True
for key in keys:
if is_stored[key]:
print('walsh transformation of {} already exists.'.format(key))
continue
print('performing walsh transformation on {}.'.format(key))
step = self.nx*self.ny*self.nz // 4
if key == 'depth':
step = self._nz
with h5py.File(self.fname,mode='a') as f:
try:
del f[key]
except KeyError:
pass
dxyz_group = f.create_group(key)
walsh_group = f['walsh_matrix']
for i in range(4):
print("\t progress {}/4".format(i))
part_walsh = walsh_group[blocks[i]][:]
if key == 'depth':
part_walsh = walsh_group[blocks[i]][:self._nz]
part_walsh = matvec_op[key](part_walsh)
with cp.cuda.Device(2):
res = cp.zeros((step,step))
j = 0
while j*step < part_walsh.shape[1]:
tmp_block_gpu = cp.asarray(part_walsh[:,j*step:(j+1)*step])
res += tmp_block_gpu @ tmp_block_gpu.T
j += 1
res = cp.asnumpy(res)
if key in self._smooth_components:
res[np.abs(res)<1.0e-1*norm_walsh] = 0.
tmp_block_gpu = None
mempool = cp.get_default_memory_pool()
pinned_mempool = cp.get_default_pinned_memory_pool()
mempool.free_all_blocks()
pinned_mempool.free_all_blocks()
dxyz_group.create_dataset(blocks[i],data=res)
if ('depth' in keys) and (not is_stored['depth']):
with h5py.File(self.fname,mode='a') as f:
try:
del f['depth_constraint']
except KeyError:
pass
dxyz_group = f['depth']
dxyz_group.create_dataset('constraint',data=self.constraints['depth'])
import cupy as cp
import numpy as np
from cupy import testing
from skimage import data
from cupyimg.skimage import color
from cupyimg.skimage.util import img_as_bool
from cupyimg.skimage.morphology import binary, grey, selem
from cupyimg.scipy import ndimage as ndi
import pytest
img = color.rgb2gray(cp.asarray(data.astronaut()))
bw_img = img > 100 / 255.0
def test_non_square_image():
strel = selem.square(3)
binary_res = binary.binary_erosion(bw_img[:100, :200], strel)
grey_res = img_as_bool(grey.erosion(bw_img[:100, :200], strel))
testing.assert_array_equal(binary_res, grey_res)
def test_binary_erosion():
strel = selem.square(3)
binary_res = binary.binary_erosion(bw_img, strel)
grey_res = img_as_bool(grey.erosion(bw_img, strel))
testing.assert_array_equal(binary_res, grey_res)
def test_binary_dilation():
def to_gpu(self, arr):
return cupy.asarray(arr)
def to_gpu(*args):
""" Upload numpy arrays to GPU and return them"""
if len(args) > 1:
return (cp.asarray(x) for x in args)
else:
return cp.asarray(args[0])
def corner_peaks(
image,
min_distance=1,
threshold_abs=None,
threshold_rel=None,
exclude_border=True,
indices=True,
num_peaks=np.inf,
footprint=None,
labels=None,
*,
num_peaks_per_label=np.inf,
p_norm=np.inf,
):
"""Find peaks in corner measure response image.
This differs from `skimage.feature.peak_local_max` in that it suppresses
multiple connected peaks with the same accumulator value.
Parameters
----------
image : ndarray
Input image.
min_distance : int, optional
The minimal allowed distance separating peaks.
* : *
See :py:meth:`skimage.feature.peak_local_max`.
p_norm : float
Which Minkowski p-norm to use. Should be in the range [1, inf].
A finite large p may cause a ValueError if overflow can occur.
``inf`` corresponds to the Chebyshev distance and 2 to the
Euclidean distance.
Returns
-------
output : ndarray or ndarray of bools
* If `indices = True` : (row, column, ...) coordinates of peaks.
* If `indices = False` : Boolean array shaped like `image`, with peaks
represented by True values.
See also
--------
skimage.feature.peak_local_max
Notes
-----
.. versionchanged:: 0.18
The default value of `threshold_rel` has changed to None, which
corresponds to letting `skimage.feature.peak_local_max` decide on the
default. This is equivalent to `threshold_rel=0`.
The `num_peaks` limit is applied before suppression of connected peaks.
To limit the number of peaks after suppression, set `num_peaks=np.inf` and
post-process the output of this function.
Examples
--------
>>> from cupyimg.skimage.feature import peak_local_max
>>> response = cp.zeros((5, 5))
>>> response[2:4, 2:4] = 1
>>> response
array([[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 1., 1., 0.],
[0., 0., 1., 1., 0.],
[0., 0., 0., 0., 0.]])
>>> peak_local_max(response)
array([[2, 2],
[2, 3],
[3, 2],
[3, 3]])
>>> corner_peaks(response)
array([[2, 2]])
"""
if cp.isinf(num_peaks):
num_peaks = None
# Get the coordinates of the detected peaks
coords = peak_local_max(
image,
min_distance=min_distance,
threshold_abs=threshold_abs,
threshold_rel=threshold_rel,
exclude_border=exclude_border,
num_peaks=np.inf,
footprint=footprint,
labels=labels,
num_peaks_per_label=num_peaks_per_label,
)
if len(coords):
# TODO: modify to do KDTree on the GPU (cuSpatial?)
coords = cp.asnumpy(coords)
# Use KDtree to find the peaks that are too close to each other
tree = spatial.cKDTree(coords)
rejected_peaks_indices = set()
for idx, point in enumerate(coords):
if idx not in rejected_peaks_indices:
candidates = tree.query_ball_point(point,
r=min_distance,
p=p_norm)
candidates.remove(idx)
rejected_peaks_indices.update(candidates)
# Remove the peaks that are too close to each other
coords = np.delete(coords, tuple(rejected_peaks_indices),
axis=0)[:num_peaks]
coords = cp.asarray(coords)
if indices:
return coords
peaks = cp.zeros_like(image, dtype=bool)
peaks[tuple(coords.T)] = True
return peaks
def deformation(self, prm):
"""
Apply 2D Gaussian and Planar deformation.
Computation is parallelized on GPU using cupy.
"""
import cupy as cp
xy_cp = cp.asarray(prm.xy)
a_cp = cp.asarray(self.a)
b_cp = cp.asarray(self.b)
c_cp = cp.asarray(self.c)
d_cp = cp.asarray(self.d)
sigma_cp = cp.asarray(self.sigma)
e_cp = cp.asarray(self.e)
f_cp = cp.asarray(self.f)
g_cp = cp.asarray(self.g)
z_cp = cp.asarray(prm.z)
func_planar = cp.ElementwiseKernel(
in_params='T x, T y, T e, T f, T g',
out_params='T z',
operation= \
'''
z = e + f*x + g*y;
''',
name='func_planar'
)
func_gauss2d = cp.ElementwiseKernel(
in_params='T x, T y, T b, T c, T d, T sigma',
out_params='T z',
operation= \
'''
z = b*expf(-(powf(x-c,2) + powf(y-d,2))/(2*powf(sigma,2)));
''',
name='func_gauss2d'
)
gauss_2d_cp = cp.zeros_like(xy_cp[:, 0])
for i in range(len(self.b)):
gauss_2d_cp += func_gauss2d(xy_cp[:, 0], xy_cp[:, 1], b_cp[i], c_cp[i], d_cp[i], sigma_cp[i])
s1_cp = a_cp + (1.5 / z_cp) * cp.outer(cp.transpose(gauss_2d_cp), z_cp)
s2_cp = func_planar(xy_cp[:, 0], xy_cp[:, 1], e_cp, f_cp, g_cp)
refl_cp = cp.asarray(self.refl)
for i in range(prm.nxy_tr):
s = s1_cp[i, :] + s2_cp[i] + z_cp
mat = cp.tile(z_cp, (len(s), 1)) - cp.tile(cp.expand_dims(s, 1), (1, len(z_cp)))
refl_cp[i, :] = cp.dot(refl_cp[i, :], cp.sinc(mat))
return np.reshape(cp.asnumpy(refl_cp), [prm.nxy_tr, prm.nz_tr])
def get_good_channels(raw_data=None, probe=None, params=None):
"""
of the channels indicated by the user as good (chanMap)
further subset those that have a mean firing rate above a certain value
(default is ops.minfr_goodchannels = 0.1Hz)
needs the same filtering parameters in ops as usual
also needs to know where to start processing batches (twind)
and how many channels there are in total (NchanTOT)
"""
fs = params.fs
fshigh = params.fshigh
fslow = params.fslow
Nbatch = get_Nbatch(raw_data, params)
NT = params.NT
spkTh = params.spkTh
nt0 = params.nt0
minfr_goodchannels = params.minfr_goodchannels
chanMap = probe.chanMap
# Nchan = probe.Nchan
NchanTOT = len(chanMap)
ich = []
k = 0
ttime = 0
# skip every 100 batches
for ibatch in tqdm(range(0, Nbatch, int(ceil(Nbatch / 100))),
desc="Finding good channels"):
i = NT * ibatch
buff = raw_data[:, i:i + NT]
assert np.isfortran(buff)
if buff.size == 0:
break
# Put on GPU.
buff = cp.asarray(buff, dtype=np.float32)
assert cp.isfortran(buff)
datr = gpufilter(buff,
chanMap=chanMap,
fs=fs,
fshigh=fshigh,
fslow=fslow)
# very basic threshold crossings calculation
s = cp.std(datr, axis=0)
datr = datr / s # standardize each channel ( but don't whiten)
mdat = my_min(
datr, 30,
0) # get local minima as min value in +/- 30-sample range
# take local minima that cross the negative threshold
xi, xj = cp.nonzero((datr < mdat + 1e-3) & (datr < spkTh))
# filtering may create transients at beginning or end. Remove those.
xj = xj[(xi >= nt0) & (xi <= NT - nt0)]
# collect the channel identities for the detected spikes
ich.append(xj)
k += xj.size
# keep track of total time where we took spikes from
ttime += datr.shape[0] / fs
ich = cp.concatenate(ich)
# count how many spikes each channel got
nc, _ = cp.histogram(ich, cp.arange(NchanTOT + 1))
# divide by total time to get firing rate
nc = nc / ttime
# keep only those channels above the preset mean firing rate
igood = cp.asnumpy(nc >= minfr_goodchannels)
logger.info('Found %d threshold crossings in %2.2f seconds of data.' %
(k, ttime))
logger.info('Found %d/%d bad channels.' % (np.sum(~igood), len(igood)))
return igood
def test_with_strides(self, dtype):
a = testing.shaped_arange((2, 3, 4), cupy, dtype).T
b = cupy.asarray(
DummyObjectWithCudaArrayInterface(a, self.ver, self.strides))
assert a.strides == b.strides
assert a.nbytes == b.data.mem.size
def get_whitening_matrix(raw_data=None, probe=None, params=None):
"""
based on a subset of the data, compute a channel whitening matrix
this requires temporal filtering first (gpufilter)
"""
Nbatch = get_Nbatch(raw_data, params)
ntbuff = params.ntbuff
NTbuff = params.NTbuff
whiteningRange = params.whiteningRange
scaleproc = params.scaleproc
NT = params.NT
fs = params.fs
fshigh = params.fshigh
nSkipCov = params.nSkipCov
xc = probe.xc
yc = probe.yc
chanMap = probe.chanMap
Nchan = probe.Nchan
chanMap = probe.chanMap
# Nchan is obtained after the bad channels have been removed
CC = cp.zeros((Nchan, Nchan))
for ibatch in tqdm(range(0, Nbatch, nSkipCov),
desc="Computing the whitening matrix"):
# WARNING: we use Fortran order, so raw_data is NchanTOT x nsamples
i = max(0, (NT - ntbuff) * ibatch - 2 * ntbuff)
buff = raw_data[:, i:i + NT - ntbuff]
nsampcurr = buff.shape[1]
if nsampcurr < NTbuff:
buff = np.concatenate(
(buff,
np.tile(buff[:, nsampcurr - 1][:, np.newaxis], (1, NTbuff))),
axis=1)
buff_g = cp.asarray(buff, dtype=np.float32)
# apply filters and median subtraction
datr = gpufilter(buff_g, fs=fs, fshigh=fshigh, chanMap=chanMap)
CC = CC + cp.dot(datr.T, datr) / NT # sample covariance
CC = CC / ceil((Nbatch - 1) / nSkipCov)
if whiteningRange < np.inf:
# if there are too many channels, a finite whiteningRange is more robust to noise
# in the estimation of the covariance
whiteningRange = min(whiteningRange, Nchan)
# this function performs the same matrix inversions as below, just on subsets of
# channels around each channel
Wrot = whiteningLocal(CC, yc, xc, whiteningRange)
else:
Wrot = whiteningFromCovariance(CC)
Wrot = Wrot * scaleproc
logger.info("Computed the whitening matrix.")
return Wrot
def test_not_copied(self, dtype):
a = testing.shaped_arange((2, 3, 4), cupy, dtype)
b = cupy.asarray(
DummyObjectWithCudaArrayInterface(a, self.ver, self.strides))
a.fill(0)
testing.assert_array_equal(a, b)
def evaluate(self):
self.load_from_file()
m1 = self.fiducial_binaries["mass_1"]
m2 = self.fiducial_binaries["mass_2"]
# Note that spins are redshift independent
spin_1x, spin_1y, spin_1z = [
self.fiducial_binaries[k]
for k in ["spin_1x", "spin_1y", "spin_1z"]
]
spin_2x, spin_2y, spin_2z = [
self.fiducial_binaries[k]
for k in ["spin_2x", "spin_2y", "spin_2z"]
]
if _GPU_ENABLED:
import cupy as xp
else:
import numpy as xp
m1 = xp.asarray(m1)
m2 = xp.asarray(m2)
spin_1x = xp.asarray(spin_1x)
spin_1y = xp.asarray(spin_1y)
spin_1z = xp.asarray(spin_1z)
spin_2x = xp.asarray(spin_2x)
spin_2y = xp.asarray(spin_2y)
spin_2z = xp.asarray(spin_2z)
pdf_spin_fiducial = xp.asarray(self.pdf_spin_fiducial)
pdf_mass_fiducial = xp.asarray(self.pdf_mass_fiducial)
pdf_spin_pop = self.spin_src_pop_model.prob({
"spin_1x": spin_1x,
"spin_1y": spin_1y,
"spin_1z": spin_1z,
"spin_2x": spin_2x,
"spin_2y": spin_2y,
"spin_2z": spin_2z,
})
weights_spin = pdf_spin_pop / pdf_spin_fiducial
for img in range(self.N_img):
self.predictions[img] = xp.asarray(self.predictions[img])
self.pdf_dLs_fiducial[img] = xp.asarray(self.pdf_dLs_fiducial[img])
self.apparent_dLs[img] = xp.asarray(self.apparent_dLs[img])
def epsilon(z_src):
det_mass_pop_dist = DetectorFrameComponentMassesFromSourceFrame(
self.mass_src_pop_model, z_src)
pdf_mass_pop = det_mass_pop_dist.prob({"mass_1": m1, "mass_2": m2})
weights_mass = pdf_mass_pop / pdf_mass_fiducial
weights_source = weights_mass * weights_spin
integrand = weights_source
for img in range(self.N_img):
pdf_dL_fiducial = self.pdf_dLs_fiducial[img]
dL_pop_dist = LuminosityDistancePriorFromAbsoluteMagnificationRedshift(
self.abs_magn_dist[img], z_src)
pdf_dL_pop = dL_pop_dist.prob(self.apparent_dLs[img])
weights_dL = pdf_dL_pop / pdf_dL_fiducial
integrand *= self.predictions[img] * weights_dL
return float(xp.sum(integrand) / float(self.N_inj))
logger = logging.getLogger(__prog__)
logger.info("Integrating over source redshift")
z_dist = LensedSourceRedshiftProbDist(
merger_rate_density=self.merger_rate_density_src_pop_model,
optical_depth=self.optical_depth)
zs = z_dist.sample(size=self.N_z)
if _GPU_ENABLED:
import cupy as cp
zs = cp.asnumpy(zs)
epsilons = []
for z in tqdm.tqdm(zs):
epsilons.append(epsilon(z))
beta = np.sum(epsilons).astype(float) / self.N_z
self.f.close()
return beta
def calc_log_prior_total_det(self):
self.log_prior_det_val = 0
self.log_total_det_val = 0
blocks = ['0','1','2','3']
prior_eigs = np.zeros(self._nx*self._ny*self._nz)
total_eigs = np.zeros(self._nx*self._ny*self._nz)
step = self._nx*self._ny*self._nz//4
try:
depth_weight = self._weights['depth']
except KeyError:
depth_weight = 1.
with h5py.File(self.fname,mode='r') as f:
if 'depth' in self._weights.keys():
depth_walsh = f['depth']['0'][:]
for i_b,block in enumerate(blocks):
tmp_block = np.zeros((step,step))
for dxyz_name in self._smooth_components:
try:
dxyz_walsh = f[dxyz_name][block][:].reshape(step//self._nz,
self._nz,
step//self._nz,
self._nz)
ein_path = np.einsum_path('mi,xiyj,jn->xmyn',
depth_walsh.T,
dxyz_walsh,
depth_walsh,
optimize='optimal')[0]
tmp_multi = np.einsum('mi,xiyj,jn->xmyn',
depth_walsh.T,
dxyz_walsh,
depth_walsh,
optimize=ein_path)
tmp_block += depth_weight*self._weights[dxyz_name]*tmp_multi.reshape(step,step)
except KeyError:
pass
if 'refer' in self._weights.keys():
tmp_multi_small = depth_walsh.T@depth_walsh
for i in range(step//self._nz):
tmp_block[i*self._nz:(i+1)*self._nz,
i*self._nz:(i+1)*self._nz] += depth_weight*self._weights['refer']*tmp_multi_small
with cp.cuda.Device(2):
tmp_block_gpu = cp.asarray(tmp_block,dtype=np.float32)
eigs = cp.linalg.eigvalsh(tmp_block_gpu)
prior_eigs[i_b*step:(i_b+1)*step] = cp.asnumpy(eigs)
self.log_prior_det_val += cp.asnumpy(cp.sum(cp.log(eigs)))
tmp_block_gpu = None
eigs = None
free_gpu()
tmp_block += self._weights['obs']*f['kernel'][block][:]
with cp.cuda.Device(2):
tmp_block_gpu = cp.asarray(tmp_block,dtype=np.float32)
eigs = cp.linalg.eigvalsh(tmp_block_gpu)
total_eigs[i_b*step:(i_b+1)*step] = cp.asnumpy(eigs)
self.log_total_det_val += cp.asnumpy(cp.sum(cp.log(eigs)))
tmp_block_gpu = None
eigs = None
free_gpu()
self.log_prior_det_val = cp.asnumpy(self.log_prior_det_val)
self.log_total_det_val = cp.asnumpy(self.log_total_det_val)
self.eigs = {'prior':prior_eigs,'total':total_eigs}
return self.log_prior_det_val,self.log_total_det_val
def to_sp_dask_array(cudf_or_array, client=None):
"""
Converts an array or cuDF to a sparse Dask array backed by sparse CuPy.
CSR matrices. Unfortunately, due to current limitations in Dask, there is
no direct path to convert a cupy.sparse.spmatrix into a CuPy backed
dask.Array without copying to host.
NOTE: Until https://github.com/cupy/cupy/issues/2655 and
https://github.com/dask/dask/issues/5604 are implemented, compute()
will not be able to be called on a Dask.array that is backed with
sparse CuPy arrays because they lack the necessary functionality
to be stacked into a single array. The array returned from this
utility will, however, still be able to be passed into functions
that can make use of sparse CuPy-backed Dask.Array (eg. Distributed
Naive Bayes).
Relevant cuML issue: https://github.com/rapidsai/cuml/issues/1387
Parameters
----------
cudf_or_array : cuDF Dataframe, array-like sparse / dense array, or
Dask DataFrame/Array
client : dask.distributed.Client (optional) Dask client
dtype : output dtype
Returns
-------
dask_array : dask.Array backed by cupy.sparse.csr_matrix
"""
client = default_client() if client is None else client
# Makes sure the MatDescriptor workaround for CuPy sparse arrays
# is loaded (since Dask lazy-loaded serialization in cuML is only
# executed when object from the cuML package needs serialization.
# This can go away once the MatDescriptor pickling bug is fixed
# in CuPy.
# Ref: https://github.com/cupy/cupy/issues/3061
from cuml.comm import serialize # NOQA
shape = cudf_or_array.shape
if isinstance(cudf_or_array, dask.dataframe.DataFrame) or \
isinstance(cudf_or_array, cudf.DataFrame):
dtypes = np.unique(cudf_or_array.dtypes)
if len(dtypes) > 1:
raise ValueError("DataFrame should contain only a single dtype")
dtype = dtypes[0]
else:
dtype = cudf_or_array.dtype
meta = cupyx.scipy.sparse.csr_matrix(rmm_cupy_ary(cp.zeros, 1))
if isinstance(cudf_or_array, dask.array.Array):
# At the time of developing this, using map_blocks will not work
# to convert a Dask.Array to CuPy sparse arrays underneath.
parts = client.sync(_extract_partitions, cudf_or_array)
cudf_or_array = [
client.submit(_conv_np_to_df, part, workers=[w])
for w, part in parts
]
cudf_or_array = to_dask_cudf(cudf_or_array)
if isinstance(cudf_or_array, dask.dataframe.DataFrame):
"""
Dask.Dataframe needs special attention since it has multiple dtypes.
Just use the first (and assume all the rest are the same)
"""
cudf_or_array = cudf_or_array.map_partitions(
_conv_df_to_sp, meta=dask.array.from_array(meta))
# This will also handle the input of dask.array.Array
return cudf_or_array
else:
if scipy.sparse.isspmatrix(cudf_or_array):
cudf_or_array = \
cupyx.scipy.sparse.csr_matrix(cudf_or_array.tocsr())
elif cupyx.scipy.sparse.isspmatrix(cudf_or_array):
pass
elif isinstance(cudf_or_array, cudf.DataFrame):
cupy_ary = cp.asarray(cudf_or_array.as_gpu_matrix(), dtype)
cudf_or_array = cupyx.scipy.sparse.csr_matrix(cupy_ary)
elif isinstance(cudf_or_array, np.ndarray):
cupy_ary = rmm_cupy_ary(cp.asarray,
cudf_or_array,
dtype=cudf_or_array.dtype)
cudf_or_array = cupyx.scipy.sparse.csr_matrix(cupy_ary)
elif isinstance(cudf_or_array, cp.core.core.ndarray):
cudf_or_array = cupyx.scipy.sparse.csr_matrix(cudf_or_array)
else:
raise ValueError("Unexpected input type %s" % type(cudf_or_array))
# Push to worker
cudf_or_array = client.scatter(cudf_or_array)
return dask.array.from_delayed(cudf_or_array, shape=shape, meta=meta)
test_x /= 255
cuda.check_cuda_available()
all_results = None
for j in range(NUM_MODELS):
model = MLP()
model.to_gpu()
optimizer = optimizers.Adam()
optimizer.setup(model)
for epoch in range(1, NUM_EPOCH + 1):
perm = np.random.permutation(NUM_TRAIN)
train_accuracy, train_loss = 0, 0
for i in range(0, NUM_TRAIN, BATCH_SIZE):
x = chainer.Variable(cp.asarray(train_x[perm[i:i + BATCH_SIZE]]), volatile='off')
t = chainer.Variable(cp.asarray(train_y[perm[i:i + BATCH_SIZE]]), volatile='off')
optimizer.update(model, x, t)
train_loss += float(model.loss.data) * len(t.data)
train_accuracy += float(model.accuracy.data) * len(t.data)
epoch_result = None
test_accuracy, test_loss = 0, 0
for i in range(0, NUM_TEST, BATCH_SIZE):
x = chainer.Variable(cp.asarray(test_x[i:i + BATCH_SIZE]), volatile='on')
t = chainer.Variable(cp.asarray(test_y[i:i + BATCH_SIZE]), volatile='on')
batch_result = model(x, t, False)
if epoch == NUM_EPOCH:
if i == 0:
def _array_to_gpu(array, device, stream):
if array is None:
return None
if isinstance(array, chainerx.ndarray):
# TODO(niboshi): Update this logic once both CuPy and ChainerX support
# the array interface.
if array.device.backend.name == 'cuda':
# Convert to cupy.ndarray on the same device as source array
array = cupy.ndarray(
array.shape,
array.dtype,
cupy.cuda.MemoryPointer(
cupy.cuda.UnownedMemory(
array.data_ptr + array.offset,
array.data_size,
array,
array.device.index),
0),
strides=array.strides)
else:
array = chainerx.to_numpy(array)
elif isinstance(array, (numpy.number, numpy.bool_)):
array = numpy.asarray(array)
elif isinstance(array, intel64.mdarray):
array = numpy.asarray(array)
if isinstance(array, ndarray):
if array.device == device:
return array
is_numpy = False
elif isinstance(array, numpy.ndarray):
is_numpy = True
else:
raise TypeError(
'The array sent to gpu must be an array or a NumPy scalar.'
'\nActual type: {0}.'.format(type(array)))
if stream is not None and stream.ptr != 0:
ret = cupy.empty_like(array)
if is_numpy:
# cpu to gpu
mem = cupy.cuda.alloc_pinned_memory(array.nbytes)
src = numpy.frombuffer(
mem, array.dtype, array.size).reshape(array.shape)
src[...] = array
ret.set(src, stream)
cupy.cuda.pinned_memory._add_to_watch_list(
stream.record(), mem)
else:
# gpu to gpu
with array.device:
src = array.copy()
event = Stream.null.record()
stream.wait_event(event)
ret.data.copy_from_device_async(
src.data, src.nbytes, stream)
# to hold a reference until the end of the asynchronous
# memcpy
stream.add_callback(lambda *x: None, (src, ret))
return ret
with device:
if is_numpy:
return cupy.asarray(array)
# Need to make a copy when an array is copied to another device
return cupy.array(array, copy=True)
def _calc_array(self, cpu_c_flat: np.ndarray) -> np.ndarray:
gpu_c_flat = cp.asarray(cpu_c_flat)
gpu_iteration_flat = self.server.compute_flat_array(gpu_c_flat)
cpu_iteration_flat = cp.asnumpy(gpu_iteration_flat)
# cpu_iteration_flat = compute_array.ComputeGpu.compute(cpu_c_flat)
return cpu_iteration_flat
def FC_to_FCWB(source, target, tftype='affine'):
source = [[(xyz[0] * + 475)/1.7, (xyz[1] -185) / -2, (xyz[2]-125) / -2.5] for xyz in source]
source = cp.asarray(source, dtype=cp.float32)
target = cp.asarray(target, dtype=cp.float32)
tf_param, _, _ = cpd.registration_cpd(source, target, tf_type_name=tftype, use_cuda=use_cuda)
return tf_param
def FC_to_FCWB_transform(object, tf_param):
object = [[(xyz[0] * + 475) / 1.7, (xyz[1] - 185) / -2, (xyz[2] - 125) / -2.5] for xyz in list(object)]
return to_cpu(tf_param.transform(cp.asarray(object, dtype=cp.float32)))
def __call__(self, loc, score,
anchor, img_size, scale=1.):
"""input should be ndarray
Propose RoIs.
Inputs :obj:`loc, score, anchor` refer to the same anchor when indexed
by the same index.
On notations, :math:`R` is the total number of anchors. This is equal
to product of the height and the width of an image and the number of
anchor bases per pixel.
Type of the output is same as the inputs.
Args:
loc (array): Predicted offsets and scaling to anchors.
Its shape is :math:`(R, 4)`.
score (array): Predicted foreground probability for anchors.
Its shape is :math:`(R,)`.
anchor (array): Coordinates of anchors. Its shape is
:math:`(R, 4)`.
img_size (tuple of ints): A tuple :obj:`height, width`,
which contains image size after scaling.
scale (float): The scaling factor used to scale an image after
reading it from a file.
Returns:
array:
An array of coordinates of proposal boxes.
Its shape is :math:`(S, 4)`. :math:`S` is less than
:obj:`self.n_test_post_nms` in test time and less than
:obj:`self.n_train_post_nms` in train time. :math:`S` depends on
the size of the predicted bounding boxes and the number of
bounding boxes discarded by NMS.
"""
# NOTE: when test, remember
# faster_rcnn.eval()
# to set self.traing = False
if self.parent_model.training:
n_pre_nms = self.n_train_pre_nms
n_post_nms = self.n_train_post_nms
else:
n_pre_nms = self.n_test_pre_nms
n_post_nms = self.n_test_post_nms
# Convert anchors into proposal via bbox transformations.
# roi = loc2bbox(anchor, loc)
roi = loc2bbox(anchor, loc)
# Clip predicted boxes to image.
roi[:, slice(0, 4, 2)] = np.clip(
roi[:, slice(0, 4, 2)], 0, img_size[0])
roi[:, slice(1, 4, 2)] = np.clip(
roi[:, slice(1, 4, 2)], 0, img_size[1])
# Remove predicted boxes with either height or width < threshold.
min_size = self.min_size * scale
hs = roi[:, 2] - roi[:, 0]
ws = roi[:, 3] - roi[:, 1]
keep = np.where((hs >= min_size) & (ws >= min_size))[0]
roi = roi[keep, :]
score = score[keep]
# Sort all (proposal, score) pairs by score from highest to lowest.
# Take top pre_nms_topN (e.g. 6000).
order = score.ravel().argsort()[::-1]
if n_pre_nms > 0:
order = order[:n_pre_nms]
roi = roi[order, :]
# Apply nms (e.g. threshold = 0.7).
# Take after_nms_topN (e.g. 300).
# unNOTE: somthing is wrong here!
# TODO: remove cuda.to_gpu
keep = non_maximum_suppression(
cp.ascontiguousarray(cp.asarray(roi)),
thresh=self.nms_thresh)
if n_post_nms > 0:
keep = keep[:n_post_nms]
roi = roi[keep]
return roi
def FCWB_to_FC_transform(object, tf_param):
return to_cpu(tf_param.transform(cp.asarray(object, dtype=cp.float32)))
def randomized_svd_gpu(M, n_components, n_oversamples=10, n_iter='auto',
transpose='auto', random_state=0, lib='cupy'):
"""Computes a truncated randomized SVD on GPU. Adapted from Sklearn.
Parameters
----------
M : ndarray or sparse matrix
Matrix to decompose
n_components : int
Number of singular values and vectors to extract.
n_oversamples : int (default is 10)
Additional number of random vectors to sample the range of M so as
to ensure proper conditioning. The total number of random vectors
used to find the range of M is n_components + n_oversamples. Smaller
number can improve speed but can negatively impact the quality of
approximation of singular vectors and singular values.
n_iter : int or 'auto' (default is 'auto')
Number of power iterations. It can be used to deal with very noisy
problems. When 'auto', it is set to 4, unless `n_components` is small
(< .1 * min(X.shape)) `n_iter` in which case is set to 7.
This improves precision with few components.
transpose : True, False or 'auto' (default)
Whether the algorithm should be applied to M.T instead of M. The
result should approximately be the same. The 'auto' mode will
trigger the transposition if M.shape[1] > M.shape[0] since this
implementation of randomized SVD tend to be a little faster in that
case.
random_state : int, RandomState instance or None, optional (default=None)
The seed of the pseudo random number generator to use when shuffling
the data. If int, random_state is the seed used by the random number
generator; If RandomState instance, random_state is the random number
generator; If None, the random number generator is the RandomState
instance used by `np.random`.
lib : {'cupy', 'pytorch'}, str optional
Chooses the GPU library to be used.
Notes
-----
This algorithm finds a (usually very good) approximate truncated
singular value decomposition using randomization to speed up the
computations. It is particularly fast on large matrices on which
you wish to extract only a small number of components. In order to
obtain further speed up, `n_iter` can be set <=2 (at the cost of
loss of precision).
References
----------
* Finding structure with randomness: Stochastic algorithms for constructing
approximate matrix decompositions
Halko, et al., 2009 http://arxiv.org/abs/arXiv:0909.4061
* A randomized algorithm for the decomposition of matrices
Per-Gunnar Martinsson, Vladimir Rokhlin and Mark Tygert
* An implementation of a randomized algorithm for principal component
analysis
A. Szlam et al. 2014
"""
random_state = check_random_state(random_state)
n_random = n_components + n_oversamples
n_samples, n_features = M.shape
if n_iter == 'auto':
# Checks if the number of iterations is explicitly specified
n_iter = 7 if n_components < .1 * min(M.shape) else 4
if transpose == 'auto':
transpose = n_samples < n_features
if transpose:
M = M.T # this implementation is a bit faster with smaller shape[1]
if lib == 'cupy':
M = cupy.array(M)
M = cupy.asarray(M)
# Generating normal random vectors with shape: (M.shape[1], n_random)
Q = random_state.normal(size=(M.shape[1], n_random))
Q = cupy.array(Q)
Q = cupy.asarray(Q)
# Perform power iterations with Q to further 'imprint' the top
# singular vectors of M in Q
for i in range(n_iter):
Q = cupy.dot(M, Q)
Q = cupy.dot(M.T, Q)
# Sample the range of M using by linear projection of Q. Extract an orthonormal basis
Q, _ = cupy.linalg.qr(cupy.dot(M, Q), mode='reduced')
# project M to the (k + p) dimensional space using the basis vectors
B = cupy.dot(Q.T, M)
B = cupy.array(B)
Q = cupy.array(Q)
# compute the SVD on the thin matrix: (k + p) wide
Uhat, s, V = cupy.linalg.svd(B, full_matrices=False, compute_uv=True)
del B
U = cupy.dot(Q, Uhat)
if transpose:
# transpose back the results according to the input convention
return V[:n_components, :].T, s[:n_components], U[:,
:n_components].T
else:
return U[:, :n_components], s[:n_components], V[:n_components, :]
elif lib == 'pytorch':
M_gpu = torch.Tensor.cuda(torch.from_numpy(M.astype('float32')))
# Generating normal random vectors with shape: (M.shape[1], n_random)
Q = torch.cuda.FloatTensor(M_gpu.shape[1], n_random).normal_()
# Perform power iterations with Q to further 'imprint' the top
# singular vectors of M in Q
for i in range(n_iter):
Q = torch.mm(M_gpu, Q)
Q = torch.mm(torch.transpose(M_gpu, 0, 1), Q)
# Sample the range of M using by linear projection of Q. Extract an orthonormal basis
Q, _ = torch.qr(torch.mm(M_gpu, Q))
# project M to the (k + p) dimensional space using the basis vectors
B = torch.mm(torch.transpose(Q, 0, 1), M_gpu)
# compute the SVD on the thin matrix: (k + p) wide
Uhat, s, V = torch.svd(B)
del B
U = torch.mm(Q, Uhat)
if transpose:
# transpose back the results according to the input convention
return (torch.transpose(V[:n_components, :], 0, 1),
s[:n_components],
torch.transpose(U[:, :n_components], 0, 1))
else:
return U[:, :n_components], s[:n_components], V[:n_components, :]
def to_cupy(self, copy=False):
self.host_to_device()
if copy:
return cp.array(self)
return cp.asarray(self)
