def rt2ct(y, axis=0):
r"""Converts a real-valued tensor to a complex-valued tensor
Converts a real-valued tensor :math:`{\bf y}` to a complex-valued tensor with FFT and conjugate symmetry.
Parameters
----------
y : Tensor
The input tensor :math:`{\bf y}\in {\mathbb C}^{2H×W}`.
axis : int
The axis for excuting FFT.
Returns
-------
Tensor
The output tensor :math:`{\bf x}\in {\mathbb R}^{H×W}` ( :attr:`axis` = 0 ), :math:`{\bf x}\in {\mathbb R}^{H×W}` ( :attr:`axis` = 1 )
"""
d = y.dim()
n = y.shape[axis]
Y = th.fft.fft(y, axis=axis)
X = Y[sl(d, axis, range(0, int(n / 2)))]
X[sl(d, axis, [[0]])].imag = Y[sl(d, axis, [[int(n / 2)]])].real
del y, Y
x = th.fft.ifft(X, axis=axis)
return x
def ct2rt(x, axis=0):
r"""Converts a complex-valued tensor to a real-valued tensor
Converts a complex-valued tensor :math:`{\bf x}` to a real-valued tensor with FFT and conjugate symmetry.
Parameters
----------
x : Tensor
The input tensor :math:`{\bf x}\in {\mathbb C}^{H×W}`.
axis : int
The axis for excuting FFT.
Returns
-------
Tensor
The output tensor :math:`{\bf y}\in {\mathbb R}^{2H×W}` ( :attr:`axis` = 0 ), :math:`{\bf y}\in {\mathbb R}^{H×2W}` ( :attr:`axis` = 1 )
"""
d = x.dim()
n = x.shape[axis]
X = th.fft.fft(x, axis=axis)
X0 = X[sl(d, axis, [[0]])]
X1 = th.conj(X[sl(d, axis, range(n - 1, 0, -1))])
Y = th.cat((X, X0.imag, X1), dim=axis)
Y[sl(d, axis, [[0]])] = X0.real + 0j
del x, X, X1
y = th.fft.ifft(Y, axis=axis)
return y
def patch2tensor(p, size=(256, 256), axis=(1, 2), mode='nfirst'):
"""merge patch to a tensor
Args:
p (Tensor): A tensor of patches.
size (tuple, optional): Merged tensor size in the dimension (the default is (256, 256)).
axis (tuple, optional): Merged axis of patch (the default is (1, 2))
mode (str, optional): Patch mode ``'nfirst'`` or ``'nlast'`` (the default is 'nfirst',
which means the first dimension is the number of patches)
Returns:
Tensor: Merged tensor.
"""
naxis = len(axis)
sizep = list(p.shape)
sizex = list(p.shape)
dimp = p.dim()
axisp = list(range(0, dimp))
npatch = []
steps = sizep.copy()
for a, s in zip(axis, size):
npatch.append(int((s * 1.) / sizep[a]))
sizex[a] = s
steps[a] = sizep[a]
if mode in ['nfirst', 'Nfirst', 'NFIRST']:
axisn = 0
N = p.shape[0]
sizex = sizex[1:]
steps = sizep[1:]
if mode in ['nlast', 'Nlast', 'NLAST']:
axisn = -1
N = p.shape[-1]
sizex = sizex[:-1]
steps = sizep[:-1]
x = th.zeros(sizex, dtype=p.dtype, device=p.device)
dimx = x.dim()
axisx = list(range(dimx))
index = []
for a, stop, step in zip(axisx, sizex, steps):
idx = np.array(range(0, stop, step))
index.append(idx)
index = arraycomb(index)
naxisx = len(axisx)
for n in range(N):
indexn = []
for a in axisx:
indexn.append(slice(index[n, a], index[n, a] + steps[a], 1))
x[sl(dimx, axisx, indexn)] = p[sl(dimp, axisn, n)]
return x
def draw_rectangle(x, rects, edgecolors=[[255, 0, 0]], linewidths=[1], fillcolors=[None], axes=(-3, -2)):
"""Draw rectangles in a tensor
Parameters
----------
x : tensor
The input with any size.
rects : list or tuple
The coordinates of the rectangles [[lefttop, rightbottom]].
edgecolors : list, optional
The color of edge.
linewidths : int, optional
The linewidths of edge.
fillcolors : int, optional
The color for filling.
axes : int, optional
The axes for drawing the rect (default [(-3, -2)]).
"""
axes = axes * len(rects) if len(axes) == 1 and len(rects) > 1 else axes
if type(x) is not th.Tensor:
x = th.tensor(x)
d = x.dim()
for rect, edgecolor, linewidth, fillcolor, axis in zip(rects, edgecolors, linewidths, fillcolors, axes):
edgecolor = th.tensor(edgecolor, dtype=x.dtype) if edgecolor is not None else None
fillcolor = th.tensor(fillcolor, dtype=x.dtype) if fillcolor is not None else None
if edgecolor is not None:
top, left, bottom, right = rect
for l in range(linewidth):
x[sl(d, axis, [slice(top, bottom + 1), [left, right]])] = edgecolor
x[sl(d, axis, [[top, bottom], slice(left, right)])] = edgecolor
top += 1
left += 1
bottom -= 1
right -= 1
if fillcolor is not None:
top, left, bottom, right = rect
top += linewidth
left += linewidth
bottom -= linewidth
right -= linewidth
x[sl(d, axis, [slice(top, bottom + 1), slice(left, right + 1)])] = fillcolor
return x
def forward(self, X):
if th.is_complex(X):
X = ((X * X.conj()).real).sqrt()
elif X.size(-1) == 2:
X = X.pow(2).sum(axis=-1).sqrt()
D = X.dim()
# compute gradients in axis direction
for a in self.axis:
d = X.size(a)
X = th.abs(X[sl(D, a, range(1, d))] - X[sl(D, a, range(0, d - 1))])
G = th.mean(X, self.axis, keepdim=True)
if self.reduction == 'mean':
V = th.mean(G)
if self.reduction == 'sum':
V = th.sum(G)
return V
def dnsampling(x,
ratio=1.,
axis=-1,
smode='uniform',
omode='discard',
seed=None,
extra=False):
"""Summary
Args:
x (Tensor): The Input tensor.
ratio (float, optional): Downsampling ratio.
axis (int, optional): Downsampling axis (default -1).
smode (str, optional): Downsampling mode: ``'uniform'``, ``'random'``, ``'random2'``.
omode (str, optional): output mode: ``'discard'`` for discarding, ``'zero'`` for zero filling.
seed (int or None, optional): seed for torch's random.
extra (bool, optional): If ``True``, also return sampling mask.
Returns:
(Tensor): Description
Raises:
TypeError: :attr:`axis`
ValueError: :attr:`ratio`, attr:`smode`, attr:`omode`
"""
nDims = x.dim()
if type(axis) is int:
if type(ratio) is not float:
raise ValueError('Downsampling ratio should be a number!')
axis = [axis]
ratio = [ratio]
elif type(axis) is list or tuple:
if len(axis) != len(ratio):
raise ValueError('You should specify the DS ratio for each axis!')
else:
raise TypeError('Wrong type of axis!')
axis, ratio = list(axis), list(ratio)
for cnt in range(len(axis)):
if axis[cnt] < 0:
axis[cnt] += nDims
# ratio[cnt] = 1. - ratio[cnt]
cnt += 1
if omode in ['discard', 'DISCARD', 'Discard']:
if smode not in ['uniform', 'UNIFORM', 'Uniform']:
raise ValueError("Only support uniform mode!")
index = [slice(None)] * nDims
for a, r in zip(axis, ratio):
sa = x.shape[a]
da = int(round(1. / r))
index[a] = slice(0, sa, da)
index = tuple(index)
if extra:
return x[index], index
else:
return x[index]
elif omode in ['zero', 'ZERO', 'Zeros']:
mshape = [1] * nDims
for a in axis:
mshape[a] = x.shape[a]
mask = th.zeros(mshape, dtype=th.uint8, device=x.device)
if smode in ['uniform', 'UNIFORM', 'Uniform']:
for a, r in zip(axis, ratio):
sa = x.shape[a]
da = int(round(1. / r))
idx = sl(nDims, a, slice(0, sa, da))
mask[idx] += 1
mask[mask < len(axis)] = 0
mask[mask >= len(axis)] = 1
elif smode in ['random', 'RANDOM', 'Random']:
setseed(seed, target='torch')
for a, r in zip(axis, ratio):
d = x.dim()
s = x.shape[a]
n = int(round(s * r))
idx = randperm(0, s, n)
idx = np.sort(idx)
idx = sl(d, a, idx)
mask[idx] += 1
mask[mask < len(axis)] = 0
mask[mask >= len(axis)] = 1
elif smode in ['random2', 'RANDOM2', 'Random2']:
setseed(seed, target='torch')
d = x.dim()
s0, s1 = x.shape[axis[0]], x.shape[axis[1]]
n0, n1 = int(round(s0 * ratio[0])), int(round(s1 * ratio[0]))
idx0 = randperm(0, s0, n0)
# idx0 = np.sort(idx0)
for i0 in idx0:
idx1 = randperm(0, s1, n1)
mask[sl(d, [axis[0], axis[1]], [[i0], idx1])] = 1
else:
raise ValueError('Not supported sampling mode: %s!' % smode)
if extra:
return x * mask, mask
else:
return x * mask
else:
raise ValueError('Not supported output mode: %s!' % omode)
def read_samples(datafiles,
keys=[['SI', 'ca', 'cr']],
nsamples=[10],
groups=[1],
mode='sequentially',
axis=0,
parts=None,
seed=None):
"""Read samples
Args:
datafiles (list): list of path strings
keys (list, optional): data keys to be read
nsamples (list, optional): number of samples for each data file
groups (list, optional): number of groups in each data file
mode (str, optional): sampling mode for all datafiles
axis (int, optional): sampling axis for all datafiles
parts (None, optional): number of parts (split samples into some parts)
seed (None, optional): the seed for random stream
Returns:
tensor: samples
Raises:
ValueError: :attr:`nsamples` should be large enough
"""
nfiles = len(datafiles)
if len(keys) == 1:
keys = keys * nfiles
if len(nsamples) == 1:
nsamples = nsamples * nfiles
if len(groups) == 1:
groups = groups * nfiles
nkeys = len(keys[0])
if parts is None:
outs = [th.tensor([])] * nkeys
else:
nparts = len(parts)
outs = [[th.tensor([])] * nparts] * nkeys
for datafile, key, n, group in zip(datafiles, keys, nsamples, groups):
if datafile[datafile.rfind('.'):] == '.mat':
data = loadmat(datafile)
if datafile[datafile.rfind('.'):] in ['.h5', '.hdf5']:
data = loadh5(datafile)
N = data[key[0]].shape[axis]
M = int(N / group) # each group has M samples
m = int(n / group) # each group has m sampled samples
if (M < m):
raise ValueError('The tensor does not has enough samples')
idx = []
if mode in ['sequentially', 'Sequentially']:
for g in range(group):
idx += list(range(int(M * g), int(M * g) + m))
if mode in ['uniformly', 'Uniformly']:
for g in range(group):
idx += list(range(int(M * g), int(M * g + M), int(M / m)))[:m]
if mode in ['randomly', 'Randomly']:
setseed(seed)
for g in range(group):
idx += randperm(int(M * g), int(M * g + M), m)
for j, k in enumerate(key):
d = np.ndim(data[k])
if parts is None:
outs[j] = th.cat(
(outs[j], th.from_numpy(data[k][sl(d, axis, [idx])])),
axis=axis)
else:
nps, npe = 0, 0
for i in range(nparts):
part = parts[i]
npe = nps + int(part * group)
outs[j][i] = th.cat(
(outs[j][i],
th.from_numpy(data[k][sl(d, axis, [idx[nps:npe]])])),
axis=axis)
nps = npe
return outs
def tensor2patch(x,
n=None,
size=(256, 256),
axis=(0, 1),
start=(0, 0),
stop=(None, None),
step=(1, 1),
shake=(0, 0),
mode='slidegrid',
seed=None):
"""sample patch from a tensor
Sample some patches from a tensor, tensor and patch can be any size.
Args:
x (Tensor): Tensor to be sampled.
n (int, optional): The number of pactches, the default is None, auto computed,
equals to the number of blocks with specified :attr:`step`
size (tuple or int, optional): The size of patch (the default is (256, 256))
axis (tuple or int, optional): The sampling axis (the default is (0, 1))
start (tuple or int, optional): Start sampling index for each axis (the default is (0, 0))
stop (tuple or int, optional): Stopp sampling index for each axis. (the default is (None, None), which [default_description])
step (tuple or int, optional): Sampling stepsize for each axis (the default is (1, 1), which [default_description])
shake (tuple or int or float, optional): float for shake rate, int for shake points (the default is (0, 0), which means no shake)
mode (str, optional): Sampling mode, ``'slidegrid'``, ``'randgrid'``, ``'randperm'`` (the default is 'slidegrid')
seed (int, optional): Random seed. (the default is None, which means no seed.)
Returns:
(Tensor): A Tensor of sampled patches.
"""
axis = [axis] if type(axis) is int else list(axis)
naxis = len(axis)
sizep = [size] * naxis if type(size) is int else list(size)
start = [start] * naxis if type(start) is int else list(start)
stop = [stop] * naxis if type(stop) is int else list(stop)
step = [step] * naxis if type(step) is int else list(step)
shake = [shake] * naxis if type(shake) is float else list(shake)
dimx = x.dim()
dimp = len(axis)
sizex = np.array(x.shape)
sizep = np.array(sizep)
npatch = []
npatch = np.uint32(sizex[axis] / sizep)
N = int(np.prod(npatch))
n = N if n is None else int(n)
yshape = list(x.shape)
for a, p in zip(axis, sizep):
yshape[a] = p
yshape = [n] + yshape
for i in range(naxis):
if stop[i] is None:
stop[i] = sizex[axis[i]]
y = th.zeros(yshape, dtype=x.dtype, device=x.device)
if mode in ['slidegrid', 'SLIDEGRID', 'SlideGrid']:
assert n <= N, ('n should be slower than ' + str(N + 1))
seppos = slidegrid(start, stop, step, shake, n)
if mode in ['randgrid', 'RANDGRID', 'RandGrid']:
assert n <= N, ('n should be slower than ' + str(N + 1))
setseed(seed, target='torch')
seppos = randgrid(start, stop, step, shake, n)
if mode in ['randperm', 'RANDPERM', 'RandPerm']:
setseed(seed, target='torch')
stop = [x - y for x, y in zip(stop, sizep)]
seppos = randgrid(start, stop, [1] * dimp, [0] * dimp, n)
for i in range(n):
indexi = []
for j in range(dimp):
indexi.append(slice(seppos[j][i], seppos[j][i] + sizep[j]))
t = x[sl(dimx, axis, indexi)]
y[i] = t
return y
def shuffle_tensor(x, axis=0, groups=1, mode='inter', seed=None, extra=False):
"""shuffle a tensor
Shuffle a tensor randomly.
Args:
x (Tensor): A torch tensor to be shuffled.
axis (int, optional): The axis to be shuffled (default 0)
groups (number, optional): The number of groups in this tensor (default 1)
mode (str, optional):
- ``'inter'``: between groups (default)
- ``'intra'``: within group
- ``'whole'``: the whole
seed (None or number, optional): random seed (the default is None)
extra (bool, optional): If ``True``, also returns the shuffle indexes, the default is ``False``.
Returns:
y (Tensor): Shuffled torch tensor.
idx (list): Shuffled indexes, if :attr:`extra` is ``True``, this will also be returned.
Example:
::
setseed(2020, 'torch')
x = th.randint(1000, (20, 3, 4))
y1, idx1 = shuffle_tensor(x, axis=0, groups=4, mode='intra', extra=True)
y2, idx2 = shuffle_tensor(x, axis=0, groups=4, mode='inter', extra=True)
y3, idx3 = shuffle_tensor(x, axis=0, groups=4, mode='whole', extra=True)
print(x.shape)
print(y1.shape)
print(y2.shape)
print(y3.shape)
print(idx1)
print(idx2)
print(idx3)
the outputs are as follows:
torch.Size([20, 3, 4])
torch.Size([20, 3, 4])
torch.Size([20, 3, 4])
torch.Size([20, 3, 4])
[1, 0, 3, 4, 2, 8, 6, 5, 9, 7, 13, 11, 12, 14, 10, 18, 15, 17, 16, 19]
[0, 1, 2, 3, 4, 10, 11, 12, 13, 14, 5, 6, 7, 8, 9, 15, 16, 17, 18, 19]
[1, 13, 12, 5, 19, 9, 11, 6, 4, 16, 17, 3, 8, 18, 7, 10, 15, 0, 14, 2]
"""
N = x.shape[axis]
M = int(N / groups) # each group has M samples
idx = []
setseed(seed, target='torch')
if mode in ['whole', 'Whole', 'WHOLE']:
idx = list(randperm(0, N, N).numpy())
if mode in ['intra', 'Intra', 'INTRA']:
for g in range(groups):
idx += list(randperm(int(M * g), int(M * g + M), M).numpy())
if mode in ['inter', 'Inter', 'INTER']:
for g in range(groups):
idx += [list(range(int(M * g), int(M * g + M)))]
groupidx = list(randperm(0, groups, groups).numpy())
iidx = idx.copy()
idx = []
for i in groupidx:
idx += iidx[i]
if extra:
return x[sl(x.dim(), axis=axis, idx=[idx])], idx
else:
return x[sl(x.dim(), axis=axis, idx=[idx])]
def sample_tensor(x,
n,
axis=0,
groups=1,
mode='sequentially',
seed=None,
extra=False):
"""sample a tensor
Sample a tensor sequentially/uniformly/randomly.
Args:
x (torch.Tensor): a torch tensor to be sampled
n (int): sample number
axis (int, optional): the axis to be sampled (the default is 0)
groups (int, optional): number of groups in this tensor (the default is 1)
mode (str, optional): - ``'sequentially'``: evenly spaced (default)
- ``'uniformly'``: [0, int(n/groups)]
- ``'randomly'``: randomly selected, non-returned sampling
seed (None or int, optional): only work for ``'randomly'`` mode (the default is None)
extra (bool, optional): If ``True``, also return the selected indexes, the default is ``False``.
Returns:
y (torch.Tensor): Sampled torch tensor.
idx (list): Sampled indexes, if :attr:`extra` is ``True``, this will also be returned.
Example:
::
setseed(2020, 'torch')
x = th.randint(1000, (20, 3, 4))
y1, idx1 = sample_tensor(x, 10, axis=0, groups=2, mode='sequentially', extra=True)
y2, idx2 = sample_tensor(x, 10, axis=0, groups=2, mode='uniformly', extra=True)
y3, idx3 = sample_tensor(x, 10, axis=0, groups=2, mode='randomly', extra=True)
print(x.shape)
print(y1.shape)
print(y2.shape)
print(y3.shape)
print(idx1)
print(idx2)
print(idx3)
the outputs are as follows:
torch.Size([20, 3, 4])
torch.Size([10, 3, 4])
torch.Size([10, 3, 4])
torch.Size([10, 3, 4])
[0, 1, 2, 3, 4, 10, 11, 12, 13, 14]
[0, 2, 4, 6, 8, 10, 12, 14, 16, 18]
[3, 1, 5, 8, 7, 17, 18, 13, 16, 10]
Raises:
ValueError: The tensor does not has enough samples.
"""
N = x.shape[axis]
M = int(N / groups) # each group has M samples
m = int(n / groups) # each group has m sampled samples
if (M < m):
raise ValueError('The tensor does not has enough samples')
idx = []
if mode in ['sequentially', 'Sequentially']:
for g in range(groups):
idx += list(range(int(M * g), int(M * g) + m))
if mode in ['uniformly', 'Uniformly']:
for g in range(groups):
idx += list(range(int(M * g), int(M * g + M), int(M / m)))[:m]
if mode in ['randomly', 'Randomly']:
setseed(seed, target='torch')
for g in range(groups):
idx += list(randperm(int(M * g), int(M * g + M), m).numpy())
if extra:
return x[sl(x.dim(), axis=axis, idx=[idx])], idx
else:
return x[sl(x.dim(), axis=axis, idx=[idx])]