def prepare_train_data(self):
batch_size = self.train_batch_size
is_neg = cp.logical_not(self._train_labels)
# Do not store verification matrix if using the negatives generation shortcut
neg_mat = None if self.use_neg_trick else cp.array(self._neg_mat)
# If there are no negative samples in the local portion of the training data, do nothing
any_neg = cp.any(is_neg)
if any_neg:
self._train_users[is_neg], self._train_items[is_neg] = generate_negatives(
self._train_users[is_neg], neg_mat, self.num_items, use_trick=self.use_neg_trick
)
shuffled_order = cp.random.permutation(self._train_users.shape[0])
self._train_users = self._train_users[shuffled_order]
self._train_items = self._train_items[shuffled_order]
self._train_labels = self._train_labels[shuffled_order]
is_neg = cp.logical_not(self._train_labels)
# Manually create batches
split_indices = np.arange(batch_size, self._train_users.shape[0], batch_size)
self.train_users_batches = np.split(self._train_users, split_indices)
self.train_items_batches = np.split(self._train_items, split_indices)
self.train_labels_batches = np.split(self._train_labels, split_indices)
def binary_hit_or_miss(input, structure1=None, structure2=None, output=None,
origin1=0, origin2=None):
"""
Multidimensional binary hit-or-miss transform.
The hit-or-miss transform finds the locations of a given pattern
inside the input image.
Args:
input (cupy.ndarray): Binary image where a pattern is to be detected.
structure1 (cupy.ndarray, optional): Part of the structuring element to
be fitted to the foreground (non-zero elements) of ``input``. If no
value is provided, a structure of square connectivity 1 is chosen.
structure2 (cupy.ndarray, optional): Second part of the structuring
element that has to miss completely the foreground. If no value is
provided, the complementary of ``structure1`` is taken.
output (cupy.ndarray, dtype or None, optional): Array of the same shape
as input, into which the output is placed. By default, a new array
is created.
origin1 (int or tuple of ints, optional): Placement of the first part
of the structuring element ``structure1``, by default 0 for a
centered structure.
origin2 (int or tuple of ints or None, optional): Placement of the
second part of the structuring element ``structure2``, by default 0
for a centered structure. If a value is provided for ``origin1``
and not for ``origin2``, then ``origin2`` is set to ``origin1``.
Returns:
cupy.ndarray: Hit-or-miss transform of ``input`` with the given
structuring element (``structure1``, ``structure2``).
.. warning::
This function may synchronize the device.
.. seealso:: :func:`scipy.ndimage.binary_hit_or_miss`
"""
if structure1 is None:
structure1 = generate_binary_structure(input.ndim, 1)
if structure2 is None:
structure2 = cupy.logical_not(structure1)
origin1 = _util._fix_sequence_arg(origin1, input.ndim, 'origin1', int)
if origin2 is None:
origin2 = origin1
else:
origin2 = _util._fix_sequence_arg(origin2, input.ndim, 'origin2', int)
tmp1 = _binary_erosion(input, structure1, 1, None, None, 0, origin1, 0,
False)
inplace = isinstance(output, cupy.ndarray)
result = _binary_erosion(input, structure2, 1, None, output, 0, origin2, 1,
False)
if inplace:
cupy.logical_not(output, output)
cupy.logical_and(tmp1, output, output)
else:
cupy.logical_not(result, result)
return cupy.logical_and(tmp1, result)
def _preprocess(labels):
label_values, inv_idx = cp.unique(labels, return_inverse=True)
if not (label_values == 0).any():
warn('Random walker only segments unlabeled areas, where '
'labels == 0. No zero valued areas in labels were '
'found. Returning provided labels.',
stacklevel=2)
return labels, None, None, None, None
# If some labeled pixels are isolated inside pruned zones, prune them
# as well and keep the labels for the final output
null_mask = labels == 0
pos_mask = labels > 0
mask = labels >= 0
fill = ndi.binary_propagation(null_mask, mask=mask)
isolated = cp.logical_and(pos_mask, cp.logical_not(fill))
pos_mask[isolated] = False
# If the array has pruned zones, be sure that no isolated pixels
# exist between pruned zones (they could not be determined)
if label_values[0] < 0 or cp.any(isolated): # synchronize!
isolated = cp.logical_and(
cp.logical_not(ndi.binary_propagation(pos_mask, mask=mask)),
null_mask)
labels[isolated] = -1
if cp.all(isolated[null_mask]):
warn('All unlabeled pixels are isolated, they could not be '
'determined by the random walker algorithm.',
stacklevel=2)
return labels, None, None, None, None
mask[isolated] = False
mask = cp.atleast_3d(mask)
else:
mask = None
# Reorder label values to have consecutive integers (no gaps)
zero_idx = cp.searchsorted(label_values, cp.array(0))
labels = cp.atleast_3d(inv_idx.reshape(labels.shape) - zero_idx)
nlabels = label_values[zero_idx + 1:].shape[0]
inds_isolated_seeds = cp.nonzero(isolated)
isolated_values = labels[inds_isolated_seeds]
return labels, nlabels, mask, inds_isolated_seeds, isolated_values
def remove_small_holes_gpu(
mask: cupy.ndarray,
area_threshold: int,
) -> None:
""" See scikit-image remove_small_holes()
N.B.
Input array must be a labeled mask.
This is a inplace operation.
"""
cupy.logical_not(mask, out=mask)
remove_small_objects_gpu(mask, area_threshold)
cupy.logical_not(mask, out=mask)
def binary_fill_holes(input, structure=None, output=None, origin=0):
"""Fill the holes in binary objects.
Args:
input (cupy.ndarray): N-D binary array with holes to be filled.
structure (cupy.ndarray, optional): Structuring element used in the
computation; large-size elements make computations faster but may
miss holes separated from the background by thin regions. The
default element (with a square connectivity equal to one) yields
the intuitive result where all holes in the input have been filled.
output (cupy.ndarray, dtype or None, optional): Array of the same shape
as input, into which the output is placed. By default, a new array
is created.
origin (int, tuple of ints, optional): Position of the structuring
element.
Returns:
cupy.ndarray: Transformation of the initial image ``input`` where holes
have been filled.
.. warning::
This function may synchronize the device.
.. seealso:: :func:`scipy.ndimage.binary_fill_holes`
"""
mask = cupy.logical_not(input)
tmp = cupy.zeros(mask.shape, bool)
inplace = isinstance(output, cupy.ndarray)
# TODO (grlee77): set brute_force=False below once implemented
if inplace:
binary_dilation(tmp,
structure,
-1,
mask,
output,
1,
origin,
brute_force=True)
cupy.logical_not(output, output)
else:
output = binary_dilation(tmp,
structure,
-1,
mask,
None,
1,
origin,
brute_force=True)
cupy.logical_not(output, output)
return output
def image_complexity(self, X, masks):
# 복잡도 기준
threshold = 25.0
# reshape
#masks = cp.reshape(masks, masks.shape[:-1])
R, G, B = X[:, :, :, 0:1], X[:, :, :, 1:2], X[:, :, :, 2:]
# compute rg = R - G
rg = cp.absolute(R - G)
# compute yb = 0.5 * (R + G) - B
yb = cp.absolute(0.5 * (R + G) - B)
_masks = cp.logical_not(masks)
# compute the mean and standard deviation of both `rg` and `yb` ()
# no masked_where on cupy
rb_masked = np.ma.masked_where(_masks, rg)
(rb_mean, rb_std) = (cp.mean(rb_masked, axis=(1, 2, 3)),
cp.std(rb_masked, axis=(1, 2, 3)))
yb_masked = np.ma.masked_where(_masks, yb)
(yb_mean, yb_std) = (cp.mean(yb_masked, axis=(1, 2, 3)),
cp.std(yb_masked, axis=(1, 2, 3)))
# combine the mean and standard deviations
std_root = cp.sqrt((rb_std**2) + (yb_std**2))
mean_root = cp.sqrt((rb_mean**2) + (yb_mean**2))
# derive the "colorfulness" metric and return it
complexity = std_root + (0.3 * mean_root)
# 수치 수정
print('image_complexity Done.')
return (complexity >= threshold).astype(cp.uint8)
def remove_small_holes_gpu(
mask: cupy.ndarray,
area_threshold: int,
) -> None:
""" See scikit-image remove_small_holes()
N.B.
Input array must be a binary mask (bool type) or
labeled mask (int type).
This is a inplace operation.
"""
_check_dtype_supported(mask)
cupy.logical_not(mask, out=mask)
remove_small_objects_gpu(mask, area_threshold)
cupy.logical_not(mask, out=mask)
def generate_negatives(neg_users, true_mat, item_range, sort=False, use_trick=False):
"""
Generate negative samples for data augmentation
"""
neg_u = []
neg_i = []
# If using the shortcut, generate negative items without checking if the associated
# user has interacted with it. Speeds up training significantly with very low impact
# on accuracy.
if use_trick:
neg_items = cp.random.randint(0, high=item_range, size=neg_users.shape[0])
return neg_users, neg_items
# Otherwise, generate negative items, check if associated user has interacted with it,
# then generate a new one if true
while len(neg_users) > 0:
neg_items = cp.random.randint(0, high=item_range, size=neg_users.shape[0])
neg_mask = true_mat[neg_users, neg_items]
neg_u.append(neg_users[neg_mask])
neg_i.append(neg_items[neg_mask])
neg_users = neg_users[cp.logical_not(neg_mask)]
neg_users = cp.concatenate(neg_u)
neg_items = cp.concatenate(neg_i)
if sort == False:
return neg_users, neg_items
sorted_users = cp.sort(neg_users)
sort_indices = cp.argsort(neg_users)
return sorted_users, neg_items[sort_indices]
def _comparison(self, other, op, op_name):
if _util.isscalarlike(other):
data = cupy.asarray(other, dtype=self.dtype).reshape(1)
if numpy.isnan(data[0]):
if op_name == '_ne_':
return csr_matrix(cupy.ones(self.shape, dtype=numpy.bool_))
else:
return csr_matrix(self.shape, dtype=numpy.bool_)
indices = cupy.zeros((1, ), dtype=numpy.int32)
indptr = cupy.arange(2, dtype=numpy.int32)
other = csr_matrix((data, indices, indptr), shape=(1, 1))
return binopt_csr(self, other, op_name)
elif _util.isdense(other):
return op(self.todense(), other)
elif isspmatrix_csr(other):
self.sum_duplicates()
other.sum_duplicates()
if op_name in ('_ne_', '_lt_', '_gt_'):
return binopt_csr(self, other, op_name)
warnings.warn(
"Comparing sparse matrices using ==, <=, and >= is "
"inefficient, try using !=, <, or > instead.",
SparseEfficiencyWarning)
if op_name == '_eq_':
opposite_op_name = '_ne_'
elif op_name == '_le_':
opposite_op_name = '_gt_'
elif op_name == '_ge_':
opposite_op_name = '_lt_'
res = binopt_csr(self, other, opposite_op_name)
out = cupy.logical_not(res.toarray())
return csr_matrix(out)
raise NotImplementedError
def transform(self, X) -> SparseCumlArray:
"""Impute all missing values in X.
Parameters
----------
X : {array-like, sparse matrix}, shape (n_samples, n_features)
The input data to complete.
"""
check_is_fitted(self)
X = self._validate_input(X, in_fit=False)
X_indicator = super()._transform_indicator(X)
statistics = self.statistics_
if X.shape[1] != statistics.shape[0]:
raise ValueError("X has %d features per sample, expected %d" %
(X.shape[1], self.statistics_.shape[0]))
# Delete the invalid columns if strategy is not constant
if self.strategy == "constant":
valid_statistics = statistics
else:
# same as np.isnan but also works for object dtypes
invalid_mask = _get_mask(statistics, np.nan)
valid_mask = np.logical_not(invalid_mask)
valid_statistics = statistics[valid_mask]
valid_statistics_indexes = np.flatnonzero(valid_mask)
if invalid_mask.any():
missing = np.arange(X.shape[1])[invalid_mask]
if self.verbose:
warnings.warn("Deleting features without "
"observed values: %s" % missing)
X = X[:, valid_statistics_indexes]
# Do actual imputation
if sparse.issparse(X):
if self.missing_values == 0:
raise ValueError("Imputation not possible when missing_values "
"== 0 and input is sparse. Provide a dense "
"array instead.")
else:
mask = _get_mask(X.data, self.missing_values)
indexes = np.repeat(np.arange(len(X.indptr) - 1, dtype=np.int),
np.diff(X.indptr).tolist())[mask]
X.data[mask] = valid_statistics[indexes].astype(X.dtype,
copy=False)
else:
mask = _get_mask(X, self.missing_values)
if self.strategy == "constant":
X[mask] = valid_statistics[0]
else:
for i, vi in enumerate(valid_statistics_indexes):
feature_idxs = np.flatnonzero(mask[:, vi])
X[feature_idxs, vi] = valid_statistics[i]
X = super()._concatenate_indicator(X, X_indicator)
return X
def test_01_01_circle(self):
"""Test that the Canny filter finds the outlines of a circle"""
i, j = cp.mgrid[-200:200, -200:200].astype(float) / 200
c = cp.abs(cp.sqrt(i * i + j * j) - 0.5) < 0.02
result = feature.canny(c.astype(float), 4, 0, 0,
cp.ones(c.shape, bool))
#
# erode and dilate the circle to get rings that should contain the
# outlines
#
# TODO: grlee77: only implemented brute_force=True, so added that to
# these tests
cd = binary_dilation(c, iterations=3, brute_force=True)
ce = binary_erosion(c, iterations=3, brute_force=True)
cde = cp.logical_and(cd, cp.logical_not(ce))
self.assertTrue(cp.all(cde[result]))
#
# The circle has a radius of 100. There are two rings here, one
# for the inside edge and one for the outside. So that's
# 100 * 2 * 2 * 3 for those places where pi is still 3.
# The edge contains both pixels if there's a tie, so we
# bump the count a little.
point_count = cp.sum(result)
self.assertTrue(point_count > 1200)
self.assertTrue(point_count < 1600)
def logical_not(x: Array, /) -> Array:
"""
Array API compatible wrapper for :py:func:`np.logical_not <numpy.logical_not>`.
See its docstring for more information.
"""
if x.dtype not in _boolean_dtypes:
raise TypeError("Only boolean dtypes are allowed in logical_not")
return Array._new(np.logical_not(x._array))
def extract_color(self, X, masks):
# reshape
masks_ = cp.reshape(masks, masks.shape[:-1]).astype(cp.bool)
# extract color
num_X = len(X)
color = cp.empty((num_X, ), dtype=np.string_)
bg_color = cp.empty((num_X, ), dtype=np.string_)
# test
for i, (x, m) in enumerate(zip(X, masks_)):
#print(self.nearest_color(colors, cp.mean(X[i][masks[i]], axis=0).astype(int)))
color[i] = self.nearest_color(
cp.mean(x[m], axis=0).astype(cp.uint8))
#print(self.nearest_color(colors, cp.mean(X[i][cp.logical_not(masks[i])], axis=0).astype(int)))
bg_color[i] = self.nearest_color(
cp.mean(x[cp.logical_not(m)], axis=0).astype(cp.uint8))
print('extract_color Done.')
return color, bg_color
def test_01_02_circle_with_noise(self):
"""Test that the Canny filter finds the circle outlines
in a noisy image"""
cp.random.seed(0)
i, j = cp.mgrid[-200:200, -200:200].astype(float) / 200
c = cp.abs(cp.sqrt(i * i + j * j) - 0.5) < 0.02
cf = c.astype(float) * 0.5 + cp.random.uniform(size=c.shape) * 0.5
result = F.canny(cf, 4, 0.1, 0.2, cp.ones(c.shape, bool))
#
# erode and dilate the circle to get rings that should contain the
# outlines
#
cd = binary_dilation(c, iterations=4, brute_force=True)
ce = binary_erosion(c, iterations=4, brute_force=True)
cde = cp.logical_and(cd, cp.logical_not(ce))
self.assertTrue(cp.all(cde[result]))
point_count = cp.sum(result)
self.assertTrue(point_count > 1200)
self.assertTrue(point_count < 1600)
def _get_anchor_positive_triplet_mask(labels):
"""Return a 2D mask where mask[a, p] is True iff a and p are distinct and have same label.
Args:
labels: xp.array with shape=(batch_size)
Returns:
mask: xp.array with shape=(batch_size, batch_size), dtype=xp.bool
"""
# Check that i and j are distinct
indices_not_equal = xp.logical_not(
xp.diag(xp.ones(labels.size, dtype=xp.bool)))
# Check if labels[i] == labels[j]
# By using broadcasting:
# Left side's shape (1, batch_size) => (*, batch_size)
# Right side's shape (batch_size, 1) => (batch_size, *)
labels_equal = xp.expand_dims(labels, axis=0) == xp.expand_dims(labels,
axis=1)
# Combine the two masks
return indices_not_equal & labels_equal
def binary_blobs(length=512, blob_size_fraction=0.1, n_dim=2,
volume_fraction=0.5, seed=None):
"""
Generate synthetic binary image with several rounded blob-like objects.
Parameters
----------
length : int, optional
Linear size of output image.
blob_size_fraction : float, optional
Typical linear size of blob, as a fraction of ``length``, should be
smaller than 1.
n_dim : int, optional
Number of dimensions of output image.
volume_fraction : float, default 0.5
Fraction of image pixels covered by the blobs (where the output is 1).
Should be in [0, 1].
seed : int, optional
Seed to initialize the random number generator.
If `None`, a random seed from the operating system is used.
Returns
-------
blobs : ndarray of bools
Output binary image
Notes
-----
Warning: CuPy does not give identical randomly generated numbers as NumPy,
so using a specific seed here will not give an identical pattern to the
scikit-image implementation.
The behavior for a given random seed may also change across CuPy major
versions.
See: https://docs.cupy.dev/en/stable/reference/random.html
Examples
--------
>>> from cucim.skimage import data
>>> # tiny size (5, 5)
>>> blobs = data.binary_blobs(length=5, blob_size_fraction=0.2, seed=1)
>>> # larger size
>>> blobs = data.binary_blobs(length=256, blob_size_fraction=0.1)
>>> # Finer structures
>>> blobs = data.binary_blobs(length=256, blob_size_fraction=0.05)
>>> # Blobs cover a smaller volume fraction of the image
>>> blobs = data.binary_blobs(length=256, volume_fraction=0.3)
"""
# filters is quite an expensive import since it imports all of scipy.signal
# We lazy import here
from ..filters import gaussian
rs = cp.random.RandomState(seed)
shape = tuple([length] * n_dim)
mask = cp.zeros(shape)
n_pts = max(int(1. / blob_size_fraction) ** n_dim, 1)
points = (length * rs.rand(n_dim, n_pts)).astype(int)
mask[tuple(indices for indices in points)] = 1
mask = gaussian(mask, sigma=0.25 * length * blob_size_fraction)
threshold = cp.percentile(mask, 100 * (1 - volume_fraction))
return cp.logical_not(mask < threshold)
def cross_correlate_masked(arr1,
arr2,
m1,
m2,
mode="full",
axes=(-2, -1),
overlap_ratio=0.3):
"""
Masked normalized cross-correlation between arrays.
Parameters
----------
arr1 : ndarray
First array.
arr2 : ndarray
Seconds array. The dimensions of `arr2` along axes that are not
transformed should be equal to that of `arr1`.
m1 : ndarray
Mask of `arr1`. The mask should evaluate to `True`
(or 1) on valid pixels. `m1` should have the same shape as `arr1`.
m2 : ndarray
Mask of `arr2`. The mask should evaluate to `True`
(or 1) on valid pixels. `m2` should have the same shape as `arr2`.
mode : {'full', 'same'}, optional
'full':
This returns the convolution at each point of overlap. At
the end-points of the convolution, the signals do not overlap
completely, and boundary effects may be seen.
'same':
The output is the same size as `arr1`, centered with respect
to the `‘full’` output. Boundary effects are less prominent.
axes : tuple of ints, optional
Axes along which to compute the cross-correlation.
overlap_ratio : float, optional
Minimum allowed overlap ratio between images. The correlation for
translations corresponding with an overlap ratio lower than this
threshold will be ignored. A lower `overlap_ratio` leads to smaller
maximum translation, while a higher `overlap_ratio` leads to greater
robustness against spurious matches due to small overlap between
masked images.
Returns
-------
out : ndarray
Masked normalized cross-correlation.
Raises
------
ValueError : if correlation `mode` is not valid, or array dimensions along
non-transformation axes are not equal.
References
----------
.. [1] Dirk Padfield. Masked Object Registration in the Fourier Domain.
IEEE Transactions on Image Processing, vol. 21(5),
pp. 2706-2718 (2012). :DOI:`10.1109/TIP.2011.2181402`
.. [2] D. Padfield. "Masked FFT registration". In Proc. Computer Vision and
Pattern Recognition, pp. 2918-2925 (2010).
:DOI:`10.1109/CVPR.2010.5540032`
"""
if mode not in {"full", "same"}:
raise ValueError("Correlation mode {} is not valid.".format(mode))
if arr1.dtype.kind == "c" or arr2.dtype.kind == "c":
raise ValueError("complex-valued arr1, arr2 are not supported")
fixed_image = cp.asarray(arr1, dtype=np.float)
fixed_mask = cp.asarray(m1, dtype=np.bool)
moving_image = cp.asarray(arr2, dtype=np.float)
moving_mask = cp.asarray(m2, dtype=np.bool)
eps = np.finfo(np.float).eps
# Array dimensions along non-transformation axes should be equal.
all_axes = set(range(fixed_image.ndim))
for axis in all_axes - set(axes):
if fixed_image.shape[axis] != moving_image.shape[axis]:
raise ValueError(
"Array shapes along non-transformation axes should be "
"equal, but dimensions along axis {a} are not".format(a=axis))
# Determine final size along transformation axes
# Note that it might be faster to compute Fourier transform in a slightly
# larger shape (`fast_shape`). Then, after all fourier transforms are done,
# we slice back to`final_shape` using `final_slice`.
final_shape = list(arr1.shape)
for axis in axes:
final_shape[axis] = (fixed_image.shape[axis] +
moving_image.shape[axis] - 1)
final_shape = tuple(final_shape)
final_slice = tuple([slice(0, int(sz)) for sz in final_shape])
# Extent transform axes to the next fast length (i.e. multiple of 3, 5, or
# 7)
fast_shape = tuple([next_fast_len(final_shape[ax]) for ax in axes])
# We use numpy.fft or the new scipy.fft because they allow leaving the
# transform axes unchanged which was not possible with scipy.fftpack's
# fftn/ifftn in older versions of SciPy.
# E.g. arr shape (2, 3, 7), transform along axes (0, 1) with shape (4, 4)
# results in arr_fft shape (4, 4, 7)
fft = partial(fftmodule.fftn, s=fast_shape, axes=axes)
ifft = partial(fftmodule.ifftn, s=fast_shape, axes=axes)
fixed_image[cp.logical_not(fixed_mask)] = 0.0
moving_image[cp.logical_not(moving_mask)] = 0.0
# N-dimensional analog to rotation by 180deg is flip over all relevant axes.
# See [1] for discussion.
rotated_moving_image = _flip(moving_image, axes=axes)
rotated_moving_mask = _flip(moving_mask, axes=axes)
fixed_fft = fft(fixed_image)
rotated_moving_fft = fft(rotated_moving_image)
fixed_mask_fft = fft(fixed_mask)
rotated_moving_mask_fft = fft(rotated_moving_mask)
# Calculate overlap of masks at every point in the convolution.
# Locations with high overlap should not be taken into account.
number_overlap_masked_px = cp.real(
ifft(rotated_moving_mask_fft * fixed_mask_fft))
number_overlap_masked_px[:] = cp.around(number_overlap_masked_px)
number_overlap_masked_px[:] = cp.fmax(number_overlap_masked_px, eps)
masked_correlated_fixed_fft = ifft(rotated_moving_mask_fft * fixed_fft)
masked_correlated_rotated_moving_fft = ifft(fixed_mask_fft *
rotated_moving_fft)
numerator = ifft(rotated_moving_fft * fixed_fft)
numerator -= (masked_correlated_fixed_fft *
masked_correlated_rotated_moving_fft /
number_overlap_masked_px)
fixed_squared_fft = fft(cp.square(fixed_image))
fixed_denom = ifft(rotated_moving_mask_fft * fixed_squared_fft)
fixed_denom -= (cp.square(masked_correlated_fixed_fft) /
number_overlap_masked_px)
fixed_denom[:] = cp.fmax(fixed_denom, 0.0)
rotated_moving_squared_fft = fft(cp.square(rotated_moving_image))
moving_denom = ifft(fixed_mask_fft * rotated_moving_squared_fft)
moving_denom -= (cp.square(masked_correlated_rotated_moving_fft) /
number_overlap_masked_px)
moving_denom[:] = cp.fmax(moving_denom, 0.0)
denom = cp.sqrt(fixed_denom * moving_denom)
# Slice back to expected convolution shape.
numerator = numerator[final_slice]
denom = denom[final_slice]
number_overlap_masked_px = number_overlap_masked_px[final_slice]
if mode == "same":
_centering = partial(_centered, newshape=fixed_image.shape, axes=axes)
denom = _centering(denom)
numerator = _centering(numerator)
number_overlap_masked_px = _centering(number_overlap_masked_px)
# Pixels where `denom` is very small will introduce large
# numbers after division. To get around this problem,
# we zero-out problematic pixels.
tol = 1e3 * eps * cp.max(cp.abs(denom), axis=axes, keepdims=True)
nonzero_indices = denom > tol
# TODO: grlee77: Added a cast to real here.
# probably it should be real earlier?
numerator = numerator.real
denom = denom.real
out = cp.zeros_like(denom)
out[nonzero_indices] = numerator[nonzero_indices] / denom[nonzero_indices]
cp.clip(out, a_min=-1, a_max=1, out=out)
# Apply overlap ratio threshold
number_px_threshold = overlap_ratio * cp.max(
number_overlap_masked_px, axis=axes, keepdims=True)
out[number_overlap_masked_px < number_px_threshold] = 0.0
return out
def evaluate_chunks(
results: [cp.ndarray, cp.ndarray,
cp.ndarray], # closest triangle, distance, projection
all_pts: cp.ndarray = None,
vertices: cp.ndarray = None,
edges: cp.ndarray = None,
edge_norms: cp.ndarray = None,
edge_normssq: cp.ndarray = None,
normals: cp.ndarray = None,
norms: cp.ndarray = None,
normssq: cp.ndarray = None,
zero_tensor: cp.ndarray = None,
one_tensor: cp.ndarray = None,
tris: cp.ndarray = None,
vertex_normals: cp.ndarray = None,
bounding_box: dict = None,
chunk_size: int = None,
num_verts: int = None) -> None:
#
# Expand vertex normals if non empty
if vertex_normals is not None:
vertex_normals = vertex_normals[tris]
vertex_normals = cp.tile(cp.expand_dims(vertex_normals, axis=2),
(1, 1, chunk_size, 1))
# begin = time.time()
#
# Load and extend the batch
num_chunks = all_pts.shape[0] // chunk_size
for i in range(num_chunks):
#
# Get subset of the query points
start_index = i * chunk_size
end_index = (i + 1) * chunk_size
pts = all_pts[start_index:end_index, :]
#
# Match the dimensions to those assumed above.
# REPEATED REPEATED
# [triangle_index, vert_index, querypoint_index, coordinates]
pts = cp.tile(cp.expand_dims(pts, axis=(0, 1)), (num_verts, 3, 1, 1))
#
# Compute the differences between
# vertices on each triangle and the
# points of interest
#
# [triangle_index, vert_index, querypoint_index, coordinates]
# ===================
# [:,0,:,:] = p - p1
# [:,1,:,:] = p - p2
# [:,2,:,:] = p - p3
diff_vectors = pts - vertices
#
# Compute alpha, beta, gamma
barycentric = cp.empty(diff_vectors.shape)
#
# gamma = u x (p - p1)
barycentric[:, 2, :, :] = cp.cross(edges[:, 0, :, :],
diff_vectors[:, 0, :, :])
# beta = (p - p1) x v
barycentric[:, 1, :, :] = cp.cross(diff_vectors[:, 0, :, :],
edges[:, 1, :, :])
# alpha = w x (p - p2)
barycentric[:, 0, :, :] = cp.cross(edges[:, 2, :, :],
diff_vectors[:, 1, :, :])
barycentric = cp.divide(
cp.sum(cp.multiply(barycentric, normals), axis=3), normssq)
#
# Test conditions
less_than_one = cp.less_equal(barycentric, one_tensor)
more_than_zero = cp.greater_equal(barycentric, zero_tensor)
#
# if 0 <= gamma and gamma <= 1
# and 0 <= beta and beta <= 1
# and 0 <= alpha and alpha <= 1:
cond1 = cp.logical_and(less_than_one, more_than_zero)
#
# if gamma <= 0:
cond2 = cp.logical_not(more_than_zero[:, 2, :])
cond2 = cp.tile(cp.expand_dims(cond2, axis=1), (1, 3, 1))
#
# if beta <= 0:
cond3 = cp.logical_not(more_than_zero[:, 1, :])
cond3 = cp.tile(cp.expand_dims(cond3, axis=1), (1, 3, 1))
#
# if alpha <= 0:
cond4 = cp.logical_not(more_than_zero[:, 0, :])
cond4 = cp.tile(cp.expand_dims(cond4, axis=1), (1, 3, 1))
#
# Get the projections for each case
xi = cp.empty(barycentric.shape)
barycentric_ext = cp.tile(cp.expand_dims(barycentric, axis=3),
(1, 1, 1, 3))
proj = cp.sum(cp.multiply(barycentric_ext, vertices), axis=1)
#
# if 0 <= gamma and gamma <= 1
# and 0 <= beta and beta <= 1
# and 0 <= alpha and alpha <= 1:
xi[cond1] = barycentric[cond1]
#
# if gamma <= 0:
# x = p - p1
# u = p2 - p1
# a = p1
# b = p2
t2 = cp.divide(
#
# u.dot(x)
cp.sum(cp.multiply(edges[:, 0, :, :], diff_vectors[:, 0, :, :]),
axis=2),
edge_normssq[:, 0])
xi2 = cp.zeros((t2.shape[0], 3, t2.shape[1]))
xi2[:, 0, :] = -t2 + 1
xi2[:, 1, :] = t2
#
t2 = cp.tile(cp.expand_dims(t2, axis=2), (1, 1, 3))
lz = cp.less(t2, cp.zeros(t2.shape))
go = cp.greater(t2, cp.ones(t2.shape))
proj2 = vertices[:, 0, :, :] + cp.multiply(t2, edges[:, 0, :, :])
proj2[lz] = vertices[:, 0, :, :][lz]
proj2[go] = vertices[:, 1, :, :][go]
#
xi[cond2] = xi2[cond2]
proj[cp.swapaxes(cond2, 1, 2)] = proj2[cp.swapaxes(cond2, 1, 2)]
#
# if beta <= 0:
# x = p - p1
# v = p3 - p1
# a = p1
# b = p3
t3 = cp.divide(
#
# v.dot(x)
cp.sum(cp.multiply(edges[:, 1, :, :], diff_vectors[:, 0, :, :]),
axis=2),
edge_normssq[:, 1])
xi3 = cp.zeros((t3.shape[0], 3, t3.shape[1]))
xi3[:, 0, :] = -t3 + 1
xi3[:, 2, :] = t3
#
t3 = cp.tile(cp.expand_dims(t3, axis=2), (1, 1, 3))
lz = cp.less(t3, cp.zeros(t3.shape))
go = cp.greater(t3, cp.ones(t3.shape))
proj3 = vertices[:, 0, :, :] + cp.multiply(t3, edges[:, 1, :, :])
proj3[lz] = vertices[:, 0, :, :][lz]
proj3[go] = vertices[:, 2, :, :][go]
#
xi[cond3] = xi3[cond3]
proj[cp.swapaxes(cond3, 1, 2)] = proj3[cp.swapaxes(cond3, 1, 2)]
#
# if alpha <= 0:
# y = p - p2
# w = p3 - p2
# a = p2
# b = p3
t4 = cp.divide(
#
# w.dot(y)
cp.sum(cp.multiply(edges[:, 2, :, :], diff_vectors[:, 1, :, :]),
axis=2),
edge_normssq[:, 2])
xi4 = cp.zeros((t4.shape[0], 3, t4.shape[1]))
xi4[:, 1, :] = -t4 + 1
xi4[:, 2, :] = t4
#
t4 = cp.tile(cp.expand_dims(t4, axis=2), (1, 1, 3))
lz = cp.less(t4, cp.zeros(t4.shape))
go = cp.greater(t4, cp.ones(t4.shape))
proj4 = vertices[:, 1, :, :] + cp.multiply(t4, edges[:, 2, :, :])
proj4[lz] = vertices[:, 1, :, :][lz]
proj4[go] = vertices[:, 2, :, :][go]
#
xi[cond4] = xi4[cond4]
proj[cp.swapaxes(cond4, 1, 2)] = proj4[cp.swapaxes(cond4, 1, 2)]
vec_to_point = pts[:, 0, :, :] - proj
distances = cp.linalg.norm(vec_to_point, axis=2)
# n = "\n"
# print(f"{pts[:,0,:,:]=}")
# print(f"{proj=}")
# print(f"{pts[:,0,:,:] - proj=}")
# print(f"{distances=}")
min_distances = cp.min(distances, axis=0)
closest_triangles = cp.argmin(distances, axis=0)
projections = proj[closest_triangles, np.arange(chunk_size), :]
#
# Distinguish close triangles
is_close = cp.isclose(distances, min_distances)
#
# Determine sign
signed_normal = normals[:, 0, :, :]
if vertex_normals is not None:
signed_normal = cp.sum(vertex_normals.transpose() * xi.transpose(),
axis=2).transpose()
is_negative = cp.less_equal(
cp.sum(cp.multiply(vec_to_point, signed_normal), axis=2), 0.)
#
# Combine
is_close_and_negative = cp.logical_and(is_close, is_negative)
#
# Determine if inside
is_inside = cp.all(cp.logical_or(is_close_and_negative,
cp.logical_not(is_close)),
axis=0)
#
# Overwrite the signs of points
# that are outside of the box
if bounding_box is not None:
#
# Extract
rotation_matrix = cp.asarray(bounding_box['rotation_matrix'])
translation_vector = cp.asarray(bounding_box['translation_vector'])
size = cp.asarray(bounding_box['size'])
#
# Transform
transformed_pts = cp.dot(
all_pts[start_index:end_index, :] - translation_vector,
rotation_matrix)
#
# Determine if outside bbox
inside_bbox = cp.all(cp.logical_and(
cp.less_equal(0., transformed_pts),
cp.less_equal(transformed_pts, size)),
axis=1)
#
# Treat points outside bbox as
# being outside of lumen
print(f"{inside_bbox=}")
is_inside = cp.logical_and(is_inside, inside_bbox)
#
# Apply sign to indicate whether the distance is
# inside or outside the mesh.
min_distances[is_inside] = -1 * min_distances[is_inside]
#
# Emplace results
# [triangle_index, vert_index, querypoint_index, coordinates]
results[0][start_index:end_index] = closest_triangles
results[1][start_index:end_index] = min_distances
results[2][start_index:end_index, :] = projections
def __init__(self, mask):
self.mask = mask
self.e = cp.array([cp.array([0, 0]) for i in range(9)])
self.w = cp.array([0.0 for i in range(9)])
for i in range(9):
if i == 0:
self.w[i] = 4 / 9
self.e[i][0] = 0
self.e[i][1] = 0
elif i == 1:
self.w[i] = 1 / 9
self.e[i][0] = 1
self.e[i][1] = 0
elif i == 2:
self.w[i] = 1 / 9
self.e[i][0] = 0
self.e[i][1] = 1
elif i == 3:
self.w[i] = 1 / 9
self.e[i][0] = -1
self.e[i][1] = 0
elif i == 4:
self.w[i] = 1 / 9
self.e[i][0] = 0
self.e[i][1] = -1
elif i == 5:
self.w[i] = 1 / 36
self.e[i][0] = 1
self.e[i][1] = 1
elif i == 6:
self.w[i] = 1 / 36
self.e[i][0] = -1
self.e[i][1] = 1
elif i == 7:
self.w[i] = 1 / 36
self.e[i][0] = -1
self.e[i][1] = -1
elif i == 8:
self.w[i] = 1 / 36
self.e[i][0] = 1
self.e[i][1] = -1
print(self.e[i], self.w[i])
self.psi = cp.full((H, W), -1.0)
self.psi[:, :10] = 1.0
self.block_mask = cp.logical_not(mask)
self.psi[self.block_mask] = psi_wall
self.left_wall = cp.full((H, 1), 1.0)
self.right_wall = cp.full((H, 1), -1.0)
self.gamma = gamma
self.top_bottom_wall = cp.full((1, W + 2), psi_wall)
self.nabla_psix = cp.zeros((H, W))
self.nabla_psiy = cp.zeros((H, W))
self.nabla_psi2 = cp.zeros((H, W))
self.rho = 1.0 - 0.05 * cp.random.rand(H, W)[mask]
#self.rho = np.ones((H, W))[mask] * rho0 # macroscopic density
self.ux = cp.zeros((H, W))[mask]
self.uy = cp.zeros((H, W))[mask]
self.p = cp.zeros((H, W))[mask]
self.f = cp.array([cp.zeros((H, W)) for i in range(9)])
self.g = cp.array([cp.zeros((H, W)) for i in range(9)])
self.feq = cp.array([cp.zeros((H, W))[mask] for i in range(9)])
self.geq = cp.array([cp.zeros((H, W))[mask] for i in range(9)])
self.mu = cp.zeros((H, W))[mask]
self.mix_tau = cp.zeros((H, W))[mask]
self.F = cp.array([cp.zeros((H, W))[mask] for i in range(9)])
self.nabla_psix = self.getNabla_psix()
self.nabla_psiy = self.getNabla_psiy()
self.nabla_psi2 = self.getNabla_psi2()
mu = self.getMu()
# self.uy = self.getUy()
self.p = self.getP()
self.mix_tau = self.getMix_tau()
for i in range(9):
self.f[i][self.mask] = self.getfeq(i)
self.g[i][self.mask] = self.getgeq(i)
def main():
rect_corner_list = []
# for rectangle block
# while True:
# if count * 20 > 360:
# break
# if flag:
# rect_corner_list.append(((count * 20, 60), ((count + 1) * 20, 80)))
# rect_corner_list.append(((count * 20, 140), ((count + 1) * 20, 160)))
# rect_corner_list.append(((count * 20, 220), ((count + 1) * 20, 240)))
# rect_corner_list.append(((count * 20, 300), ((count + 1) * 20, 320)))
# flag = False
# count += 2
# continue
# elif not flag:
# rect_corner_list.append(((count * 20, 20), ((count + 1) * 20, 40)))
# rect_corner_list.append(((count * 20, 100), ((count + 1) * 20, 120)))
# rect_corner_list.append(((count * 20, 180), ((count + 1) * 20, 200)))
# rect_corner_list.append(((count * 20, 260), ((count + 1) * 20, 280)))
# rect_corner_list.append(((count * 20, 340), ((count + 1) * 20, 360)))
# flag = True
# count += 2
# block_psi_all, corner_list = setblock(rect_corner_list)
cr = Createblock(H, W)
bb = Bounce_back(H, W)
circle_list = [[(int(W / 2 - W / 3), int(H / 2)), 30]]
circle_list = []
# circle_list.append(((int(W/2), int(H/2)), 30))
r = 13
xx = 78
count = 2
flag = True
mabiki = MAX_T // 150
# circle_list.append(((count * 2 * r, xx + r), r))
# circle_list.append(((count * 2 * r, 2 * xx + 3 * r), r))
# circle_list.append(((count * 2 * r, 3 * xx + 5 * r + 1), r + 1))
# circle_list.append(((count * 2 * r, 3 * xx + 5 * r), r))
# while True:
# if count * 2 * r > 380:
# break
# if flag:
# circle_list.append(((count * 2 * r, 3 * r), r + 5))
# circle_list.append(((count * 2 * r, xx + 5 * r), r + 5))
# circle_list.append(((count * 2 * r, 2 * xx + 7 * r), r + 5))
# circle_list.append(((count * 2 * r, 3 * xx + 9 * r), r + 5))
# print(3 * xx + 5 * r)
# print(2 * xx + 3 * r)
# flag = False
# count += 2
# continue
# elif not flag:
# circle_list.append(((count * 2 * r, xx + r), r + 5))
# circle_list.append(((count * 2 * r, 2 * xx + 3 * r), r + 5))
# circle_list.append(((count * 2 * r, 3 * xx + 5 * r), r + 5))
# flag = True
# count += 2
# while True:
# if count * r > 380:
# break
# circle_list.append(((count * r, 2 * r), r))
# circle_list.append(((count * r, 6 * r), r))
# circle_list.append(((count * r, 10 * r), r))
# circle_list.append(((count * r, 14 * r), r))
# circle_list.append(((count * r, 18 * r), r))
# count += 4
block_psi_all, side_list, concave_list, convex_list = cr.setCirleblock(circle_list)
block_mask = cp.where(block_psi_all == 1, True, False)
mask = cp.logical_not(block_mask)
cm = Compute(mask)
cc = cp.array([cm.psi])
for i in range(MAX_T):
for j in range(9):
cm.F[j] = cm.getLarge_F(j)
cm.feq[j] = cm.getfeq(j)
cm.geq[j] = cm.getgeq(j)
cm.f[j][mask] = cm.getF(j)
cm.g[j][mask] = cm.getG(j)
if i % mabiki == 0:
cc = cp.append(cc, cp.array([cm.psi]), axis=0)
print("timestep:{}".format(i))
f_behind = copy.deepcopy(cm.f)
g_behind = copy.deepcopy(cm.g)
stream(cm.f, cm.g)
bb.halfway_bounceback_circle(side_list, concave_list, convex_list, f_behind, g_behind, cm.f, cm.g)
# halfway_bounceback(corner_list, f_behind, g_behind, cm.f, cm.g)
# bottom_top_wall(f_behind[:, 1:-1], g_behind[:, 1:-1], cm.f[:, 1:-1], cm.g[:, 1:-1])
cm.zou_he_boundary_inlet()
cm.zou_he_boundary_outlet()
cm.rho = cm.getRho()
cm.udpatePsi()
cm.nabla_psix = cm.getNabla_psix()
cm.nabla_psiy = cm.getNabla_psiy()
cm.nabla_psi2 = cm.getNabla_psi2()
cm.mu = cm.getMu()
cm.ux = cm.getUx()
cm.uy = cm.getUy()
cm.p = cm.getP()
cm.mix_tau = cm.getMix_tau()
y = [i for i in range(H)]
x = [i for i in range(W)]
# fig = plt.figure()
# plt.colorbar(plt.pcolor(x, y, cc[0], cmap='RdBu'))
# ani = animation.FuncAnimation(fig, update, fargs=(x, y, cc), frames=int(len(cc)))
# ani.save('../movies/MAX_T{}_Pe{}_M{}_Ca{}_wall{}.mp4'.format(MAX_T, Pe, M, Ca, psi_wall), fps=10)
plt.figure()
plt.pcolor(x, y, cm.psi, label='MAX_T{}_Pe{}_M{}_Ca{}_wall{}'.format(MAX_T, Pe, M, Ca, psi_wall), cmap='RdBu')
plt.colorbar()
plt.legend()
# plt.grid()
plt.show()
def approximate_polygon(coords, tolerance):
"""Approximate a polygonal chain with the specified tolerance.
It is based on the Douglas-Peucker algorithm.
Note that the approximated polygon is always within the convex hull of the
original polygon.
Parameters
----------
coords : (N, 2) array
Coordinate array.
tolerance : float
Maximum distance from original points of polygon to approximated
polygonal chain. If tolerance is 0, the original coordinate array
is returned.
Returns
-------
coords : (M, 2) array
Approximated polygonal chain where M <= N.
References
----------
.. [1] https://en.wikipedia.org/wiki/Ramer-Douglas-Peucker_algorithm
"""
if tolerance <= 0:
return coords
chain = cp.zeros(coords.shape[0], "bool")
# pre-allocate distance array for all points
dists = cp.zeros(coords.shape[0])
chain[0] = True
chain[-1] = True
pos_stack = [(0, chain.shape[0] - 1)]
end_of_chain = False
while not end_of_chain:
start, end = pos_stack.pop()
# determine properties of current line segment
r0, c0 = cp.asnumpy(coords[start, :])
r1, c1 = cp.asnumpy(coords[end, :])
dr = r1 - r0
dc = c1 - c0
segment_angle = -cp.arctan2(dr, dc)
segment_dist = c0 * cp.sin(segment_angle) + r0 * cp.cos(segment_angle)
# select points in-between line segment
segment_coords = coords[start + 1:end, :]
segment_dists = dists[start + 1:end]
# check whether to take perpendicular or euclidean distance with
# inner product of vectors
# vectors from points -> start and end
dr0 = segment_coords[:, 0] - r0
dc0 = segment_coords[:, 1] - c0
dr1 = segment_coords[:, 0] - r1
dc1 = segment_coords[:, 1] - c1
# vectors points -> start and end projected on start -> end vector
projected_lengths0 = dr0 * dr + dc0 * dc
projected_lengths1 = -dr1 * dr - dc1 * dc
perp = cp.logical_and(projected_lengths0 > 0, projected_lengths1 > 0)
eucl = cp.logical_not(perp)
segment_dists[perp] = cp.abs(
segment_coords[perp, 0] * cp.cos(segment_angle) +
segment_coords[perp, 1] * cp.sin(segment_angle) - segment_dist)
segment_dists[eucl] = cp.minimum(
# distance to start point
cp.sqrt(dc0[eucl]**2 + dr0[eucl]**2),
# distance to end point
cp.sqrt(dc1[eucl]**2 + dr1[eucl]**2),
)
if cp.any(segment_dists > tolerance):
# select point with maximum distance to line
new_end = start + cp.argmax(segment_dists) + 1
pos_stack.append((new_end, end))
pos_stack.append((start, new_end))
chain[new_end] = True
if len(pos_stack) == 0:
end_of_chain = True
return coords[chain, :]
def remove_small_holes(ar, area_threshold=64, connectivity=1, in_place=False):
"""Remove contiguous holes smaller than the specified size.
Parameters
----------
ar : ndarray (arbitrary shape, int or bool type)
The array containing the connected components of interest.
area_threshold : int, optional (default: 64)
The maximum area, in pixels, of a contiguous hole that will be filled.
Replaces `min_size`.
connectivity : int, {1, 2, ..., ar.ndim}, optional (default: 1)
The connectivity defining the neighborhood of a pixel.
in_place : bool, optional (default: False)
If `True`, remove the connected components in the input array itself.
Otherwise, make a copy.
Raises
------
TypeError
If the input array is of an invalid type, such as float or string.
ValueError
If the input array contains negative values.
Returns
-------
out : ndarray, same shape and type as input `ar`
The input array with small holes within connected components removed.
Examples
--------
>>> import cupy as cp
>>> from cupyimg.skimage import morphology
>>> a = cp.array([[1, 1, 1, 1, 1, 0],
... [1, 1, 1, 0, 1, 0],
... [1, 0, 0, 1, 1, 0],
... [1, 1, 1, 1, 1, 0]], bool)
>>> b = morphology.remove_small_holes(a, 2)
>>> b
array([[ True, True, True, True, True, False],
[ True, True, True, True, True, False],
[ True, False, False, True, True, False],
[ True, True, True, True, True, False]])
>>> c = morphology.remove_small_holes(a, 2, connectivity=2)
>>> c
array([[ True, True, True, True, True, False],
[ True, True, True, False, True, False],
[ True, False, False, True, True, False],
[ True, True, True, True, True, False]])
>>> d = morphology.remove_small_holes(a, 2, in_place=True)
>>> d is a
True
Notes
-----
If the array type is int, it is assumed that it contains already-labeled
objects. The labels are not kept in the output image (this function always
outputs a bool image). It is suggested that labeling is completed after
using this function.
"""
_check_dtype_supported(ar)
# Creates warning if image is an integer image
if ar.dtype != bool:
warn(
"Any labeled images will be returned as a boolean array. "
"Did you mean to use a boolean array?",
UserWarning,
)
if in_place:
out = ar
else:
out = ar.copy()
# Creating the inverse of ar
if in_place:
cp.logical_not(out, out=out)
else:
out = cp.logical_not(out)
# removing small objects from the inverse of ar
out = remove_small_objects(out, area_threshold, connectivity, in_place)
if in_place:
cp.logical_not(out, out=out)
else:
out = cp.logical_not(out)
return out
def train(source_bpe, target_bpe, source_glove, target_glove, chunk_length,
batch_size, warmup_steps, save_decimation, num_steps, gpu_id, out,
log_level):
if not os.path.exists(out):
os.makedirs(out)
ll = getattr(logging, log_level)
stream_handler = logging.StreamHandler(sys.stdout)
stream_handler.setLevel(ll)
stream_handler.setFormatter(logging.Formatter('%(message)s'))
file_handler = logging.FileHandler(filename=os.path.join(
out, 'training.log'),
mode='a')
file_handler.setLevel(ll)
file_handler.setFormatter(logging.Formatter('%(message)s'))
logger.addHandler(stream_handler)
logger.addHandler(file_handler)
logger.setLevel(ll)
gpu_id = gpu_id if gpu_id is not None else -1
device_name = '@intel64'
if gpu_id >= 0:
device_name = f'@cupy:{gpu_id}'
with chainer.using_device(device_name):
source_vocab = make_vocab(source_glove)
target_vocab = make_vocab(target_glove)
output_model_dim = target_vocab.embedding_size
dataset = make_dataset(source_bpe, target_bpe, source_vocab,
target_vocab, chunk_length)
iterator = MultithreadIterator(dataset, batch_size)
state = TrainingState()
model = Transformer(source_vocab, target_vocab)
model.to_gpu(gpu_id)
optimizer = Adam(beta1=0.99, beta2=0.98, eps=1e-9).setup(model)
load_training(out, model, optimizer, state)
try:
for n, batch in enumerate(iterator):
if n >= num_steps:
break
if (n + 1) % save_decimation == 0:
save_training(out, model, optimizer, state)
model.cleargrads()
gc.collect()
source, target = stack_nested(batch)
source.token_ids.to_gpu(gpu_id)
source.masks.to_gpu(gpu_id)
target.token_ids.to_gpu(gpu_id)
target.masks.to_gpu(gpu_id)
output_probs = model.train_forward(source.token_ids,
target.token_ids,
input_masks=source.masks,
output_masks=target.masks)
unnormalized_loss = F.softmax_cross_entropy(
F.reshape(output_probs,
(output_probs.shape[0] * output_probs.shape[1],
output_probs.shape[2])),
F.reshape(target.token_ids, (target.token_ids.shape[0] *
target.token_ids.shape[1], )),
reduce='no')
loss_mask = xp.reshape(
xp.logical_not(target.masks.array).astype(xp.float32),
(target.masks.shape[0] * target.masks.shape[1], ))
loss = F.sum(unnormalized_loss * loss_mask) / F.sum(loss_mask)
loss.backward()
learning_rate = (output_model_dim**-0.5) * min(
(state.step**-0.5), state.step * (warmup_steps**-1.5))
optimizer.alpha = learning_rate
optimizer.update()
logger.info(
f'time = {int(time.time())} | step = {state.step} | loss = {float(loss.array)} | lr = {learning_rate}'
)
state.step += 1
finally:
save_training(out, model, optimizer, state)
def _unique_update_mask_equal_nan(mask, x0):
mask1 = cupy.logical_not(cupy.isnan(x0))
mask[:] = cupy.logical_and(mask, mask1)