def diag(v, k=0): if not isinstance(v, np.ndarray) and not isinstance(v, Array): raise TypeError( f"v must be a dask array or numpy array, got {type(v)}") name = "diag-" + tokenize(v, k) meta = meta_from_array(v, 2 if v.ndim == 1 else 1) if isinstance(v, np.ndarray) or (hasattr(v, "__array_function__") and not isinstance(v, Array)): if v.ndim == 1: m = abs(k) chunks = ((v.shape[0] + m, ), (v.shape[0] + m, )) dsk = {(name, 0, 0): (np.diag, v, k)} elif v.ndim == 2: kdiag_row_start = max(0, -k) kdiag_row_stop = min(v.shape[0], v.shape[1] - k) len_kdiag = kdiag_row_stop - kdiag_row_start chunks = ((0, ), ) if len_kdiag <= 0 else ((len_kdiag, ), ) dsk = {(name, 0): (np.diag, v, k)} else: raise ValueError("Array must be 1d or 2d only") return Array(dsk, name, chunks, meta=meta) if v.ndim != 1: if v.ndim != 2: raise ValueError("Array must be 1d or 2d only") if k == 0 and v.chunks[0] == v.chunks[1]: dsk = {(name, i): (np.diag, row[i]) for i, row in enumerate(v.__dask_keys__())} graph = HighLevelGraph.from_collections(name, dsk, dependencies=[v]) return Array(graph, name, (v.chunks[0], ), meta=meta) else: return diagonal(v, k) if k == 0: chunks_1d = v.chunks[0] blocks = v.__dask_keys__() dsk = {} for i, m in enumerate(chunks_1d): for j, n in enumerate(chunks_1d): key = (name, i, j) if i == j: dsk[key] = (np.diag, blocks[i]) else: dsk[key] = (np.zeros, (m, n)) dsk[key] = (partial(np.zeros_like, shape=(m, n)), meta) graph = HighLevelGraph.from_collections(name, dsk, dependencies=[v]) return Array(graph, name, (chunks_1d, chunks_1d), meta=meta) elif k > 0: return pad(diag(v), [[0, k], [k, 0]], mode="constant") elif k < 0: return pad(diag(v), [[-k, 0], [0, -k]], mode="constant")
def wrap_func_like(func, *args, **kwargs): """ Transform np creation function into blocked version """ x = args[0] meta = meta_from_array(x) shape = kwargs.get("shape", x.shape) parsed = _parse_wrap_args(func, args, kwargs, shape) shape = parsed["shape"] dtype = parsed["dtype"] chunks = parsed["chunks"] name = parsed["name"] kwargs = parsed["kwargs"] keys = product([name], *[range(len(bd)) for bd in chunks]) shapes = product(*chunks) shapes = list(shapes) kw = [kwargs for _ in shapes] for i, s in enumerate(list(shapes)): kw[i]["shape"] = s vals = ((partial(func, dtype=dtype, **k), ) + args for (k, s) in zip(kw, shapes)) dsk = dict(zip(keys, vals)) return Array(dsk, name, chunks, meta=meta.astype(dtype))
def _dask_unique(x, return_index=True): from dask.array.core import Array from dask import sharedict from numpy import cumsum, concatenate, unique from numpy.testing import assert_ assert_(return_index) name = "unique-" + x.name def _unique(x): return unique(x, return_index=return_index) dsk = dict(((name, i), (_unique, key)) for i, key in enumerate(x._keys())) parts = Array._get(sharedict.merge((name, dsk), x.dask), list(dsk.keys())) arrs = [a[0] for a in parts] chunks = x.chunks[0] offset = cumsum((0, ) + chunks)[:-1] idxs = [parts[i][1] + offset[i] for i in range(len(parts))] arr = concatenate(arrs) idx = concatenate(idxs) u, i = unique(arr, return_index=True) return u, idx[i]
def overlap_internal(x, axes): """Share boundaries between neighboring blocks Parameters ---------- x: da.Array A dask array axes: dict The size of the shared boundary per axis The axes input informs how many cells to overlap between neighboring blocks {0: 2, 2: 5} means share two cells in 0 axis, 5 cells in 2 axis """ token = tokenize(x, axes) name = "overlap-" + token graph = ArrayOverlapLayer( name=x.name, axes=axes, chunks=x.chunks, numblocks=x.numblocks, token=token, ) graph = HighLevelGraph.from_collections(name, graph, dependencies=[x]) chunks = _overlap_internal_chunks(x.chunks, axes) return Array(graph, name, chunks, meta=x)
def read(filename, shape, chunks): from dask.highlevelgraph import HighLevelGraph from dask.array.core import normalize_chunks, Array from itertools import product from ...tunable import delayed from numpy import prod, dtype import xmltodict records = scan_file(filename) records = {r["lime_type"]: r for r in records} data_record = records["ildg-binary-data"] data_offset = data_record["pos"] info = xmltodict.parse(records["ildg-format"]["data"])["ildgFormat"] dtype = dtype("complex%d" % (int(info["precision"]) * 2)) assert data_record["data_length"] == prod(shape) * dtype.itemsize normal_chunks = normalize_chunks(chunks, shape=shape) chunks_id = list(product(*[range(len(bd)) for bd in normal_chunks])) reads = [ delayed(read_chunk)(filename, shape, dtype, data_offset, chunks, chunk_id) for chunk_id in chunks_id ] keys = [(filename, *chunk_id) for chunk_id in chunks_id] vals = [read.key for read in reads] dsk = dict(zip(keys, vals)) graph = HighLevelGraph.from_collections(filename, dsk, dependencies=reads) return Array(graph, filename, normal_chunks, dtype=dtype)
def _compute_rechunk(x, chunks): """Compute the rechunk of *x* to the given *chunks*.""" if x.size == 0: # Special case for empty array, as the algorithm below does not behave correctly return empty(x.shape, chunks=chunks, dtype=x.dtype) ndim = x.ndim crossed = intersect_chunks(x.chunks, chunks) x2 = dict() intermediates = dict() token = tokenize(x, chunks) merge_name = "rechunk-merge-" + token split_name = "rechunk-split-" + token split_name_suffixes = count() # Pre-allocate old block references, to allow re-use and reduce the # graph's memory footprint a bit. old_blocks = np.empty([len(c) for c in x.chunks], dtype="O") for index in np.ndindex(old_blocks.shape): old_blocks[index] = (x.name, ) + index # Iterate over all new blocks new_index = product(*(range(len(c)) for c in chunks)) for new_idx, cross1 in zip(new_index, crossed): key = (merge_name, ) + new_idx old_block_indices = [[cr[i][0] for cr in cross1] for i in range(ndim)] subdims1 = [len(set(old_block_indices[i])) for i in range(ndim)] rec_cat_arg = np.empty(subdims1, dtype="O") rec_cat_arg_flat = rec_cat_arg.flat # Iterate over the old blocks required to build the new block for rec_cat_index, ind_slices in enumerate(cross1): old_block_index, slices = zip(*ind_slices) name = (split_name, next(split_name_suffixes)) old_index = old_blocks[old_block_index][1:] if all(slc.start == 0 and slc.stop == x.chunks[i][ind] for i, (slc, ind) in enumerate(zip(slices, old_index))): rec_cat_arg_flat[rec_cat_index] = old_blocks[old_block_index] else: intermediates[name] = (getitem, old_blocks[old_block_index], slices) rec_cat_arg_flat[rec_cat_index] = name assert rec_cat_index == rec_cat_arg.size - 1 # New block is formed by concatenation of sliced old blocks if all(d == 1 for d in rec_cat_arg.shape): x2[key] = rec_cat_arg.flat[0] else: x2[key] = (concatenate3, rec_cat_arg.tolist()) del old_blocks, new_index layer = toolz.merge(x2, intermediates) graph = HighLevelGraph.from_collections(merge_name, layer, dependencies=[x]) return Array(graph, merge_name, chunks, meta=x)
def modf(x): # Not actually object dtype, just need to specify something tmp = elemwise(np.modf, x, dtype=object) left = "modf1-" + tmp.name right = "modf2-" + tmp.name ldsk = {(left, ) + key[1:]: (getitem, key, 0) for key in core.flatten(tmp.__dask_keys__())} rdsk = {(right, ) + key[1:]: (getitem, key, 1) for key in core.flatten(tmp.__dask_keys__())} a = np.empty_like(getattr(x, "_meta", x), shape=(1, ) * x.ndim, dtype=x.dtype) l, r = np.modf(a) graph = HighLevelGraph.from_collections(left, ldsk, dependencies=[tmp]) L = Array(graph, left, chunks=tmp.chunks, meta=l) graph = HighLevelGraph.from_collections(right, rdsk, dependencies=[tmp]) R = Array(graph, right, chunks=tmp.chunks, meta=r) return L, R
def imread(filename, imread=None, preprocess=None): """Read a stack of images into a dask array Parameters ---------- filename: string A globstring like 'myfile.*.png' imread: function (optional) Optionally provide custom imread function. Function should expect a filename and produce a numpy array. Defaults to ``skimage.io.imread``. preprocess: function (optional) Optionally provide custom function to preprocess the image. Function should expect a numpy array for a single image. Examples -------- >>> from dask.array.image import imread >>> im = imread('2015-*-*.png') # doctest: +SKIP >>> im.shape # doctest: +SKIP (365, 1000, 1000, 3) Returns ------- Dask array of all images stacked along the first dimension. Each separate image file will be treated as an individual chunk. """ imread = imread or sk_imread filenames = sorted(glob(filename)) if not filenames: raise ValueError("No files found under name %s" % filename) name = "imread-%s" % tokenize(filenames, map(os.path.getmtime, filenames)) sample = imread(filenames[0]) if preprocess: sample = preprocess(sample) keys = [(name, i) + (0, ) * len(sample.shape) for i in range(len(filenames))] if preprocess: values = [(add_leading_dimension, (preprocess, (imread, fn))) for fn in filenames] else: values = [(add_leading_dimension, (imread, fn)) for fn in filenames] dsk = dict(zip(keys, values)) chunks = ((1, ) * len(filenames), ) + tuple((d, ) for d in sample.shape) return Array(dsk, name, chunks, sample.dtype)
def linspace(start, stop, num=50, chunks=None, dtype=None, endpoint=True): """ Return `num` evenly spaced values over the closed interval [`start`, `stop`]. TODO: implement the `endpoint`, `restep`, and `dtype` keyword args Parameters ---------- start : scalar The starting value of the sequence. stop : scalar The last value of the sequence. num : int, optional Number of samples to include in the returned dask array, including the endpoints. chunks : int The number of samples on each block. Note that the last block will have fewer samples if `num % blocksize != 0` Returns ------- samples : dask array """ num = int(num) if endpoint == False: num = num + 1 if chunks is None: raise ValueError("Must supply a chunks= keyword argument") chunks = normalize_chunks(chunks, (num, )) range_ = stop - start space = float(range_) / (num - 1) name = 'linspace-' + tokenize((start, stop, num, chunks, dtype, endpoint)) dsk = {} blockstart = start for i, bs in enumerate(chunks[0]): blockstop = blockstart + ((bs - 1) * space) task = (partial(np.linspace, dtype=dtype), blockstart, blockstop, bs) blockstart = blockstart + (space * bs) dsk[(name, i)] = task return Array(dsk, name, chunks, dtype=dtype)
def wrap_func_shape_as_first_arg(func, *args, **kwargs): """ Transform np creation function into blocked version """ if "shape" not in kwargs: shape, args = args[0], args[1:] else: shape = kwargs.pop("shape") if isinstance(shape, Array): raise TypeError("Dask array input not supported. " "Please use tuple, list, or a 1D numpy array instead.") parsed = _parse_wrap_args(func, args, kwargs, shape) shape = parsed["shape"] dtype = parsed["dtype"] chunks = parsed["chunks"] name = parsed["name"] kwargs = parsed["kwargs"] func = partial(func, dtype=dtype, **kwargs) out_ind = dep_ind = tuple(range(len(shape))) graph = core_blockwise( func, name, out_ind, ArrayChunkShapeDep(chunks), dep_ind, numblocks={}, ) return Array(graph, name, chunks, dtype=dtype, meta=kwargs.get("meta", None))
def reblock(x, blockdims=None, blockshape=None): """ Convert blocks in dask array x for new blockdims. reblock(x, blockdims=None, blockshape=None ) >>> import dask.array as da >>> a = np.random.uniform(0, 1, 7**4).reshape((7,) * 4) >>> x = da.from_array(a, blockdims=((2, 3, 2),)*4) >>> x.blockdims ((2, 3, 2), (2, 3, 2), (2, 3, 2), (2, 3, 2)) >>> y = reblock(x, blockdims=((2, 4, 1), (4, 2, 1), (4, 3), (7,))) >>> y.blockdims ((2, 4, 1), (4, 2, 1), (4, 3), (7,)) blockdims/blockshape also accept dict arguments mapping axis to blockshape >>> y = reblock(x, blockshape={1: 2}) # reblock axis 1 with blockshape 2 Parameters ---------- x: dask array blockdims: the new block dimensions to create blockshape: the new blockshape to create Provide one of blockdims or blockshape. """ if isinstance(blockdims, dict): blockdims = blockdims_dict_to_tuple(x.blockdims, blockdims) elif isinstance(blockshape, dict): blockdims = blockshape_dict_to_tuple(x.blockdims, blockshape) elif not blockdims: blockdims = blockdims_from_blockshape(x.shape, blockshape) crossed = intersect_blockdims(x.blockdims, blockdims) x2 = dict() temp_name = next(reblock_names) new_index = tuple(product(*(tuple(range(len(n))) for n in blockdims))) for flat_idx, cross1 in enumerate(crossed): new_idx = new_index[flat_idx] key = (temp_name, ) + new_idx cr2 = iter(cross1) old_blocks = tuple(tuple(ind for ind, _ in cr) for cr in cross1) subdims = tuple( len(set(ss[i] for ss in old_blocks)) for i in range(x.ndim)) rec_cat_arg = np.empty(subdims).tolist() inds_in_block = product(*(range(s) for s in subdims)) for old_block in old_blocks: ind_slics = next(cr2) old_inds = tuple( tuple(s[0] for s in ind_slics) for i in range(x.ndim)) # list of nd slices slic = tuple(tuple(s[1] for s in ind_slics) for i in range(x.ndim)) ind_in_blk = next(inds_in_block) temp = rec_cat_arg for i in range(x.ndim - 1): temp = getitem(temp, ind_in_blk[i]) for ind, slc in zip(old_inds, slic): temp[ind_in_blk[-1]] = (getitem, (x.name, ) + ind, slc) x2[key] = (rec_concatenate, rec_cat_arg) x2 = merge(x.dask, x2) return Array(x2, temp_name, blockdims=blockdims, dtype=x.dtype)
def _wrap(self, funcname, *args, size=None, chunks="auto", extra_chunks=(), **kwargs): """Wrap numpy random function to produce dask.array random function extra_chunks should be a chunks tuple to append to the end of chunks """ if size is not None and not isinstance(size, (tuple, list)): size = (size, ) shapes = list({ ar.shape for ar in chain(args, kwargs.values()) if isinstance(ar, (Array, np.ndarray)) }) if size is not None: shapes.append(size) # broadcast to the final size(shape) size = broadcast_shapes(*shapes) chunks = normalize_chunks( chunks, size, # ideally would use dtype here dtype=kwargs.get("dtype", np.float64), ) slices = slices_from_chunks(chunks) def _broadcast_any(ar, shape, chunks): if isinstance(ar, Array): return broadcast_to(ar, shape).rechunk(chunks) if isinstance(ar, np.ndarray): return np.ascontiguousarray(np.broadcast_to(ar, shape)) # Broadcast all arguments, get tiny versions as well # Start adding the relevant bits to the graph dsk = {} lookup = {} small_args = [] dependencies = [] for i, ar in enumerate(args): if isinstance(ar, (np.ndarray, Array)): res = _broadcast_any(ar, size, chunks) if isinstance(res, Array): dependencies.append(res) lookup[i] = res.name elif isinstance(res, np.ndarray): name = f"array-{tokenize(res)}" lookup[i] = name dsk[name] = res small_args.append(ar[tuple(0 for _ in ar.shape)]) else: small_args.append(ar) small_kwargs = {} for key, ar in kwargs.items(): if isinstance(ar, (np.ndarray, Array)): res = _broadcast_any(ar, size, chunks) if isinstance(res, Array): dependencies.append(res) lookup[key] = res.name elif isinstance(res, np.ndarray): name = f"array-{tokenize(res)}" lookup[key] = name dsk[name] = res small_kwargs[key] = ar[tuple(0 for _ in ar.shape)] else: small_kwargs[key] = ar sizes = list(product(*chunks)) seeds = random_state_data(len(sizes), self._numpy_state) token = tokenize(seeds, size, chunks, args, kwargs) name = f"{funcname}-{token}" keys = product([name], *([range(len(bd)) for bd in chunks] + [[0]] * len(extra_chunks))) blocks = product(*[range(len(bd)) for bd in chunks]) vals = [] for seed, size, slc, block in zip(seeds, sizes, slices, blocks): arg = [] for i, ar in enumerate(args): if i not in lookup: arg.append(ar) else: if isinstance(ar, Array): arg.append((lookup[i], ) + block) else: # np.ndarray arg.append((getitem, lookup[i], slc)) kwrg = {} for k, ar in kwargs.items(): if k not in lookup: kwrg[k] = ar else: if isinstance(ar, Array): kwrg[k] = (lookup[k], ) + block else: # np.ndarray kwrg[k] = (getitem, lookup[k], slc) vals.append((_apply_random, self._RandomState, funcname, seed, size, arg, kwrg)) meta = _apply_random( self._RandomState, funcname, seed, (0, ) * len(size), small_args, small_kwargs, ) dsk.update(dict(zip(keys, vals))) graph = HighLevelGraph.from_collections(name, dsk, dependencies=dependencies) return Array(graph, name, chunks + extra_chunks, meta=meta)
def choice(self, a, size=None, replace=True, p=None, chunks="auto"): dependencies = [] # Normalize and validate `a` if isinstance(a, Integral): # On windows the output dtype differs if p is provided or # absent, see https://github.com/numpy/numpy/issues/9867 dummy_p = np.array([1]) if p is not None else p dtype = np.random.choice(1, size=(), p=dummy_p).dtype len_a = a if a < 0: raise ValueError("a must be greater than 0") else: a = asarray(a) a = a.rechunk(a.shape) dtype = a.dtype if a.ndim != 1: raise ValueError("a must be one dimensional") len_a = len(a) dependencies.append(a) a = a.__dask_keys__()[0] # Normalize and validate `p` if p is not None: if not isinstance(p, Array): # If p is not a dask array, first check the sum is close # to 1 before converting. p = np.asarray(p) if not np.isclose(p.sum(), 1, rtol=1e-7, atol=0): raise ValueError("probabilities do not sum to 1") p = asarray(p) else: p = p.rechunk(p.shape) if p.ndim != 1: raise ValueError("p must be one dimensional") if len(p) != len_a: raise ValueError("a and p must have the same size") dependencies.append(p) p = p.__dask_keys__()[0] if size is None: size = () elif not isinstance(size, (tuple, list)): size = (size, ) chunks = normalize_chunks(chunks, size, dtype=np.float64) if not replace and len(chunks[0]) > 1: err_msg = ("replace=False is not currently supported for " "dask.array.choice with multi-chunk output " "arrays") raise NotImplementedError(err_msg) sizes = list(product(*chunks)) state_data = random_state_data(len(sizes), self._numpy_state) name = "da.random.choice-%s" % tokenize(state_data, size, chunks, a, replace, p) keys = product([name], *(range(len(bd)) for bd in chunks)) dsk = { k: (_choice, state, a, size, replace, p) for k, state, size in zip(keys, state_data, sizes) } graph = HighLevelGraph.from_collections(name, dsk, dependencies=dependencies) return Array(graph, name, chunks, dtype=dtype)
def arange(*args, **kwargs): """ Return evenly spaced values from `start` to `stop` with step size `step`. The values are half-open [start, stop), so including start and excluding stop. This is basically the same as python's range function but for dask arrays. Parameters ---------- start : int, optional The starting value of the sequence. The default is 0. stop : int The end of the interval, this value is excluded from the interval. step : int, optional The spacing between the values. The default is 1 when not specified. The last value of the sequence. chunks : int The number of samples on each block. Note that the last block will have fewer samples if ``len(array) % chunks != 0``. Returns ------- samples : dask array """ args = tuple(common.to_nptype(x) for x in args) if len(args) == 1: start = 0 stop = args[0] step = np.int_(1) elif len(args) == 2: start = args[0] stop = args[1] step = np.int_(1) elif len(args) == 3: start, stop, step = args else: raise TypeError(''' arange takes 3 positional arguments: arange([start], stop, [step]) ''') if step is None: step = np.int_(1) if start is None: start = np.int_(0) if 'chunks' not in kwargs: raise ValueError("Must supply a chunks= keyword argument") chunks = kwargs['chunks'] dtype = kwargs.get('dtype', None) if dtype is None: dtype = np.arange(0, 1, step).dtype range_ = stop - start if hasattr(step, 'dtype') and step.dtype.kind == 'm': num = int(abs(range_ / step)) else: num = int(np.round(abs(range_ / step))) chunks = normalize_chunks(chunks, (num, )) name = 'arange-' + tokenize((start, stop, step, chunks, num)) dsk = {} elem_count = 0 # Correct arange for non-integer steps. def fix_arange(start, stop, step, dtype): x0 = int(np.round(start / step)) x1 = int(np.round(stop / step)) return (np.arange(x0, x1) * step).astype(dtype) for i, bs in enumerate(chunks[0]): blockstart = start + (elem_count * step) blockstop = start + ((elem_count + bs) * step) task = (fix_arange, blockstart, blockstop, step, dtype) dsk[(name, i)] = task elem_count += bs return Array(dsk, name, chunks, dtype=dtype)
def reshape(x, shape, merge_chunks=True, limit=None): """Reshape array to new shape Parameters ---------- shape : int or tuple of ints The new shape should be compatible with the original shape. If an integer, then the result will be a 1-D array of that length. One shape dimension can be -1. In this case, the value is inferred from the length of the array and remaining dimensions. merge_chunks : bool, default True Whether to merge chunks using the logic in :meth:`dask.array.rechunk` when communication is necessary given the input array chunking and the output shape. With ``merge_chunks==False``, the input array will be rechunked to a chunksize of 1, which can create very many tasks. limit: int (optional) The maximum block size to target in bytes. If no limit is provided, it defaults to using the ``array.chunk-size`` Dask config value. Notes ----- This is a parallelized version of the ``np.reshape`` function with the following limitations: 1. It assumes that the array is stored in `row-major order`_ 2. It only allows for reshapings that collapse or merge dimensions like ``(1, 2, 3, 4) -> (1, 6, 4)`` or ``(64,) -> (4, 4, 4)`` .. _`row-major order`: https://en.wikipedia.org/wiki/Row-_and_column-major_order When communication is necessary this algorithm depends on the logic within rechunk. It endeavors to keep chunk sizes roughly the same when possible. See :ref:`array-chunks.reshaping` for a discussion the tradeoffs of ``merge_chunks``. See Also -------- dask.array.rechunk numpy.reshape """ # Sanitize inputs, look for -1 in shape from dask.array.core import PerformanceWarning from dask.array.slicing import sanitize_index shape = tuple(map(sanitize_index, shape)) known_sizes = [s for s in shape if s != -1] if len(known_sizes) < len(shape): if len(shape) - len(known_sizes) > 1: raise ValueError("can only specify one unknown dimension") # Fastpath for x.reshape(-1) on 1D arrays, allows unknown shape in x # for this case only. if len(shape) == 1 and x.ndim == 1: return x missing_size = sanitize_index(x.size / reduce(mul, known_sizes, 1)) shape = tuple(missing_size if s == -1 else s for s in shape) if np.isnan(sum(x.shape)): raise ValueError("Array chunk size or shape is unknown. shape: %s\n\n" "Possible solution with x.compute_chunk_sizes()" % str(x.shape)) if reduce(mul, shape, 1) != x.size: raise ValueError("total size of new array must be unchanged") if x.shape == shape: return x meta = meta_from_array(x, len(shape)) name = "reshape-" + tokenize(x, shape) if x.npartitions == 1: key = next(flatten(x.__dask_keys__())) dsk = {(name, ) + (0, ) * len(shape): (M.reshape, key, shape)} chunks = tuple((d, ) for d in shape) graph = HighLevelGraph.from_collections(name, dsk, dependencies=[x]) return Array(graph, name, chunks, meta=meta) # Logic or how to rechunk din = len(x.shape) dout = len(shape) if not merge_chunks and din > dout: x = x.rechunk({i: 1 for i in range(din - dout)}) inchunks, outchunks = reshape_rechunk(x.shape, shape, x.chunks) # Check output chunks are not too large max_chunksize_in_bytes = reduce( mul, [max(i) for i in outchunks]) * x.dtype.itemsize if limit is None: limit = parse_bytes(config.get("array.chunk-size")) split = config.get("array.slicing.split-large-chunks", None) else: limit = parse_bytes(limit) split = True if max_chunksize_in_bytes > limit: if split is None: msg = ( "Reshaping is producing a large chunk. To accept the large\n" "chunk and silence this warning, set the option\n" " >>> with dask.config.set(**{'array.slicing.split_large_chunks': False}):\n" " ... array.reshape(shape)\n\n" "To avoid creating the large chunks, set the option\n" " >>> with dask.config.set(**{'array.slicing.split_large_chunks': True}):\n" " ... array.reshape(shape)" "Explictly passing ``limit`` to ``reshape`` will also silence this warning\n" " >>> array.reshape(shape, limit='128 MiB')") warnings.warn(msg, PerformanceWarning, stacklevel=6) elif split: # Leave chunk sizes unaltered where possible matching_chunks = Counter(inchunks) & Counter(outchunks) chunk_plan = [] for out in outchunks: if matching_chunks[out] > 0: chunk_plan.append(out) matching_chunks[out] -= 1 else: chunk_plan.append("auto") outchunks = normalize_chunks( chunk_plan, shape=shape, limit=limit, dtype=x.dtype, previous_chunks=inchunks, ) x2 = x.rechunk(inchunks) # Construct graph in_keys = list(product([x2.name], *[range(len(c)) for c in inchunks])) out_keys = list(product([name], *[range(len(c)) for c in outchunks])) shapes = list(product(*outchunks)) dsk = { a: (M.reshape, b, shape) for a, b, shape in zip(out_keys, in_keys, shapes) } graph = HighLevelGraph.from_collections(name, dsk, dependencies=[x2]) return Array(graph, name, outchunks, meta=meta)
def percentile(a, q, method="linear", internal_method="default", **kwargs): """Approximate percentile of 1-D array Parameters ---------- a : Array q : array_like of float Percentile or sequence of percentiles to compute, which must be between 0 and 100 inclusive. method : {'linear', 'lower', 'higher', 'midpoint', 'nearest'}, optional The interpolation method to use when the desired percentile lies between two data points ``i < j``. Only valid for ``method='dask'``. - 'linear': ``i + (j - i) * fraction``, where ``fraction`` is the fractional part of the index surrounded by ``i`` and ``j``. - 'lower': ``i``. - 'higher': ``j``. - 'nearest': ``i`` or ``j``, whichever is nearest. - 'midpoint': ``(i + j) / 2``. .. versionchanged:: 2022.1.0 This argument was previously called "interpolation" internal_method : {'default', 'dask', 'tdigest'}, optional What internal method to use. By default will use dask's internal custom algorithm (``'dask'``). If set to ``'tdigest'`` will use tdigest for floats and ints and fallback to the ``'dask'`` otherwise. .. versionchanged:: 2022.1.0 This argument was previously called “method”. interpolation : str, optional Deprecated name for the method keyword argument. .. deprecated:: 2022.1.0 See Also -------- numpy.percentile : Numpy's equivalent Percentile function """ from dask.array.dispatch import percentile_lookup as _percentile from dask.array.utils import array_safe, meta_from_array allowed_internal_methods = ["default", "dask", "tdigest"] if method in allowed_internal_methods: warnings.warn( "In Dask 2022.1.0, the `method=` argument was renamed to `internal_method=`", FutureWarning, ) internal_method = method if "interpolation" in kwargs: if _numpy_122: warnings.warn( "In Dask 2022.1.0, the `interpolation=` argument to percentile was renamed to " "`method= ` ", FutureWarning, ) method = kwargs.pop("interpolation") if kwargs: raise TypeError( f"percentile() got an unexpected keyword argument {kwargs.keys()}") if not a.ndim == 1: raise NotImplementedError( "Percentiles only implemented for 1-d arrays") if isinstance(q, Number): q = [q] q = array_safe(q, like=meta_from_array(a)) token = tokenize(a, q, method) dtype = a.dtype if np.issubdtype(dtype, np.integer): dtype = (array_safe([], dtype=dtype, like=meta_from_array(a)) / 0.5).dtype meta = meta_from_array(a, dtype=dtype) if internal_method not in allowed_internal_methods: raise ValueError( f"`internal_method=` must be one of {allowed_internal_methods}") # Allow using t-digest if method is allowed and dtype is of floating or integer type if (internal_method == "tdigest" and method == "linear" and (np.issubdtype(dtype, np.floating) or np.issubdtype(dtype, np.integer))): from dask.utils import import_required import_required( "crick", "crick is a required dependency for using the t-digest method.") name = "percentile_tdigest_chunk-" + token dsk = {(name, i): (_tdigest_chunk, key) for i, key in enumerate(a.__dask_keys__())} name2 = "percentile_tdigest-" + token dsk2 = {(name2, 0): (_percentiles_from_tdigest, q, sorted(dsk))} # Otherwise use the custom percentile algorithm else: # Add 0 and 100 during calculation for more robust behavior (hopefully) calc_q = np.pad(q, 1, mode="constant") calc_q[-1] = 100 name = "percentile_chunk-" + token dsk = {(name, i): (_percentile, key, calc_q, method) for i, key in enumerate(a.__dask_keys__())} name2 = "percentile-" + token dsk2 = { (name2, 0): ( merge_percentiles, q, [calc_q] * len(a.chunks[0]), sorted(dsk), method, ) } dsk = merge(dsk, dsk2) graph = HighLevelGraph.from_collections(name2, dsk, dependencies=[a]) return Array(graph, name2, chunks=((len(q), ), ), meta=meta)
def apply_gufunc( func, signature, *args, axes=None, axis=None, keepdims=False, output_dtypes=None, output_sizes=None, vectorize=None, allow_rechunk=False, meta=None, **kwargs, ): """ Apply a generalized ufunc or similar python function to arrays. ``signature`` determines if the function consumes or produces core dimensions. The remaining dimensions in given input arrays (``*args``) are considered loop dimensions and are required to broadcast naturally against each other. In other terms, this function is like ``np.vectorize``, but for the blocks of dask arrays. If the function itself shall also be vectorized use ``vectorize=True`` for convenience. Parameters ---------- func : callable Function to call like ``func(*args, **kwargs)`` on input arrays (``*args``) that returns an array or tuple of arrays. If multiple arguments with non-matching dimensions are supplied, this function is expected to vectorize (broadcast) over axes of positional arguments in the style of NumPy universal functions [1]_ (if this is not the case, set ``vectorize=True``). If this function returns multiple outputs, ``output_core_dims`` has to be set as well. signature: string Specifies what core dimensions are consumed and produced by ``func``. According to the specification of numpy.gufunc signature [2]_ *args : numeric Input arrays or scalars to the callable function. axes: List of tuples, optional, keyword only A list of tuples with indices of axes a generalized ufunc should operate on. For instance, for a signature of ``"(i,j),(j,k)->(i,k)"`` appropriate for matrix multiplication, the base elements are two-dimensional matrices and these are taken to be stored in the two last axes of each argument. The corresponding axes keyword would be ``[(-2, -1), (-2, -1), (-2, -1)]``. For simplicity, for generalized ufuncs that operate on 1-dimensional arrays (vectors), a single integer is accepted instead of a single-element tuple, and for generalized ufuncs for which all outputs are scalars, the output tuples can be omitted. axis: int, optional, keyword only A single axis over which a generalized ufunc should operate. This is a short-cut for ufuncs that operate over a single, shared core dimension, equivalent to passing in axes with entries of (axis,) for each single-core-dimension argument and ``()`` for all others. For instance, for a signature ``"(i),(i)->()"``, it is equivalent to passing in ``axes=[(axis,), (axis,), ()]``. keepdims: bool, optional, keyword only If this is set to True, axes which are reduced over will be left in the result as a dimension with size one, so that the result will broadcast correctly against the inputs. This option can only be used for generalized ufuncs that operate on inputs that all have the same number of core dimensions and with outputs that have no core dimensions , i.e., with signatures like ``"(i),(i)->()"`` or ``"(m,m)->()"``. If used, the location of the dimensions in the output can be controlled with axes and axis. output_dtypes : Optional, dtype or list of dtypes, keyword only Valid numpy dtype specification or list thereof. If not given, a call of ``func`` with a small set of data is performed in order to try to automatically determine the output dtypes. output_sizes : dict, optional, keyword only Optional mapping from dimension names to sizes for outputs. Only used if new core dimensions (not found on inputs) appear on outputs. vectorize: bool, keyword only If set to ``True``, ``np.vectorize`` is applied to ``func`` for convenience. Defaults to ``False``. allow_rechunk: Optional, bool, keyword only Allows rechunking, otherwise chunk sizes need to match and core dimensions are to consist only of one chunk. Warning: enabling this can increase memory usage significantly. Defaults to ``False``. meta: Optional, tuple, keyword only tuple of empty ndarrays describing the shape and dtype of the output of the gufunc. Defaults to ``None``. **kwargs : dict Extra keyword arguments to pass to `func` Returns ------- Single dask.array.Array or tuple of dask.array.Array Examples -------- >>> import dask.array as da >>> import numpy as np >>> def stats(x): ... return np.mean(x, axis=-1), np.std(x, axis=-1) >>> a = da.random.normal(size=(10,20,30), chunks=(5, 10, 30)) >>> mean, std = da.apply_gufunc(stats, "(i)->(),()", a) >>> mean.compute().shape (10, 20) >>> def outer_product(x, y): ... return np.einsum("i,j->ij", x, y) >>> a = da.random.normal(size=( 20,30), chunks=(10, 30)) >>> b = da.random.normal(size=(10, 1,40), chunks=(5, 1, 40)) >>> c = da.apply_gufunc(outer_product, "(i),(j)->(i,j)", a, b, vectorize=True) >>> c.compute().shape (10, 20, 30, 40) References ---------- .. [1] https://docs.scipy.org/doc/numpy/reference/ufuncs.html .. [2] https://docs.scipy.org/doc/numpy/reference/c-api/generalized-ufuncs.html """ # Input processing: ## Signature if not isinstance(signature, str): raise TypeError("`signature` has to be of type string") # NumPy versions before https://github.com/numpy/numpy/pull/19627 # would not ignore whitespace characters in `signature` like they # are supposed to. We remove the whitespace here as a workaround. signature = re.sub(r"\s+", "", signature) input_coredimss, output_coredimss = _parse_gufunc_signature(signature) ## Determine nout: nout = None for functions of one direct return; nout = int for return tuples nout = None if not isinstance(output_coredimss, list) else len(output_coredimss) ## Consolidate onto `meta` if meta is not None and output_dtypes is not None: raise ValueError( "Only one of `meta` and `output_dtypes` should be given (`meta` is preferred)." ) if meta is None: if output_dtypes is None: ## Infer `output_dtypes` if vectorize: tempfunc = np.vectorize(func, signature=signature) else: tempfunc = func output_dtypes = apply_infer_dtype(tempfunc, args, kwargs, "apply_gufunc", "output_dtypes", nout) ## Turn `output_dtypes` into `meta` if (nout is None and isinstance(output_dtypes, (tuple, list)) and len(output_dtypes) == 1): output_dtypes = output_dtypes[0] sample = args[0] if args else None if nout is None: meta = meta_from_array(sample, dtype=output_dtypes) else: meta = tuple( meta_from_array(sample, dtype=odt) for odt in output_dtypes) ## Normalize `meta` format meta = meta_from_array(meta) if isinstance(meta, list): meta = tuple(meta) ## Validate `meta` if nout is None: if isinstance(meta, tuple): if len(meta) == 1: meta = meta[0] else: raise ValueError( "For a function with one output, must give a single item for `output_dtypes`/`meta`, " "not a tuple or list.") else: if not isinstance(meta, tuple): raise ValueError( f"For a function with {nout} outputs, must give a tuple or list for `output_dtypes`/`meta`, " "not a single item.") if len(meta) != nout: raise ValueError( f"For a function with {nout} outputs, must give a tuple or list of {nout} items for " f"`output_dtypes`/`meta`, not {len(meta)}.") ## Vectorize function, if required if vectorize: otypes = [x.dtype for x in meta] if isinstance(meta, tuple) else [meta.dtype] func = np.vectorize(func, signature=signature, otypes=otypes) ## Miscellaneous if output_sizes is None: output_sizes = {} ## Axes input_axes, output_axes = _validate_normalize_axes(axes, axis, keepdims, input_coredimss, output_coredimss) # Main code: ## Cast all input arrays to dask args = [asarray(a) for a in args] if len(input_coredimss) != len(args): raise ValueError( "According to `signature`, `func` requires %d arguments, but %s given" % (len(input_coredimss), len(args))) ## Axes: transpose input arguments transposed_args = [] for arg, iax, input_coredims in zip(args, input_axes, input_coredimss): shape = arg.shape iax = tuple(a if a < 0 else a - len(shape) for a in iax) tidc = tuple(i for i in range(-len(shape) + 0, 0) if i not in iax) + iax transposed_arg = arg.transpose(tidc) transposed_args.append(transposed_arg) args = transposed_args ## Assess input args for loop dims input_shapes = [a.shape for a in args] input_chunkss = [a.chunks for a in args] num_loopdims = [ len(s) - len(cd) for s, cd in zip(input_shapes, input_coredimss) ] max_loopdims = max(num_loopdims) if num_loopdims else None core_input_shapes = [ dict(zip(icd, s[n:])) for s, n, icd in zip(input_shapes, num_loopdims, input_coredimss) ] core_shapes = merge(*core_input_shapes) core_shapes.update(output_sizes) loop_input_dimss = [ tuple("__loopdim%d__" % d for d in range(max_loopdims - n, max_loopdims)) for n in num_loopdims ] input_dimss = [l + c for l, c in zip(loop_input_dimss, input_coredimss)] loop_output_dims = max(loop_input_dimss, key=len) if loop_input_dimss else tuple() ## Assess input args for same size and chunk sizes ### Collect sizes and chunksizes of all dims in all arrays dimsizess = {} chunksizess = {} for dims, shape, chunksizes in zip(input_dimss, input_shapes, input_chunkss): for dim, size, chunksize in zip(dims, shape, chunksizes): dimsizes = dimsizess.get(dim, []) dimsizes.append(size) dimsizess[dim] = dimsizes chunksizes_ = chunksizess.get(dim, []) chunksizes_.append(chunksize) chunksizess[dim] = chunksizes_ ### Assert correct partitioning, for case: for dim, sizes in dimsizess.items(): #### Check that the arrays have same length for same dimensions or dimension `1` if set(sizes) | {1} != {1, max(sizes)}: raise ValueError( f"Dimension `'{dim}'` with different lengths in arrays") if not allow_rechunk: chunksizes = chunksizess[dim] #### Check if core dimensions consist of only one chunk if (dim in core_shapes) and (chunksizes[0][0] < core_shapes[dim]): raise ValueError( "Core dimension `'{}'` consists of multiple chunks. To fix, rechunk into a single \ chunk along this dimension or set `allow_rechunk=True`, but beware that this may increase memory usage \ significantly.".format(dim)) #### Check if loop dimensions consist of same chunksizes, when they have sizes > 1 relevant_chunksizes = list( unique(c for s, c in zip(sizes, chunksizes) if s > 1)) if len(relevant_chunksizes) > 1: raise ValueError( f"Dimension `'{dim}'` with different chunksize present") ## Apply function - use blockwise here arginds = list(concat(zip(args, input_dimss))) ### Use existing `blockwise` but only with loopdims to enforce ### concatenation for coredims that appear also at the output ### Modifying `blockwise` could improve things here. tmp = blockwise(func, loop_output_dims, *arginds, concatenate=True, meta=meta, **kwargs) # NOTE: we likely could just use `meta` instead of `tmp._meta`, # but we use it and validate it anyway just to be sure nothing odd has happened. metas = tmp._meta if nout is None: assert not isinstance( metas, (list, tuple) ), f"meta changed from single output to multiple output during blockwise: {meta} -> {metas}" metas = (metas, ) else: assert isinstance( metas, (list, tuple) ), f"meta changed from multiple output to single output during blockwise: {meta} -> {metas}" assert ( len(metas) == nout ), f"Number of outputs changed from {nout} to {len(metas)} during blockwise" ## Prepare output shapes loop_output_shape = tmp.shape loop_output_chunks = tmp.chunks keys = list(flatten(tmp.__dask_keys__())) name, token = keys[0][0].split("-") ### *) Treat direct output if nout is None: output_coredimss = [output_coredimss] ## Split output leaf_arrs = [] for i, (ocd, oax, meta) in enumerate(zip(output_coredimss, output_axes, metas)): core_output_shape = tuple(core_shapes[d] for d in ocd) core_chunkinds = len(ocd) * (0, ) output_shape = loop_output_shape + core_output_shape output_chunks = loop_output_chunks + core_output_shape leaf_name = "%s_%d-%s" % (name, i, token) leaf_dsk = {(leaf_name, ) + key[1:] + core_chunkinds: ((getitem, key, i) if nout else key) for key in keys} graph = HighLevelGraph.from_collections(leaf_name, leaf_dsk, dependencies=[tmp]) meta = meta_from_array(meta, len(output_shape)) leaf_arr = Array(graph, leaf_name, chunks=output_chunks, shape=output_shape, meta=meta) ### Axes: if keepdims: slices = len( leaf_arr.shape) * (slice(None), ) + len(oax) * (np.newaxis, ) leaf_arr = leaf_arr[slices] tidcs = [None] * len(leaf_arr.shape) for ii, oa in zip(range(-len(oax), 0), oax): tidcs[oa] = ii j = 0 for ii in range(len(tidcs)): if tidcs[ii] is None: tidcs[ii] = j j += 1 leaf_arr = leaf_arr.transpose(tidcs) leaf_arrs.append(leaf_arr) return (*leaf_arrs, ) if nout else leaf_arrs[0] # Undo *) from above
def linspace(start, stop, num=50, endpoint=True, retstep=False, chunks="auto", dtype=None): """ Return `num` evenly spaced values over the closed interval [`start`, `stop`]. Parameters ---------- start : scalar The starting value of the sequence. stop : scalar The last value of the sequence. num : int, optional Number of samples to include in the returned dask array, including the endpoints. Default is 50. endpoint : bool, optional If True, ``stop`` is the last sample. Otherwise, it is not included. Default is True. retstep : bool, optional If True, return (samples, step), where step is the spacing between samples. Default is False. chunks : int The number of samples on each block. Note that the last block will have fewer samples if `num % blocksize != 0` dtype : dtype, optional The type of the output array. Returns ------- samples : dask array step : float, optional Only returned if ``retstep`` is True. Size of spacing between samples. See Also -------- dask.array.arange """ num = int(num) if dtype is None: dtype = np.linspace(0, 1, 1).dtype chunks = normalize_chunks(chunks, (num, ), dtype=dtype) range_ = stop - start div = (num - 1) if endpoint else num if div == 0: div = 1 step = float(range_) / div name = "linspace-" + tokenize((start, stop, num, endpoint, chunks, dtype)) dsk = {} blockstart = start for i, bs in enumerate(chunks[0]): bs_space = bs - 1 if endpoint else bs blockstop = blockstart + (bs_space * step) task = ( partial(chunk.linspace, endpoint=endpoint, dtype=dtype), blockstart, blockstop, bs, ) blockstart = blockstart + (step * bs) dsk[(name, i)] = task if retstep: return Array(dsk, name, chunks, dtype=dtype), step else: return Array(dsk, name, chunks, dtype=dtype)
def choice(self, a, size=None, replace=True, p=None, chunks=None): dsks = [] # Normalize and validate `a` if isinstance(a, Integral): # On windows the output dtype differs if p is provided or # absent, see https://github.com/numpy/numpy/issues/9867 dummy_p = np.array([1]) if p is not None else p dtype = np.random.choice(1, size=(), p=dummy_p).dtype len_a = a if a < 0: raise ValueError("a must be greater than 0") else: a = asarray(a).rechunk(a.shape) dtype = a.dtype if a.ndim != 1: raise ValueError("a must be one dimensional") len_a = len(a) dsks.append(a.dask) a = a.__dask_keys__()[0] # Normalize and validate `p` if p is not None: if not isinstance(p, Array): # If p is not a dask array, first check the sum is close # to 1 before converting. p = np.asarray(p) if not np.isclose(p.sum(), 1, rtol=1e-7, atol=0): raise ValueError("probabilities do not sum to 1") p = asarray(p) else: p = p.rechunk(p.shape) if p.ndim != 1: raise ValueError("p must be one dimensional") if len(p) != len_a: raise ValueError("a and p must have the same size") dsks.append(p.dask) p = p.__dask_keys__()[0] if size is None: size = () elif not isinstance(size, (tuple, list)): size = (size, ) chunks = normalize_chunks(chunks, size) sizes = list(product(*chunks)) state_data = random_state_data(len(sizes), self._numpy_state) name = 'da.random.choice-%s' % tokenize(state_data, size, chunks, a, replace, p) keys = product([name], *(range(len(bd)) for bd in chunks)) dsk = { k: (_choice, state, a, size, replace, p) for k, state, size in zip(keys, state_data, sizes) } return Array(sharedict.merge((name, dsk), *dsks), name, chunks, dtype=dtype)
def eye(N, chunks="auto", M=None, k=0, dtype=float): """ Return a 2-D Array with ones on the diagonal and zeros elsewhere. Parameters ---------- N : int Number of rows in the output. chunks : int, str How to chunk the array. Must be one of the following forms: - A blocksize like 1000. - A size in bytes, like "100 MiB" which will choose a uniform block-like shape - The word "auto" which acts like the above, but uses a configuration value ``array.chunk-size`` for the chunk size M : int, optional Number of columns in the output. If None, defaults to `N`. k : int, optional Index of the diagonal: 0 (the default) refers to the main diagonal, a positive value refers to an upper diagonal, and a negative value to a lower diagonal. dtype : data-type, optional Data-type of the returned array. Returns ------- I : Array of shape (N,M) An array where all elements are equal to zero, except for the `k`-th diagonal, whose values are equal to one. """ eye = {} if M is None: M = N if dtype is None: dtype = float if not isinstance(chunks, (int, str)): raise ValueError("chunks must be an int or string") vchunks, hchunks = normalize_chunks(chunks, shape=(N, M), dtype=dtype) chunks = vchunks[0] token = tokenize(N, chunks, M, k, dtype) name_eye = "eye-" + token for i, vchunk in enumerate(vchunks): for j, hchunk in enumerate(hchunks): if (j - i - 1) * chunks <= k <= (j - i + 1) * chunks: eye[name_eye, i, j] = ( np.eye, vchunk, hchunk, k - (j - i) * chunks, dtype, ) else: eye[name_eye, i, j] = (np.zeros, (vchunk, hchunk), dtype) return Array(eye, name_eye, shape=(N, M), chunks=(chunks, chunks), dtype=dtype)
def arange(*args, chunks="auto", like=None, dtype=None, **kwargs): """ Return evenly spaced values from `start` to `stop` with step size `step`. The values are half-open [start, stop), so including start and excluding stop. This is basically the same as python's range function but for dask arrays. When using a non-integer step, such as 0.1, the results will often not be consistent. It is better to use linspace for these cases. Parameters ---------- start : int, optional The starting value of the sequence. The default is 0. stop : int The end of the interval, this value is excluded from the interval. step : int, optional The spacing between the values. The default is 1 when not specified. The last value of the sequence. chunks : int The number of samples on each block. Note that the last block will have fewer samples if ``len(array) % chunks != 0``. Defaults to "auto" which will automatically determine chunk sizes. dtype : numpy.dtype Output dtype. Omit to infer it from start, stop, step Defaults to ``None``. like : array type or ``None`` Array to extract meta from. Defaults to ``None``. Returns ------- samples : dask array See Also -------- dask.array.linspace """ if len(args) == 1: start = 0 stop = args[0] step = 1 elif len(args) == 2: start = args[0] stop = args[1] step = 1 elif len(args) == 3: start, stop, step = args else: raise TypeError(""" arange takes 3 positional arguments: arange([start], stop, [step]) """) num = int(max(np.ceil((stop - start) / step), 0)) meta = meta_from_array(like) if like is not None else None if dtype is None: dtype = np.arange(start, stop, step * num if num else step).dtype chunks = normalize_chunks(chunks, (num, ), dtype=dtype) if kwargs: raise TypeError("Unexpected keyword argument(s): %s" % ",".join(kwargs.keys())) name = "arange-" + tokenize((start, stop, step, chunks, dtype)) dsk = {} elem_count = 0 for i, bs in enumerate(chunks[0]): blockstart = start + (elem_count * step) blockstop = start + ((elem_count + bs) * step) task = ( partial(chunk.arange, like=like), blockstart, blockstop, step, bs, dtype, ) dsk[(name, i)] = task elem_count += bs return Array(dsk, name, chunks, dtype=dtype, meta=meta)
def _wrap(self, func, *args, **kwargs): """ Wrap numpy random function to produce dask.array random function extra_chunks should be a chunks tuple to append to the end of chunks """ size = kwargs.pop('size', None) chunks = kwargs.pop('chunks') extra_chunks = kwargs.pop('extra_chunks', ()) if size is not None and not isinstance(size, (tuple, list)): size = (size, ) args_shapes = { ar.shape for ar in args if isinstance(ar, (Array, np.ndarray)) } args_shapes.union({ ar.shape for ar in kwargs.values() if isinstance(ar, (Array, np.ndarray)) }) shapes = list(args_shapes) if size is not None: shapes += [size] # broadcast to the final size(shape) size = broadcast_shapes(*shapes) chunks = normalize_chunks(chunks, size) slices = slices_from_chunks(chunks) def _broadcast_any(ar, shape, chunks): if isinstance(ar, Array): return broadcast_to(ar, shape).rechunk(chunks) if isinstance(ar, np.ndarray): return np.ascontiguousarray(np.broadcast_to(ar, shape)) # Broadcast all arguments, get tiny versions as well # Start adding the relevant bits to the graph dsk = {} dsks = [] lookup = {} small_args = [] for i, ar in enumerate(args): if isinstance(ar, (np.ndarray, Array)): res = _broadcast_any(ar, size, chunks) if isinstance(res, Array): dsks.append(res.dask) lookup[i] = res.name elif isinstance(res, np.ndarray): name = 'array-{}'.format(tokenize(res)) lookup[i] = name dsk[name] = res small_args.append(ar[tuple(0 for _ in ar.shape)]) else: small_args.append(ar) small_kwargs = {} for key, ar in kwargs.items(): if isinstance(ar, (np.ndarray, Array)): res = _broadcast_any(ar, size, chunks) if isinstance(res, Array): dsks.append(res.dask) lookup[key] = res.name elif isinstance(res, np.ndarray): name = 'array-{}'.format(tokenize(res)) lookup[key] = name dsk[name] = res small_kwargs[key] = ar[tuple(0 for _ in ar.shape)] else: small_kwargs[key] = ar # Get dtype small_kwargs['size'] = (0, ) dtype = func(xoroshiro128plus.RandomState(), *small_args, **small_kwargs).dtype sizes = list(product(*chunks)) state_data = random_state_data(len(sizes), self._numpy_state) token = tokenize(state_data, size, chunks, args, kwargs) name = 'da.random.{0}-{1}'.format(func.__name__, token) keys = product([name], *([range(len(bd)) for bd in chunks] + [[0]] * len(extra_chunks))) blocks = product(*[range(len(bd)) for bd in chunks]) vals = [] for state, size, slc, block in zip(state_data, sizes, slices, blocks): arg = [] for i, ar in enumerate(args): if i not in lookup: arg.append(ar) else: if isinstance(ar, Array): arg.append((lookup[i], ) + block) else: # np.ndarray arg.append((getitem, lookup[i], slc)) kwrg = {} for k, ar in kwargs.items(): if k not in lookup: kwrg[k] = ar else: if isinstance(ar, Array): kwrg[k] = (lookup[k], ) + block else: # np.ndarray kwrg[k] = (getitem, lookup[k], slc) vals.append((_apply_random, func.__name__, state, size, arg, kwrg)) dsk.update(dict(zip(keys, vals))) dsk = sharedict.merge((name, dsk), *dsks) return Array(dsk, name, chunks + extra_chunks, dtype=dtype)
def diagonal(a, offset=0, axis1=0, axis2=1): name = "diagonal-" + tokenize(a, offset, axis1, axis2) if a.ndim < 2: # NumPy uses `diag` as we do here. raise ValueError("diag requires an array of at least two dimensions") def _axis_fmt(axis, name, ndim): if axis < 0: t = ndim + axis if t < 0: msg = "{}: axis {} is out of bounds for array of dimension {}" raise np.AxisError(msg.format(name, axis, ndim)) axis = t return axis def pop_axes(chunks, axis1, axis2): chunks = list(chunks) chunks.pop(axis2) chunks.pop(axis1) return tuple(chunks) axis1 = _axis_fmt(axis1, "axis1", a.ndim) axis2 = _axis_fmt(axis2, "axis2", a.ndim) if axis1 == axis2: raise ValueError("axis1 and axis2 cannot be the same") a = asarray(a) k = offset if axis1 > axis2: axis1, axis2 = axis2, axis1 k = -offset free_axes = set(range(a.ndim)) - {axis1, axis2} free_indices = list(product(*(range(a.numblocks[i]) for i in free_axes))) ndims_free = len(free_axes) # equation of diagonal: i = j - k kdiag_row_start = max(0, -k) kdiag_col_start = max(0, k) kdiag_row_stop = min(a.shape[axis1], a.shape[axis2] - k) len_kdiag = kdiag_row_stop - kdiag_row_start if len_kdiag <= 0: xp = np if is_cupy_type(a._meta): import cupy xp = cupy out_chunks = pop_axes(a.chunks, axis1, axis2) + ((0, ), ) dsk = dict() for free_idx in free_indices: shape = tuple(out_chunks[axis][free_idx[axis]] for axis in range(ndims_free)) dsk[(name, ) + free_idx + (0, )] = ( partial(xp.empty, dtype=a.dtype), shape + (0, ), ) meta = meta_from_array(a, ndims_free + 1) return Array(dsk, name, out_chunks, meta=meta) # compute row index ranges for chunks along axis1: row_stops_ = np.cumsum(a.chunks[axis1]) row_starts = np.roll(row_stops_, 1) row_starts[0] = 0 # compute column index ranges for chunks along axis2: col_stops_ = np.cumsum(a.chunks[axis2]) col_starts = np.roll(col_stops_, 1) col_starts[0] = 0 # locate first chunk containing diagonal: row_blockid = np.arange(a.numblocks[axis1]) col_blockid = np.arange(a.numblocks[axis2]) row_filter = (row_starts <= kdiag_row_start) & (kdiag_row_start < row_stops_) col_filter = (col_starts <= kdiag_col_start) & (kdiag_col_start < col_stops_) (I, ) = row_blockid[row_filter] (J, ) = col_blockid[col_filter] # follow k-diagonal through chunks while constructing dask graph: dsk = dict() i = 0 kdiag_chunks = () while kdiag_row_start < a.shape[axis1] and kdiag_col_start < a.shape[axis2]: # localize block info: nrows, ncols = a.chunks[axis1][I], a.chunks[axis2][J] kdiag_row_start -= row_starts[I] kdiag_col_start -= col_starts[J] k = -kdiag_row_start if kdiag_row_start > 0 else kdiag_col_start kdiag_row_end = min(nrows, ncols - k) kdiag_len = kdiag_row_end - kdiag_row_start # increment dask graph: for free_idx in free_indices: input_idx = (free_idx[:axis1] + (I, ) + free_idx[axis1:axis2 - 1] + (J, ) + free_idx[axis2 - 1:]) output_idx = free_idx + (i, ) dsk[(name, ) + output_idx] = ( np.diagonal, (a.name, ) + input_idx, k, axis1, axis2, ) kdiag_chunks += (kdiag_len, ) # prepare for next iteration: i += 1 kdiag_row_start = kdiag_row_end + row_starts[I] kdiag_col_start = min(ncols, nrows + k) + col_starts[J] I = I + 1 if kdiag_row_start == row_stops_[I] else I J = J + 1 if kdiag_col_start == col_stops_[J] else J out_chunks = pop_axes(a.chunks, axis1, axis2) + (kdiag_chunks, ) graph = HighLevelGraph.from_collections(name, dsk, dependencies=[a]) meta = meta_from_array(a, ndims_free + 1) return Array(graph, name, out_chunks, meta=meta)