def test_add():
df = pd.DataFrame(np.identity(12))
df2 = df.copy()
df2.index += 1
sf1 = sp.SparseFrame(df)
sf2 = sp.SparseFrame(df2)
correct = sf1.add(sf2).todense()
dsf = dsp.from_pandas(df, npartitions=4)
dsf2 = dsp.from_pandas(df2, npartitions=4)
res = dsf.add(dsf2).compute().todense()
pdt.assert_frame_equal(res, correct)
def from_pandas(df, npartitions=None, chunksize=None, name=None):
"""
Parameters
----------
df : pandas.DataFrame or pandas.Series
The DataFrame/Series with which to construct a Dask DataFrame/Series
npartitions : int, optional
The number of partitions of the index to create. Note that depending on
the size and index of the dataframe, the output may have fewer
partitions than requested.
chunksize : int, optional
The size of the partitions of the index.
name: string, optional
An optional keyname for the dataframe. Define when dataframe large.
Defaults to hashing the input. Hashing takes a lot of time on large df.
"""
nrows = df.shape[0]
if chunksize is None:
chunksize = int(ceil(nrows / npartitions))
else:
npartitions = int(ceil(nrows / chunksize))
if not df.index.is_monotonic_increasing:
df = df.sort_index()
divisions, locations = sorted_division_locations(df.index,
chunksize=chunksize)
name = name or 'from_pandas-{}'.format(tokenize(df, npartitions))
dsk = dict(
((name, i), sp.SparseFrame(df.iloc[start:stop]))
for i, (start, stop) in enumerate(zip(locations[:-1], locations[1:])))
meta = _make_meta(df)
return SparseFrame(dsk, name, meta, divisions)
def one_hot_encode(ddf, column, categories, index_col):
"""
Sparse one hot encoding of dask.DataFrame
Convert a dask.DataFrame into a series of SparseFrames. By one hot
encoding a single column
Parameters
----------
ddf: dask.DataFrame
e.g. the clickstream
column: str
column name to one hot encode in with SparseFrame
categories: iterable
possible category values
index_col: str, iterable
which columns to use as index
Returns
-------
sparse_one_hot: dask.Series
"""
idx_meta = ddf._meta.reset_index().set_index(index_col).index[:0] \
if index_col else ddf._meta.index
meta = sp.SparseFrame(np.array([]), columns=categories, index=idx_meta)
dsf = ddf.map_partitions(sparse_one_hot,
column=column,
categories=categories,
index_col=index_col,
meta=object)
return SparseFrame(dsf.dask, dsf._name, meta, dsf.divisions)
def _make_meta(inp):
if isinstance(inp, sp.SparseFrame):
return inp.iloc[:0]
else:
meta = dd_make_meta(inp)
if isinstance(meta, pd.core.generic.NDFrame):
return sp.SparseFrame(meta)
return meta
def test_read_npz():
sf = sp.SparseFrame(np.identity(100))
with tmpdir() as tmp:
sf.iloc[:25].to_npz(os.path.join(tmp, '1'))
sf.iloc[25:50].to_npz(os.path.join(tmp, '2'))
sf.iloc[50:75].to_npz(os.path.join(tmp, '3'))
sf.iloc[75:].to_npz(os.path.join(tmp, '4'))
dsf = dsp.read_npz(os.path.join(tmp, '*.npz'))
sf = dsf.compute()
assert np.all(sf.data.toarray() == np.identity(100))
def read_npz(path, read_divisions=False, storage_options=None):
"""
Read SparseFrame from npz archives
Parameters
----------
path: str
path to load files from can contain '*' to
reference multiple files
read_divisions: bool
if the files are sorted read the index for each file
to obtain divions. If files are not sorted this will
raise and error.
Returns
-------
dsf: dask.SparseFrame
"""
dsk = {}
name = 'read_npz-{}'.format(tokenize(path))
loader = None
divisions = None
try:
loader = _open_npz_archive(
path.split('*')[0] + 'metadata.npz', storage_options)
divisions = loader['divisions']
_paths = loader['partitions']
except FileNotFoundError:
_paths = _sorted(list(glob(path)))
finally:
if loader:
loader.close()
archive = _open_npz_archive(_paths[0], storage_options)
meta_idx, meta_cols = archive['frame_index'], archive['frame_columns']
meta = sp.SparseFrame(np.empty(shape=(0, len(meta_cols))),
index=meta_idx[:0],
columns=meta_cols)
for i, p in enumerate(_paths):
dsk[name, i] = (sp.SparseFrame.read_npz, p, storage_options)
if divisions is None and read_divisions:
level = 0 if isinstance(meta_idx, pd.MultiIndex) else None
divisions = _npz_read_divisions(_paths, level=level)
elif divisions is None:
divisions = [None] * (len(_paths) + 1)
return SparseFrame(dsk, name, meta, divisions=divisions)
def __init__(self, dsk, name, meta, divisions=None):
if isinstance(meta, SparseFrame):
# TODO: remove this case once we subclass from dask._Frame
meta = meta._meta
if not isinstance(meta, sp.SparseFrame):
meta = sp.SparseFrame(meta)
self.dask = dsk
self._name = name
self._meta = make_meta(meta)
self.divisions = tuple(divisions)
self.ndim = 2
self.loc = _LocIndexer(self)
def test_distributed_join_shortcut(how):
left = pd.DataFrame(np.identity(10),
index=np.arange(10),
columns=list('ABCDEFGHIJ'))
right = pd.DataFrame(np.identity(10),
index=np.arange(5, 15),
columns=list('KLMNOPQRST'))
correct = left.join(right, how=how).fillna(0)
d_left = dsp.from_pandas(left, npartitions=2)
d_right = sp.SparseFrame(right)
joined = d_left.join(d_right, how=how)
res = joined.compute().todense()
pdt.assert_frame_equal(correct, res)
def _construct_item_features(self, item_features, item_ids):
"""Create item features during predict."""
# align feature names
if self.indicator_setting in ['both', 'items']:
item_indicator = sp.SparseFrame(self._item_indicator,
index=self.iid_map.index)
item_indicator = item_indicator.reindex(item_ids).data
else:
item_indicator = None
if self.item_feature_names is None:
return item_indicator
item_feat_csr = item_features\
.loc[:, self.item_feature_names]\
.reindex(item_ids, axis=0)\
.data
if item_indicator is not None:
item_feat_csr = sparse.hstack([item_feat_csr, item_indicator])
return item_feat_csr
def read_npz(path, sorted=False):
"""
Read SparseFrame from npz archives
Parameters
----------
path: str
path to load files from can contain '*' to
reference multiple files
sorted: bool
if the files are sorted read the index for each file
to obtain divions
Returns
-------
dsf: dask.SparseFrame
"""
dsk = {}
name = 'read_npz-{}'.format(tokenize(path))
_paths = _sorted(list(glob(path)))
archive = np.load(_paths[0])
meta_idx, meta_cols = archive['frame_index'], archive['frame_columns']
meta = sp.SparseFrame(np.empty(shape=(0, len(meta_cols))),
index=meta_idx[:0],
columns=meta_cols)
for i, p in enumerate(_paths):
dsk[name, i] = (sp.SparseFrame.read_npz, p)
if sorted:
level = 0 if isinstance(meta_idx, pd.MultiIndex) else None
divisions = _npz_read_divisions(_paths, level=level)
else:
divisions = [None] * (len(_paths) + 1)
return SparseFrame(dsk, name, meta, divisions=divisions)
def foo(sf, x, y):
return sp.SparseFrame(sf.data * x * y,
index=sf.index,
columns=sf.columns)
def meta_nonempty_sparsity(x):
idx = _nonempty_index(x.index)
return sp.SparseFrame(sparse.csr_matrix((len(idx), len(x.columns))),
index=idx,
columns=x.columns)
def one_hot_encode(ddf, column=None, categories=None, index_col=None,
order=None, prefixes=False, sep='_',
ignore_cat_order_mismatch=False):
"""
Sparse one hot encoding of dask.DataFrame.
Convert a dask.DataFrame into a series of SparseFrames by one-hot
encoding specified columns.
Parameters
----------
ddf: dask.DataFrame
e.g. the clickstream
categories: dict
Maps ``column name`` to specification on how to treat this column.
Specification can be:
- iterable of possible category values;
- ``None`` if this column is already of categorical dtype;
- ``False`` if this column should not be one-hot-encoded - it will be
included in the result untouched.
This argument decides which column(s) will be processed by this
function. See description of `order` and `ignore_cat_order_mismatch`.
By default, try to ohe-hot-encode all categorical columns and include
all the other columns untouched.
index_col: str | iterable
which columns to use as index
order: iterable
Specify order in which one-hot encoded columns should be aligned.
Must have the same elements as keys of ``categories``.
If `order = [col_name1, col_name2]`
and `categories = {col_name1: ['A', 'B'], col_name2: ['C', 'D']}`,
then the resulting SparseFrame will have columns
`['A', 'B', 'C', 'D']`.
If you don't specify order, then output columns' order depends on
iteration over `categories` dictionary. You can pass `categories`
as an OrderedDict instead of providing `order` explicitly.
prefixes: bool
If False, column names will be the same as categories,
so that new columns will be named like:
[cat11, cat12, cat21, cat22, ...].
If True, original column name followed by a separator will be added
in front of each category name, so that new columns will be named like:
[col1_cat11, col1_cat12, col2_cat21, col2_cat22, ...].
See ``sep`` argument.
sep: str
Separator used when ``prefixes`` is True.
column: DEPRECATED
Kept only for backward compatibility.
ignore_cat_order_mismatch: bool
If a column being one-hot encoded is of categorical dtype, it has
its categories already predefined, so we don't need to explicitly pass
them in `categories` argument (see this argument's description).
However, if we pass them, they may be different than ones defined in
column.cat.categories. In such a situation, a ValueError will be
raised. However, if only orders of categories are different (but sets
of elements are same), you may specify ignore_cat_order_mismatch=True
to suppress this error. In such a situation, column's predefined
categories will be used.
Returns
-------
sparse_one_hot: sparsity.dask.SparseFrame
"""
idx_meta = ddf._meta.reset_index().set_index(index_col).index[:0] \
if index_col else ddf._meta.index
columns = sparse_one_hot(ddf._meta,
column=column,
categories=categories,
index_col=index_col,
order=order,
prefixes=prefixes,
sep=sep,
ignore_cat_order_mismatch=ignore_cat_order_mismatch
).columns
meta = sp.SparseFrame(np.empty(shape=(0, len(columns))), columns=columns,
index=idx_meta)
dsf = ddf.map_partitions(sparse_one_hot,
column=column,
categories=categories,
index_col=index_col,
order=order,
prefixes=prefixes,
sep=sep,
ignore_cat_order_mismatch=ignore_cat_order_mismatch,
meta=object)
return SparseFrame(dsf.dask, dsf._name, meta, dsf.divisions)
