static-frame / static-frame Goto Github PK

Support creating a Series from multiple Series, assuming non-overlapping indices.

Support assigning types for multiple columns at once on TypeBlocks, so as to support consolidating (or isolating) blocks in TypeBlocks.

Possibly support getitem style selections on TypeBlocks on Frame, i.e., in a manner similar to assign():

>>> frame.astype[['b', 'g,' ,'f']](float)
>>> frame.astype['g':](bool)

By returning the keys of the Index, we can get a set-like object from the Index without creating a new container, useful in places where we, in a loop, OR multiple indices.

I want to make a Series with an IndexHierarchy index using Series.from_items. (I like from_items esp. for hard-coded data because it puts the indices directly next to their respective values, which is much easier to keep track of when reading & debugging code.)

For example, I want to do this:

series = sf.Series.from_items([
    (('one', 1), 'hello'),
    (('two', 2), 'cya'),
])

What I get is this series:

<Series>
<Index>
('one', 1) hello
('two', 2) cya
<object>   <<U5>

But what I want is this series:

<Series>
<IndexHierarchy>
one              1        hello
two              2        cya
<object>         <object> <<U5>

So:

1. Is there an existing method for doing this?
2. If not, is there a method that would convert from the series I made into the series I want? (Something like the opposite of reindex_flat?) If there isn't, would it be worth implementing?
3. If "no" to all of the above, could you change from_items to interpret tuples as entries for an IndexHierarchy, without breaking anything else?

Permit rolling on one or both axis simultaneously. Implementation will be very similar to TypeBlocks._drop_blocks()

On the c branch:

• Update setup.py with name="static-frame" (it's currently "C-SF", for testing).
• Update .travis.yml to deploy off of the master branch.
• Update .travis.yml to use the "real" PyPI.
• Update .travis.yml with @flexatone's PyPI credentials (see me for this).
• Optionally, update .travis.yml to deploy to PyPI only on tags. I'd rather just deploy whenever the version changes in setup.py (which is what it does now), but we can chat about that.

On the master branch:

• Make a new minor release.
• Merge c.
• Use dev releases to make the necessary changes to our conda-forge recipe. This will be the last tricky part, but I'm pretty sure I've figured it out!

I listened to the TalkPython episode today that covered this library and I realized that this is the library I was looking for about a year ago. I have implemented parts of this library several times trying to address the pain points it tries to solve, so thank you!

Is the Apache Arrow project on your radar at all? It is very much aligned with the goals of this library and would provide several other features for free.

Permit modifying an IndexHierarchy at a specified, integer depth.

Analog to Frame.drop(); implement with np.delete().

Explore options for exporting to xarray.

For examples:

sf.Frame.from_records([
    ('one', 1, 'hello'),
    ('two', 2, 'cya'),
], columns=['name', 'val', 'msg']).set_index_hierarchy(['name', 'val'], drop=True)

(correctly afaik) produces this frame:

<Frame>
<Index>                       msg   <<U4>
<Index...
one                  1        hello
two                  2        cya
<object>             <object> <<U5>

But if I remove the second row, like so:

sf.Frame.from_records([
    ('one', 1, 'hello'),
], columns=['name', 'val', 'msg']).set_index_hierarchy(['name', 'val'], drop=True)

this fails with the following traceback:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.5/dist-packages/static_frame/core/frame.py", line 2083, in set_index_hierarchy
    index = IndexHierarchy.from_labels(index_labels, name=column_name)
  File "/usr/local/lib/python3.5/dist-packages/static_frame/core/index_hierarchy.py", line 586, in from_labels
    for d, v in enumerate(label):
TypeError: 'int' object is not iterable

One detail to note is that it seems to fail on the first non-iterable set_index_hierarchy column value; it just so happens that the first is 'one' which is iterable.

Is this in some way an intended behavior?

Support pivot style transformation of Frame.

Similar to DF.add_prefix, DF.add_suffix.

Optimized implementation of drop implemented on TypeBlocks, permitting Frame.set_index() and Frame.set_index_hierarchy() better dropping.

New (or reused) implementations with superior performance.

When using Frame.from_records, I found myself wishing I could pass a mapping of column name to dtype for the dtypes` kwarg.

Permit independent index, columns arguments, like Frame.reindex(); implement with TypeBlocks._drop_blocks().

A minimal example:

from io import StringIO

import static_frame as sf

INPUT_STR = '''
196412	0.0
196501	0.0
196502	0.0
196503	0.0
196504	0.0
196505	0.0'''

if __name__ == '__main__':
    works_but_wrong_headers = sf.Frame.from_tsv(StringIO(INPUT_STR), index_column=0)
    desired_but_fails = sf.Frame.from_tsv(StringIO(INPUT_STR), index_column=0, header_is_columns=False)

The exception I observe:

Traceback (most recent call last):
  File "/home/<username>/sf_test_case.py", line 15, in <module>
    desired_but_fails = sf.Frame.from_tsv(StringIO(INPUT_STR), index_column=0, header_is_columns=False)
  File "/home/<username>/src/rapc/static_frame/static_frame/core/frame.py", line 713, in from_tsv
    return cls.from_csv(fp, delimiter='\t', **kwargs)
  File "/home/<username>/src/rapc/static_frame/static_frame/core/frame.py", line 696, in from_csv
    missing_values={''},
  File "/home/<username>/.env37/lib/python3.7/site-packages/numpy/lib/npyio.py", line 1771, in genfromtxt
    names = list(names)
TypeError: 'bool' object is not iterable

I have a static_frame.Series, like so:

import static_frame as sf
import numpy as np
    
colors = ('red', 'green')
shapes = ('square', 'circle', 'triangle')
    
series = sf.Series(np.arange(6), index=sf.IndexHierarchy.from_product(shapes, colors))

I want to be able to perform some "groupby" function, like summing all like-color values together into a new Series that looks something like this:

<IndexHierarchy>       <Series>
red   6
green 9
<<U8> <int64>

With a pandas.Series, I could do this:

agg_series = series.groupby(level=1).sum()

However as far as I can tell, there's no way to do this with static_frame objects (without building the result manually). The closest functionality I can find is the group iterator functions, but those don't let you group by index hierarchy levels.

I understand that static_frame is not intended as a drop-in replacement for pandas, but this seems like a desirable feature for any data-crunching library.

This must follow the addition of a name attribute, as given only a Series we need a column name; for now, extend_items() can be used.

Since static_frame objects are immutable, is there any reason they can't (or shouldn't) implement __hash__?

I nominate mypy, since it's another project I'm becoming more involved with. I believe this will prove useful as StaticFrame begins interfacing with C code internally. Ideally, we'd use it as a CI job too.

The following modules pass mypy --strict:

• doc/doc_build.py
• doc/source/conf.py
• setup.py
• static_frame/__init__.py
• static_frame/__main__.py
• static_frame/core/__init__.py
• static_frame/core/array_go.py
• static_frame/core/display.py
• static_frame/core/display_color.py
• static_frame/core/display_html_datatables.py
• static_frame/core/doc_str.py
• static_frame/core/frame.py
• static_frame/core/hloc.py
• static_frame/core/index.py
• static_frame/core/index_base.py
• static_frame/core/index_hierarchy.py
• static_frame/core/index_level.py
• static_frame/core/iter_node.py
• static_frame/core/operator_delegate.py
• static_frame/core/series.py
• static_frame/core/type_blocks.py
• static_frame/core/util.py
• static_frame/performance/__init__.py
• static_frame/performance/core.py
• static_frame/performance/main.py
• static_frame/performance/perf_test.py
• static_frame/performance/pydata_2018.py
• static_frame/test/__init__.py
• static_frame/test/property/__init__.py
• static_frame/test/property/strategies.py
• static_frame/test/property/test_index.py
• static_frame/test/property/test_strategies.py
• static_frame/test/property/test_type_blocks.py
• static_frame/test/property/test_util.py
• static_frame/test/test_case.py
• static_frame/test/unit/__init__.py
• static_frame/test/unit/test_array_go.py
• static_frame/test/unit/test_display.py
• static_frame/test/unit/test_display_color.py
• static_frame/test/unit/test_doc.py
• static_frame/test/unit/test_frame.py
• static_frame/test/unit/test_index.py
• static_frame/test/unit/test_index_hierarchy.py
• static_frame/test/unit/test_index_level.py
• static_frame/test/unit/test_main.py
• static_frame/test/unit/test_series.py
• static_frame/test/unit/test_type_blocks.py
• static_frame/test/unit/test_util.py

Family of functions for more refined NaN filling.

Analog to Series.clip(), but with boundary objects will need to be Frames, not Series.

Implement dropping from root side, in addition from leaves.
Reverse notation and default to drop from root side.

Hi, got my interest peaked by your interview with Micheal Kennedy at Talk Python to Me. The library looks very interesting, especially the focus on removing ambiguity.

Most of my pandas work involve handling time-series data. Specifically of sensor data where the time between isn’t at all constant. I do lots of grouping by various time intervals, and time-aware interpolations.

After an admittedly cursory glance I didn’t see how I would do that kind of work with static-frame. Am I missing something, or haven’t you had reason/opportunity to consider that kind of use case?

While basic pickling appears to work, post-pickle-loaded NumPy arrays have flags.writeable set to True. Might need appropriate pickle handlers defined for copyreg moduie.

Permit creating a hierarchical index from columns already found in a Frame. There is no analog for Series, and have specialized method is better than overloading the behavior of Frame.set_index().

Extending Display to support HTML output, implement _repr_html_ methods on containers.

Note that the index name is lost when saving to csv in the below example.

f = sf.Frame([1, 2, 3], columns=['a'], index=sf.Index(range(3), name='Important Name'))
f
<Frame>
<Index>                 a       <<U1>
<Index: Important Name>
0                       1
1                       2
2                       3
<int64>                 <int64>

f.to_csv('/tmp/tmp.txt')
sf.Frame.from_csv('/tmp/tmp.txt')
<Frame>
<Index> index   a       <<U5>
<Index>
0       0       1
1       1       2
2       2       3
<int64> <int64> <int64>

MVCE:

import static_frame as sf
import numpy as np

f = sf.Series(np.nan, index=sf.IndexHierarchy.from_product(['A', 'B'], [1, 2]))
f.dropna()

The output will be the unmodified series, with NaNs still present. Surprisingly, Series.fillna() works correctly as expected.