]>
Commit | Line | Data |
---|---|---|
1d09f67e TL |
1 | # Licensed to the Apache Software Foundation (ASF) under one |
2 | # or more contributor license agreements. See the NOTICE file | |
3 | # distributed with this work for additional information | |
4 | # regarding copyright ownership. The ASF licenses this file | |
5 | # to you under the Apache License, Version 2.0 (the | |
6 | # "License"); you may not use this file except in compliance | |
7 | # with the License. You may obtain a copy of the License at | |
8 | # | |
9 | # http://www.apache.org/licenses/LICENSE-2.0 | |
10 | # | |
11 | # Unless required by applicable law or agreed to in writing, | |
12 | # software distributed under the License is distributed on an | |
13 | # "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY | |
14 | # KIND, either express or implied. See the License for the | |
15 | # specific language governing permissions and limitations | |
16 | # under the License. | |
17 | ||
18 | import gc | |
19 | import decimal | |
20 | import json | |
21 | import multiprocessing as mp | |
22 | import sys | |
23 | ||
24 | from collections import OrderedDict | |
25 | from datetime import date, datetime, time, timedelta, timezone | |
26 | ||
27 | import hypothesis as h | |
28 | import hypothesis.extra.pytz as tzst | |
29 | import hypothesis.strategies as st | |
30 | import numpy as np | |
31 | import numpy.testing as npt | |
32 | import pytest | |
33 | import pytz | |
34 | ||
35 | from pyarrow.pandas_compat import get_logical_type, _pandas_api | |
36 | from pyarrow.tests.util import invoke_script, random_ascii, rands | |
37 | import pyarrow.tests.strategies as past | |
38 | from pyarrow.vendored.version import Version | |
39 | ||
40 | import pyarrow as pa | |
41 | try: | |
42 | from pyarrow import parquet as pq | |
43 | except ImportError: | |
44 | pass | |
45 | ||
46 | try: | |
47 | import pandas as pd | |
48 | import pandas.testing as tm | |
49 | from .pandas_examples import dataframe_with_arrays, dataframe_with_lists | |
50 | except ImportError: | |
51 | pass | |
52 | ||
53 | ||
54 | # Marks all of the tests in this module | |
55 | pytestmark = pytest.mark.pandas | |
56 | ||
57 | ||
58 | def _alltypes_example(size=100): | |
59 | return pd.DataFrame({ | |
60 | 'uint8': np.arange(size, dtype=np.uint8), | |
61 | 'uint16': np.arange(size, dtype=np.uint16), | |
62 | 'uint32': np.arange(size, dtype=np.uint32), | |
63 | 'uint64': np.arange(size, dtype=np.uint64), | |
64 | 'int8': np.arange(size, dtype=np.int16), | |
65 | 'int16': np.arange(size, dtype=np.int16), | |
66 | 'int32': np.arange(size, dtype=np.int32), | |
67 | 'int64': np.arange(size, dtype=np.int64), | |
68 | 'float32': np.arange(size, dtype=np.float32), | |
69 | 'float64': np.arange(size, dtype=np.float64), | |
70 | 'bool': np.random.randn(size) > 0, | |
71 | # TODO(wesm): Pandas only support ns resolution, Arrow supports s, ms, | |
72 | # us, ns | |
73 | 'datetime': np.arange("2016-01-01T00:00:00.001", size, | |
74 | dtype='datetime64[ms]'), | |
75 | 'str': [str(x) for x in range(size)], | |
76 | 'str_with_nulls': [None] + [str(x) for x in range(size - 2)] + [None], | |
77 | 'empty_str': [''] * size | |
78 | }) | |
79 | ||
80 | ||
81 | def _check_pandas_roundtrip(df, expected=None, use_threads=False, | |
82 | expected_schema=None, | |
83 | check_dtype=True, schema=None, | |
84 | preserve_index=False, | |
85 | as_batch=False): | |
86 | klass = pa.RecordBatch if as_batch else pa.Table | |
87 | table = klass.from_pandas(df, schema=schema, | |
88 | preserve_index=preserve_index, | |
89 | nthreads=2 if use_threads else 1) | |
90 | result = table.to_pandas(use_threads=use_threads) | |
91 | ||
92 | if expected_schema: | |
93 | # all occurrences of _check_pandas_roundtrip passes expected_schema | |
94 | # without the pandas generated key-value metadata | |
95 | assert table.schema.equals(expected_schema) | |
96 | ||
97 | if expected is None: | |
98 | expected = df | |
99 | ||
100 | tm.assert_frame_equal(result, expected, check_dtype=check_dtype, | |
101 | check_index_type=('equiv' if preserve_index | |
102 | else False)) | |
103 | ||
104 | ||
105 | def _check_series_roundtrip(s, type_=None, expected_pa_type=None): | |
106 | arr = pa.array(s, from_pandas=True, type=type_) | |
107 | ||
108 | if type_ is not None and expected_pa_type is None: | |
109 | expected_pa_type = type_ | |
110 | ||
111 | if expected_pa_type is not None: | |
112 | assert arr.type == expected_pa_type | |
113 | ||
114 | result = pd.Series(arr.to_pandas(), name=s.name) | |
115 | tm.assert_series_equal(s, result) | |
116 | ||
117 | ||
118 | def _check_array_roundtrip(values, expected=None, mask=None, | |
119 | type=None): | |
120 | arr = pa.array(values, from_pandas=True, mask=mask, type=type) | |
121 | result = arr.to_pandas() | |
122 | ||
123 | values_nulls = pd.isnull(values) | |
124 | if mask is None: | |
125 | assert arr.null_count == values_nulls.sum() | |
126 | else: | |
127 | assert arr.null_count == (mask | values_nulls).sum() | |
128 | ||
129 | if expected is None: | |
130 | if mask is None: | |
131 | expected = pd.Series(values) | |
132 | else: | |
133 | expected = pd.Series(np.ma.masked_array(values, mask=mask)) | |
134 | ||
135 | tm.assert_series_equal(pd.Series(result), expected, check_names=False) | |
136 | ||
137 | ||
138 | def _check_array_from_pandas_roundtrip(np_array, type=None): | |
139 | arr = pa.array(np_array, from_pandas=True, type=type) | |
140 | result = arr.to_pandas() | |
141 | npt.assert_array_equal(result, np_array) | |
142 | ||
143 | ||
144 | class TestConvertMetadata: | |
145 | """ | |
146 | Conversion tests for Pandas metadata & indices. | |
147 | """ | |
148 | ||
149 | def test_non_string_columns(self): | |
150 | df = pd.DataFrame({0: [1, 2, 3]}) | |
151 | table = pa.Table.from_pandas(df) | |
152 | assert table.field(0).name == '0' | |
153 | ||
154 | def test_from_pandas_with_columns(self): | |
155 | df = pd.DataFrame({0: [1, 2, 3], 1: [1, 3, 3], 2: [2, 4, 5]}, | |
156 | columns=[1, 0]) | |
157 | ||
158 | table = pa.Table.from_pandas(df, columns=[0, 1]) | |
159 | expected = pa.Table.from_pandas(df[[0, 1]]) | |
160 | assert expected.equals(table) | |
161 | ||
162 | record_batch_table = pa.RecordBatch.from_pandas(df, columns=[0, 1]) | |
163 | record_batch_expected = pa.RecordBatch.from_pandas(df[[0, 1]]) | |
164 | assert record_batch_expected.equals(record_batch_table) | |
165 | ||
166 | def test_column_index_names_are_preserved(self): | |
167 | df = pd.DataFrame({'data': [1, 2, 3]}) | |
168 | df.columns.names = ['a'] | |
169 | _check_pandas_roundtrip(df, preserve_index=True) | |
170 | ||
171 | def test_range_index_shortcut(self): | |
172 | # ARROW-1639 | |
173 | index_name = 'foo' | |
174 | df = pd.DataFrame({'a': [1, 2, 3, 4]}, | |
175 | index=pd.RangeIndex(0, 8, step=2, name=index_name)) | |
176 | ||
177 | df2 = pd.DataFrame({'a': [4, 5, 6, 7]}, | |
178 | index=pd.RangeIndex(0, 4)) | |
179 | ||
180 | table = pa.Table.from_pandas(df) | |
181 | table_no_index_name = pa.Table.from_pandas(df2) | |
182 | ||
183 | # The RangeIndex is tracked in the metadata only | |
184 | assert len(table.schema) == 1 | |
185 | ||
186 | result = table.to_pandas() | |
187 | tm.assert_frame_equal(result, df) | |
188 | assert isinstance(result.index, pd.RangeIndex) | |
189 | assert _pandas_api.get_rangeindex_attribute(result.index, 'step') == 2 | |
190 | assert result.index.name == index_name | |
191 | ||
192 | result2 = table_no_index_name.to_pandas() | |
193 | tm.assert_frame_equal(result2, df2) | |
194 | assert isinstance(result2.index, pd.RangeIndex) | |
195 | assert _pandas_api.get_rangeindex_attribute(result2.index, 'step') == 1 | |
196 | assert result2.index.name is None | |
197 | ||
198 | def test_range_index_force_serialization(self): | |
199 | # ARROW-5427: preserve_index=True will force the RangeIndex to | |
200 | # be serialized as a column rather than tracked more | |
201 | # efficiently as metadata | |
202 | df = pd.DataFrame({'a': [1, 2, 3, 4]}, | |
203 | index=pd.RangeIndex(0, 8, step=2, name='foo')) | |
204 | ||
205 | table = pa.Table.from_pandas(df, preserve_index=True) | |
206 | assert table.num_columns == 2 | |
207 | assert 'foo' in table.column_names | |
208 | ||
209 | restored = table.to_pandas() | |
210 | tm.assert_frame_equal(restored, df) | |
211 | ||
212 | def test_rangeindex_doesnt_warn(self): | |
213 | # ARROW-5606: pandas 0.25 deprecated private _start/stop/step | |
214 | # attributes -> can be removed if support < pd 0.25 is dropped | |
215 | df = pd.DataFrame(np.random.randn(4, 2), columns=['a', 'b']) | |
216 | ||
217 | with pytest.warns(None) as record: | |
218 | _check_pandas_roundtrip(df, preserve_index=True) | |
219 | ||
220 | assert len(record) == 0 | |
221 | ||
222 | def test_multiindex_columns(self): | |
223 | columns = pd.MultiIndex.from_arrays([ | |
224 | ['one', 'two'], ['X', 'Y'] | |
225 | ]) | |
226 | df = pd.DataFrame([(1, 'a'), (2, 'b'), (3, 'c')], columns=columns) | |
227 | _check_pandas_roundtrip(df, preserve_index=True) | |
228 | ||
229 | def test_multiindex_columns_with_dtypes(self): | |
230 | columns = pd.MultiIndex.from_arrays( | |
231 | [ | |
232 | ['one', 'two'], | |
233 | pd.DatetimeIndex(['2017-08-01', '2017-08-02']), | |
234 | ], | |
235 | names=['level_1', 'level_2'], | |
236 | ) | |
237 | df = pd.DataFrame([(1, 'a'), (2, 'b'), (3, 'c')], columns=columns) | |
238 | _check_pandas_roundtrip(df, preserve_index=True) | |
239 | ||
240 | def test_multiindex_with_column_dtype_object(self): | |
241 | # ARROW-3651 & ARROW-9096 | |
242 | # Bug when dtype of the columns is object. | |
243 | ||
244 | # uinderlying dtype: integer | |
245 | df = pd.DataFrame([1], columns=pd.Index([1], dtype=object)) | |
246 | _check_pandas_roundtrip(df, preserve_index=True) | |
247 | ||
248 | # underlying dtype: floating | |
249 | df = pd.DataFrame([1], columns=pd.Index([1.1], dtype=object)) | |
250 | _check_pandas_roundtrip(df, preserve_index=True) | |
251 | ||
252 | # underlying dtype: datetime | |
253 | # ARROW-9096: a simple roundtrip now works | |
254 | df = pd.DataFrame([1], columns=pd.Index( | |
255 | [datetime(2018, 1, 1)], dtype="object")) | |
256 | _check_pandas_roundtrip(df, preserve_index=True) | |
257 | ||
258 | def test_multiindex_columns_unicode(self): | |
259 | columns = pd.MultiIndex.from_arrays([['あ', 'い'], ['X', 'Y']]) | |
260 | df = pd.DataFrame([(1, 'a'), (2, 'b'), (3, 'c')], columns=columns) | |
261 | _check_pandas_roundtrip(df, preserve_index=True) | |
262 | ||
263 | def test_multiindex_doesnt_warn(self): | |
264 | # ARROW-3953: pandas 0.24 rename of MultiIndex labels to codes | |
265 | columns = pd.MultiIndex.from_arrays([['one', 'two'], ['X', 'Y']]) | |
266 | df = pd.DataFrame([(1, 'a'), (2, 'b'), (3, 'c')], columns=columns) | |
267 | ||
268 | with pytest.warns(None) as record: | |
269 | _check_pandas_roundtrip(df, preserve_index=True) | |
270 | ||
271 | assert len(record) == 0 | |
272 | ||
273 | def test_integer_index_column(self): | |
274 | df = pd.DataFrame([(1, 'a'), (2, 'b'), (3, 'c')]) | |
275 | _check_pandas_roundtrip(df, preserve_index=True) | |
276 | ||
277 | def test_index_metadata_field_name(self): | |
278 | # test None case, and strangely named non-index columns | |
279 | df = pd.DataFrame( | |
280 | [(1, 'a', 3.1), (2, 'b', 2.2), (3, 'c', 1.3)], | |
281 | index=pd.MultiIndex.from_arrays( | |
282 | [['c', 'b', 'a'], [3, 2, 1]], | |
283 | names=[None, 'foo'] | |
284 | ), | |
285 | columns=['a', None, '__index_level_0__'], | |
286 | ) | |
287 | with pytest.warns(UserWarning): | |
288 | t = pa.Table.from_pandas(df, preserve_index=True) | |
289 | js = t.schema.pandas_metadata | |
290 | ||
291 | col1, col2, col3, idx0, foo = js['columns'] | |
292 | ||
293 | assert col1['name'] == 'a' | |
294 | assert col1['name'] == col1['field_name'] | |
295 | ||
296 | assert col2['name'] is None | |
297 | assert col2['field_name'] == 'None' | |
298 | ||
299 | assert col3['name'] == '__index_level_0__' | |
300 | assert col3['name'] == col3['field_name'] | |
301 | ||
302 | idx0_descr, foo_descr = js['index_columns'] | |
303 | assert idx0_descr == '__index_level_0__' | |
304 | assert idx0['field_name'] == idx0_descr | |
305 | assert idx0['name'] is None | |
306 | ||
307 | assert foo_descr == 'foo' | |
308 | assert foo['field_name'] == foo_descr | |
309 | assert foo['name'] == foo_descr | |
310 | ||
311 | def test_categorical_column_index(self): | |
312 | df = pd.DataFrame( | |
313 | [(1, 'a', 2.0), (2, 'b', 3.0), (3, 'c', 4.0)], | |
314 | columns=pd.Index(list('def'), dtype='category') | |
315 | ) | |
316 | t = pa.Table.from_pandas(df, preserve_index=True) | |
317 | js = t.schema.pandas_metadata | |
318 | ||
319 | column_indexes, = js['column_indexes'] | |
320 | assert column_indexes['name'] is None | |
321 | assert column_indexes['pandas_type'] == 'categorical' | |
322 | assert column_indexes['numpy_type'] == 'int8' | |
323 | ||
324 | md = column_indexes['metadata'] | |
325 | assert md['num_categories'] == 3 | |
326 | assert md['ordered'] is False | |
327 | ||
328 | def test_string_column_index(self): | |
329 | df = pd.DataFrame( | |
330 | [(1, 'a', 2.0), (2, 'b', 3.0), (3, 'c', 4.0)], | |
331 | columns=pd.Index(list('def'), name='stringz') | |
332 | ) | |
333 | t = pa.Table.from_pandas(df, preserve_index=True) | |
334 | js = t.schema.pandas_metadata | |
335 | ||
336 | column_indexes, = js['column_indexes'] | |
337 | assert column_indexes['name'] == 'stringz' | |
338 | assert column_indexes['name'] == column_indexes['field_name'] | |
339 | assert column_indexes['numpy_type'] == 'object' | |
340 | assert column_indexes['pandas_type'] == 'unicode' | |
341 | ||
342 | md = column_indexes['metadata'] | |
343 | ||
344 | assert len(md) == 1 | |
345 | assert md['encoding'] == 'UTF-8' | |
346 | ||
347 | def test_datetimetz_column_index(self): | |
348 | df = pd.DataFrame( | |
349 | [(1, 'a', 2.0), (2, 'b', 3.0), (3, 'c', 4.0)], | |
350 | columns=pd.date_range( | |
351 | start='2017-01-01', periods=3, tz='America/New_York' | |
352 | ) | |
353 | ) | |
354 | t = pa.Table.from_pandas(df, preserve_index=True) | |
355 | js = t.schema.pandas_metadata | |
356 | ||
357 | column_indexes, = js['column_indexes'] | |
358 | assert column_indexes['name'] is None | |
359 | assert column_indexes['pandas_type'] == 'datetimetz' | |
360 | assert column_indexes['numpy_type'] == 'datetime64[ns]' | |
361 | ||
362 | md = column_indexes['metadata'] | |
363 | assert md['timezone'] == 'America/New_York' | |
364 | ||
365 | def test_datetimetz_row_index(self): | |
366 | df = pd.DataFrame({ | |
367 | 'a': pd.date_range( | |
368 | start='2017-01-01', periods=3, tz='America/New_York' | |
369 | ) | |
370 | }) | |
371 | df = df.set_index('a') | |
372 | ||
373 | _check_pandas_roundtrip(df, preserve_index=True) | |
374 | ||
375 | def test_categorical_row_index(self): | |
376 | df = pd.DataFrame({'a': [1, 2, 3], 'b': [1, 2, 3]}) | |
377 | df['a'] = df.a.astype('category') | |
378 | df = df.set_index('a') | |
379 | ||
380 | _check_pandas_roundtrip(df, preserve_index=True) | |
381 | ||
382 | def test_duplicate_column_names_does_not_crash(self): | |
383 | df = pd.DataFrame([(1, 'a'), (2, 'b')], columns=list('aa')) | |
384 | with pytest.raises(ValueError): | |
385 | pa.Table.from_pandas(df) | |
386 | ||
387 | def test_dictionary_indices_boundscheck(self): | |
388 | # ARROW-1658. No validation of indices leads to segfaults in pandas | |
389 | indices = [[0, 1], [0, -1]] | |
390 | ||
391 | for inds in indices: | |
392 | arr = pa.DictionaryArray.from_arrays(inds, ['a'], safe=False) | |
393 | batch = pa.RecordBatch.from_arrays([arr], ['foo']) | |
394 | table = pa.Table.from_batches([batch, batch, batch]) | |
395 | ||
396 | with pytest.raises(IndexError): | |
397 | arr.to_pandas() | |
398 | ||
399 | with pytest.raises(IndexError): | |
400 | table.to_pandas() | |
401 | ||
402 | def test_unicode_with_unicode_column_and_index(self): | |
403 | df = pd.DataFrame({'あ': ['い']}, index=['う']) | |
404 | ||
405 | _check_pandas_roundtrip(df, preserve_index=True) | |
406 | ||
407 | def test_mixed_column_names(self): | |
408 | # mixed type column names are not reconstructed exactly | |
409 | df = pd.DataFrame({'a': [1, 2], 'b': [3, 4]}) | |
410 | ||
411 | for cols in [['あ', b'a'], [1, '2'], [1, 1.5]]: | |
412 | df.columns = pd.Index(cols, dtype=object) | |
413 | ||
414 | # assert that the from_pandas raises the warning | |
415 | with pytest.warns(UserWarning): | |
416 | pa.Table.from_pandas(df) | |
417 | ||
418 | expected = df.copy() | |
419 | expected.columns = df.columns.values.astype(str) | |
420 | with pytest.warns(UserWarning): | |
421 | _check_pandas_roundtrip(df, expected=expected, | |
422 | preserve_index=True) | |
423 | ||
424 | def test_binary_column_name(self): | |
425 | column_data = ['い'] | |
426 | key = 'あ'.encode() | |
427 | data = {key: column_data} | |
428 | df = pd.DataFrame(data) | |
429 | ||
430 | # we can't use _check_pandas_roundtrip here because our metadata | |
431 | # is always decoded as utf8: even if binary goes in, utf8 comes out | |
432 | t = pa.Table.from_pandas(df, preserve_index=True) | |
433 | df2 = t.to_pandas() | |
434 | assert df.values[0] == df2.values[0] | |
435 | assert df.index.values[0] == df2.index.values[0] | |
436 | assert df.columns[0] == key | |
437 | ||
438 | def test_multiindex_duplicate_values(self): | |
439 | num_rows = 3 | |
440 | numbers = list(range(num_rows)) | |
441 | index = pd.MultiIndex.from_arrays( | |
442 | [['foo', 'foo', 'bar'], numbers], | |
443 | names=['foobar', 'some_numbers'], | |
444 | ) | |
445 | ||
446 | df = pd.DataFrame({'numbers': numbers}, index=index) | |
447 | ||
448 | _check_pandas_roundtrip(df, preserve_index=True) | |
449 | ||
450 | def test_metadata_with_mixed_types(self): | |
451 | df = pd.DataFrame({'data': [b'some_bytes', 'some_unicode']}) | |
452 | table = pa.Table.from_pandas(df) | |
453 | js = table.schema.pandas_metadata | |
454 | assert 'mixed' not in js | |
455 | data_column = js['columns'][0] | |
456 | assert data_column['pandas_type'] == 'bytes' | |
457 | assert data_column['numpy_type'] == 'object' | |
458 | ||
459 | def test_ignore_metadata(self): | |
460 | df = pd.DataFrame({'a': [1, 2, 3], 'b': ['foo', 'bar', 'baz']}, | |
461 | index=['one', 'two', 'three']) | |
462 | table = pa.Table.from_pandas(df) | |
463 | ||
464 | result = table.to_pandas(ignore_metadata=True) | |
465 | expected = (table.cast(table.schema.remove_metadata()) | |
466 | .to_pandas()) | |
467 | ||
468 | tm.assert_frame_equal(result, expected) | |
469 | ||
470 | def test_list_metadata(self): | |
471 | df = pd.DataFrame({'data': [[1], [2, 3, 4], [5] * 7]}) | |
472 | schema = pa.schema([pa.field('data', type=pa.list_(pa.int64()))]) | |
473 | table = pa.Table.from_pandas(df, schema=schema) | |
474 | js = table.schema.pandas_metadata | |
475 | assert 'mixed' not in js | |
476 | data_column = js['columns'][0] | |
477 | assert data_column['pandas_type'] == 'list[int64]' | |
478 | assert data_column['numpy_type'] == 'object' | |
479 | ||
480 | def test_struct_metadata(self): | |
481 | df = pd.DataFrame({'dicts': [{'a': 1, 'b': 2}, {'a': 3, 'b': 4}]}) | |
482 | table = pa.Table.from_pandas(df) | |
483 | pandas_metadata = table.schema.pandas_metadata | |
484 | assert pandas_metadata['columns'][0]['pandas_type'] == 'object' | |
485 | ||
486 | def test_decimal_metadata(self): | |
487 | expected = pd.DataFrame({ | |
488 | 'decimals': [ | |
489 | decimal.Decimal('394092382910493.12341234678'), | |
490 | -decimal.Decimal('314292388910493.12343437128'), | |
491 | ] | |
492 | }) | |
493 | table = pa.Table.from_pandas(expected) | |
494 | js = table.schema.pandas_metadata | |
495 | assert 'mixed' not in js | |
496 | data_column = js['columns'][0] | |
497 | assert data_column['pandas_type'] == 'decimal' | |
498 | assert data_column['numpy_type'] == 'object' | |
499 | assert data_column['metadata'] == {'precision': 26, 'scale': 11} | |
500 | ||
501 | def test_table_column_subset_metadata(self): | |
502 | # ARROW-1883 | |
503 | # non-default index | |
504 | for index in [ | |
505 | pd.Index(['a', 'b', 'c'], name='index'), | |
506 | pd.date_range("2017-01-01", periods=3, tz='Europe/Brussels')]: | |
507 | df = pd.DataFrame({'a': [1, 2, 3], | |
508 | 'b': [.1, .2, .3]}, index=index) | |
509 | table = pa.Table.from_pandas(df) | |
510 | ||
511 | table_subset = table.remove_column(1) | |
512 | result = table_subset.to_pandas() | |
513 | expected = df[['a']] | |
514 | if isinstance(df.index, pd.DatetimeIndex): | |
515 | df.index.freq = None | |
516 | tm.assert_frame_equal(result, expected) | |
517 | ||
518 | table_subset2 = table_subset.remove_column(1) | |
519 | result = table_subset2.to_pandas() | |
520 | tm.assert_frame_equal(result, df[['a']].reset_index(drop=True)) | |
521 | ||
522 | def test_to_pandas_column_subset_multiindex(self): | |
523 | # ARROW-10122 | |
524 | df = pd.DataFrame( | |
525 | {"first": list(range(5)), | |
526 | "second": list(range(5)), | |
527 | "value": np.arange(5)} | |
528 | ) | |
529 | table = pa.Table.from_pandas(df.set_index(["first", "second"])) | |
530 | ||
531 | subset = table.select(["first", "value"]) | |
532 | result = subset.to_pandas() | |
533 | expected = df[["first", "value"]].set_index("first") | |
534 | tm.assert_frame_equal(result, expected) | |
535 | ||
536 | def test_empty_list_metadata(self): | |
537 | # Create table with array of empty lists, forced to have type | |
538 | # list(string) in pyarrow | |
539 | c1 = [["test"], ["a", "b"], None] | |
540 | c2 = [[], [], []] | |
541 | arrays = OrderedDict([ | |
542 | ('c1', pa.array(c1, type=pa.list_(pa.string()))), | |
543 | ('c2', pa.array(c2, type=pa.list_(pa.string()))), | |
544 | ]) | |
545 | rb = pa.RecordBatch.from_arrays( | |
546 | list(arrays.values()), | |
547 | list(arrays.keys()) | |
548 | ) | |
549 | tbl = pa.Table.from_batches([rb]) | |
550 | ||
551 | # First roundtrip changes schema, because pandas cannot preserve the | |
552 | # type of empty lists | |
553 | df = tbl.to_pandas() | |
554 | tbl2 = pa.Table.from_pandas(df) | |
555 | md2 = tbl2.schema.pandas_metadata | |
556 | ||
557 | # Second roundtrip | |
558 | df2 = tbl2.to_pandas() | |
559 | expected = pd.DataFrame(OrderedDict([('c1', c1), ('c2', c2)])) | |
560 | ||
561 | tm.assert_frame_equal(df2, expected) | |
562 | ||
563 | assert md2['columns'] == [ | |
564 | { | |
565 | 'name': 'c1', | |
566 | 'field_name': 'c1', | |
567 | 'metadata': None, | |
568 | 'numpy_type': 'object', | |
569 | 'pandas_type': 'list[unicode]', | |
570 | }, | |
571 | { | |
572 | 'name': 'c2', | |
573 | 'field_name': 'c2', | |
574 | 'metadata': None, | |
575 | 'numpy_type': 'object', | |
576 | 'pandas_type': 'list[empty]', | |
577 | } | |
578 | ] | |
579 | ||
580 | def test_metadata_pandas_version(self): | |
581 | df = pd.DataFrame({'a': [1, 2, 3], 'b': [1, 2, 3]}) | |
582 | table = pa.Table.from_pandas(df) | |
583 | assert table.schema.pandas_metadata['pandas_version'] is not None | |
584 | ||
585 | def test_mismatch_metadata_schema(self): | |
586 | # ARROW-10511 | |
587 | # It is possible that the metadata and actual schema is not fully | |
588 | # matching (eg no timezone information for tz-aware column) | |
589 | # -> to_pandas() conversion should not fail on that | |
590 | df = pd.DataFrame({"datetime": pd.date_range("2020-01-01", periods=3)}) | |
591 | ||
592 | # OPTION 1: casting after conversion | |
593 | table = pa.Table.from_pandas(df) | |
594 | # cast the "datetime" column to be tz-aware | |
595 | new_col = table["datetime"].cast(pa.timestamp('ns', tz="UTC")) | |
596 | new_table1 = table.set_column( | |
597 | 0, pa.field("datetime", new_col.type), new_col | |
598 | ) | |
599 | ||
600 | # OPTION 2: specify schema during conversion | |
601 | schema = pa.schema([("datetime", pa.timestamp('ns', tz="UTC"))]) | |
602 | new_table2 = pa.Table.from_pandas(df, schema=schema) | |
603 | ||
604 | expected = df.copy() | |
605 | expected["datetime"] = expected["datetime"].dt.tz_localize("UTC") | |
606 | ||
607 | for new_table in [new_table1, new_table2]: | |
608 | # ensure the new table still has the pandas metadata | |
609 | assert new_table.schema.pandas_metadata is not None | |
610 | # convert to pandas | |
611 | result = new_table.to_pandas() | |
612 | tm.assert_frame_equal(result, expected) | |
613 | ||
614 | ||
615 | class TestConvertPrimitiveTypes: | |
616 | """ | |
617 | Conversion tests for primitive (e.g. numeric) types. | |
618 | """ | |
619 | ||
620 | def test_float_no_nulls(self): | |
621 | data = {} | |
622 | fields = [] | |
623 | dtypes = [('f2', pa.float16()), | |
624 | ('f4', pa.float32()), | |
625 | ('f8', pa.float64())] | |
626 | num_values = 100 | |
627 | ||
628 | for numpy_dtype, arrow_dtype in dtypes: | |
629 | values = np.random.randn(num_values) | |
630 | data[numpy_dtype] = values.astype(numpy_dtype) | |
631 | fields.append(pa.field(numpy_dtype, arrow_dtype)) | |
632 | ||
633 | df = pd.DataFrame(data) | |
634 | schema = pa.schema(fields) | |
635 | _check_pandas_roundtrip(df, expected_schema=schema) | |
636 | ||
637 | def test_float_nulls(self): | |
638 | num_values = 100 | |
639 | ||
640 | null_mask = np.random.randint(0, 10, size=num_values) < 3 | |
641 | dtypes = [('f2', pa.float16()), | |
642 | ('f4', pa.float32()), | |
643 | ('f8', pa.float64())] | |
644 | names = ['f2', 'f4', 'f8'] | |
645 | expected_cols = [] | |
646 | ||
647 | arrays = [] | |
648 | fields = [] | |
649 | for name, arrow_dtype in dtypes: | |
650 | values = np.random.randn(num_values).astype(name) | |
651 | ||
652 | arr = pa.array(values, from_pandas=True, mask=null_mask) | |
653 | arrays.append(arr) | |
654 | fields.append(pa.field(name, arrow_dtype)) | |
655 | values[null_mask] = np.nan | |
656 | ||
657 | expected_cols.append(values) | |
658 | ||
659 | ex_frame = pd.DataFrame(dict(zip(names, expected_cols)), | |
660 | columns=names) | |
661 | ||
662 | table = pa.Table.from_arrays(arrays, names) | |
663 | assert table.schema.equals(pa.schema(fields)) | |
664 | result = table.to_pandas() | |
665 | tm.assert_frame_equal(result, ex_frame) | |
666 | ||
667 | def test_float_nulls_to_ints(self): | |
668 | # ARROW-2135 | |
669 | df = pd.DataFrame({"a": [1.0, 2.0, np.NaN]}) | |
670 | schema = pa.schema([pa.field("a", pa.int16(), nullable=True)]) | |
671 | table = pa.Table.from_pandas(df, schema=schema, safe=False) | |
672 | assert table[0].to_pylist() == [1, 2, None] | |
673 | tm.assert_frame_equal(df, table.to_pandas()) | |
674 | ||
675 | def test_float_nulls_to_boolean(self): | |
676 | s = pd.Series([0.0, 1.0, 2.0, None, -3.0]) | |
677 | expected = pd.Series([False, True, True, None, True]) | |
678 | _check_array_roundtrip(s, expected=expected, type=pa.bool_()) | |
679 | ||
680 | def test_series_from_pandas_false_respected(self): | |
681 | # Check that explicit from_pandas=False is respected | |
682 | s = pd.Series([0.0, np.nan]) | |
683 | arr = pa.array(s, from_pandas=False) | |
684 | assert arr.null_count == 0 | |
685 | assert np.isnan(arr[1].as_py()) | |
686 | ||
687 | def test_integer_no_nulls(self): | |
688 | data = OrderedDict() | |
689 | fields = [] | |
690 | ||
691 | numpy_dtypes = [ | |
692 | ('i1', pa.int8()), ('i2', pa.int16()), | |
693 | ('i4', pa.int32()), ('i8', pa.int64()), | |
694 | ('u1', pa.uint8()), ('u2', pa.uint16()), | |
695 | ('u4', pa.uint32()), ('u8', pa.uint64()), | |
696 | ('longlong', pa.int64()), ('ulonglong', pa.uint64()) | |
697 | ] | |
698 | num_values = 100 | |
699 | ||
700 | for dtype, arrow_dtype in numpy_dtypes: | |
701 | info = np.iinfo(dtype) | |
702 | values = np.random.randint(max(info.min, np.iinfo(np.int_).min), | |
703 | min(info.max, np.iinfo(np.int_).max), | |
704 | size=num_values) | |
705 | data[dtype] = values.astype(dtype) | |
706 | fields.append(pa.field(dtype, arrow_dtype)) | |
707 | ||
708 | df = pd.DataFrame(data) | |
709 | schema = pa.schema(fields) | |
710 | _check_pandas_roundtrip(df, expected_schema=schema) | |
711 | ||
712 | def test_all_integer_types(self): | |
713 | # Test all Numpy integer aliases | |
714 | data = OrderedDict() | |
715 | numpy_dtypes = ['i1', 'i2', 'i4', 'i8', 'u1', 'u2', 'u4', 'u8', | |
716 | 'byte', 'ubyte', 'short', 'ushort', 'intc', 'uintc', | |
717 | 'int_', 'uint', 'longlong', 'ulonglong'] | |
718 | for dtype in numpy_dtypes: | |
719 | data[dtype] = np.arange(12, dtype=dtype) | |
720 | df = pd.DataFrame(data) | |
721 | _check_pandas_roundtrip(df) | |
722 | ||
723 | # Do the same with pa.array() | |
724 | # (for some reason, it doesn't use the same code paths at all) | |
725 | for np_arr in data.values(): | |
726 | arr = pa.array(np_arr) | |
727 | assert arr.to_pylist() == np_arr.tolist() | |
728 | ||
729 | def test_integer_byteorder(self): | |
730 | # Byteswapped arrays are not supported yet | |
731 | int_dtypes = ['i1', 'i2', 'i4', 'i8', 'u1', 'u2', 'u4', 'u8'] | |
732 | for dt in int_dtypes: | |
733 | for order in '=<>': | |
734 | data = np.array([1, 2, 42], dtype=order + dt) | |
735 | for np_arr in (data, data[::2]): | |
736 | if data.dtype.isnative: | |
737 | arr = pa.array(data) | |
738 | assert arr.to_pylist() == data.tolist() | |
739 | else: | |
740 | with pytest.raises(NotImplementedError): | |
741 | arr = pa.array(data) | |
742 | ||
743 | def test_integer_with_nulls(self): | |
744 | # pandas requires upcast to float dtype | |
745 | ||
746 | int_dtypes = ['i1', 'i2', 'i4', 'i8', 'u1', 'u2', 'u4', 'u8'] | |
747 | num_values = 100 | |
748 | ||
749 | null_mask = np.random.randint(0, 10, size=num_values) < 3 | |
750 | ||
751 | expected_cols = [] | |
752 | arrays = [] | |
753 | for name in int_dtypes: | |
754 | values = np.random.randint(0, 100, size=num_values) | |
755 | ||
756 | arr = pa.array(values, mask=null_mask) | |
757 | arrays.append(arr) | |
758 | ||
759 | expected = values.astype('f8') | |
760 | expected[null_mask] = np.nan | |
761 | ||
762 | expected_cols.append(expected) | |
763 | ||
764 | ex_frame = pd.DataFrame(dict(zip(int_dtypes, expected_cols)), | |
765 | columns=int_dtypes) | |
766 | ||
767 | table = pa.Table.from_arrays(arrays, int_dtypes) | |
768 | result = table.to_pandas() | |
769 | ||
770 | tm.assert_frame_equal(result, ex_frame) | |
771 | ||
772 | def test_array_from_pandas_type_cast(self): | |
773 | arr = np.arange(10, dtype='int64') | |
774 | ||
775 | target_type = pa.int8() | |
776 | ||
777 | result = pa.array(arr, type=target_type) | |
778 | expected = pa.array(arr.astype('int8')) | |
779 | assert result.equals(expected) | |
780 | ||
781 | def test_boolean_no_nulls(self): | |
782 | num_values = 100 | |
783 | ||
784 | np.random.seed(0) | |
785 | ||
786 | df = pd.DataFrame({'bools': np.random.randn(num_values) > 0}) | |
787 | field = pa.field('bools', pa.bool_()) | |
788 | schema = pa.schema([field]) | |
789 | _check_pandas_roundtrip(df, expected_schema=schema) | |
790 | ||
791 | def test_boolean_nulls(self): | |
792 | # pandas requires upcast to object dtype | |
793 | num_values = 100 | |
794 | np.random.seed(0) | |
795 | ||
796 | mask = np.random.randint(0, 10, size=num_values) < 3 | |
797 | values = np.random.randint(0, 10, size=num_values) < 5 | |
798 | ||
799 | arr = pa.array(values, mask=mask) | |
800 | ||
801 | expected = values.astype(object) | |
802 | expected[mask] = None | |
803 | ||
804 | field = pa.field('bools', pa.bool_()) | |
805 | schema = pa.schema([field]) | |
806 | ex_frame = pd.DataFrame({'bools': expected}) | |
807 | ||
808 | table = pa.Table.from_arrays([arr], ['bools']) | |
809 | assert table.schema.equals(schema) | |
810 | result = table.to_pandas() | |
811 | ||
812 | tm.assert_frame_equal(result, ex_frame) | |
813 | ||
814 | def test_boolean_to_int(self): | |
815 | # test from dtype=bool | |
816 | s = pd.Series([True, True, False, True, True] * 2) | |
817 | expected = pd.Series([1, 1, 0, 1, 1] * 2) | |
818 | _check_array_roundtrip(s, expected=expected, type=pa.int64()) | |
819 | ||
820 | def test_boolean_objects_to_int(self): | |
821 | # test from dtype=object | |
822 | s = pd.Series([True, True, False, True, True] * 2, dtype=object) | |
823 | expected = pd.Series([1, 1, 0, 1, 1] * 2) | |
824 | expected_msg = 'Expected integer, got bool' | |
825 | with pytest.raises(pa.ArrowTypeError, match=expected_msg): | |
826 | _check_array_roundtrip(s, expected=expected, type=pa.int64()) | |
827 | ||
828 | def test_boolean_nulls_to_float(self): | |
829 | # test from dtype=object | |
830 | s = pd.Series([True, True, False, None, True] * 2) | |
831 | expected = pd.Series([1.0, 1.0, 0.0, None, 1.0] * 2) | |
832 | _check_array_roundtrip(s, expected=expected, type=pa.float64()) | |
833 | ||
834 | def test_boolean_multiple_columns(self): | |
835 | # ARROW-6325 (multiple columns resulting in strided conversion) | |
836 | df = pd.DataFrame(np.ones((3, 2), dtype='bool'), columns=['a', 'b']) | |
837 | _check_pandas_roundtrip(df) | |
838 | ||
839 | def test_float_object_nulls(self): | |
840 | arr = np.array([None, 1.5, np.float64(3.5)] * 5, dtype=object) | |
841 | df = pd.DataFrame({'floats': arr}) | |
842 | expected = pd.DataFrame({'floats': pd.to_numeric(arr)}) | |
843 | field = pa.field('floats', pa.float64()) | |
844 | schema = pa.schema([field]) | |
845 | _check_pandas_roundtrip(df, expected=expected, | |
846 | expected_schema=schema) | |
847 | ||
848 | def test_float_with_null_as_integer(self): | |
849 | # ARROW-2298 | |
850 | s = pd.Series([np.nan, 1., 2., np.nan]) | |
851 | ||
852 | types = [pa.int8(), pa.int16(), pa.int32(), pa.int64(), | |
853 | pa.uint8(), pa.uint16(), pa.uint32(), pa.uint64()] | |
854 | for ty in types: | |
855 | result = pa.array(s, type=ty) | |
856 | expected = pa.array([None, 1, 2, None], type=ty) | |
857 | assert result.equals(expected) | |
858 | ||
859 | df = pd.DataFrame({'has_nulls': s}) | |
860 | schema = pa.schema([pa.field('has_nulls', ty)]) | |
861 | result = pa.Table.from_pandas(df, schema=schema, | |
862 | preserve_index=False) | |
863 | assert result[0].chunk(0).equals(expected) | |
864 | ||
865 | def test_int_object_nulls(self): | |
866 | arr = np.array([None, 1, np.int64(3)] * 5, dtype=object) | |
867 | df = pd.DataFrame({'ints': arr}) | |
868 | expected = pd.DataFrame({'ints': pd.to_numeric(arr)}) | |
869 | field = pa.field('ints', pa.int64()) | |
870 | schema = pa.schema([field]) | |
871 | _check_pandas_roundtrip(df, expected=expected, | |
872 | expected_schema=schema) | |
873 | ||
874 | def test_boolean_object_nulls(self): | |
875 | arr = np.array([False, None, True] * 100, dtype=object) | |
876 | df = pd.DataFrame({'bools': arr}) | |
877 | field = pa.field('bools', pa.bool_()) | |
878 | schema = pa.schema([field]) | |
879 | _check_pandas_roundtrip(df, expected_schema=schema) | |
880 | ||
881 | def test_all_nulls_cast_numeric(self): | |
882 | arr = np.array([None], dtype=object) | |
883 | ||
884 | def _check_type(t): | |
885 | a2 = pa.array(arr, type=t) | |
886 | assert a2.type == t | |
887 | assert a2[0].as_py() is None | |
888 | ||
889 | _check_type(pa.int32()) | |
890 | _check_type(pa.float64()) | |
891 | ||
892 | def test_half_floats_from_numpy(self): | |
893 | arr = np.array([1.5, np.nan], dtype=np.float16) | |
894 | a = pa.array(arr, type=pa.float16()) | |
895 | x, y = a.to_pylist() | |
896 | assert isinstance(x, np.float16) | |
897 | assert x == 1.5 | |
898 | assert isinstance(y, np.float16) | |
899 | assert np.isnan(y) | |
900 | ||
901 | a = pa.array(arr, type=pa.float16(), from_pandas=True) | |
902 | x, y = a.to_pylist() | |
903 | assert isinstance(x, np.float16) | |
904 | assert x == 1.5 | |
905 | assert y is None | |
906 | ||
907 | ||
908 | @pytest.mark.parametrize('dtype', | |
909 | ['i1', 'i2', 'i4', 'i8', 'u1', 'u2', 'u4', 'u8']) | |
910 | def test_array_integer_object_nulls_option(dtype): | |
911 | num_values = 100 | |
912 | ||
913 | null_mask = np.random.randint(0, 10, size=num_values) < 3 | |
914 | values = np.random.randint(0, 100, size=num_values, dtype=dtype) | |
915 | ||
916 | array = pa.array(values, mask=null_mask) | |
917 | ||
918 | if null_mask.any(): | |
919 | expected = values.astype('O') | |
920 | expected[null_mask] = None | |
921 | else: | |
922 | expected = values | |
923 | ||
924 | result = array.to_pandas(integer_object_nulls=True) | |
925 | ||
926 | np.testing.assert_equal(result, expected) | |
927 | ||
928 | ||
929 | @pytest.mark.parametrize('dtype', | |
930 | ['i1', 'i2', 'i4', 'i8', 'u1', 'u2', 'u4', 'u8']) | |
931 | def test_table_integer_object_nulls_option(dtype): | |
932 | num_values = 100 | |
933 | ||
934 | null_mask = np.random.randint(0, 10, size=num_values) < 3 | |
935 | values = np.random.randint(0, 100, size=num_values, dtype=dtype) | |
936 | ||
937 | array = pa.array(values, mask=null_mask) | |
938 | ||
939 | if null_mask.any(): | |
940 | expected = values.astype('O') | |
941 | expected[null_mask] = None | |
942 | else: | |
943 | expected = values | |
944 | ||
945 | expected = pd.DataFrame({dtype: expected}) | |
946 | ||
947 | table = pa.Table.from_arrays([array], [dtype]) | |
948 | result = table.to_pandas(integer_object_nulls=True) | |
949 | ||
950 | tm.assert_frame_equal(result, expected) | |
951 | ||
952 | ||
953 | class TestConvertDateTimeLikeTypes: | |
954 | """ | |
955 | Conversion tests for datetime- and timestamp-like types (date64, etc.). | |
956 | """ | |
957 | ||
958 | def test_timestamps_notimezone_no_nulls(self): | |
959 | df = pd.DataFrame({ | |
960 | 'datetime64': np.array([ | |
961 | '2007-07-13T01:23:34.123456789', | |
962 | '2006-01-13T12:34:56.432539784', | |
963 | '2010-08-13T05:46:57.437699912'], | |
964 | dtype='datetime64[ns]') | |
965 | }) | |
966 | field = pa.field('datetime64', pa.timestamp('ns')) | |
967 | schema = pa.schema([field]) | |
968 | _check_pandas_roundtrip( | |
969 | df, | |
970 | expected_schema=schema, | |
971 | ) | |
972 | ||
973 | def test_timestamps_notimezone_nulls(self): | |
974 | df = pd.DataFrame({ | |
975 | 'datetime64': np.array([ | |
976 | '2007-07-13T01:23:34.123456789', | |
977 | None, | |
978 | '2010-08-13T05:46:57.437699912'], | |
979 | dtype='datetime64[ns]') | |
980 | }) | |
981 | field = pa.field('datetime64', pa.timestamp('ns')) | |
982 | schema = pa.schema([field]) | |
983 | _check_pandas_roundtrip( | |
984 | df, | |
985 | expected_schema=schema, | |
986 | ) | |
987 | ||
988 | def test_timestamps_with_timezone(self): | |
989 | df = pd.DataFrame({ | |
990 | 'datetime64': np.array([ | |
991 | '2007-07-13T01:23:34.123', | |
992 | '2006-01-13T12:34:56.432', | |
993 | '2010-08-13T05:46:57.437'], | |
994 | dtype='datetime64[ms]') | |
995 | }) | |
996 | df['datetime64'] = df['datetime64'].dt.tz_localize('US/Eastern') | |
997 | _check_pandas_roundtrip(df) | |
998 | ||
999 | _check_series_roundtrip(df['datetime64']) | |
1000 | ||
1001 | # drop-in a null and ns instead of ms | |
1002 | df = pd.DataFrame({ | |
1003 | 'datetime64': np.array([ | |
1004 | '2007-07-13T01:23:34.123456789', | |
1005 | None, | |
1006 | '2006-01-13T12:34:56.432539784', | |
1007 | '2010-08-13T05:46:57.437699912'], | |
1008 | dtype='datetime64[ns]') | |
1009 | }) | |
1010 | df['datetime64'] = df['datetime64'].dt.tz_localize('US/Eastern') | |
1011 | ||
1012 | _check_pandas_roundtrip(df) | |
1013 | ||
1014 | def test_python_datetime(self): | |
1015 | # ARROW-2106 | |
1016 | date_array = [datetime.today() + timedelta(days=x) for x in range(10)] | |
1017 | df = pd.DataFrame({ | |
1018 | 'datetime': pd.Series(date_array, dtype=object) | |
1019 | }) | |
1020 | ||
1021 | table = pa.Table.from_pandas(df) | |
1022 | assert isinstance(table[0].chunk(0), pa.TimestampArray) | |
1023 | ||
1024 | result = table.to_pandas() | |
1025 | expected_df = pd.DataFrame({ | |
1026 | 'datetime': date_array | |
1027 | }) | |
1028 | tm.assert_frame_equal(expected_df, result) | |
1029 | ||
1030 | def test_python_datetime_with_pytz_tzinfo(self): | |
1031 | for tz in [pytz.utc, pytz.timezone('US/Eastern'), pytz.FixedOffset(1)]: | |
1032 | values = [datetime(2018, 1, 1, 12, 23, 45, tzinfo=tz)] | |
1033 | df = pd.DataFrame({'datetime': values}) | |
1034 | _check_pandas_roundtrip(df) | |
1035 | ||
1036 | @h.given(st.none() | tzst.timezones()) | |
1037 | def test_python_datetime_with_pytz_timezone(self, tz): | |
1038 | values = [datetime(2018, 1, 1, 12, 23, 45, tzinfo=tz)] | |
1039 | df = pd.DataFrame({'datetime': values}) | |
1040 | _check_pandas_roundtrip(df) | |
1041 | ||
1042 | def test_python_datetime_with_timezone_tzinfo(self): | |
1043 | from datetime import timezone | |
1044 | ||
1045 | if Version(pd.__version__) > Version("0.25.0"): | |
1046 | # older pandas versions fail on datetime.timezone.utc (as in input) | |
1047 | # vs pytz.UTC (as in result) | |
1048 | values = [datetime(2018, 1, 1, 12, 23, 45, tzinfo=timezone.utc)] | |
1049 | # also test with index to ensure both paths roundtrip (ARROW-9962) | |
1050 | df = pd.DataFrame({'datetime': values}, index=values) | |
1051 | _check_pandas_roundtrip(df, preserve_index=True) | |
1052 | ||
1053 | # datetime.timezone is going to be pytz.FixedOffset | |
1054 | hours = 1 | |
1055 | tz_timezone = timezone(timedelta(hours=hours)) | |
1056 | tz_pytz = pytz.FixedOffset(hours * 60) | |
1057 | values = [datetime(2018, 1, 1, 12, 23, 45, tzinfo=tz_timezone)] | |
1058 | values_exp = [datetime(2018, 1, 1, 12, 23, 45, tzinfo=tz_pytz)] | |
1059 | df = pd.DataFrame({'datetime': values}, index=values) | |
1060 | df_exp = pd.DataFrame({'datetime': values_exp}, index=values_exp) | |
1061 | _check_pandas_roundtrip(df, expected=df_exp, preserve_index=True) | |
1062 | ||
1063 | def test_python_datetime_subclass(self): | |
1064 | ||
1065 | class MyDatetime(datetime): | |
1066 | # see https://github.com/pandas-dev/pandas/issues/21142 | |
1067 | nanosecond = 0.0 | |
1068 | ||
1069 | date_array = [MyDatetime(2000, 1, 1, 1, 1, 1)] | |
1070 | df = pd.DataFrame({"datetime": pd.Series(date_array, dtype=object)}) | |
1071 | ||
1072 | table = pa.Table.from_pandas(df) | |
1073 | assert isinstance(table[0].chunk(0), pa.TimestampArray) | |
1074 | ||
1075 | result = table.to_pandas() | |
1076 | expected_df = pd.DataFrame({"datetime": date_array}) | |
1077 | ||
1078 | # https://github.com/pandas-dev/pandas/issues/21142 | |
1079 | expected_df["datetime"] = pd.to_datetime(expected_df["datetime"]) | |
1080 | ||
1081 | tm.assert_frame_equal(expected_df, result) | |
1082 | ||
1083 | def test_python_date_subclass(self): | |
1084 | ||
1085 | class MyDate(date): | |
1086 | pass | |
1087 | ||
1088 | date_array = [MyDate(2000, 1, 1)] | |
1089 | df = pd.DataFrame({"date": pd.Series(date_array, dtype=object)}) | |
1090 | ||
1091 | table = pa.Table.from_pandas(df) | |
1092 | assert isinstance(table[0].chunk(0), pa.Date32Array) | |
1093 | ||
1094 | result = table.to_pandas() | |
1095 | expected_df = pd.DataFrame( | |
1096 | {"date": np.array([date(2000, 1, 1)], dtype=object)} | |
1097 | ) | |
1098 | tm.assert_frame_equal(expected_df, result) | |
1099 | ||
1100 | def test_datetime64_to_date32(self): | |
1101 | # ARROW-1718 | |
1102 | arr = pa.array([date(2017, 10, 23), None]) | |
1103 | c = pa.chunked_array([arr]) | |
1104 | s = c.to_pandas() | |
1105 | ||
1106 | arr2 = pa.Array.from_pandas(s, type=pa.date32()) | |
1107 | ||
1108 | assert arr2.equals(arr.cast('date32')) | |
1109 | ||
1110 | @pytest.mark.parametrize('mask', [ | |
1111 | None, | |
1112 | np.array([True, False, False, True, False, False]), | |
1113 | ]) | |
1114 | def test_pandas_datetime_to_date64(self, mask): | |
1115 | s = pd.to_datetime([ | |
1116 | '2018-05-10T00:00:00', | |
1117 | '2018-05-11T00:00:00', | |
1118 | '2018-05-12T00:00:00', | |
1119 | '2018-05-10T10:24:01', | |
1120 | '2018-05-11T10:24:01', | |
1121 | '2018-05-12T10:24:01', | |
1122 | ]) | |
1123 | arr = pa.Array.from_pandas(s, type=pa.date64(), mask=mask) | |
1124 | ||
1125 | data = np.array([ | |
1126 | date(2018, 5, 10), | |
1127 | date(2018, 5, 11), | |
1128 | date(2018, 5, 12), | |
1129 | date(2018, 5, 10), | |
1130 | date(2018, 5, 11), | |
1131 | date(2018, 5, 12), | |
1132 | ]) | |
1133 | expected = pa.array(data, mask=mask, type=pa.date64()) | |
1134 | ||
1135 | assert arr.equals(expected) | |
1136 | ||
1137 | def test_array_types_date_as_object(self): | |
1138 | data = [date(2000, 1, 1), | |
1139 | None, | |
1140 | date(1970, 1, 1), | |
1141 | date(2040, 2, 26)] | |
1142 | expected_d = np.array(['2000-01-01', None, '1970-01-01', | |
1143 | '2040-02-26'], dtype='datetime64[D]') | |
1144 | ||
1145 | expected_ns = np.array(['2000-01-01', None, '1970-01-01', | |
1146 | '2040-02-26'], dtype='datetime64[ns]') | |
1147 | ||
1148 | objects = [pa.array(data), | |
1149 | pa.chunked_array([data])] | |
1150 | ||
1151 | for obj in objects: | |
1152 | result = obj.to_pandas() | |
1153 | expected_obj = expected_d.astype(object) | |
1154 | assert result.dtype == expected_obj.dtype | |
1155 | npt.assert_array_equal(result, expected_obj) | |
1156 | ||
1157 | result = obj.to_pandas(date_as_object=False) | |
1158 | assert result.dtype == expected_ns.dtype | |
1159 | npt.assert_array_equal(result, expected_ns) | |
1160 | ||
1161 | def test_table_convert_date_as_object(self): | |
1162 | df = pd.DataFrame({ | |
1163 | 'date': [date(2000, 1, 1), | |
1164 | None, | |
1165 | date(1970, 1, 1), | |
1166 | date(2040, 2, 26)]}) | |
1167 | ||
1168 | table = pa.Table.from_pandas(df, preserve_index=False) | |
1169 | ||
1170 | df_datetime = table.to_pandas(date_as_object=False) | |
1171 | df_object = table.to_pandas() | |
1172 | ||
1173 | tm.assert_frame_equal(df.astype('datetime64[ns]'), df_datetime, | |
1174 | check_dtype=True) | |
1175 | tm.assert_frame_equal(df, df_object, check_dtype=True) | |
1176 | ||
1177 | def test_date_infer(self): | |
1178 | df = pd.DataFrame({ | |
1179 | 'date': [date(2000, 1, 1), | |
1180 | None, | |
1181 | date(1970, 1, 1), | |
1182 | date(2040, 2, 26)]}) | |
1183 | table = pa.Table.from_pandas(df, preserve_index=False) | |
1184 | field = pa.field('date', pa.date32()) | |
1185 | ||
1186 | # schema's metadata is generated by from_pandas conversion | |
1187 | expected_schema = pa.schema([field], metadata=table.schema.metadata) | |
1188 | assert table.schema.equals(expected_schema) | |
1189 | ||
1190 | result = table.to_pandas() | |
1191 | tm.assert_frame_equal(result, df) | |
1192 | ||
1193 | def test_date_mask(self): | |
1194 | arr = np.array([date(2017, 4, 3), date(2017, 4, 4)], | |
1195 | dtype='datetime64[D]') | |
1196 | mask = [True, False] | |
1197 | result = pa.array(arr, mask=np.array(mask)) | |
1198 | expected = np.array([None, date(2017, 4, 4)], dtype='datetime64[D]') | |
1199 | expected = pa.array(expected, from_pandas=True) | |
1200 | assert expected.equals(result) | |
1201 | ||
1202 | def test_date_objects_typed(self): | |
1203 | arr = np.array([ | |
1204 | date(2017, 4, 3), | |
1205 | None, | |
1206 | date(2017, 4, 4), | |
1207 | date(2017, 4, 5)], dtype=object) | |
1208 | ||
1209 | arr_i4 = np.array([17259, -1, 17260, 17261], dtype='int32') | |
1210 | arr_i8 = arr_i4.astype('int64') * 86400000 | |
1211 | mask = np.array([False, True, False, False]) | |
1212 | ||
1213 | t32 = pa.date32() | |
1214 | t64 = pa.date64() | |
1215 | ||
1216 | a32 = pa.array(arr, type=t32) | |
1217 | a64 = pa.array(arr, type=t64) | |
1218 | ||
1219 | a32_expected = pa.array(arr_i4, mask=mask, type=t32) | |
1220 | a64_expected = pa.array(arr_i8, mask=mask, type=t64) | |
1221 | ||
1222 | assert a32.equals(a32_expected) | |
1223 | assert a64.equals(a64_expected) | |
1224 | ||
1225 | # Test converting back to pandas | |
1226 | colnames = ['date32', 'date64'] | |
1227 | table = pa.Table.from_arrays([a32, a64], colnames) | |
1228 | ||
1229 | ex_values = (np.array(['2017-04-03', '2017-04-04', '2017-04-04', | |
1230 | '2017-04-05'], | |
1231 | dtype='datetime64[D]')) | |
1232 | ex_values[1] = pd.NaT.value | |
1233 | ||
1234 | ex_datetime64ns = ex_values.astype('datetime64[ns]') | |
1235 | expected_pandas = pd.DataFrame({'date32': ex_datetime64ns, | |
1236 | 'date64': ex_datetime64ns}, | |
1237 | columns=colnames) | |
1238 | table_pandas = table.to_pandas(date_as_object=False) | |
1239 | tm.assert_frame_equal(table_pandas, expected_pandas) | |
1240 | ||
1241 | table_pandas_objects = table.to_pandas() | |
1242 | ex_objects = ex_values.astype('object') | |
1243 | expected_pandas_objects = pd.DataFrame({'date32': ex_objects, | |
1244 | 'date64': ex_objects}, | |
1245 | columns=colnames) | |
1246 | tm.assert_frame_equal(table_pandas_objects, | |
1247 | expected_pandas_objects) | |
1248 | ||
1249 | def test_pandas_null_values(self): | |
1250 | # ARROW-842 | |
1251 | pd_NA = getattr(pd, 'NA', None) | |
1252 | values = np.array([datetime(2000, 1, 1), pd.NaT, pd_NA], dtype=object) | |
1253 | values_with_none = np.array([datetime(2000, 1, 1), None, None], | |
1254 | dtype=object) | |
1255 | result = pa.array(values, from_pandas=True) | |
1256 | expected = pa.array(values_with_none, from_pandas=True) | |
1257 | assert result.equals(expected) | |
1258 | assert result.null_count == 2 | |
1259 | ||
1260 | # ARROW-9407 | |
1261 | assert pa.array([pd.NaT], from_pandas=True).type == pa.null() | |
1262 | assert pa.array([pd_NA], from_pandas=True).type == pa.null() | |
1263 | ||
1264 | def test_dates_from_integers(self): | |
1265 | t1 = pa.date32() | |
1266 | t2 = pa.date64() | |
1267 | ||
1268 | arr = np.array([17259, 17260, 17261], dtype='int32') | |
1269 | arr2 = arr.astype('int64') * 86400000 | |
1270 | ||
1271 | a1 = pa.array(arr, type=t1) | |
1272 | a2 = pa.array(arr2, type=t2) | |
1273 | ||
1274 | expected = date(2017, 4, 3) | |
1275 | assert a1[0].as_py() == expected | |
1276 | assert a2[0].as_py() == expected | |
1277 | ||
1278 | def test_pytime_from_pandas(self): | |
1279 | pytimes = [time(1, 2, 3, 1356), | |
1280 | time(4, 5, 6, 1356)] | |
1281 | ||
1282 | # microseconds | |
1283 | t1 = pa.time64('us') | |
1284 | ||
1285 | aobjs = np.array(pytimes + [None], dtype=object) | |
1286 | parr = pa.array(aobjs) | |
1287 | assert parr.type == t1 | |
1288 | assert parr[0].as_py() == pytimes[0] | |
1289 | assert parr[1].as_py() == pytimes[1] | |
1290 | assert parr[2].as_py() is None | |
1291 | ||
1292 | # DataFrame | |
1293 | df = pd.DataFrame({'times': aobjs}) | |
1294 | batch = pa.RecordBatch.from_pandas(df) | |
1295 | assert batch[0].equals(parr) | |
1296 | ||
1297 | # Test ndarray of int64 values | |
1298 | arr = np.array([_pytime_to_micros(v) for v in pytimes], | |
1299 | dtype='int64') | |
1300 | ||
1301 | a1 = pa.array(arr, type=pa.time64('us')) | |
1302 | assert a1[0].as_py() == pytimes[0] | |
1303 | ||
1304 | a2 = pa.array(arr * 1000, type=pa.time64('ns')) | |
1305 | assert a2[0].as_py() == pytimes[0] | |
1306 | ||
1307 | a3 = pa.array((arr / 1000).astype('i4'), | |
1308 | type=pa.time32('ms')) | |
1309 | assert a3[0].as_py() == pytimes[0].replace(microsecond=1000) | |
1310 | ||
1311 | a4 = pa.array((arr / 1000000).astype('i4'), | |
1312 | type=pa.time32('s')) | |
1313 | assert a4[0].as_py() == pytimes[0].replace(microsecond=0) | |
1314 | ||
1315 | def test_arrow_time_to_pandas(self): | |
1316 | pytimes = [time(1, 2, 3, 1356), | |
1317 | time(4, 5, 6, 1356), | |
1318 | time(0, 0, 0)] | |
1319 | ||
1320 | expected = np.array(pytimes[:2] + [None]) | |
1321 | expected_ms = np.array([x.replace(microsecond=1000) | |
1322 | for x in pytimes[:2]] + | |
1323 | [None]) | |
1324 | expected_s = np.array([x.replace(microsecond=0) | |
1325 | for x in pytimes[:2]] + | |
1326 | [None]) | |
1327 | ||
1328 | arr = np.array([_pytime_to_micros(v) for v in pytimes], | |
1329 | dtype='int64') | |
1330 | arr = np.array([_pytime_to_micros(v) for v in pytimes], | |
1331 | dtype='int64') | |
1332 | ||
1333 | null_mask = np.array([False, False, True], dtype=bool) | |
1334 | ||
1335 | a1 = pa.array(arr, mask=null_mask, type=pa.time64('us')) | |
1336 | a2 = pa.array(arr * 1000, mask=null_mask, | |
1337 | type=pa.time64('ns')) | |
1338 | ||
1339 | a3 = pa.array((arr / 1000).astype('i4'), mask=null_mask, | |
1340 | type=pa.time32('ms')) | |
1341 | a4 = pa.array((arr / 1000000).astype('i4'), mask=null_mask, | |
1342 | type=pa.time32('s')) | |
1343 | ||
1344 | names = ['time64[us]', 'time64[ns]', 'time32[ms]', 'time32[s]'] | |
1345 | batch = pa.RecordBatch.from_arrays([a1, a2, a3, a4], names) | |
1346 | ||
1347 | for arr, expected_values in [(a1, expected), | |
1348 | (a2, expected), | |
1349 | (a3, expected_ms), | |
1350 | (a4, expected_s)]: | |
1351 | result_pandas = arr.to_pandas() | |
1352 | assert (result_pandas.values == expected_values).all() | |
1353 | ||
1354 | df = batch.to_pandas() | |
1355 | expected_df = pd.DataFrame({'time64[us]': expected, | |
1356 | 'time64[ns]': expected, | |
1357 | 'time32[ms]': expected_ms, | |
1358 | 'time32[s]': expected_s}, | |
1359 | columns=names) | |
1360 | ||
1361 | tm.assert_frame_equal(df, expected_df) | |
1362 | ||
1363 | def test_numpy_datetime64_columns(self): | |
1364 | datetime64_ns = np.array([ | |
1365 | '2007-07-13T01:23:34.123456789', | |
1366 | None, | |
1367 | '2006-01-13T12:34:56.432539784', | |
1368 | '2010-08-13T05:46:57.437699912'], | |
1369 | dtype='datetime64[ns]') | |
1370 | _check_array_from_pandas_roundtrip(datetime64_ns) | |
1371 | ||
1372 | datetime64_us = np.array([ | |
1373 | '2007-07-13T01:23:34.123456', | |
1374 | None, | |
1375 | '2006-01-13T12:34:56.432539', | |
1376 | '2010-08-13T05:46:57.437699'], | |
1377 | dtype='datetime64[us]') | |
1378 | _check_array_from_pandas_roundtrip(datetime64_us) | |
1379 | ||
1380 | datetime64_ms = np.array([ | |
1381 | '2007-07-13T01:23:34.123', | |
1382 | None, | |
1383 | '2006-01-13T12:34:56.432', | |
1384 | '2010-08-13T05:46:57.437'], | |
1385 | dtype='datetime64[ms]') | |
1386 | _check_array_from_pandas_roundtrip(datetime64_ms) | |
1387 | ||
1388 | datetime64_s = np.array([ | |
1389 | '2007-07-13T01:23:34', | |
1390 | None, | |
1391 | '2006-01-13T12:34:56', | |
1392 | '2010-08-13T05:46:57'], | |
1393 | dtype='datetime64[s]') | |
1394 | _check_array_from_pandas_roundtrip(datetime64_s) | |
1395 | ||
1396 | def test_timestamp_to_pandas_ns(self): | |
1397 | # non-ns timestamp gets cast to ns on conversion to pandas | |
1398 | arr = pa.array([1, 2, 3], pa.timestamp('ms')) | |
1399 | expected = pd.Series(pd.to_datetime([1, 2, 3], unit='ms')) | |
1400 | s = arr.to_pandas() | |
1401 | tm.assert_series_equal(s, expected) | |
1402 | arr = pa.chunked_array([arr]) | |
1403 | s = arr.to_pandas() | |
1404 | tm.assert_series_equal(s, expected) | |
1405 | ||
1406 | def test_timestamp_to_pandas_out_of_bounds(self): | |
1407 | # ARROW-7758 check for out of bounds timestamps for non-ns timestamps | |
1408 | ||
1409 | for unit in ['s', 'ms', 'us']: | |
1410 | for tz in [None, 'America/New_York']: | |
1411 | arr = pa.array([datetime(1, 1, 1)], pa.timestamp(unit, tz=tz)) | |
1412 | table = pa.table({'a': arr}) | |
1413 | ||
1414 | msg = "would result in out of bounds timestamp" | |
1415 | with pytest.raises(ValueError, match=msg): | |
1416 | arr.to_pandas() | |
1417 | ||
1418 | with pytest.raises(ValueError, match=msg): | |
1419 | table.to_pandas() | |
1420 | ||
1421 | with pytest.raises(ValueError, match=msg): | |
1422 | # chunked array | |
1423 | table.column('a').to_pandas() | |
1424 | ||
1425 | # just ensure those don't give an error, but do not | |
1426 | # check actual garbage output | |
1427 | arr.to_pandas(safe=False) | |
1428 | table.to_pandas(safe=False) | |
1429 | table.column('a').to_pandas(safe=False) | |
1430 | ||
1431 | def test_timestamp_to_pandas_empty_chunked(self): | |
1432 | # ARROW-7907 table with chunked array with 0 chunks | |
1433 | table = pa.table({'a': pa.chunked_array([], type=pa.timestamp('us'))}) | |
1434 | result = table.to_pandas() | |
1435 | expected = pd.DataFrame({'a': pd.Series([], dtype="datetime64[ns]")}) | |
1436 | tm.assert_frame_equal(result, expected) | |
1437 | ||
1438 | @pytest.mark.parametrize('dtype', [pa.date32(), pa.date64()]) | |
1439 | def test_numpy_datetime64_day_unit(self, dtype): | |
1440 | datetime64_d = np.array([ | |
1441 | '2007-07-13', | |
1442 | None, | |
1443 | '2006-01-15', | |
1444 | '2010-08-19'], | |
1445 | dtype='datetime64[D]') | |
1446 | _check_array_from_pandas_roundtrip(datetime64_d, type=dtype) | |
1447 | ||
1448 | def test_array_from_pandas_date_with_mask(self): | |
1449 | m = np.array([True, False, True]) | |
1450 | data = pd.Series([ | |
1451 | date(1990, 1, 1), | |
1452 | date(1991, 1, 1), | |
1453 | date(1992, 1, 1) | |
1454 | ]) | |
1455 | ||
1456 | result = pa.Array.from_pandas(data, mask=m) | |
1457 | ||
1458 | expected = pd.Series([None, date(1991, 1, 1), None]) | |
1459 | assert pa.Array.from_pandas(expected).equals(result) | |
1460 | ||
1461 | @pytest.mark.skipif( | |
1462 | Version('1.16.0') <= Version(np.__version__) < Version('1.16.1'), | |
1463 | reason='Until numpy/numpy#12745 is resolved') | |
1464 | def test_fixed_offset_timezone(self): | |
1465 | df = pd.DataFrame({ | |
1466 | 'a': [ | |
1467 | pd.Timestamp('2012-11-11 00:00:00+01:00'), | |
1468 | pd.NaT | |
1469 | ] | |
1470 | }) | |
1471 | _check_pandas_roundtrip(df) | |
1472 | _check_serialize_components_roundtrip(df) | |
1473 | ||
1474 | def test_timedeltas_no_nulls(self): | |
1475 | df = pd.DataFrame({ | |
1476 | 'timedelta64': np.array([0, 3600000000000, 7200000000000], | |
1477 | dtype='timedelta64[ns]') | |
1478 | }) | |
1479 | field = pa.field('timedelta64', pa.duration('ns')) | |
1480 | schema = pa.schema([field]) | |
1481 | _check_pandas_roundtrip( | |
1482 | df, | |
1483 | expected_schema=schema, | |
1484 | ) | |
1485 | ||
1486 | def test_timedeltas_nulls(self): | |
1487 | df = pd.DataFrame({ | |
1488 | 'timedelta64': np.array([0, None, 7200000000000], | |
1489 | dtype='timedelta64[ns]') | |
1490 | }) | |
1491 | field = pa.field('timedelta64', pa.duration('ns')) | |
1492 | schema = pa.schema([field]) | |
1493 | _check_pandas_roundtrip( | |
1494 | df, | |
1495 | expected_schema=schema, | |
1496 | ) | |
1497 | ||
1498 | def test_month_day_nano_interval(self): | |
1499 | from pandas.tseries.offsets import DateOffset | |
1500 | df = pd.DataFrame({ | |
1501 | 'date_offset': [None, | |
1502 | DateOffset(days=3600, months=3600, microseconds=3, | |
1503 | nanoseconds=600)] | |
1504 | }) | |
1505 | schema = pa.schema([('date_offset', pa.month_day_nano_interval())]) | |
1506 | _check_pandas_roundtrip( | |
1507 | df, | |
1508 | expected_schema=schema) | |
1509 | ||
1510 | ||
1511 | # ---------------------------------------------------------------------- | |
1512 | # Conversion tests for string and binary types. | |
1513 | ||
1514 | ||
1515 | class TestConvertStringLikeTypes: | |
1516 | ||
1517 | def test_pandas_unicode(self): | |
1518 | repeats = 1000 | |
1519 | values = ['foo', None, 'bar', 'mañana', np.nan] | |
1520 | df = pd.DataFrame({'strings': values * repeats}) | |
1521 | field = pa.field('strings', pa.string()) | |
1522 | schema = pa.schema([field]) | |
1523 | ||
1524 | _check_pandas_roundtrip(df, expected_schema=schema) | |
1525 | ||
1526 | def test_bytes_to_binary(self): | |
1527 | values = ['qux', b'foo', None, bytearray(b'barz'), 'qux', np.nan] | |
1528 | df = pd.DataFrame({'strings': values}) | |
1529 | ||
1530 | table = pa.Table.from_pandas(df) | |
1531 | assert table[0].type == pa.binary() | |
1532 | ||
1533 | values2 = [b'qux', b'foo', None, b'barz', b'qux', np.nan] | |
1534 | expected = pd.DataFrame({'strings': values2}) | |
1535 | _check_pandas_roundtrip(df, expected) | |
1536 | ||
1537 | @pytest.mark.large_memory | |
1538 | def test_bytes_exceed_2gb(self): | |
1539 | v1 = b'x' * 100000000 | |
1540 | v2 = b'x' * 147483646 | |
1541 | ||
1542 | # ARROW-2227, hit exactly 2GB on the nose | |
1543 | df = pd.DataFrame({ | |
1544 | 'strings': [v1] * 20 + [v2] + ['x'] * 20 | |
1545 | }) | |
1546 | arr = pa.array(df['strings']) | |
1547 | assert isinstance(arr, pa.ChunkedArray) | |
1548 | assert arr.num_chunks == 2 | |
1549 | arr = None | |
1550 | ||
1551 | table = pa.Table.from_pandas(df) | |
1552 | assert table[0].num_chunks == 2 | |
1553 | ||
1554 | @pytest.mark.large_memory | |
1555 | @pytest.mark.parametrize('char', ['x', b'x']) | |
1556 | def test_auto_chunking_pandas_series_of_strings(self, char): | |
1557 | # ARROW-2367 | |
1558 | v1 = char * 100000000 | |
1559 | v2 = char * 147483646 | |
1560 | ||
1561 | df = pd.DataFrame({ | |
1562 | 'strings': [[v1]] * 20 + [[v2]] + [[b'x']] | |
1563 | }) | |
1564 | arr = pa.array(df['strings'], from_pandas=True) | |
1565 | assert isinstance(arr, pa.ChunkedArray) | |
1566 | assert arr.num_chunks == 2 | |
1567 | assert len(arr.chunk(0)) == 21 | |
1568 | assert len(arr.chunk(1)) == 1 | |
1569 | ||
1570 | def test_fixed_size_bytes(self): | |
1571 | values = [b'foo', None, bytearray(b'bar'), None, None, b'hey'] | |
1572 | df = pd.DataFrame({'strings': values}) | |
1573 | schema = pa.schema([pa.field('strings', pa.binary(3))]) | |
1574 | table = pa.Table.from_pandas(df, schema=schema) | |
1575 | assert table.schema[0].type == schema[0].type | |
1576 | assert table.schema[0].name == schema[0].name | |
1577 | result = table.to_pandas() | |
1578 | tm.assert_frame_equal(result, df) | |
1579 | ||
1580 | def test_fixed_size_bytes_does_not_accept_varying_lengths(self): | |
1581 | values = [b'foo', None, b'ba', None, None, b'hey'] | |
1582 | df = pd.DataFrame({'strings': values}) | |
1583 | schema = pa.schema([pa.field('strings', pa.binary(3))]) | |
1584 | with pytest.raises(pa.ArrowInvalid): | |
1585 | pa.Table.from_pandas(df, schema=schema) | |
1586 | ||
1587 | def test_variable_size_bytes(self): | |
1588 | s = pd.Series([b'123', b'', b'a', None]) | |
1589 | _check_series_roundtrip(s, type_=pa.binary()) | |
1590 | ||
1591 | def test_binary_from_bytearray(self): | |
1592 | s = pd.Series([bytearray(b'123'), bytearray(b''), bytearray(b'a'), | |
1593 | None]) | |
1594 | # Explicitly set type | |
1595 | _check_series_roundtrip(s, type_=pa.binary()) | |
1596 | # Infer type from bytearrays | |
1597 | _check_series_roundtrip(s, expected_pa_type=pa.binary()) | |
1598 | ||
1599 | def test_large_binary(self): | |
1600 | s = pd.Series([b'123', b'', b'a', None]) | |
1601 | _check_series_roundtrip(s, type_=pa.large_binary()) | |
1602 | df = pd.DataFrame({'a': s}) | |
1603 | _check_pandas_roundtrip( | |
1604 | df, schema=pa.schema([('a', pa.large_binary())])) | |
1605 | ||
1606 | def test_large_string(self): | |
1607 | s = pd.Series(['123', '', 'a', None]) | |
1608 | _check_series_roundtrip(s, type_=pa.large_string()) | |
1609 | df = pd.DataFrame({'a': s}) | |
1610 | _check_pandas_roundtrip( | |
1611 | df, schema=pa.schema([('a', pa.large_string())])) | |
1612 | ||
1613 | def test_table_empty_str(self): | |
1614 | values = ['', '', '', '', ''] | |
1615 | df = pd.DataFrame({'strings': values}) | |
1616 | field = pa.field('strings', pa.string()) | |
1617 | schema = pa.schema([field]) | |
1618 | table = pa.Table.from_pandas(df, schema=schema) | |
1619 | ||
1620 | result1 = table.to_pandas(strings_to_categorical=False) | |
1621 | expected1 = pd.DataFrame({'strings': values}) | |
1622 | tm.assert_frame_equal(result1, expected1, check_dtype=True) | |
1623 | ||
1624 | result2 = table.to_pandas(strings_to_categorical=True) | |
1625 | expected2 = pd.DataFrame({'strings': pd.Categorical(values)}) | |
1626 | tm.assert_frame_equal(result2, expected2, check_dtype=True) | |
1627 | ||
1628 | def test_selective_categoricals(self): | |
1629 | values = ['', '', '', '', ''] | |
1630 | df = pd.DataFrame({'strings': values}) | |
1631 | field = pa.field('strings', pa.string()) | |
1632 | schema = pa.schema([field]) | |
1633 | table = pa.Table.from_pandas(df, schema=schema) | |
1634 | expected_str = pd.DataFrame({'strings': values}) | |
1635 | expected_cat = pd.DataFrame({'strings': pd.Categorical(values)}) | |
1636 | ||
1637 | result1 = table.to_pandas(categories=['strings']) | |
1638 | tm.assert_frame_equal(result1, expected_cat, check_dtype=True) | |
1639 | result2 = table.to_pandas(categories=[]) | |
1640 | tm.assert_frame_equal(result2, expected_str, check_dtype=True) | |
1641 | result3 = table.to_pandas(categories=('strings',)) | |
1642 | tm.assert_frame_equal(result3, expected_cat, check_dtype=True) | |
1643 | result4 = table.to_pandas(categories=tuple()) | |
1644 | tm.assert_frame_equal(result4, expected_str, check_dtype=True) | |
1645 | ||
1646 | def test_to_pandas_categorical_zero_length(self): | |
1647 | # ARROW-3586 | |
1648 | array = pa.array([], type=pa.int32()) | |
1649 | table = pa.Table.from_arrays(arrays=[array], names=['col']) | |
1650 | # This would segfault under 0.11.0 | |
1651 | table.to_pandas(categories=['col']) | |
1652 | ||
1653 | def test_to_pandas_categories_already_dictionary(self): | |
1654 | # Showed up in ARROW-6434, ARROW-6435 | |
1655 | array = pa.array(['foo', 'foo', 'foo', 'bar']).dictionary_encode() | |
1656 | table = pa.Table.from_arrays(arrays=[array], names=['col']) | |
1657 | result = table.to_pandas(categories=['col']) | |
1658 | assert table.to_pandas().equals(result) | |
1659 | ||
1660 | def test_table_str_to_categorical_without_na(self): | |
1661 | values = ['a', 'a', 'b', 'b', 'c'] | |
1662 | df = pd.DataFrame({'strings': values}) | |
1663 | field = pa.field('strings', pa.string()) | |
1664 | schema = pa.schema([field]) | |
1665 | table = pa.Table.from_pandas(df, schema=schema) | |
1666 | ||
1667 | result = table.to_pandas(strings_to_categorical=True) | |
1668 | expected = pd.DataFrame({'strings': pd.Categorical(values)}) | |
1669 | tm.assert_frame_equal(result, expected, check_dtype=True) | |
1670 | ||
1671 | with pytest.raises(pa.ArrowInvalid): | |
1672 | table.to_pandas(strings_to_categorical=True, | |
1673 | zero_copy_only=True) | |
1674 | ||
1675 | def test_table_str_to_categorical_with_na(self): | |
1676 | values = [None, 'a', 'b', np.nan] | |
1677 | df = pd.DataFrame({'strings': values}) | |
1678 | field = pa.field('strings', pa.string()) | |
1679 | schema = pa.schema([field]) | |
1680 | table = pa.Table.from_pandas(df, schema=schema) | |
1681 | ||
1682 | result = table.to_pandas(strings_to_categorical=True) | |
1683 | expected = pd.DataFrame({'strings': pd.Categorical(values)}) | |
1684 | tm.assert_frame_equal(result, expected, check_dtype=True) | |
1685 | ||
1686 | with pytest.raises(pa.ArrowInvalid): | |
1687 | table.to_pandas(strings_to_categorical=True, | |
1688 | zero_copy_only=True) | |
1689 | ||
1690 | # Regression test for ARROW-2101 | |
1691 | def test_array_of_bytes_to_strings(self): | |
1692 | converted = pa.array(np.array([b'x'], dtype=object), pa.string()) | |
1693 | assert converted.type == pa.string() | |
1694 | ||
1695 | # Make sure that if an ndarray of bytes is passed to the array | |
1696 | # constructor and the type is string, it will fail if those bytes | |
1697 | # cannot be converted to utf-8 | |
1698 | def test_array_of_bytes_to_strings_bad_data(self): | |
1699 | with pytest.raises( | |
1700 | pa.lib.ArrowInvalid, | |
1701 | match="was not a utf8 string"): | |
1702 | pa.array(np.array([b'\x80\x81'], dtype=object), pa.string()) | |
1703 | ||
1704 | def test_numpy_string_array_to_fixed_size_binary(self): | |
1705 | arr = np.array([b'foo', b'bar', b'baz'], dtype='|S3') | |
1706 | ||
1707 | converted = pa.array(arr, type=pa.binary(3)) | |
1708 | expected = pa.array(list(arr), type=pa.binary(3)) | |
1709 | assert converted.equals(expected) | |
1710 | ||
1711 | mask = np.array([False, True, False]) | |
1712 | converted = pa.array(arr, type=pa.binary(3), mask=mask) | |
1713 | expected = pa.array([b'foo', None, b'baz'], type=pa.binary(3)) | |
1714 | assert converted.equals(expected) | |
1715 | ||
1716 | with pytest.raises(pa.lib.ArrowInvalid, | |
1717 | match=r'Got bytestring of length 3 \(expected 4\)'): | |
1718 | arr = np.array([b'foo', b'bar', b'baz'], dtype='|S3') | |
1719 | pa.array(arr, type=pa.binary(4)) | |
1720 | ||
1721 | with pytest.raises( | |
1722 | pa.lib.ArrowInvalid, | |
1723 | match=r'Got bytestring of length 12 \(expected 3\)'): | |
1724 | arr = np.array([b'foo', b'bar', b'baz'], dtype='|U3') | |
1725 | pa.array(arr, type=pa.binary(3)) | |
1726 | ||
1727 | ||
1728 | class TestConvertDecimalTypes: | |
1729 | """ | |
1730 | Conversion test for decimal types. | |
1731 | """ | |
1732 | decimal32 = [ | |
1733 | decimal.Decimal('-1234.123'), | |
1734 | decimal.Decimal('1234.439') | |
1735 | ] | |
1736 | decimal64 = [ | |
1737 | decimal.Decimal('-129934.123331'), | |
1738 | decimal.Decimal('129534.123731') | |
1739 | ] | |
1740 | decimal128 = [ | |
1741 | decimal.Decimal('394092382910493.12341234678'), | |
1742 | decimal.Decimal('-314292388910493.12343437128') | |
1743 | ] | |
1744 | ||
1745 | @pytest.mark.parametrize(('values', 'expected_type'), [ | |
1746 | pytest.param(decimal32, pa.decimal128(7, 3), id='decimal32'), | |
1747 | pytest.param(decimal64, pa.decimal128(12, 6), id='decimal64'), | |
1748 | pytest.param(decimal128, pa.decimal128(26, 11), id='decimal128') | |
1749 | ]) | |
1750 | def test_decimal_from_pandas(self, values, expected_type): | |
1751 | expected = pd.DataFrame({'decimals': values}) | |
1752 | table = pa.Table.from_pandas(expected, preserve_index=False) | |
1753 | field = pa.field('decimals', expected_type) | |
1754 | ||
1755 | # schema's metadata is generated by from_pandas conversion | |
1756 | expected_schema = pa.schema([field], metadata=table.schema.metadata) | |
1757 | assert table.schema.equals(expected_schema) | |
1758 | ||
1759 | @pytest.mark.parametrize('values', [ | |
1760 | pytest.param(decimal32, id='decimal32'), | |
1761 | pytest.param(decimal64, id='decimal64'), | |
1762 | pytest.param(decimal128, id='decimal128') | |
1763 | ]) | |
1764 | def test_decimal_to_pandas(self, values): | |
1765 | expected = pd.DataFrame({'decimals': values}) | |
1766 | converted = pa.Table.from_pandas(expected) | |
1767 | df = converted.to_pandas() | |
1768 | tm.assert_frame_equal(df, expected) | |
1769 | ||
1770 | def test_decimal_fails_with_truncation(self): | |
1771 | data1 = [decimal.Decimal('1.234')] | |
1772 | type1 = pa.decimal128(10, 2) | |
1773 | with pytest.raises(pa.ArrowInvalid): | |
1774 | pa.array(data1, type=type1) | |
1775 | ||
1776 | data2 = [decimal.Decimal('1.2345')] | |
1777 | type2 = pa.decimal128(10, 3) | |
1778 | with pytest.raises(pa.ArrowInvalid): | |
1779 | pa.array(data2, type=type2) | |
1780 | ||
1781 | def test_decimal_with_different_precisions(self): | |
1782 | data = [ | |
1783 | decimal.Decimal('0.01'), | |
1784 | decimal.Decimal('0.001'), | |
1785 | ] | |
1786 | series = pd.Series(data) | |
1787 | array = pa.array(series) | |
1788 | assert array.to_pylist() == data | |
1789 | assert array.type == pa.decimal128(3, 3) | |
1790 | ||
1791 | array = pa.array(data, type=pa.decimal128(12, 5)) | |
1792 | expected = [decimal.Decimal('0.01000'), decimal.Decimal('0.00100')] | |
1793 | assert array.to_pylist() == expected | |
1794 | ||
1795 | def test_decimal_with_None_explicit_type(self): | |
1796 | series = pd.Series([decimal.Decimal('3.14'), None]) | |
1797 | _check_series_roundtrip(series, type_=pa.decimal128(12, 5)) | |
1798 | ||
1799 | # Test that having all None values still produces decimal array | |
1800 | series = pd.Series([None] * 2) | |
1801 | _check_series_roundtrip(series, type_=pa.decimal128(12, 5)) | |
1802 | ||
1803 | def test_decimal_with_None_infer_type(self): | |
1804 | series = pd.Series([decimal.Decimal('3.14'), None]) | |
1805 | _check_series_roundtrip(series, expected_pa_type=pa.decimal128(3, 2)) | |
1806 | ||
1807 | def test_strided_objects(self, tmpdir): | |
1808 | # see ARROW-3053 | |
1809 | data = { | |
1810 | 'a': {0: 'a'}, | |
1811 | 'b': {0: decimal.Decimal('0.0')} | |
1812 | } | |
1813 | ||
1814 | # This yields strided objects | |
1815 | df = pd.DataFrame.from_dict(data) | |
1816 | _check_pandas_roundtrip(df) | |
1817 | ||
1818 | ||
1819 | class TestConvertListTypes: | |
1820 | """ | |
1821 | Conversion tests for list<> types. | |
1822 | """ | |
1823 | ||
1824 | def test_column_of_arrays(self): | |
1825 | df, schema = dataframe_with_arrays() | |
1826 | _check_pandas_roundtrip(df, schema=schema, expected_schema=schema) | |
1827 | table = pa.Table.from_pandas(df, schema=schema, preserve_index=False) | |
1828 | ||
1829 | # schema's metadata is generated by from_pandas conversion | |
1830 | expected_schema = schema.with_metadata(table.schema.metadata) | |
1831 | assert table.schema.equals(expected_schema) | |
1832 | ||
1833 | for column in df.columns: | |
1834 | field = schema.field(column) | |
1835 | _check_array_roundtrip(df[column], type=field.type) | |
1836 | ||
1837 | def test_column_of_arrays_to_py(self): | |
1838 | # Test regression in ARROW-1199 not caught in above test | |
1839 | dtype = 'i1' | |
1840 | arr = np.array([ | |
1841 | np.arange(10, dtype=dtype), | |
1842 | np.arange(5, dtype=dtype), | |
1843 | None, | |
1844 | np.arange(1, dtype=dtype) | |
1845 | ], dtype=object) | |
1846 | type_ = pa.list_(pa.int8()) | |
1847 | parr = pa.array(arr, type=type_) | |
1848 | ||
1849 | assert parr[0].as_py() == list(range(10)) | |
1850 | assert parr[1].as_py() == list(range(5)) | |
1851 | assert parr[2].as_py() is None | |
1852 | assert parr[3].as_py() == [0] | |
1853 | ||
1854 | def test_column_of_boolean_list(self): | |
1855 | # ARROW-4370: Table to pandas conversion fails for list of bool | |
1856 | array = pa.array([[True, False], [True]], type=pa.list_(pa.bool_())) | |
1857 | table = pa.Table.from_arrays([array], names=['col1']) | |
1858 | df = table.to_pandas() | |
1859 | ||
1860 | expected_df = pd.DataFrame({'col1': [[True, False], [True]]}) | |
1861 | tm.assert_frame_equal(df, expected_df) | |
1862 | ||
1863 | s = table[0].to_pandas() | |
1864 | tm.assert_series_equal(pd.Series(s), df['col1'], check_names=False) | |
1865 | ||
1866 | def test_column_of_decimal_list(self): | |
1867 | array = pa.array([[decimal.Decimal('1'), decimal.Decimal('2')], | |
1868 | [decimal.Decimal('3.3')]], | |
1869 | type=pa.list_(pa.decimal128(2, 1))) | |
1870 | table = pa.Table.from_arrays([array], names=['col1']) | |
1871 | df = table.to_pandas() | |
1872 | ||
1873 | expected_df = pd.DataFrame( | |
1874 | {'col1': [[decimal.Decimal('1'), decimal.Decimal('2')], | |
1875 | [decimal.Decimal('3.3')]]}) | |
1876 | tm.assert_frame_equal(df, expected_df) | |
1877 | ||
1878 | def test_nested_types_from_ndarray_null_entries(self): | |
1879 | # Root cause of ARROW-6435 | |
1880 | s = pd.Series(np.array([np.nan, np.nan], dtype=object)) | |
1881 | ||
1882 | for ty in [pa.list_(pa.int64()), | |
1883 | pa.large_list(pa.int64()), | |
1884 | pa.struct([pa.field('f0', 'int32')])]: | |
1885 | result = pa.array(s, type=ty) | |
1886 | expected = pa.array([None, None], type=ty) | |
1887 | assert result.equals(expected) | |
1888 | ||
1889 | with pytest.raises(TypeError): | |
1890 | pa.array(s.values, type=ty) | |
1891 | ||
1892 | def test_column_of_lists(self): | |
1893 | df, schema = dataframe_with_lists() | |
1894 | _check_pandas_roundtrip(df, schema=schema, expected_schema=schema) | |
1895 | table = pa.Table.from_pandas(df, schema=schema, preserve_index=False) | |
1896 | ||
1897 | # schema's metadata is generated by from_pandas conversion | |
1898 | expected_schema = schema.with_metadata(table.schema.metadata) | |
1899 | assert table.schema.equals(expected_schema) | |
1900 | ||
1901 | for column in df.columns: | |
1902 | field = schema.field(column) | |
1903 | _check_array_roundtrip(df[column], type=field.type) | |
1904 | ||
1905 | def test_column_of_lists_first_empty(self): | |
1906 | # ARROW-2124 | |
1907 | num_lists = [[], [2, 3, 4], [3, 6, 7, 8], [], [2]] | |
1908 | series = pd.Series([np.array(s, dtype=float) for s in num_lists]) | |
1909 | arr = pa.array(series) | |
1910 | result = pd.Series(arr.to_pandas()) | |
1911 | tm.assert_series_equal(result, series) | |
1912 | ||
1913 | def test_column_of_lists_chunked(self): | |
1914 | # ARROW-1357 | |
1915 | df = pd.DataFrame({ | |
1916 | 'lists': np.array([ | |
1917 | [1, 2], | |
1918 | None, | |
1919 | [2, 3], | |
1920 | [4, 5], | |
1921 | [6, 7], | |
1922 | [8, 9] | |
1923 | ], dtype=object) | |
1924 | }) | |
1925 | ||
1926 | schema = pa.schema([ | |
1927 | pa.field('lists', pa.list_(pa.int64())) | |
1928 | ]) | |
1929 | ||
1930 | t1 = pa.Table.from_pandas(df[:2], schema=schema) | |
1931 | t2 = pa.Table.from_pandas(df[2:], schema=schema) | |
1932 | ||
1933 | table = pa.concat_tables([t1, t2]) | |
1934 | result = table.to_pandas() | |
1935 | ||
1936 | tm.assert_frame_equal(result, df) | |
1937 | ||
1938 | def test_empty_column_of_lists_chunked(self): | |
1939 | df = pd.DataFrame({ | |
1940 | 'lists': np.array([], dtype=object) | |
1941 | }) | |
1942 | ||
1943 | schema = pa.schema([ | |
1944 | pa.field('lists', pa.list_(pa.int64())) | |
1945 | ]) | |
1946 | ||
1947 | table = pa.Table.from_pandas(df, schema=schema) | |
1948 | result = table.to_pandas() | |
1949 | ||
1950 | tm.assert_frame_equal(result, df) | |
1951 | ||
1952 | def test_column_of_lists_chunked2(self): | |
1953 | data1 = [[0, 1], [2, 3], [4, 5], [6, 7], [10, 11], | |
1954 | [12, 13], [14, 15], [16, 17]] | |
1955 | data2 = [[8, 9], [18, 19]] | |
1956 | ||
1957 | a1 = pa.array(data1) | |
1958 | a2 = pa.array(data2) | |
1959 | ||
1960 | t1 = pa.Table.from_arrays([a1], names=['a']) | |
1961 | t2 = pa.Table.from_arrays([a2], names=['a']) | |
1962 | ||
1963 | concatenated = pa.concat_tables([t1, t2]) | |
1964 | ||
1965 | result = concatenated.to_pandas() | |
1966 | expected = pd.DataFrame({'a': data1 + data2}) | |
1967 | ||
1968 | tm.assert_frame_equal(result, expected) | |
1969 | ||
1970 | def test_column_of_lists_strided(self): | |
1971 | df, schema = dataframe_with_lists() | |
1972 | df = pd.concat([df] * 6, ignore_index=True) | |
1973 | ||
1974 | arr = df['int64'].values[::3] | |
1975 | assert arr.strides[0] != 8 | |
1976 | ||
1977 | _check_array_roundtrip(arr) | |
1978 | ||
1979 | def test_nested_lists_all_none(self): | |
1980 | data = np.array([[None, None], None], dtype=object) | |
1981 | ||
1982 | arr = pa.array(data) | |
1983 | expected = pa.array(list(data)) | |
1984 | assert arr.equals(expected) | |
1985 | assert arr.type == pa.list_(pa.null()) | |
1986 | ||
1987 | data2 = np.array([None, None, [None, None], | |
1988 | np.array([None, None], dtype=object)], | |
1989 | dtype=object) | |
1990 | arr = pa.array(data2) | |
1991 | expected = pa.array([None, None, [None, None], [None, None]]) | |
1992 | assert arr.equals(expected) | |
1993 | ||
1994 | def test_nested_lists_all_empty(self): | |
1995 | # ARROW-2128 | |
1996 | data = pd.Series([[], [], []]) | |
1997 | arr = pa.array(data) | |
1998 | expected = pa.array(list(data)) | |
1999 | assert arr.equals(expected) | |
2000 | assert arr.type == pa.list_(pa.null()) | |
2001 | ||
2002 | def test_nested_list_first_empty(self): | |
2003 | # ARROW-2711 | |
2004 | data = pd.Series([[], ["a"]]) | |
2005 | arr = pa.array(data) | |
2006 | expected = pa.array(list(data)) | |
2007 | assert arr.equals(expected) | |
2008 | assert arr.type == pa.list_(pa.string()) | |
2009 | ||
2010 | def test_nested_smaller_ints(self): | |
2011 | # ARROW-1345, ARROW-2008, there were some type inference bugs happening | |
2012 | # before | |
2013 | data = pd.Series([np.array([1, 2, 3], dtype='i1'), None]) | |
2014 | result = pa.array(data) | |
2015 | result2 = pa.array(data.values) | |
2016 | expected = pa.array([[1, 2, 3], None], type=pa.list_(pa.int8())) | |
2017 | assert result.equals(expected) | |
2018 | assert result2.equals(expected) | |
2019 | ||
2020 | data3 = pd.Series([np.array([1, 2, 3], dtype='f4'), None]) | |
2021 | result3 = pa.array(data3) | |
2022 | expected3 = pa.array([[1, 2, 3], None], type=pa.list_(pa.float32())) | |
2023 | assert result3.equals(expected3) | |
2024 | ||
2025 | def test_infer_lists(self): | |
2026 | data = OrderedDict([ | |
2027 | ('nan_ints', [[None, 1], [2, 3]]), | |
2028 | ('ints', [[0, 1], [2, 3]]), | |
2029 | ('strs', [[None, 'b'], ['c', 'd']]), | |
2030 | ('nested_strs', [[[None, 'b'], ['c', 'd']], None]) | |
2031 | ]) | |
2032 | df = pd.DataFrame(data) | |
2033 | ||
2034 | expected_schema = pa.schema([ | |
2035 | pa.field('nan_ints', pa.list_(pa.int64())), | |
2036 | pa.field('ints', pa.list_(pa.int64())), | |
2037 | pa.field('strs', pa.list_(pa.string())), | |
2038 | pa.field('nested_strs', pa.list_(pa.list_(pa.string()))) | |
2039 | ]) | |
2040 | ||
2041 | _check_pandas_roundtrip(df, expected_schema=expected_schema) | |
2042 | ||
2043 | def test_fixed_size_list(self): | |
2044 | # ARROW-7365 | |
2045 | fixed_ty = pa.list_(pa.int64(), list_size=4) | |
2046 | variable_ty = pa.list_(pa.int64()) | |
2047 | ||
2048 | data = [[0, 1, 2, 3], None, [4, 5, 6, 7], [8, 9, 10, 11]] | |
2049 | fixed_arr = pa.array(data, type=fixed_ty) | |
2050 | variable_arr = pa.array(data, type=variable_ty) | |
2051 | ||
2052 | result = fixed_arr.to_pandas() | |
2053 | expected = variable_arr.to_pandas() | |
2054 | ||
2055 | for left, right in zip(result, expected): | |
2056 | if left is None: | |
2057 | assert right is None | |
2058 | npt.assert_array_equal(left, right) | |
2059 | ||
2060 | def test_infer_numpy_array(self): | |
2061 | data = OrderedDict([ | |
2062 | ('ints', [ | |
2063 | np.array([0, 1], dtype=np.int64), | |
2064 | np.array([2, 3], dtype=np.int64) | |
2065 | ]) | |
2066 | ]) | |
2067 | df = pd.DataFrame(data) | |
2068 | expected_schema = pa.schema([ | |
2069 | pa.field('ints', pa.list_(pa.int64())) | |
2070 | ]) | |
2071 | ||
2072 | _check_pandas_roundtrip(df, expected_schema=expected_schema) | |
2073 | ||
2074 | def test_to_list_of_structs_pandas(self): | |
2075 | ints = pa.array([1, 2, 3], pa.int32()) | |
2076 | strings = pa.array([['a', 'b'], ['c', 'd'], ['e', 'f']], | |
2077 | pa.list_(pa.string())) | |
2078 | structs = pa.StructArray.from_arrays([ints, strings], ['f1', 'f2']) | |
2079 | data = pa.ListArray.from_arrays([0, 1, 3], structs) | |
2080 | ||
2081 | expected = pd.Series([ | |
2082 | [{'f1': 1, 'f2': ['a', 'b']}], | |
2083 | [{'f1': 2, 'f2': ['c', 'd']}, | |
2084 | {'f1': 3, 'f2': ['e', 'f']}] | |
2085 | ]) | |
2086 | ||
2087 | series = pd.Series(data.to_pandas()) | |
2088 | tm.assert_series_equal(series, expected) | |
2089 | ||
2090 | @pytest.mark.parametrize('t,data,expected', [ | |
2091 | ( | |
2092 | pa.int64, | |
2093 | [[1, 2], [3], None], | |
2094 | [None, [3], None] | |
2095 | ), | |
2096 | ( | |
2097 | pa.string, | |
2098 | [['aaa', 'bb'], ['c'], None], | |
2099 | [None, ['c'], None] | |
2100 | ), | |
2101 | ( | |
2102 | pa.null, | |
2103 | [[None, None], [None], None], | |
2104 | [None, [None], None] | |
2105 | ) | |
2106 | ]) | |
2107 | def test_array_from_pandas_typed_array_with_mask(self, t, data, expected): | |
2108 | m = np.array([True, False, True]) | |
2109 | ||
2110 | s = pd.Series(data) | |
2111 | result = pa.Array.from_pandas(s, mask=m, type=pa.list_(t())) | |
2112 | ||
2113 | assert pa.Array.from_pandas(expected, | |
2114 | type=pa.list_(t())).equals(result) | |
2115 | ||
2116 | def test_empty_list_roundtrip(self): | |
2117 | empty_list_array = np.empty((3,), dtype=object) | |
2118 | empty_list_array.fill([]) | |
2119 | ||
2120 | df = pd.DataFrame({'a': np.array(['1', '2', '3']), | |
2121 | 'b': empty_list_array}) | |
2122 | tbl = pa.Table.from_pandas(df) | |
2123 | ||
2124 | result = tbl.to_pandas() | |
2125 | ||
2126 | tm.assert_frame_equal(result, df) | |
2127 | ||
2128 | def test_array_from_nested_arrays(self): | |
2129 | df, schema = dataframe_with_arrays() | |
2130 | for field in schema: | |
2131 | arr = df[field.name].values | |
2132 | expected = pa.array(list(arr), type=field.type) | |
2133 | result = pa.array(arr) | |
2134 | assert result.type == field.type # == list<scalar> | |
2135 | assert result.equals(expected) | |
2136 | ||
2137 | def test_nested_large_list(self): | |
2138 | s = (pa.array([[[1, 2, 3], [4]], None], | |
2139 | type=pa.large_list(pa.large_list(pa.int64()))) | |
2140 | .to_pandas()) | |
2141 | tm.assert_series_equal( | |
2142 | s, pd.Series([[[1, 2, 3], [4]], None], dtype=object), | |
2143 | check_names=False) | |
2144 | ||
2145 | def test_large_binary_list(self): | |
2146 | for list_type_factory in (pa.list_, pa.large_list): | |
2147 | s = (pa.array([["aa", "bb"], None, ["cc"], []], | |
2148 | type=list_type_factory(pa.large_binary())) | |
2149 | .to_pandas()) | |
2150 | tm.assert_series_equal( | |
2151 | s, pd.Series([[b"aa", b"bb"], None, [b"cc"], []]), | |
2152 | check_names=False) | |
2153 | s = (pa.array([["aa", "bb"], None, ["cc"], []], | |
2154 | type=list_type_factory(pa.large_string())) | |
2155 | .to_pandas()) | |
2156 | tm.assert_series_equal( | |
2157 | s, pd.Series([["aa", "bb"], None, ["cc"], []]), | |
2158 | check_names=False) | |
2159 | ||
2160 | def test_list_of_dictionary(self): | |
2161 | child = pa.array(["foo", "bar", None, "foo"]).dictionary_encode() | |
2162 | arr = pa.ListArray.from_arrays([0, 1, 3, 3, 4], child) | |
2163 | ||
2164 | # Expected a Series of lists | |
2165 | expected = pd.Series(arr.to_pylist()) | |
2166 | tm.assert_series_equal(arr.to_pandas(), expected) | |
2167 | ||
2168 | # Same but with nulls | |
2169 | arr = arr.take([0, 1, None, 3]) | |
2170 | expected[2] = None | |
2171 | tm.assert_series_equal(arr.to_pandas(), expected) | |
2172 | ||
2173 | @pytest.mark.large_memory | |
2174 | def test_auto_chunking_on_list_overflow(self): | |
2175 | # ARROW-9976 | |
2176 | n = 2**21 | |
2177 | df = pd.DataFrame.from_dict({ | |
2178 | "a": list(np.zeros((n, 2**10), dtype='uint8')), | |
2179 | "b": range(n) | |
2180 | }) | |
2181 | table = pa.Table.from_pandas(df) | |
2182 | ||
2183 | column_a = table[0] | |
2184 | assert column_a.num_chunks == 2 | |
2185 | assert len(column_a.chunk(0)) == 2**21 - 1 | |
2186 | assert len(column_a.chunk(1)) == 1 | |
2187 | ||
2188 | def test_map_array_roundtrip(self): | |
2189 | data = [[(b'a', 1), (b'b', 2)], | |
2190 | [(b'c', 3)], | |
2191 | [(b'd', 4), (b'e', 5), (b'f', 6)], | |
2192 | [(b'g', 7)]] | |
2193 | ||
2194 | df = pd.DataFrame({"map": data}) | |
2195 | schema = pa.schema([("map", pa.map_(pa.binary(), pa.int32()))]) | |
2196 | ||
2197 | _check_pandas_roundtrip(df, schema=schema) | |
2198 | ||
2199 | def test_map_array_chunked(self): | |
2200 | data1 = [[(b'a', 1), (b'b', 2)], | |
2201 | [(b'c', 3)], | |
2202 | [(b'd', 4), (b'e', 5), (b'f', 6)], | |
2203 | [(b'g', 7)]] | |
2204 | data2 = [[(k, v * 2) for k, v in row] for row in data1] | |
2205 | ||
2206 | arr1 = pa.array(data1, type=pa.map_(pa.binary(), pa.int32())) | |
2207 | arr2 = pa.array(data2, type=pa.map_(pa.binary(), pa.int32())) | |
2208 | arr = pa.chunked_array([arr1, arr2]) | |
2209 | ||
2210 | expected = pd.Series(data1 + data2) | |
2211 | actual = arr.to_pandas() | |
2212 | tm.assert_series_equal(actual, expected, check_names=False) | |
2213 | ||
2214 | def test_map_array_with_nulls(self): | |
2215 | data = [[(b'a', 1), (b'b', 2)], | |
2216 | None, | |
2217 | [(b'd', 4), (b'e', 5), (b'f', None)], | |
2218 | [(b'g', 7)]] | |
2219 | ||
2220 | # None value in item array causes upcast to float | |
2221 | expected = [[(k, float(v) if v is not None else None) for k, v in row] | |
2222 | if row is not None else None for row in data] | |
2223 | expected = pd.Series(expected) | |
2224 | ||
2225 | arr = pa.array(data, type=pa.map_(pa.binary(), pa.int32())) | |
2226 | actual = arr.to_pandas() | |
2227 | tm.assert_series_equal(actual, expected, check_names=False) | |
2228 | ||
2229 | def test_map_array_dictionary_encoded(self): | |
2230 | offsets = pa.array([0, 3, 5]) | |
2231 | items = pa.array(['a', 'b', 'c', 'a', 'd']).dictionary_encode() | |
2232 | keys = pa.array(list(range(len(items)))) | |
2233 | arr = pa.MapArray.from_arrays(offsets, keys, items) | |
2234 | ||
2235 | # Dictionary encoded values converted to dense | |
2236 | expected = pd.Series( | |
2237 | [[(0, 'a'), (1, 'b'), (2, 'c')], [(3, 'a'), (4, 'd')]]) | |
2238 | ||
2239 | actual = arr.to_pandas() | |
2240 | tm.assert_series_equal(actual, expected, check_names=False) | |
2241 | ||
2242 | ||
2243 | class TestConvertStructTypes: | |
2244 | """ | |
2245 | Conversion tests for struct types. | |
2246 | """ | |
2247 | ||
2248 | def test_pandas_roundtrip(self): | |
2249 | df = pd.DataFrame({'dicts': [{'a': 1, 'b': 2}, {'a': 3, 'b': 4}]}) | |
2250 | ||
2251 | expected_schema = pa.schema([ | |
2252 | ('dicts', pa.struct([('a', pa.int64()), ('b', pa.int64())])), | |
2253 | ]) | |
2254 | ||
2255 | _check_pandas_roundtrip(df, expected_schema=expected_schema) | |
2256 | ||
2257 | # specifying schema explicitly in from_pandas | |
2258 | _check_pandas_roundtrip( | |
2259 | df, schema=expected_schema, expected_schema=expected_schema) | |
2260 | ||
2261 | def test_to_pandas(self): | |
2262 | ints = pa.array([None, 2, 3], type=pa.int64()) | |
2263 | strs = pa.array(['a', None, 'c'], type=pa.string()) | |
2264 | bools = pa.array([True, False, None], type=pa.bool_()) | |
2265 | arr = pa.StructArray.from_arrays( | |
2266 | [ints, strs, bools], | |
2267 | ['ints', 'strs', 'bools']) | |
2268 | ||
2269 | expected = pd.Series([ | |
2270 | {'ints': None, 'strs': 'a', 'bools': True}, | |
2271 | {'ints': 2, 'strs': None, 'bools': False}, | |
2272 | {'ints': 3, 'strs': 'c', 'bools': None}, | |
2273 | ]) | |
2274 | ||
2275 | series = pd.Series(arr.to_pandas()) | |
2276 | tm.assert_series_equal(series, expected) | |
2277 | ||
2278 | def test_to_pandas_multiple_chunks(self): | |
2279 | # ARROW-11855 | |
2280 | gc.collect() | |
2281 | bytes_start = pa.total_allocated_bytes() | |
2282 | ints1 = pa.array([1], type=pa.int64()) | |
2283 | ints2 = pa.array([2], type=pa.int64()) | |
2284 | arr1 = pa.StructArray.from_arrays([ints1], ['ints']) | |
2285 | arr2 = pa.StructArray.from_arrays([ints2], ['ints']) | |
2286 | arr = pa.chunked_array([arr1, arr2]) | |
2287 | ||
2288 | expected = pd.Series([ | |
2289 | {'ints': 1}, | |
2290 | {'ints': 2} | |
2291 | ]) | |
2292 | ||
2293 | series = pd.Series(arr.to_pandas()) | |
2294 | tm.assert_series_equal(series, expected) | |
2295 | ||
2296 | del series | |
2297 | del arr | |
2298 | del arr1 | |
2299 | del arr2 | |
2300 | del ints1 | |
2301 | del ints2 | |
2302 | bytes_end = pa.total_allocated_bytes() | |
2303 | assert bytes_end == bytes_start | |
2304 | ||
2305 | def test_from_numpy(self): | |
2306 | dt = np.dtype([('x', np.int32), | |
2307 | (('y_title', 'y'), np.bool_)]) | |
2308 | ty = pa.struct([pa.field('x', pa.int32()), | |
2309 | pa.field('y', pa.bool_())]) | |
2310 | ||
2311 | data = np.array([], dtype=dt) | |
2312 | arr = pa.array(data, type=ty) | |
2313 | assert arr.to_pylist() == [] | |
2314 | ||
2315 | data = np.array([(42, True), (43, False)], dtype=dt) | |
2316 | arr = pa.array(data, type=ty) | |
2317 | assert arr.to_pylist() == [{'x': 42, 'y': True}, | |
2318 | {'x': 43, 'y': False}] | |
2319 | ||
2320 | # With mask | |
2321 | arr = pa.array(data, mask=np.bool_([False, True]), type=ty) | |
2322 | assert arr.to_pylist() == [{'x': 42, 'y': True}, None] | |
2323 | ||
2324 | # Trivial struct type | |
2325 | dt = np.dtype([]) | |
2326 | ty = pa.struct([]) | |
2327 | ||
2328 | data = np.array([], dtype=dt) | |
2329 | arr = pa.array(data, type=ty) | |
2330 | assert arr.to_pylist() == [] | |
2331 | ||
2332 | data = np.array([(), ()], dtype=dt) | |
2333 | arr = pa.array(data, type=ty) | |
2334 | assert arr.to_pylist() == [{}, {}] | |
2335 | ||
2336 | def test_from_numpy_nested(self): | |
2337 | # Note: an object field inside a struct | |
2338 | dt = np.dtype([('x', np.dtype([('xx', np.int8), | |
2339 | ('yy', np.bool_)])), | |
2340 | ('y', np.int16), | |
2341 | ('z', np.object_)]) | |
2342 | # Note: itemsize is not a multiple of sizeof(object) | |
2343 | assert dt.itemsize == 12 | |
2344 | ty = pa.struct([pa.field('x', pa.struct([pa.field('xx', pa.int8()), | |
2345 | pa.field('yy', pa.bool_())])), | |
2346 | pa.field('y', pa.int16()), | |
2347 | pa.field('z', pa.string())]) | |
2348 | ||
2349 | data = np.array([], dtype=dt) | |
2350 | arr = pa.array(data, type=ty) | |
2351 | assert arr.to_pylist() == [] | |
2352 | ||
2353 | data = np.array([ | |
2354 | ((1, True), 2, 'foo'), | |
2355 | ((3, False), 4, 'bar')], dtype=dt) | |
2356 | arr = pa.array(data, type=ty) | |
2357 | assert arr.to_pylist() == [ | |
2358 | {'x': {'xx': 1, 'yy': True}, 'y': 2, 'z': 'foo'}, | |
2359 | {'x': {'xx': 3, 'yy': False}, 'y': 4, 'z': 'bar'}] | |
2360 | ||
2361 | @pytest.mark.slow | |
2362 | @pytest.mark.large_memory | |
2363 | def test_from_numpy_large(self): | |
2364 | # Exercise rechunking + nulls | |
2365 | target_size = 3 * 1024**3 # 4GB | |
2366 | dt = np.dtype([('x', np.float64), ('y', 'object')]) | |
2367 | bs = 65536 - dt.itemsize | |
2368 | block = b'.' * bs | |
2369 | n = target_size // (bs + dt.itemsize) | |
2370 | data = np.zeros(n, dtype=dt) | |
2371 | data['x'] = np.random.random_sample(n) | |
2372 | data['y'] = block | |
2373 | # Add implicit nulls | |
2374 | data['x'][data['x'] < 0.2] = np.nan | |
2375 | ||
2376 | ty = pa.struct([pa.field('x', pa.float64()), | |
2377 | pa.field('y', pa.binary())]) | |
2378 | arr = pa.array(data, type=ty, from_pandas=True) | |
2379 | assert arr.num_chunks == 2 | |
2380 | ||
2381 | def iter_chunked_array(arr): | |
2382 | for chunk in arr.iterchunks(): | |
2383 | yield from chunk | |
2384 | ||
2385 | def check(arr, data, mask=None): | |
2386 | assert len(arr) == len(data) | |
2387 | xs = data['x'] | |
2388 | ys = data['y'] | |
2389 | for i, obj in enumerate(iter_chunked_array(arr)): | |
2390 | try: | |
2391 | d = obj.as_py() | |
2392 | if mask is not None and mask[i]: | |
2393 | assert d is None | |
2394 | else: | |
2395 | x = xs[i] | |
2396 | if np.isnan(x): | |
2397 | assert d['x'] is None | |
2398 | else: | |
2399 | assert d['x'] == x | |
2400 | assert d['y'] == ys[i] | |
2401 | except Exception: | |
2402 | print("Failed at index", i) | |
2403 | raise | |
2404 | ||
2405 | check(arr, data) | |
2406 | del arr | |
2407 | ||
2408 | # Now with explicit mask | |
2409 | mask = np.random.random_sample(n) < 0.2 | |
2410 | arr = pa.array(data, type=ty, mask=mask, from_pandas=True) | |
2411 | assert arr.num_chunks == 2 | |
2412 | ||
2413 | check(arr, data, mask) | |
2414 | del arr | |
2415 | ||
2416 | def test_from_numpy_bad_input(self): | |
2417 | ty = pa.struct([pa.field('x', pa.int32()), | |
2418 | pa.field('y', pa.bool_())]) | |
2419 | dt = np.dtype([('x', np.int32), | |
2420 | ('z', np.bool_)]) | |
2421 | ||
2422 | data = np.array([], dtype=dt) | |
2423 | with pytest.raises(ValueError, | |
2424 | match="Missing field 'y'"): | |
2425 | pa.array(data, type=ty) | |
2426 | data = np.int32([]) | |
2427 | with pytest.raises(TypeError, | |
2428 | match="Expected struct array"): | |
2429 | pa.array(data, type=ty) | |
2430 | ||
2431 | def test_from_tuples(self): | |
2432 | df = pd.DataFrame({'tuples': [(1, 2), (3, 4)]}) | |
2433 | expected_df = pd.DataFrame( | |
2434 | {'tuples': [{'a': 1, 'b': 2}, {'a': 3, 'b': 4}]}) | |
2435 | ||
2436 | # conversion from tuples works when specifying expected struct type | |
2437 | struct_type = pa.struct([('a', pa.int64()), ('b', pa.int64())]) | |
2438 | ||
2439 | arr = np.asarray(df['tuples']) | |
2440 | _check_array_roundtrip( | |
2441 | arr, expected=expected_df['tuples'], type=struct_type) | |
2442 | ||
2443 | expected_schema = pa.schema([('tuples', struct_type)]) | |
2444 | _check_pandas_roundtrip( | |
2445 | df, expected=expected_df, schema=expected_schema, | |
2446 | expected_schema=expected_schema) | |
2447 | ||
2448 | def test_struct_of_dictionary(self): | |
2449 | names = ['ints', 'strs'] | |
2450 | children = [pa.array([456, 789, 456]).dictionary_encode(), | |
2451 | pa.array(["foo", "foo", None]).dictionary_encode()] | |
2452 | arr = pa.StructArray.from_arrays(children, names=names) | |
2453 | ||
2454 | # Expected a Series of {field name: field value} dicts | |
2455 | rows_as_tuples = zip(*(child.to_pylist() for child in children)) | |
2456 | rows_as_dicts = [dict(zip(names, row)) for row in rows_as_tuples] | |
2457 | ||
2458 | expected = pd.Series(rows_as_dicts) | |
2459 | tm.assert_series_equal(arr.to_pandas(), expected) | |
2460 | ||
2461 | # Same but with nulls | |
2462 | arr = arr.take([0, None, 2]) | |
2463 | expected[1] = None | |
2464 | tm.assert_series_equal(arr.to_pandas(), expected) | |
2465 | ||
2466 | ||
2467 | class TestZeroCopyConversion: | |
2468 | """ | |
2469 | Tests that zero-copy conversion works with some types. | |
2470 | """ | |
2471 | ||
2472 | def test_zero_copy_success(self): | |
2473 | result = pa.array([0, 1, 2]).to_pandas(zero_copy_only=True) | |
2474 | npt.assert_array_equal(result, [0, 1, 2]) | |
2475 | ||
2476 | def test_zero_copy_dictionaries(self): | |
2477 | arr = pa.DictionaryArray.from_arrays( | |
2478 | np.array([0, 0]), | |
2479 | np.array([5])) | |
2480 | ||
2481 | result = arr.to_pandas(zero_copy_only=True) | |
2482 | values = pd.Categorical([5, 5]) | |
2483 | ||
2484 | tm.assert_series_equal(pd.Series(result), pd.Series(values), | |
2485 | check_names=False) | |
2486 | ||
2487 | def test_zero_copy_timestamp(self): | |
2488 | arr = np.array(['2007-07-13'], dtype='datetime64[ns]') | |
2489 | result = pa.array(arr).to_pandas(zero_copy_only=True) | |
2490 | npt.assert_array_equal(result, arr) | |
2491 | ||
2492 | def test_zero_copy_duration(self): | |
2493 | arr = np.array([1], dtype='timedelta64[ns]') | |
2494 | result = pa.array(arr).to_pandas(zero_copy_only=True) | |
2495 | npt.assert_array_equal(result, arr) | |
2496 | ||
2497 | def check_zero_copy_failure(self, arr): | |
2498 | with pytest.raises(pa.ArrowInvalid): | |
2499 | arr.to_pandas(zero_copy_only=True) | |
2500 | ||
2501 | def test_zero_copy_failure_on_object_types(self): | |
2502 | self.check_zero_copy_failure(pa.array(['A', 'B', 'C'])) | |
2503 | ||
2504 | def test_zero_copy_failure_with_int_when_nulls(self): | |
2505 | self.check_zero_copy_failure(pa.array([0, 1, None])) | |
2506 | ||
2507 | def test_zero_copy_failure_with_float_when_nulls(self): | |
2508 | self.check_zero_copy_failure(pa.array([0.0, 1.0, None])) | |
2509 | ||
2510 | def test_zero_copy_failure_on_bool_types(self): | |
2511 | self.check_zero_copy_failure(pa.array([True, False])) | |
2512 | ||
2513 | def test_zero_copy_failure_on_list_types(self): | |
2514 | arr = pa.array([[1, 2], [8, 9]], type=pa.list_(pa.int64())) | |
2515 | self.check_zero_copy_failure(arr) | |
2516 | ||
2517 | def test_zero_copy_failure_on_timestamp_with_nulls(self): | |
2518 | arr = np.array([1, None], dtype='datetime64[ns]') | |
2519 | self.check_zero_copy_failure(pa.array(arr)) | |
2520 | ||
2521 | def test_zero_copy_failure_on_duration_with_nulls(self): | |
2522 | arr = np.array([1, None], dtype='timedelta64[ns]') | |
2523 | self.check_zero_copy_failure(pa.array(arr)) | |
2524 | ||
2525 | ||
2526 | def _non_threaded_conversion(): | |
2527 | df = _alltypes_example() | |
2528 | _check_pandas_roundtrip(df, use_threads=False) | |
2529 | _check_pandas_roundtrip(df, use_threads=False, as_batch=True) | |
2530 | ||
2531 | ||
2532 | def _threaded_conversion(): | |
2533 | df = _alltypes_example() | |
2534 | _check_pandas_roundtrip(df, use_threads=True) | |
2535 | _check_pandas_roundtrip(df, use_threads=True, as_batch=True) | |
2536 | ||
2537 | ||
2538 | class TestConvertMisc: | |
2539 | """ | |
2540 | Miscellaneous conversion tests. | |
2541 | """ | |
2542 | ||
2543 | type_pairs = [ | |
2544 | (np.int8, pa.int8()), | |
2545 | (np.int16, pa.int16()), | |
2546 | (np.int32, pa.int32()), | |
2547 | (np.int64, pa.int64()), | |
2548 | (np.uint8, pa.uint8()), | |
2549 | (np.uint16, pa.uint16()), | |
2550 | (np.uint32, pa.uint32()), | |
2551 | (np.uint64, pa.uint64()), | |
2552 | (np.float16, pa.float16()), | |
2553 | (np.float32, pa.float32()), | |
2554 | (np.float64, pa.float64()), | |
2555 | # XXX unsupported | |
2556 | # (np.dtype([('a', 'i2')]), pa.struct([pa.field('a', pa.int16())])), | |
2557 | (np.object_, pa.string()), | |
2558 | (np.object_, pa.binary()), | |
2559 | (np.object_, pa.binary(10)), | |
2560 | (np.object_, pa.list_(pa.int64())), | |
2561 | ] | |
2562 | ||
2563 | def test_all_none_objects(self): | |
2564 | df = pd.DataFrame({'a': [None, None, None]}) | |
2565 | _check_pandas_roundtrip(df) | |
2566 | ||
2567 | def test_all_none_category(self): | |
2568 | df = pd.DataFrame({'a': [None, None, None]}) | |
2569 | df['a'] = df['a'].astype('category') | |
2570 | _check_pandas_roundtrip(df) | |
2571 | ||
2572 | def test_empty_arrays(self): | |
2573 | for dtype, pa_type in self.type_pairs: | |
2574 | arr = np.array([], dtype=dtype) | |
2575 | _check_array_roundtrip(arr, type=pa_type) | |
2576 | ||
2577 | def test_non_threaded_conversion(self): | |
2578 | _non_threaded_conversion() | |
2579 | ||
2580 | def test_threaded_conversion_multiprocess(self): | |
2581 | # Parallel conversion should work from child processes too (ARROW-2963) | |
2582 | pool = mp.Pool(2) | |
2583 | try: | |
2584 | pool.apply(_threaded_conversion) | |
2585 | finally: | |
2586 | pool.close() | |
2587 | pool.join() | |
2588 | ||
2589 | def test_category(self): | |
2590 | repeats = 5 | |
2591 | v1 = ['foo', None, 'bar', 'qux', np.nan] | |
2592 | v2 = [4, 5, 6, 7, 8] | |
2593 | v3 = [b'foo', None, b'bar', b'qux', np.nan] | |
2594 | ||
2595 | arrays = { | |
2596 | 'cat_strings': pd.Categorical(v1 * repeats), | |
2597 | 'cat_strings_with_na': pd.Categorical(v1 * repeats, | |
2598 | categories=['foo', 'bar']), | |
2599 | 'cat_ints': pd.Categorical(v2 * repeats), | |
2600 | 'cat_binary': pd.Categorical(v3 * repeats), | |
2601 | 'cat_strings_ordered': pd.Categorical( | |
2602 | v1 * repeats, categories=['bar', 'qux', 'foo'], | |
2603 | ordered=True), | |
2604 | 'ints': v2 * repeats, | |
2605 | 'ints2': v2 * repeats, | |
2606 | 'strings': v1 * repeats, | |
2607 | 'strings2': v1 * repeats, | |
2608 | 'strings3': v3 * repeats} | |
2609 | df = pd.DataFrame(arrays) | |
2610 | _check_pandas_roundtrip(df) | |
2611 | ||
2612 | for k in arrays: | |
2613 | _check_array_roundtrip(arrays[k]) | |
2614 | ||
2615 | def test_category_implicit_from_pandas(self): | |
2616 | # ARROW-3374 | |
2617 | def _check(v): | |
2618 | arr = pa.array(v) | |
2619 | result = arr.to_pandas() | |
2620 | tm.assert_series_equal(pd.Series(result), pd.Series(v)) | |
2621 | ||
2622 | arrays = [ | |
2623 | pd.Categorical(['a', 'b', 'c'], categories=['a', 'b']), | |
2624 | pd.Categorical(['a', 'b', 'c'], categories=['a', 'b'], | |
2625 | ordered=True) | |
2626 | ] | |
2627 | for arr in arrays: | |
2628 | _check(arr) | |
2629 | ||
2630 | def test_empty_category(self): | |
2631 | # ARROW-2443 | |
2632 | df = pd.DataFrame({'cat': pd.Categorical([])}) | |
2633 | _check_pandas_roundtrip(df) | |
2634 | ||
2635 | def test_category_zero_chunks(self): | |
2636 | # ARROW-5952 | |
2637 | for pa_type, dtype in [(pa.string(), 'object'), (pa.int64(), 'int64')]: | |
2638 | a = pa.chunked_array([], pa.dictionary(pa.int8(), pa_type)) | |
2639 | result = a.to_pandas() | |
2640 | expected = pd.Categorical([], categories=np.array([], dtype=dtype)) | |
2641 | tm.assert_series_equal(pd.Series(result), pd.Series(expected)) | |
2642 | ||
2643 | table = pa.table({'a': a}) | |
2644 | result = table.to_pandas() | |
2645 | expected = pd.DataFrame({'a': expected}) | |
2646 | tm.assert_frame_equal(result, expected) | |
2647 | ||
2648 | @pytest.mark.parametrize( | |
2649 | "data,error_type", | |
2650 | [ | |
2651 | ({"a": ["a", 1, 2.0]}, pa.ArrowTypeError), | |
2652 | ({"a": ["a", 1, 2.0]}, pa.ArrowTypeError), | |
2653 | ({"a": [1, True]}, pa.ArrowTypeError), | |
2654 | ({"a": [True, "a"]}, pa.ArrowInvalid), | |
2655 | ({"a": [1, "a"]}, pa.ArrowInvalid), | |
2656 | ({"a": [1.0, "a"]}, pa.ArrowInvalid), | |
2657 | ], | |
2658 | ) | |
2659 | def test_mixed_types_fails(self, data, error_type): | |
2660 | df = pd.DataFrame(data) | |
2661 | msg = "Conversion failed for column a with type object" | |
2662 | with pytest.raises(error_type, match=msg): | |
2663 | pa.Table.from_pandas(df) | |
2664 | ||
2665 | def test_strided_data_import(self): | |
2666 | cases = [] | |
2667 | ||
2668 | columns = ['a', 'b', 'c'] | |
2669 | N, K = 100, 3 | |
2670 | random_numbers = np.random.randn(N, K).copy() * 100 | |
2671 | ||
2672 | numeric_dtypes = ['i1', 'i2', 'i4', 'i8', 'u1', 'u2', 'u4', 'u8', | |
2673 | 'f4', 'f8'] | |
2674 | ||
2675 | for type_name in numeric_dtypes: | |
2676 | cases.append(random_numbers.astype(type_name)) | |
2677 | ||
2678 | # strings | |
2679 | cases.append(np.array([random_ascii(10) for i in range(N * K)], | |
2680 | dtype=object) | |
2681 | .reshape(N, K).copy()) | |
2682 | ||
2683 | # booleans | |
2684 | boolean_objects = (np.array([True, False, True] * N, dtype=object) | |
2685 | .reshape(N, K).copy()) | |
2686 | ||
2687 | # add some nulls, so dtype comes back as objects | |
2688 | boolean_objects[5] = None | |
2689 | cases.append(boolean_objects) | |
2690 | ||
2691 | cases.append(np.arange("2016-01-01T00:00:00.001", N * K, | |
2692 | dtype='datetime64[ms]') | |
2693 | .reshape(N, K).copy()) | |
2694 | ||
2695 | strided_mask = (random_numbers > 0).astype(bool)[:, 0] | |
2696 | ||
2697 | for case in cases: | |
2698 | df = pd.DataFrame(case, columns=columns) | |
2699 | col = df['a'] | |
2700 | ||
2701 | _check_pandas_roundtrip(df) | |
2702 | _check_array_roundtrip(col) | |
2703 | _check_array_roundtrip(col, mask=strided_mask) | |
2704 | ||
2705 | def test_all_nones(self): | |
2706 | def _check_series(s): | |
2707 | converted = pa.array(s) | |
2708 | assert isinstance(converted, pa.NullArray) | |
2709 | assert len(converted) == 3 | |
2710 | assert converted.null_count == 3 | |
2711 | for item in converted: | |
2712 | assert item is pa.NA | |
2713 | ||
2714 | _check_series(pd.Series([None] * 3, dtype=object)) | |
2715 | _check_series(pd.Series([np.nan] * 3, dtype=object)) | |
2716 | _check_series(pd.Series([None, np.nan, None], dtype=object)) | |
2717 | ||
2718 | def test_partial_schema(self): | |
2719 | data = OrderedDict([ | |
2720 | ('a', [0, 1, 2, 3, 4]), | |
2721 | ('b', np.array([-10, -5, 0, 5, 10], dtype=np.int32)), | |
2722 | ('c', [-10, -5, 0, 5, 10]) | |
2723 | ]) | |
2724 | df = pd.DataFrame(data) | |
2725 | ||
2726 | partial_schema = pa.schema([ | |
2727 | pa.field('c', pa.int64()), | |
2728 | pa.field('a', pa.int64()) | |
2729 | ]) | |
2730 | ||
2731 | _check_pandas_roundtrip(df, schema=partial_schema, | |
2732 | expected=df[['c', 'a']], | |
2733 | expected_schema=partial_schema) | |
2734 | ||
2735 | def test_table_batch_empty_dataframe(self): | |
2736 | df = pd.DataFrame({}) | |
2737 | _check_pandas_roundtrip(df) | |
2738 | _check_pandas_roundtrip(df, as_batch=True) | |
2739 | ||
2740 | df2 = pd.DataFrame({}, index=[0, 1, 2]) | |
2741 | _check_pandas_roundtrip(df2, preserve_index=True) | |
2742 | _check_pandas_roundtrip(df2, as_batch=True, preserve_index=True) | |
2743 | ||
2744 | def test_convert_empty_table(self): | |
2745 | arr = pa.array([], type=pa.int64()) | |
2746 | empty_objects = pd.Series(np.array([], dtype=object)) | |
2747 | tm.assert_series_equal(arr.to_pandas(), | |
2748 | pd.Series(np.array([], dtype=np.int64))) | |
2749 | arr = pa.array([], type=pa.string()) | |
2750 | tm.assert_series_equal(arr.to_pandas(), empty_objects) | |
2751 | arr = pa.array([], type=pa.list_(pa.int64())) | |
2752 | tm.assert_series_equal(arr.to_pandas(), empty_objects) | |
2753 | arr = pa.array([], type=pa.struct([pa.field('a', pa.int64())])) | |
2754 | tm.assert_series_equal(arr.to_pandas(), empty_objects) | |
2755 | ||
2756 | def test_non_natural_stride(self): | |
2757 | """ | |
2758 | ARROW-2172: converting from a Numpy array with a stride that's | |
2759 | not a multiple of itemsize. | |
2760 | """ | |
2761 | dtype = np.dtype([('x', np.int32), ('y', np.int16)]) | |
2762 | data = np.array([(42, -1), (-43, 2)], dtype=dtype) | |
2763 | assert data.strides == (6,) | |
2764 | arr = pa.array(data['x'], type=pa.int32()) | |
2765 | assert arr.to_pylist() == [42, -43] | |
2766 | arr = pa.array(data['y'], type=pa.int16()) | |
2767 | assert arr.to_pylist() == [-1, 2] | |
2768 | ||
2769 | def test_array_from_strided_numpy_array(self): | |
2770 | # ARROW-5651 | |
2771 | np_arr = np.arange(0, 10, dtype=np.float32)[1:-1:2] | |
2772 | pa_arr = pa.array(np_arr, type=pa.float64()) | |
2773 | expected = pa.array([1.0, 3.0, 5.0, 7.0], type=pa.float64()) | |
2774 | pa_arr.equals(expected) | |
2775 | ||
2776 | def test_safe_unsafe_casts(self): | |
2777 | # ARROW-2799 | |
2778 | df = pd.DataFrame({ | |
2779 | 'A': list('abc'), | |
2780 | 'B': np.linspace(0, 1, 3) | |
2781 | }) | |
2782 | ||
2783 | schema = pa.schema([ | |
2784 | pa.field('A', pa.string()), | |
2785 | pa.field('B', pa.int32()) | |
2786 | ]) | |
2787 | ||
2788 | with pytest.raises(ValueError): | |
2789 | pa.Table.from_pandas(df, schema=schema) | |
2790 | ||
2791 | table = pa.Table.from_pandas(df, schema=schema, safe=False) | |
2792 | assert table.column('B').type == pa.int32() | |
2793 | ||
2794 | def test_error_sparse(self): | |
2795 | # ARROW-2818 | |
2796 | try: | |
2797 | df = pd.DataFrame({'a': pd.arrays.SparseArray([1, np.nan, 3])}) | |
2798 | except AttributeError: | |
2799 | # pandas.arrays module introduced in pandas 0.24 | |
2800 | df = pd.DataFrame({'a': pd.SparseArray([1, np.nan, 3])}) | |
2801 | with pytest.raises(TypeError, match="Sparse pandas data"): | |
2802 | pa.Table.from_pandas(df) | |
2803 | ||
2804 | ||
2805 | def test_safe_cast_from_float_with_nans_to_int(): | |
2806 | # TODO(kszucs): write tests for creating Date32 and Date64 arrays, see | |
2807 | # ARROW-4258 and https://github.com/apache/arrow/pull/3395 | |
2808 | values = pd.Series([1, 2, None, 4]) | |
2809 | arr = pa.Array.from_pandas(values, type=pa.int32(), safe=True) | |
2810 | expected = pa.array([1, 2, None, 4], type=pa.int32()) | |
2811 | assert arr.equals(expected) | |
2812 | ||
2813 | ||
2814 | def _fully_loaded_dataframe_example(): | |
2815 | index = pd.MultiIndex.from_arrays([ | |
2816 | pd.date_range('2000-01-01', periods=5).repeat(2), | |
2817 | np.tile(np.array(['foo', 'bar'], dtype=object), 5) | |
2818 | ]) | |
2819 | ||
2820 | c1 = pd.date_range('2000-01-01', periods=10) | |
2821 | data = { | |
2822 | 0: c1, | |
2823 | 1: c1.tz_localize('utc'), | |
2824 | 2: c1.tz_localize('US/Eastern'), | |
2825 | 3: c1[::2].tz_localize('utc').repeat(2).astype('category'), | |
2826 | 4: ['foo', 'bar'] * 5, | |
2827 | 5: pd.Series(['foo', 'bar'] * 5).astype('category').values, | |
2828 | 6: [True, False] * 5, | |
2829 | 7: np.random.randn(10), | |
2830 | 8: np.random.randint(0, 100, size=10), | |
2831 | 9: pd.period_range('2013', periods=10, freq='M') | |
2832 | } | |
2833 | ||
2834 | if Version(pd.__version__) >= Version('0.21'): | |
2835 | # There is an issue with pickling IntervalIndex in pandas 0.20.x | |
2836 | data[10] = pd.interval_range(start=1, freq=1, periods=10) | |
2837 | ||
2838 | return pd.DataFrame(data, index=index) | |
2839 | ||
2840 | ||
2841 | @pytest.mark.parametrize('columns', ([b'foo'], ['foo'])) | |
2842 | def test_roundtrip_with_bytes_unicode(columns): | |
2843 | df = pd.DataFrame(columns=columns) | |
2844 | table1 = pa.Table.from_pandas(df) | |
2845 | table2 = pa.Table.from_pandas(table1.to_pandas()) | |
2846 | assert table1.equals(table2) | |
2847 | assert table1.schema.equals(table2.schema) | |
2848 | assert table1.schema.metadata == table2.schema.metadata | |
2849 | ||
2850 | ||
2851 | def _check_serialize_components_roundtrip(pd_obj): | |
2852 | with pytest.warns(FutureWarning): | |
2853 | ctx = pa.default_serialization_context() | |
2854 | ||
2855 | with pytest.warns(FutureWarning): | |
2856 | components = ctx.serialize(pd_obj).to_components() | |
2857 | with pytest.warns(FutureWarning): | |
2858 | deserialized = ctx.deserialize_components(components) | |
2859 | ||
2860 | if isinstance(pd_obj, pd.DataFrame): | |
2861 | tm.assert_frame_equal(pd_obj, deserialized) | |
2862 | else: | |
2863 | tm.assert_series_equal(pd_obj, deserialized) | |
2864 | ||
2865 | ||
2866 | @pytest.mark.skipif( | |
2867 | Version('1.16.0') <= Version(np.__version__) < Version('1.16.1'), | |
2868 | reason='Until numpy/numpy#12745 is resolved') | |
2869 | def test_serialize_deserialize_pandas(): | |
2870 | # ARROW-1784, serialize and deserialize DataFrame by decomposing | |
2871 | # BlockManager | |
2872 | df = _fully_loaded_dataframe_example() | |
2873 | _check_serialize_components_roundtrip(df) | |
2874 | ||
2875 | ||
2876 | def test_serialize_deserialize_empty_pandas(): | |
2877 | # ARROW-7996, serialize and deserialize empty pandas objects | |
2878 | df = pd.DataFrame({'col1': [], 'col2': [], 'col3': []}) | |
2879 | _check_serialize_components_roundtrip(df) | |
2880 | ||
2881 | series = pd.Series([], dtype=np.float32, name='col') | |
2882 | _check_serialize_components_roundtrip(series) | |
2883 | ||
2884 | ||
2885 | def _pytime_from_micros(val): | |
2886 | microseconds = val % 1000000 | |
2887 | val //= 1000000 | |
2888 | seconds = val % 60 | |
2889 | val //= 60 | |
2890 | minutes = val % 60 | |
2891 | hours = val // 60 | |
2892 | return time(hours, minutes, seconds, microseconds) | |
2893 | ||
2894 | ||
2895 | def _pytime_to_micros(pytime): | |
2896 | return (pytime.hour * 3600000000 + | |
2897 | pytime.minute * 60000000 + | |
2898 | pytime.second * 1000000 + | |
2899 | pytime.microsecond) | |
2900 | ||
2901 | ||
2902 | def test_convert_unsupported_type_error_message(): | |
2903 | # ARROW-1454 | |
2904 | ||
2905 | # custom python objects | |
2906 | class A: | |
2907 | pass | |
2908 | ||
2909 | df = pd.DataFrame({'a': [A(), A()]}) | |
2910 | ||
2911 | msg = 'Conversion failed for column a with type object' | |
2912 | with pytest.raises(ValueError, match=msg): | |
2913 | pa.Table.from_pandas(df) | |
2914 | ||
2915 | # period unsupported for pandas <= 0.25 | |
2916 | if Version(pd.__version__) <= Version('0.25'): | |
2917 | df = pd.DataFrame({ | |
2918 | 'a': pd.period_range('2000-01-01', periods=20), | |
2919 | }) | |
2920 | ||
2921 | msg = 'Conversion failed for column a with type (period|object)' | |
2922 | with pytest.raises((TypeError, ValueError), match=msg): | |
2923 | pa.Table.from_pandas(df) | |
2924 | ||
2925 | ||
2926 | # ---------------------------------------------------------------------- | |
2927 | # Hypothesis tests | |
2928 | ||
2929 | ||
2930 | @h.given(past.arrays(past.pandas_compatible_types)) | |
2931 | def test_array_to_pandas_roundtrip(arr): | |
2932 | s = arr.to_pandas() | |
2933 | restored = pa.array(s, type=arr.type, from_pandas=True) | |
2934 | assert restored.equals(arr) | |
2935 | ||
2936 | ||
2937 | # ---------------------------------------------------------------------- | |
2938 | # Test object deduplication in to_pandas | |
2939 | ||
2940 | ||
2941 | def _generate_dedup_example(nunique, repeats): | |
2942 | unique_values = [rands(10) for i in range(nunique)] | |
2943 | return unique_values * repeats | |
2944 | ||
2945 | ||
2946 | def _assert_nunique(obj, expected): | |
2947 | assert len({id(x) for x in obj}) == expected | |
2948 | ||
2949 | ||
2950 | def test_to_pandas_deduplicate_strings_array_types(): | |
2951 | nunique = 100 | |
2952 | repeats = 10 | |
2953 | values = _generate_dedup_example(nunique, repeats) | |
2954 | ||
2955 | for arr in [pa.array(values, type=pa.binary()), | |
2956 | pa.array(values, type=pa.utf8()), | |
2957 | pa.chunked_array([values, values])]: | |
2958 | _assert_nunique(arr.to_pandas(), nunique) | |
2959 | _assert_nunique(arr.to_pandas(deduplicate_objects=False), len(arr)) | |
2960 | ||
2961 | ||
2962 | def test_to_pandas_deduplicate_strings_table_types(): | |
2963 | nunique = 100 | |
2964 | repeats = 10 | |
2965 | values = _generate_dedup_example(nunique, repeats) | |
2966 | ||
2967 | arr = pa.array(values) | |
2968 | rb = pa.RecordBatch.from_arrays([arr], ['foo']) | |
2969 | tbl = pa.Table.from_batches([rb]) | |
2970 | ||
2971 | for obj in [rb, tbl]: | |
2972 | _assert_nunique(obj.to_pandas()['foo'], nunique) | |
2973 | _assert_nunique(obj.to_pandas(deduplicate_objects=False)['foo'], | |
2974 | len(obj)) | |
2975 | ||
2976 | ||
2977 | def test_to_pandas_deduplicate_integers_as_objects(): | |
2978 | nunique = 100 | |
2979 | repeats = 10 | |
2980 | ||
2981 | # Python automatically interns smaller integers | |
2982 | unique_values = list(np.random.randint(10000000, 1000000000, size=nunique)) | |
2983 | unique_values[nunique // 2] = None | |
2984 | ||
2985 | arr = pa.array(unique_values * repeats) | |
2986 | ||
2987 | _assert_nunique(arr.to_pandas(integer_object_nulls=True), nunique) | |
2988 | _assert_nunique(arr.to_pandas(integer_object_nulls=True, | |
2989 | deduplicate_objects=False), | |
2990 | # Account for None | |
2991 | (nunique - 1) * repeats + 1) | |
2992 | ||
2993 | ||
2994 | def test_to_pandas_deduplicate_date_time(): | |
2995 | nunique = 100 | |
2996 | repeats = 10 | |
2997 | ||
2998 | unique_values = list(range(nunique)) | |
2999 | ||
3000 | cases = [ | |
3001 | # raw type, array type, to_pandas options | |
3002 | ('int32', 'date32', {'date_as_object': True}), | |
3003 | ('int64', 'date64', {'date_as_object': True}), | |
3004 | ('int32', 'time32[ms]', {}), | |
3005 | ('int64', 'time64[us]', {}) | |
3006 | ] | |
3007 | ||
3008 | for raw_type, array_type, pandas_options in cases: | |
3009 | raw_arr = pa.array(unique_values * repeats, type=raw_type) | |
3010 | casted_arr = raw_arr.cast(array_type) | |
3011 | ||
3012 | _assert_nunique(casted_arr.to_pandas(**pandas_options), | |
3013 | nunique) | |
3014 | _assert_nunique(casted_arr.to_pandas(deduplicate_objects=False, | |
3015 | **pandas_options), | |
3016 | len(casted_arr)) | |
3017 | ||
3018 | ||
3019 | # --------------------------------------------------------------------- | |
3020 | ||
3021 | def test_table_from_pandas_checks_field_nullability(): | |
3022 | # ARROW-2136 | |
3023 | df = pd.DataFrame({'a': [1.2, 2.1, 3.1], | |
3024 | 'b': [np.nan, 'string', 'foo']}) | |
3025 | schema = pa.schema([pa.field('a', pa.float64(), nullable=False), | |
3026 | pa.field('b', pa.utf8(), nullable=False)]) | |
3027 | ||
3028 | with pytest.raises(ValueError): | |
3029 | pa.Table.from_pandas(df, schema=schema) | |
3030 | ||
3031 | ||
3032 | def test_table_from_pandas_keeps_column_order_of_dataframe(): | |
3033 | df1 = pd.DataFrame(OrderedDict([ | |
3034 | ('partition', [0, 0, 1, 1]), | |
3035 | ('arrays', [[0, 1, 2], [3, 4], None, None]), | |
3036 | ('floats', [None, None, 1.1, 3.3]) | |
3037 | ])) | |
3038 | df2 = df1[['floats', 'partition', 'arrays']] | |
3039 | ||
3040 | schema1 = pa.schema([ | |
3041 | ('partition', pa.int64()), | |
3042 | ('arrays', pa.list_(pa.int64())), | |
3043 | ('floats', pa.float64()), | |
3044 | ]) | |
3045 | schema2 = pa.schema([ | |
3046 | ('floats', pa.float64()), | |
3047 | ('partition', pa.int64()), | |
3048 | ('arrays', pa.list_(pa.int64())) | |
3049 | ]) | |
3050 | ||
3051 | table1 = pa.Table.from_pandas(df1, preserve_index=False) | |
3052 | table2 = pa.Table.from_pandas(df2, preserve_index=False) | |
3053 | ||
3054 | assert table1.schema.equals(schema1) | |
3055 | assert table2.schema.equals(schema2) | |
3056 | ||
3057 | ||
3058 | def test_table_from_pandas_keeps_column_order_of_schema(): | |
3059 | # ARROW-3766 | |
3060 | df = pd.DataFrame(OrderedDict([ | |
3061 | ('partition', [0, 0, 1, 1]), | |
3062 | ('arrays', [[0, 1, 2], [3, 4], None, None]), | |
3063 | ('floats', [None, None, 1.1, 3.3]) | |
3064 | ])) | |
3065 | ||
3066 | schema = pa.schema([ | |
3067 | ('floats', pa.float64()), | |
3068 | ('arrays', pa.list_(pa.int32())), | |
3069 | ('partition', pa.int32()) | |
3070 | ]) | |
3071 | ||
3072 | df1 = df[df.partition == 0] | |
3073 | df2 = df[df.partition == 1][['floats', 'partition', 'arrays']] | |
3074 | ||
3075 | table1 = pa.Table.from_pandas(df1, schema=schema, preserve_index=False) | |
3076 | table2 = pa.Table.from_pandas(df2, schema=schema, preserve_index=False) | |
3077 | ||
3078 | assert table1.schema.equals(schema) | |
3079 | assert table1.schema.equals(table2.schema) | |
3080 | ||
3081 | ||
3082 | def test_table_from_pandas_columns_argument_only_does_filtering(): | |
3083 | df = pd.DataFrame(OrderedDict([ | |
3084 | ('partition', [0, 0, 1, 1]), | |
3085 | ('arrays', [[0, 1, 2], [3, 4], None, None]), | |
3086 | ('floats', [None, None, 1.1, 3.3]) | |
3087 | ])) | |
3088 | ||
3089 | columns1 = ['arrays', 'floats', 'partition'] | |
3090 | schema1 = pa.schema([ | |
3091 | ('arrays', pa.list_(pa.int64())), | |
3092 | ('floats', pa.float64()), | |
3093 | ('partition', pa.int64()) | |
3094 | ]) | |
3095 | ||
3096 | columns2 = ['floats', 'partition'] | |
3097 | schema2 = pa.schema([ | |
3098 | ('floats', pa.float64()), | |
3099 | ('partition', pa.int64()) | |
3100 | ]) | |
3101 | ||
3102 | table1 = pa.Table.from_pandas(df, columns=columns1, preserve_index=False) | |
3103 | table2 = pa.Table.from_pandas(df, columns=columns2, preserve_index=False) | |
3104 | ||
3105 | assert table1.schema.equals(schema1) | |
3106 | assert table2.schema.equals(schema2) | |
3107 | ||
3108 | ||
3109 | def test_table_from_pandas_columns_and_schema_are_mutually_exclusive(): | |
3110 | df = pd.DataFrame(OrderedDict([ | |
3111 | ('partition', [0, 0, 1, 1]), | |
3112 | ('arrays', [[0, 1, 2], [3, 4], None, None]), | |
3113 | ('floats', [None, None, 1.1, 3.3]) | |
3114 | ])) | |
3115 | schema = pa.schema([ | |
3116 | ('partition', pa.int32()), | |
3117 | ('arrays', pa.list_(pa.int32())), | |
3118 | ('floats', pa.float64()), | |
3119 | ]) | |
3120 | columns = ['arrays', 'floats'] | |
3121 | ||
3122 | with pytest.raises(ValueError): | |
3123 | pa.Table.from_pandas(df, schema=schema, columns=columns) | |
3124 | ||
3125 | ||
3126 | def test_table_from_pandas_keeps_schema_nullability(): | |
3127 | # ARROW-5169 | |
3128 | df = pd.DataFrame({'a': [1, 2, 3, 4]}) | |
3129 | ||
3130 | schema = pa.schema([ | |
3131 | pa.field('a', pa.int64(), nullable=False), | |
3132 | ]) | |
3133 | ||
3134 | table = pa.Table.from_pandas(df) | |
3135 | assert table.schema.field('a').nullable is True | |
3136 | table = pa.Table.from_pandas(df, schema=schema) | |
3137 | assert table.schema.field('a').nullable is False | |
3138 | ||
3139 | ||
3140 | def test_table_from_pandas_schema_index_columns(): | |
3141 | # ARROW-5220 | |
3142 | df = pd.DataFrame({'a': [1, 2, 3], 'b': [0.1, 0.2, 0.3]}) | |
3143 | ||
3144 | schema = pa.schema([ | |
3145 | ('a', pa.int64()), | |
3146 | ('b', pa.float64()), | |
3147 | ('index', pa.int32()), | |
3148 | ]) | |
3149 | ||
3150 | # schema includes index with name not in dataframe | |
3151 | with pytest.raises(KeyError, match="name 'index' present in the"): | |
3152 | pa.Table.from_pandas(df, schema=schema) | |
3153 | ||
3154 | df.index.name = 'index' | |
3155 | ||
3156 | # schema includes correct index name -> roundtrip works | |
3157 | _check_pandas_roundtrip(df, schema=schema, preserve_index=True, | |
3158 | expected_schema=schema) | |
3159 | ||
3160 | # schema includes correct index name but preserve_index=False | |
3161 | with pytest.raises(ValueError, match="'preserve_index=False' was"): | |
3162 | pa.Table.from_pandas(df, schema=schema, preserve_index=False) | |
3163 | ||
3164 | # in case of preserve_index=None -> RangeIndex serialized as metadata | |
3165 | # clashes with the index in the schema | |
3166 | with pytest.raises(ValueError, match="name 'index' is present in the " | |
3167 | "schema, but it is a RangeIndex"): | |
3168 | pa.Table.from_pandas(df, schema=schema, preserve_index=None) | |
3169 | ||
3170 | df.index = pd.Index([0, 1, 2], name='index') | |
3171 | ||
3172 | # for non-RangeIndex, both preserve_index=None and True work | |
3173 | _check_pandas_roundtrip(df, schema=schema, preserve_index=None, | |
3174 | expected_schema=schema) | |
3175 | _check_pandas_roundtrip(df, schema=schema, preserve_index=True, | |
3176 | expected_schema=schema) | |
3177 | ||
3178 | # schema has different order (index column not at the end) | |
3179 | schema = pa.schema([ | |
3180 | ('index', pa.int32()), | |
3181 | ('a', pa.int64()), | |
3182 | ('b', pa.float64()), | |
3183 | ]) | |
3184 | _check_pandas_roundtrip(df, schema=schema, preserve_index=None, | |
3185 | expected_schema=schema) | |
3186 | _check_pandas_roundtrip(df, schema=schema, preserve_index=True, | |
3187 | expected_schema=schema) | |
3188 | ||
3189 | # schema does not include the index -> index is not included as column | |
3190 | # even though preserve_index=True/None | |
3191 | schema = pa.schema([ | |
3192 | ('a', pa.int64()), | |
3193 | ('b', pa.float64()), | |
3194 | ]) | |
3195 | expected = df.copy() | |
3196 | expected = expected.reset_index(drop=True) | |
3197 | _check_pandas_roundtrip(df, schema=schema, preserve_index=None, | |
3198 | expected_schema=schema, expected=expected) | |
3199 | _check_pandas_roundtrip(df, schema=schema, preserve_index=True, | |
3200 | expected_schema=schema, expected=expected) | |
3201 | ||
3202 | # dataframe with a MultiIndex | |
3203 | df.index = pd.MultiIndex.from_tuples([('a', 1), ('a', 2), ('b', 1)], | |
3204 | names=['level1', 'level2']) | |
3205 | schema = pa.schema([ | |
3206 | ('level1', pa.string()), | |
3207 | ('level2', pa.int64()), | |
3208 | ('a', pa.int64()), | |
3209 | ('b', pa.float64()), | |
3210 | ]) | |
3211 | _check_pandas_roundtrip(df, schema=schema, preserve_index=True, | |
3212 | expected_schema=schema) | |
3213 | _check_pandas_roundtrip(df, schema=schema, preserve_index=None, | |
3214 | expected_schema=schema) | |
3215 | ||
3216 | # only one of the levels of the MultiIndex is included | |
3217 | schema = pa.schema([ | |
3218 | ('level2', pa.int64()), | |
3219 | ('a', pa.int64()), | |
3220 | ('b', pa.float64()), | |
3221 | ]) | |
3222 | expected = df.copy() | |
3223 | expected = expected.reset_index('level1', drop=True) | |
3224 | _check_pandas_roundtrip(df, schema=schema, preserve_index=True, | |
3225 | expected_schema=schema, expected=expected) | |
3226 | _check_pandas_roundtrip(df, schema=schema, preserve_index=None, | |
3227 | expected_schema=schema, expected=expected) | |
3228 | ||
3229 | ||
3230 | def test_table_from_pandas_schema_index_columns__unnamed_index(): | |
3231 | # ARROW-6999 - unnamed indices in specified schema | |
3232 | df = pd.DataFrame({'a': [1, 2, 3], 'b': [0.1, 0.2, 0.3]}) | |
3233 | ||
3234 | expected_schema = pa.schema([ | |
3235 | ('a', pa.int64()), | |
3236 | ('b', pa.float64()), | |
3237 | ('__index_level_0__', pa.int64()), | |
3238 | ]) | |
3239 | ||
3240 | schema = pa.Schema.from_pandas(df, preserve_index=True) | |
3241 | table = pa.Table.from_pandas(df, preserve_index=True, schema=schema) | |
3242 | assert table.schema.remove_metadata().equals(expected_schema) | |
3243 | ||
3244 | # non-RangeIndex (preserved by default) | |
3245 | df = pd.DataFrame({'a': [1, 2, 3], 'b': [0.1, 0.2, 0.3]}, index=[0, 1, 2]) | |
3246 | schema = pa.Schema.from_pandas(df) | |
3247 | table = pa.Table.from_pandas(df, schema=schema) | |
3248 | assert table.schema.remove_metadata().equals(expected_schema) | |
3249 | ||
3250 | ||
3251 | def test_table_from_pandas_schema_with_custom_metadata(): | |
3252 | # ARROW-7087 - metadata disappear from pandas | |
3253 | df = pd.DataFrame() | |
3254 | schema = pa.Schema.from_pandas(df).with_metadata({'meta': 'True'}) | |
3255 | table = pa.Table.from_pandas(df, schema=schema) | |
3256 | assert table.schema.metadata.get(b'meta') == b'True' | |
3257 | ||
3258 | ||
3259 | def test_table_from_pandas_schema_field_order_metadat(): | |
3260 | # ARROW-10532 | |
3261 | # ensure that a different field order in specified schema doesn't | |
3262 | # mangle metadata | |
3263 | df = pd.DataFrame({ | |
3264 | "datetime": pd.date_range("2020-01-01T00:00:00Z", freq="H", periods=2), | |
3265 | "float": np.random.randn(2) | |
3266 | }) | |
3267 | ||
3268 | schema = pa.schema([ | |
3269 | pa.field("float", pa.float32(), nullable=True), | |
3270 | pa.field("datetime", pa.timestamp("s", tz="UTC"), nullable=False) | |
3271 | ]) | |
3272 | ||
3273 | table = pa.Table.from_pandas(df, schema=schema) | |
3274 | assert table.schema.equals(schema) | |
3275 | metadata_float = table.schema.pandas_metadata["columns"][0] | |
3276 | assert metadata_float["name"] == "float" | |
3277 | assert metadata_float["metadata"] is None | |
3278 | metadata_datetime = table.schema.pandas_metadata["columns"][1] | |
3279 | assert metadata_datetime["name"] == "datetime" | |
3280 | assert metadata_datetime["metadata"] == {'timezone': 'UTC'} | |
3281 | ||
3282 | result = table.to_pandas() | |
3283 | expected = df[["float", "datetime"]].astype({"float": "float32"}) | |
3284 | tm.assert_frame_equal(result, expected) | |
3285 | ||
3286 | ||
3287 | # ---------------------------------------------------------------------- | |
3288 | # RecordBatch, Table | |
3289 | ||
3290 | ||
3291 | def test_recordbatch_from_to_pandas(): | |
3292 | data = pd.DataFrame({ | |
3293 | 'c1': np.array([1, 2, 3, 4, 5], dtype='int64'), | |
3294 | 'c2': np.array([1, 2, 3, 4, 5], dtype='uint32'), | |
3295 | 'c3': np.random.randn(5), | |
3296 | 'c4': ['foo', 'bar', None, 'baz', 'qux'], | |
3297 | 'c5': [False, True, False, True, False] | |
3298 | }) | |
3299 | ||
3300 | batch = pa.RecordBatch.from_pandas(data) | |
3301 | result = batch.to_pandas() | |
3302 | tm.assert_frame_equal(data, result) | |
3303 | ||
3304 | ||
3305 | def test_recordbatchlist_to_pandas(): | |
3306 | data1 = pd.DataFrame({ | |
3307 | 'c1': np.array([1, 1, 2], dtype='uint32'), | |
3308 | 'c2': np.array([1.0, 2.0, 3.0], dtype='float64'), | |
3309 | 'c3': [True, None, False], | |
3310 | 'c4': ['foo', 'bar', None] | |
3311 | }) | |
3312 | ||
3313 | data2 = pd.DataFrame({ | |
3314 | 'c1': np.array([3, 5], dtype='uint32'), | |
3315 | 'c2': np.array([4.0, 5.0], dtype='float64'), | |
3316 | 'c3': [True, True], | |
3317 | 'c4': ['baz', 'qux'] | |
3318 | }) | |
3319 | ||
3320 | batch1 = pa.RecordBatch.from_pandas(data1) | |
3321 | batch2 = pa.RecordBatch.from_pandas(data2) | |
3322 | ||
3323 | table = pa.Table.from_batches([batch1, batch2]) | |
3324 | result = table.to_pandas() | |
3325 | data = pd.concat([data1, data2]).reset_index(drop=True) | |
3326 | tm.assert_frame_equal(data, result) | |
3327 | ||
3328 | ||
3329 | def test_recordbatch_table_pass_name_to_pandas(): | |
3330 | rb = pa.record_batch([pa.array([1, 2, 3, 4])], names=['a0']) | |
3331 | t = pa.table([pa.array([1, 2, 3, 4])], names=['a0']) | |
3332 | assert rb[0].to_pandas().name == 'a0' | |
3333 | assert t[0].to_pandas().name == 'a0' | |
3334 | ||
3335 | ||
3336 | # ---------------------------------------------------------------------- | |
3337 | # Metadata serialization | |
3338 | ||
3339 | ||
3340 | @pytest.mark.parametrize( | |
3341 | ('type', 'expected'), | |
3342 | [ | |
3343 | (pa.null(), 'empty'), | |
3344 | (pa.bool_(), 'bool'), | |
3345 | (pa.int8(), 'int8'), | |
3346 | (pa.int16(), 'int16'), | |
3347 | (pa.int32(), 'int32'), | |
3348 | (pa.int64(), 'int64'), | |
3349 | (pa.uint8(), 'uint8'), | |
3350 | (pa.uint16(), 'uint16'), | |
3351 | (pa.uint32(), 'uint32'), | |
3352 | (pa.uint64(), 'uint64'), | |
3353 | (pa.float16(), 'float16'), | |
3354 | (pa.float32(), 'float32'), | |
3355 | (pa.float64(), 'float64'), | |
3356 | (pa.date32(), 'date'), | |
3357 | (pa.date64(), 'date'), | |
3358 | (pa.binary(), 'bytes'), | |
3359 | (pa.binary(length=4), 'bytes'), | |
3360 | (pa.string(), 'unicode'), | |
3361 | (pa.list_(pa.list_(pa.int16())), 'list[list[int16]]'), | |
3362 | (pa.decimal128(18, 3), 'decimal'), | |
3363 | (pa.timestamp('ms'), 'datetime'), | |
3364 | (pa.timestamp('us', 'UTC'), 'datetimetz'), | |
3365 | (pa.time32('s'), 'time'), | |
3366 | (pa.time64('us'), 'time') | |
3367 | ] | |
3368 | ) | |
3369 | def test_logical_type(type, expected): | |
3370 | assert get_logical_type(type) == expected | |
3371 | ||
3372 | ||
3373 | # ---------------------------------------------------------------------- | |
3374 | # to_pandas uses MemoryPool | |
3375 | ||
3376 | def test_array_uses_memory_pool(): | |
3377 | # ARROW-6570 | |
3378 | N = 10000 | |
3379 | arr = pa.array(np.arange(N, dtype=np.int64), | |
3380 | mask=np.random.randint(0, 2, size=N).astype(np.bool_)) | |
3381 | ||
3382 | # In the case the gc is caught loafing | |
3383 | gc.collect() | |
3384 | ||
3385 | prior_allocation = pa.total_allocated_bytes() | |
3386 | ||
3387 | x = arr.to_pandas() | |
3388 | assert pa.total_allocated_bytes() == (prior_allocation + N * 8) | |
3389 | x = None # noqa | |
3390 | gc.collect() | |
3391 | ||
3392 | assert pa.total_allocated_bytes() == prior_allocation | |
3393 | ||
3394 | # zero copy does not allocate memory | |
3395 | arr = pa.array(np.arange(N, dtype=np.int64)) | |
3396 | ||
3397 | prior_allocation = pa.total_allocated_bytes() | |
3398 | x = arr.to_pandas() # noqa | |
3399 | assert pa.total_allocated_bytes() == prior_allocation | |
3400 | ||
3401 | ||
3402 | def test_singleton_blocks_zero_copy(): | |
3403 | # Part of ARROW-3789 | |
3404 | t = pa.table([pa.array(np.arange(1000, dtype=np.int64))], ['f0']) | |
3405 | ||
3406 | # Zero copy if split_blocks=True | |
3407 | _check_to_pandas_memory_unchanged(t, split_blocks=True) | |
3408 | ||
3409 | prior_allocation = pa.total_allocated_bytes() | |
3410 | result = t.to_pandas() | |
3411 | assert result['f0'].values.flags.writeable | |
3412 | assert pa.total_allocated_bytes() > prior_allocation | |
3413 | ||
3414 | ||
3415 | def _check_to_pandas_memory_unchanged(obj, **kwargs): | |
3416 | prior_allocation = pa.total_allocated_bytes() | |
3417 | x = obj.to_pandas(**kwargs) # noqa | |
3418 | ||
3419 | # Memory allocation unchanged -- either zero copy or self-destructing | |
3420 | assert pa.total_allocated_bytes() == prior_allocation | |
3421 | ||
3422 | ||
3423 | def test_to_pandas_split_blocks(): | |
3424 | # ARROW-3789 | |
3425 | t = pa.table([ | |
3426 | pa.array([1, 2, 3, 4, 5], type='i1'), | |
3427 | pa.array([1, 2, 3, 4, 5], type='i4'), | |
3428 | pa.array([1, 2, 3, 4, 5], type='i8'), | |
3429 | pa.array([1, 2, 3, 4, 5], type='f4'), | |
3430 | pa.array([1, 2, 3, 4, 5], type='f8'), | |
3431 | pa.array([1, 2, 3, 4, 5], type='f8'), | |
3432 | pa.array([1, 2, 3, 4, 5], type='f8'), | |
3433 | pa.array([1, 2, 3, 4, 5], type='f8'), | |
3434 | ], ['f{}'.format(i) for i in range(8)]) | |
3435 | ||
3436 | _check_blocks_created(t, 8) | |
3437 | _check_to_pandas_memory_unchanged(t, split_blocks=True) | |
3438 | ||
3439 | ||
3440 | def _check_blocks_created(t, number): | |
3441 | x = t.to_pandas(split_blocks=True) | |
3442 | assert len(x._data.blocks) == number | |
3443 | ||
3444 | ||
3445 | def test_to_pandas_self_destruct(): | |
3446 | K = 50 | |
3447 | ||
3448 | def _make_table(): | |
3449 | return pa.table([ | |
3450 | # Slice to force a copy | |
3451 | pa.array(np.random.randn(10000)[::2]) | |
3452 | for i in range(K) | |
3453 | ], ['f{}'.format(i) for i in range(K)]) | |
3454 | ||
3455 | t = _make_table() | |
3456 | _check_to_pandas_memory_unchanged(t, split_blocks=True, self_destruct=True) | |
3457 | ||
3458 | # Check non-split-block behavior | |
3459 | t = _make_table() | |
3460 | _check_to_pandas_memory_unchanged(t, self_destruct=True) | |
3461 | ||
3462 | ||
3463 | def test_table_uses_memory_pool(): | |
3464 | N = 10000 | |
3465 | arr = pa.array(np.arange(N, dtype=np.int64)) | |
3466 | t = pa.table([arr, arr, arr], ['f0', 'f1', 'f2']) | |
3467 | ||
3468 | prior_allocation = pa.total_allocated_bytes() | |
3469 | x = t.to_pandas() | |
3470 | ||
3471 | assert pa.total_allocated_bytes() == (prior_allocation + 3 * N * 8) | |
3472 | ||
3473 | # Check successful garbage collection | |
3474 | x = None # noqa | |
3475 | gc.collect() | |
3476 | assert pa.total_allocated_bytes() == prior_allocation | |
3477 | ||
3478 | ||
3479 | def test_object_leak_in_numpy_array(): | |
3480 | # ARROW-6876 | |
3481 | arr = pa.array([{'a': 1}]) | |
3482 | np_arr = arr.to_pandas() | |
3483 | assert np_arr.dtype == np.dtype('object') | |
3484 | obj = np_arr[0] | |
3485 | refcount = sys.getrefcount(obj) | |
3486 | assert sys.getrefcount(obj) == refcount | |
3487 | del np_arr | |
3488 | assert sys.getrefcount(obj) == refcount - 1 | |
3489 | ||
3490 | ||
3491 | def test_object_leak_in_dataframe(): | |
3492 | # ARROW-6876 | |
3493 | arr = pa.array([{'a': 1}]) | |
3494 | table = pa.table([arr], ['f0']) | |
3495 | col = table.to_pandas()['f0'] | |
3496 | assert col.dtype == np.dtype('object') | |
3497 | obj = col[0] | |
3498 | refcount = sys.getrefcount(obj) | |
3499 | assert sys.getrefcount(obj) == refcount | |
3500 | del col | |
3501 | assert sys.getrefcount(obj) == refcount - 1 | |
3502 | ||
3503 | ||
3504 | # ---------------------------------------------------------------------- | |
3505 | # Some nested array tests array tests | |
3506 | ||
3507 | ||
3508 | def test_array_from_py_float32(): | |
3509 | data = [[1.2, 3.4], [9.0, 42.0]] | |
3510 | ||
3511 | t = pa.float32() | |
3512 | ||
3513 | arr1 = pa.array(data[0], type=t) | |
3514 | arr2 = pa.array(data, type=pa.list_(t)) | |
3515 | ||
3516 | expected1 = np.array(data[0], dtype=np.float32) | |
3517 | expected2 = pd.Series([np.array(data[0], dtype=np.float32), | |
3518 | np.array(data[1], dtype=np.float32)]) | |
3519 | ||
3520 | assert arr1.type == t | |
3521 | assert arr1.equals(pa.array(expected1)) | |
3522 | assert arr2.equals(pa.array(expected2)) | |
3523 | ||
3524 | ||
3525 | # ---------------------------------------------------------------------- | |
3526 | # Timestamp tests | |
3527 | ||
3528 | ||
3529 | def test_cast_timestamp_unit(): | |
3530 | # ARROW-1680 | |
3531 | val = datetime.now() | |
3532 | s = pd.Series([val]) | |
3533 | s_nyc = s.dt.tz_localize('tzlocal()').dt.tz_convert('America/New_York') | |
3534 | ||
3535 | us_with_tz = pa.timestamp('us', tz='America/New_York') | |
3536 | ||
3537 | arr = pa.Array.from_pandas(s_nyc, type=us_with_tz) | |
3538 | ||
3539 | # ARROW-1906 | |
3540 | assert arr.type == us_with_tz | |
3541 | ||
3542 | arr2 = pa.Array.from_pandas(s, type=pa.timestamp('us')) | |
3543 | ||
3544 | assert arr[0].as_py() == s_nyc[0].to_pydatetime() | |
3545 | assert arr2[0].as_py() == s[0].to_pydatetime() | |
3546 | ||
3547 | # Disallow truncation | |
3548 | arr = pa.array([123123], type='int64').cast(pa.timestamp('ms')) | |
3549 | expected = pa.array([123], type='int64').cast(pa.timestamp('s')) | |
3550 | ||
3551 | # sanity check that the cast worked right | |
3552 | assert arr.type == pa.timestamp('ms') | |
3553 | ||
3554 | target = pa.timestamp('s') | |
3555 | with pytest.raises(ValueError): | |
3556 | arr.cast(target) | |
3557 | ||
3558 | result = arr.cast(target, safe=False) | |
3559 | assert result.equals(expected) | |
3560 | ||
3561 | # ARROW-1949 | |
3562 | series = pd.Series([pd.Timestamp(1), pd.Timestamp(10), pd.Timestamp(1000)]) | |
3563 | expected = pa.array([0, 0, 1], type=pa.timestamp('us')) | |
3564 | ||
3565 | with pytest.raises(ValueError): | |
3566 | pa.array(series, type=pa.timestamp('us')) | |
3567 | ||
3568 | with pytest.raises(ValueError): | |
3569 | pa.Array.from_pandas(series, type=pa.timestamp('us')) | |
3570 | ||
3571 | result = pa.Array.from_pandas(series, type=pa.timestamp('us'), safe=False) | |
3572 | assert result.equals(expected) | |
3573 | ||
3574 | result = pa.array(series, type=pa.timestamp('us'), safe=False) | |
3575 | assert result.equals(expected) | |
3576 | ||
3577 | ||
3578 | def test_nested_with_timestamp_tz_round_trip(): | |
3579 | ts = pd.Timestamp.now() | |
3580 | ts_dt = ts.to_pydatetime() | |
3581 | arr = pa.array([ts_dt], type=pa.timestamp('us', tz='America/New_York')) | |
3582 | struct = pa.StructArray.from_arrays([arr, arr], ['start', 'stop']) | |
3583 | ||
3584 | result = struct.to_pandas() | |
3585 | restored = pa.array(result) | |
3586 | assert restored.equals(struct) | |
3587 | ||
3588 | ||
3589 | def test_nested_with_timestamp_tz(): | |
3590 | # ARROW-7723 | |
3591 | ts = pd.Timestamp.now() | |
3592 | ts_dt = ts.to_pydatetime() | |
3593 | ||
3594 | # XXX: Ensure that this data does not get promoted to nanoseconds (and thus | |
3595 | # integers) to preserve behavior in 0.15.1 | |
3596 | for unit in ['s', 'ms', 'us']: | |
3597 | if unit in ['s', 'ms']: | |
3598 | # This is used for verifying timezone conversion to micros are not | |
3599 | # important | |
3600 | def truncate(x): return x.replace(microsecond=0) | |
3601 | else: | |
3602 | def truncate(x): return x | |
3603 | arr = pa.array([ts], type=pa.timestamp(unit)) | |
3604 | arr2 = pa.array([ts], type=pa.timestamp(unit, tz='America/New_York')) | |
3605 | ||
3606 | arr3 = pa.StructArray.from_arrays([arr, arr], ['start', 'stop']) | |
3607 | arr4 = pa.StructArray.from_arrays([arr2, arr2], ['start', 'stop']) | |
3608 | ||
3609 | result = arr3.to_pandas() | |
3610 | assert isinstance(result[0]['start'], datetime) | |
3611 | assert result[0]['start'].tzinfo is None | |
3612 | assert isinstance(result[0]['stop'], datetime) | |
3613 | assert result[0]['stop'].tzinfo is None | |
3614 | ||
3615 | result = arr4.to_pandas() | |
3616 | assert isinstance(result[0]['start'], datetime) | |
3617 | assert result[0]['start'].tzinfo is not None | |
3618 | utc_dt = result[0]['start'].astimezone(timezone.utc) | |
3619 | assert truncate(utc_dt).replace(tzinfo=None) == truncate(ts_dt) | |
3620 | assert isinstance(result[0]['stop'], datetime) | |
3621 | assert result[0]['stop'].tzinfo is not None | |
3622 | ||
3623 | # same conversion for table | |
3624 | result = pa.table({'a': arr3}).to_pandas() | |
3625 | assert isinstance(result['a'][0]['start'], datetime) | |
3626 | assert result['a'][0]['start'].tzinfo is None | |
3627 | assert isinstance(result['a'][0]['stop'], datetime) | |
3628 | assert result['a'][0]['stop'].tzinfo is None | |
3629 | ||
3630 | result = pa.table({'a': arr4}).to_pandas() | |
3631 | assert isinstance(result['a'][0]['start'], datetime) | |
3632 | assert result['a'][0]['start'].tzinfo is not None | |
3633 | assert isinstance(result['a'][0]['stop'], datetime) | |
3634 | assert result['a'][0]['stop'].tzinfo is not None | |
3635 | ||
3636 | ||
3637 | # ---------------------------------------------------------------------- | |
3638 | # DictionaryArray tests | |
3639 | ||
3640 | ||
3641 | def test_dictionary_with_pandas(): | |
3642 | src_indices = np.repeat([0, 1, 2], 2) | |
3643 | dictionary = np.array(['foo', 'bar', 'baz'], dtype=object) | |
3644 | mask = np.array([False, False, True, False, False, False]) | |
3645 | ||
3646 | for index_type in ['uint8', 'int8', 'uint16', 'int16', 'uint32', 'int32', | |
3647 | 'uint64', 'int64']: | |
3648 | indices = src_indices.astype(index_type) | |
3649 | d1 = pa.DictionaryArray.from_arrays(indices, dictionary) | |
3650 | d2 = pa.DictionaryArray.from_arrays(indices, dictionary, mask=mask) | |
3651 | ||
3652 | if index_type[0] == 'u': | |
3653 | # TODO: unsigned dictionary indices to pandas | |
3654 | with pytest.raises(TypeError): | |
3655 | d1.to_pandas() | |
3656 | continue | |
3657 | ||
3658 | pandas1 = d1.to_pandas() | |
3659 | ex_pandas1 = pd.Categorical.from_codes(indices, categories=dictionary) | |
3660 | ||
3661 | tm.assert_series_equal(pd.Series(pandas1), pd.Series(ex_pandas1)) | |
3662 | ||
3663 | pandas2 = d2.to_pandas() | |
3664 | assert pandas2.isnull().sum() == 1 | |
3665 | ||
3666 | # Unsigned integers converted to signed | |
3667 | signed_indices = indices | |
3668 | if index_type[0] == 'u': | |
3669 | signed_indices = indices.astype(index_type[1:]) | |
3670 | ex_pandas2 = pd.Categorical.from_codes(np.where(mask, -1, | |
3671 | signed_indices), | |
3672 | categories=dictionary) | |
3673 | ||
3674 | tm.assert_series_equal(pd.Series(pandas2), pd.Series(ex_pandas2)) | |
3675 | ||
3676 | ||
3677 | def random_strings(n, item_size, pct_null=0, dictionary=None): | |
3678 | if dictionary is not None: | |
3679 | result = dictionary[np.random.randint(0, len(dictionary), size=n)] | |
3680 | else: | |
3681 | result = np.array([random_ascii(item_size) for i in range(n)], | |
3682 | dtype=object) | |
3683 | ||
3684 | if pct_null > 0: | |
3685 | result[np.random.rand(n) < pct_null] = None | |
3686 | ||
3687 | return result | |
3688 | ||
3689 | ||
3690 | def test_variable_dictionary_to_pandas(): | |
3691 | np.random.seed(12345) | |
3692 | ||
3693 | d1 = pa.array(random_strings(100, 32), type='string') | |
3694 | d2 = pa.array(random_strings(100, 16), type='string') | |
3695 | d3 = pa.array(random_strings(10000, 10), type='string') | |
3696 | ||
3697 | a1 = pa.DictionaryArray.from_arrays( | |
3698 | np.random.randint(0, len(d1), size=1000, dtype='i4'), | |
3699 | d1 | |
3700 | ) | |
3701 | a2 = pa.DictionaryArray.from_arrays( | |
3702 | np.random.randint(0, len(d2), size=1000, dtype='i4'), | |
3703 | d2 | |
3704 | ) | |
3705 | ||
3706 | # With some nulls | |
3707 | a3 = pa.DictionaryArray.from_arrays( | |
3708 | np.random.randint(0, len(d3), size=1000, dtype='i4'), d3) | |
3709 | ||
3710 | i4 = pa.array( | |
3711 | np.random.randint(0, len(d3), size=1000, dtype='i4'), | |
3712 | mask=np.random.rand(1000) < 0.1 | |
3713 | ) | |
3714 | a4 = pa.DictionaryArray.from_arrays(i4, d3) | |
3715 | ||
3716 | expected_dict = pa.concat_arrays([d1, d2, d3]) | |
3717 | ||
3718 | a = pa.chunked_array([a1, a2, a3, a4]) | |
3719 | a_dense = pa.chunked_array([a1.cast('string'), | |
3720 | a2.cast('string'), | |
3721 | a3.cast('string'), | |
3722 | a4.cast('string')]) | |
3723 | ||
3724 | result = a.to_pandas() | |
3725 | result_dense = a_dense.to_pandas() | |
3726 | ||
3727 | assert (result.cat.categories == expected_dict.to_pandas()).all() | |
3728 | ||
3729 | expected_dense = result.astype('str') | |
3730 | expected_dense[result_dense.isnull()] = None | |
3731 | tm.assert_series_equal(result_dense, expected_dense) | |
3732 | ||
3733 | ||
3734 | def test_dictionary_encoded_nested_to_pandas(): | |
3735 | # ARROW-6899 | |
3736 | child = pa.array(['a', 'a', 'a', 'b', 'b']).dictionary_encode() | |
3737 | ||
3738 | arr = pa.ListArray.from_arrays([0, 3, 5], child) | |
3739 | ||
3740 | result = arr.to_pandas() | |
3741 | expected = pd.Series([np.array(['a', 'a', 'a'], dtype=object), | |
3742 | np.array(['b', 'b'], dtype=object)]) | |
3743 | ||
3744 | tm.assert_series_equal(result, expected) | |
3745 | ||
3746 | ||
3747 | def test_dictionary_from_pandas(): | |
3748 | cat = pd.Categorical(['a', 'b', 'a']) | |
3749 | expected_type = pa.dictionary(pa.int8(), pa.string()) | |
3750 | ||
3751 | result = pa.array(cat) | |
3752 | assert result.to_pylist() == ['a', 'b', 'a'] | |
3753 | assert result.type.equals(expected_type) | |
3754 | ||
3755 | # with missing values in categorical | |
3756 | cat = pd.Categorical(['a', 'b', None, 'a']) | |
3757 | ||
3758 | result = pa.array(cat) | |
3759 | assert result.to_pylist() == ['a', 'b', None, 'a'] | |
3760 | assert result.type.equals(expected_type) | |
3761 | ||
3762 | # with additional mask | |
3763 | result = pa.array(cat, mask=np.array([False, False, False, True])) | |
3764 | assert result.to_pylist() == ['a', 'b', None, None] | |
3765 | assert result.type.equals(expected_type) | |
3766 | ||
3767 | ||
3768 | def test_dictionary_from_pandas_specified_type(): | |
3769 | # ARROW-7168 - ensure specified type is always respected | |
3770 | ||
3771 | # the same as cat = pd.Categorical(['a', 'b']) but explicit about dtypes | |
3772 | cat = pd.Categorical.from_codes( | |
3773 | np.array([0, 1], dtype='int8'), np.array(['a', 'b'], dtype=object)) | |
3774 | ||
3775 | # different index type -> allow this | |
3776 | # (the type of the 'codes' in pandas is not part of the data type) | |
3777 | typ = pa.dictionary(index_type=pa.int16(), value_type=pa.string()) | |
3778 | result = pa.array(cat, type=typ) | |
3779 | assert result.type.equals(typ) | |
3780 | assert result.to_pylist() == ['a', 'b'] | |
3781 | ||
3782 | # mismatching values type -> raise error | |
3783 | typ = pa.dictionary(index_type=pa.int8(), value_type=pa.int64()) | |
3784 | with pytest.raises(pa.ArrowInvalid): | |
3785 | result = pa.array(cat, type=typ) | |
3786 | ||
3787 | # mismatching order -> raise error (for now a deprecation warning) | |
3788 | typ = pa.dictionary( | |
3789 | index_type=pa.int8(), value_type=pa.string(), ordered=True) | |
3790 | with pytest.warns(FutureWarning, match="The 'ordered' flag of the passed"): | |
3791 | result = pa.array(cat, type=typ) | |
3792 | assert result.to_pylist() == ['a', 'b'] | |
3793 | ||
3794 | # with mask | |
3795 | typ = pa.dictionary(index_type=pa.int16(), value_type=pa.string()) | |
3796 | result = pa.array(cat, type=typ, mask=np.array([False, True])) | |
3797 | assert result.type.equals(typ) | |
3798 | assert result.to_pylist() == ['a', None] | |
3799 | ||
3800 | # empty categorical -> be flexible in values type to allow | |
3801 | cat = pd.Categorical([]) | |
3802 | ||
3803 | typ = pa.dictionary(index_type=pa.int8(), value_type=pa.string()) | |
3804 | result = pa.array(cat, type=typ) | |
3805 | assert result.type.equals(typ) | |
3806 | assert result.to_pylist() == [] | |
3807 | typ = pa.dictionary(index_type=pa.int8(), value_type=pa.int64()) | |
3808 | result = pa.array(cat, type=typ) | |
3809 | assert result.type.equals(typ) | |
3810 | assert result.to_pylist() == [] | |
3811 | ||
3812 | # passing non-dictionary type | |
3813 | cat = pd.Categorical(['a', 'b']) | |
3814 | result = pa.array(cat, type=pa.string()) | |
3815 | expected = pa.array(['a', 'b'], type=pa.string()) | |
3816 | assert result.equals(expected) | |
3817 | assert result.to_pylist() == ['a', 'b'] | |
3818 | ||
3819 | ||
3820 | # ---------------------------------------------------------------------- | |
3821 | # Array protocol in pandas conversions tests | |
3822 | ||
3823 | ||
3824 | def test_array_protocol(): | |
3825 | if Version(pd.__version__) < Version('0.24.0'): | |
3826 | pytest.skip('IntegerArray only introduced in 0.24') | |
3827 | ||
3828 | df = pd.DataFrame({'a': pd.Series([1, 2, None], dtype='Int64')}) | |
3829 | ||
3830 | if Version(pd.__version__) < Version('0.26.0.dev'): | |
3831 | # with pandas<=0.25, trying to convert nullable integer errors | |
3832 | with pytest.raises(TypeError): | |
3833 | pa.table(df) | |
3834 | else: | |
3835 | # __arrow_array__ added to pandas IntegerArray in 0.26.0.dev | |
3836 | ||
3837 | # default conversion | |
3838 | result = pa.table(df) | |
3839 | expected = pa.array([1, 2, None], pa.int64()) | |
3840 | assert result[0].chunk(0).equals(expected) | |
3841 | ||
3842 | # with specifying schema | |
3843 | schema = pa.schema([('a', pa.float64())]) | |
3844 | result = pa.table(df, schema=schema) | |
3845 | expected2 = pa.array([1, 2, None], pa.float64()) | |
3846 | assert result[0].chunk(0).equals(expected2) | |
3847 | ||
3848 | # pass Series to pa.array | |
3849 | result = pa.array(df['a']) | |
3850 | assert result.equals(expected) | |
3851 | result = pa.array(df['a'], type=pa.float64()) | |
3852 | assert result.equals(expected2) | |
3853 | ||
3854 | # pass actual ExtensionArray to pa.array | |
3855 | result = pa.array(df['a'].values) | |
3856 | assert result.equals(expected) | |
3857 | result = pa.array(df['a'].values, type=pa.float64()) | |
3858 | assert result.equals(expected2) | |
3859 | ||
3860 | ||
3861 | class DummyExtensionType(pa.PyExtensionType): | |
3862 | ||
3863 | def __init__(self): | |
3864 | pa.PyExtensionType.__init__(self, pa.int64()) | |
3865 | ||
3866 | def __reduce__(self): | |
3867 | return DummyExtensionType, () | |
3868 | ||
3869 | ||
3870 | def PandasArray__arrow_array__(self, type=None): | |
3871 | # hardcode dummy return regardless of self - we only want to check that | |
3872 | # this method is correctly called | |
3873 | storage = pa.array([1, 2, 3], type=pa.int64()) | |
3874 | return pa.ExtensionArray.from_storage(DummyExtensionType(), storage) | |
3875 | ||
3876 | ||
3877 | def test_array_protocol_pandas_extension_types(monkeypatch): | |
3878 | # ARROW-7022 - ensure protocol works for Period / Interval extension dtypes | |
3879 | ||
3880 | if Version(pd.__version__) < Version('0.24.0'): | |
3881 | pytest.skip('Period/IntervalArray only introduced in 0.24') | |
3882 | ||
3883 | storage = pa.array([1, 2, 3], type=pa.int64()) | |
3884 | expected = pa.ExtensionArray.from_storage(DummyExtensionType(), storage) | |
3885 | ||
3886 | monkeypatch.setattr(pd.arrays.PeriodArray, "__arrow_array__", | |
3887 | PandasArray__arrow_array__, raising=False) | |
3888 | monkeypatch.setattr(pd.arrays.IntervalArray, "__arrow_array__", | |
3889 | PandasArray__arrow_array__, raising=False) | |
3890 | for arr in [pd.period_range("2012-01-01", periods=3, freq="D").array, | |
3891 | pd.interval_range(1, 4).array]: | |
3892 | result = pa.array(arr) | |
3893 | assert result.equals(expected) | |
3894 | result = pa.array(pd.Series(arr)) | |
3895 | assert result.equals(expected) | |
3896 | result = pa.array(pd.Index(arr)) | |
3897 | assert result.equals(expected) | |
3898 | result = pa.table(pd.DataFrame({'a': arr})).column('a').chunk(0) | |
3899 | assert result.equals(expected) | |
3900 | ||
3901 | ||
3902 | # ---------------------------------------------------------------------- | |
3903 | # Pandas ExtensionArray support | |
3904 | ||
3905 | ||
3906 | def _Int64Dtype__from_arrow__(self, array): | |
3907 | # for test only deal with single chunk for now | |
3908 | # TODO: do we require handling of chunked arrays in the protocol? | |
3909 | if isinstance(array, pa.Array): | |
3910 | arr = array | |
3911 | else: | |
3912 | # ChunkedArray - here only deal with a single chunk for the test | |
3913 | arr = array.chunk(0) | |
3914 | buflist = arr.buffers() | |
3915 | data = np.frombuffer(buflist[-1], dtype='int64')[ | |
3916 | arr.offset:arr.offset + len(arr)] | |
3917 | bitmask = buflist[0] | |
3918 | if bitmask is not None: | |
3919 | mask = pa.BooleanArray.from_buffers( | |
3920 | pa.bool_(), len(arr), [None, bitmask]) | |
3921 | mask = np.asarray(mask) | |
3922 | else: | |
3923 | mask = np.ones(len(arr), dtype=bool) | |
3924 | int_arr = pd.arrays.IntegerArray(data.copy(), ~mask, copy=False) | |
3925 | return int_arr | |
3926 | ||
3927 | ||
3928 | def test_convert_to_extension_array(monkeypatch): | |
3929 | if Version(pd.__version__) < Version("0.26.0.dev"): | |
3930 | pytest.skip("Conversion from IntegerArray to arrow not yet supported") | |
3931 | ||
3932 | import pandas.core.internals as _int | |
3933 | ||
3934 | # table converted from dataframe with extension types (so pandas_metadata | |
3935 | # has this information) | |
3936 | df = pd.DataFrame( | |
3937 | {'a': [1, 2, 3], 'b': pd.array([2, 3, 4], dtype='Int64'), | |
3938 | 'c': [4, 5, 6]}) | |
3939 | table = pa.table(df) | |
3940 | ||
3941 | # Int64Dtype is recognized -> convert to extension block by default | |
3942 | # for a proper roundtrip | |
3943 | result = table.to_pandas() | |
3944 | assert not isinstance(result._data.blocks[0], _int.ExtensionBlock) | |
3945 | assert result._data.blocks[0].values.dtype == np.dtype("int64") | |
3946 | assert isinstance(result._data.blocks[1], _int.ExtensionBlock) | |
3947 | tm.assert_frame_equal(result, df) | |
3948 | ||
3949 | # test with missing values | |
3950 | df2 = pd.DataFrame({'a': pd.array([1, 2, None], dtype='Int64')}) | |
3951 | table2 = pa.table(df2) | |
3952 | result = table2.to_pandas() | |
3953 | assert isinstance(result._data.blocks[0], _int.ExtensionBlock) | |
3954 | tm.assert_frame_equal(result, df2) | |
3955 | ||
3956 | # monkeypatch pandas Int64Dtype to *not* have the protocol method | |
3957 | if Version(pd.__version__) < Version("1.3.0.dev"): | |
3958 | monkeypatch.delattr( | |
3959 | pd.core.arrays.integer._IntegerDtype, "__from_arrow__") | |
3960 | else: | |
3961 | monkeypatch.delattr( | |
3962 | pd.core.arrays.integer.NumericDtype, "__from_arrow__") | |
3963 | # Int64Dtype has no __from_arrow__ -> use normal conversion | |
3964 | result = table.to_pandas() | |
3965 | assert len(result._data.blocks) == 1 | |
3966 | assert not isinstance(result._data.blocks[0], _int.ExtensionBlock) | |
3967 | ||
3968 | ||
3969 | class MyCustomIntegerType(pa.PyExtensionType): | |
3970 | ||
3971 | def __init__(self): | |
3972 | pa.PyExtensionType.__init__(self, pa.int64()) | |
3973 | ||
3974 | def __reduce__(self): | |
3975 | return MyCustomIntegerType, () | |
3976 | ||
3977 | def to_pandas_dtype(self): | |
3978 | return pd.Int64Dtype() | |
3979 | ||
3980 | ||
3981 | def test_conversion_extensiontype_to_extensionarray(monkeypatch): | |
3982 | # converting extension type to linked pandas ExtensionDtype/Array | |
3983 | import pandas.core.internals as _int | |
3984 | ||
3985 | if Version(pd.__version__) < Version("0.24.0"): | |
3986 | pytest.skip("ExtensionDtype introduced in pandas 0.24") | |
3987 | ||
3988 | storage = pa.array([1, 2, 3, 4], pa.int64()) | |
3989 | arr = pa.ExtensionArray.from_storage(MyCustomIntegerType(), storage) | |
3990 | table = pa.table({'a': arr}) | |
3991 | ||
3992 | if Version(pd.__version__) < Version("0.26.0.dev"): | |
3993 | # ensure pandas Int64Dtype has the protocol method (for older pandas) | |
3994 | monkeypatch.setattr( | |
3995 | pd.Int64Dtype, '__from_arrow__', _Int64Dtype__from_arrow__, | |
3996 | raising=False) | |
3997 | ||
3998 | # extension type points to Int64Dtype, which knows how to create a | |
3999 | # pandas ExtensionArray | |
4000 | result = arr.to_pandas() | |
4001 | assert isinstance(result._data.blocks[0], _int.ExtensionBlock) | |
4002 | expected = pd.Series([1, 2, 3, 4], dtype='Int64') | |
4003 | tm.assert_series_equal(result, expected) | |
4004 | ||
4005 | result = table.to_pandas() | |
4006 | assert isinstance(result._data.blocks[0], _int.ExtensionBlock) | |
4007 | expected = pd.DataFrame({'a': pd.array([1, 2, 3, 4], dtype='Int64')}) | |
4008 | tm.assert_frame_equal(result, expected) | |
4009 | ||
4010 | # monkeypatch pandas Int64Dtype to *not* have the protocol method | |
4011 | # (remove the version added above and the actual version for recent pandas) | |
4012 | if Version(pd.__version__) < Version("0.26.0.dev"): | |
4013 | monkeypatch.delattr(pd.Int64Dtype, "__from_arrow__") | |
4014 | elif Version(pd.__version__) < Version("1.3.0.dev"): | |
4015 | monkeypatch.delattr( | |
4016 | pd.core.arrays.integer._IntegerDtype, "__from_arrow__") | |
4017 | else: | |
4018 | monkeypatch.delattr( | |
4019 | pd.core.arrays.integer.NumericDtype, "__from_arrow__") | |
4020 | ||
4021 | result = arr.to_pandas() | |
4022 | assert not isinstance(result._data.blocks[0], _int.ExtensionBlock) | |
4023 | expected = pd.Series([1, 2, 3, 4]) | |
4024 | tm.assert_series_equal(result, expected) | |
4025 | ||
4026 | with pytest.raises(ValueError): | |
4027 | table.to_pandas() | |
4028 | ||
4029 | ||
4030 | def test_to_pandas_extension_dtypes_mapping(): | |
4031 | if Version(pd.__version__) < Version("0.26.0.dev"): | |
4032 | pytest.skip("Conversion to pandas IntegerArray not yet supported") | |
4033 | ||
4034 | table = pa.table({'a': pa.array([1, 2, 3], pa.int64())}) | |
4035 | ||
4036 | # default use numpy dtype | |
4037 | result = table.to_pandas() | |
4038 | assert result['a'].dtype == np.dtype('int64') | |
4039 | ||
4040 | # specify to override the default | |
4041 | result = table.to_pandas(types_mapper={pa.int64(): pd.Int64Dtype()}.get) | |
4042 | assert isinstance(result['a'].dtype, pd.Int64Dtype) | |
4043 | ||
4044 | # types that return None in function get normal conversion | |
4045 | table = pa.table({'a': pa.array([1, 2, 3], pa.int32())}) | |
4046 | result = table.to_pandas(types_mapper={pa.int64(): pd.Int64Dtype()}.get) | |
4047 | assert result['a'].dtype == np.dtype('int32') | |
4048 | ||
4049 | # `types_mapper` overrules the pandas metadata | |
4050 | table = pa.table(pd.DataFrame({'a': pd.array([1, 2, 3], dtype="Int64")})) | |
4051 | result = table.to_pandas() | |
4052 | assert isinstance(result['a'].dtype, pd.Int64Dtype) | |
4053 | result = table.to_pandas( | |
4054 | types_mapper={pa.int64(): pd.PeriodDtype('D')}.get) | |
4055 | assert isinstance(result['a'].dtype, pd.PeriodDtype) | |
4056 | ||
4057 | ||
4058 | def test_array_to_pandas(): | |
4059 | if Version(pd.__version__) < Version("1.1"): | |
4060 | pytest.skip("ExtensionDtype to_pandas method missing") | |
4061 | ||
4062 | for arr in [pd.period_range("2012-01-01", periods=3, freq="D").array, | |
4063 | pd.interval_range(1, 4).array]: | |
4064 | result = pa.array(arr).to_pandas() | |
4065 | expected = pd.Series(arr) | |
4066 | tm.assert_series_equal(result, expected) | |
4067 | ||
4068 | # TODO implement proper conversion for chunked array | |
4069 | # result = pa.table({"col": arr})["col"].to_pandas() | |
4070 | # expected = pd.Series(arr, name="col") | |
4071 | # tm.assert_series_equal(result, expected) | |
4072 | ||
4073 | ||
4074 | # ---------------------------------------------------------------------- | |
4075 | # Legacy metadata compatibility tests | |
4076 | ||
4077 | ||
4078 | def test_metadata_compat_range_index_pre_0_12(): | |
4079 | # Forward compatibility for metadata created from pandas.RangeIndex | |
4080 | # prior to pyarrow 0.13.0 | |
4081 | a_values = ['foo', 'bar', None, 'baz'] | |
4082 | b_values = ['a', 'a', 'b', 'b'] | |
4083 | a_arrow = pa.array(a_values, type='utf8') | |
4084 | b_arrow = pa.array(b_values, type='utf8') | |
4085 | ||
4086 | rng_index_arrow = pa.array([0, 2, 4, 6], type='int64') | |
4087 | ||
4088 | gen_name_0 = '__index_level_0__' | |
4089 | gen_name_1 = '__index_level_1__' | |
4090 | ||
4091 | # Case 1: named RangeIndex | |
4092 | e1 = pd.DataFrame({ | |
4093 | 'a': a_values | |
4094 | }, index=pd.RangeIndex(0, 8, step=2, name='qux')) | |
4095 | t1 = pa.Table.from_arrays([a_arrow, rng_index_arrow], | |
4096 | names=['a', 'qux']) | |
4097 | t1 = t1.replace_schema_metadata({ | |
4098 | b'pandas': json.dumps( | |
4099 | {'index_columns': ['qux'], | |
4100 | 'column_indexes': [{'name': None, | |
4101 | 'field_name': None, | |
4102 | 'pandas_type': 'unicode', | |
4103 | 'numpy_type': 'object', | |
4104 | 'metadata': {'encoding': 'UTF-8'}}], | |
4105 | 'columns': [{'name': 'a', | |
4106 | 'field_name': 'a', | |
4107 | 'pandas_type': 'unicode', | |
4108 | 'numpy_type': 'object', | |
4109 | 'metadata': None}, | |
4110 | {'name': 'qux', | |
4111 | 'field_name': 'qux', | |
4112 | 'pandas_type': 'int64', | |
4113 | 'numpy_type': 'int64', | |
4114 | 'metadata': None}], | |
4115 | 'pandas_version': '0.23.4'} | |
4116 | )}) | |
4117 | r1 = t1.to_pandas() | |
4118 | tm.assert_frame_equal(r1, e1) | |
4119 | ||
4120 | # Case 2: named RangeIndex, but conflicts with an actual column | |
4121 | e2 = pd.DataFrame({ | |
4122 | 'qux': a_values | |
4123 | }, index=pd.RangeIndex(0, 8, step=2, name='qux')) | |
4124 | t2 = pa.Table.from_arrays([a_arrow, rng_index_arrow], | |
4125 | names=['qux', gen_name_0]) | |
4126 | t2 = t2.replace_schema_metadata({ | |
4127 | b'pandas': json.dumps( | |
4128 | {'index_columns': [gen_name_0], | |
4129 | 'column_indexes': [{'name': None, | |
4130 | 'field_name': None, | |
4131 | 'pandas_type': 'unicode', | |
4132 | 'numpy_type': 'object', | |
4133 | 'metadata': {'encoding': 'UTF-8'}}], | |
4134 | 'columns': [{'name': 'a', | |
4135 | 'field_name': 'a', | |
4136 | 'pandas_type': 'unicode', | |
4137 | 'numpy_type': 'object', | |
4138 | 'metadata': None}, | |
4139 | {'name': 'qux', | |
4140 | 'field_name': gen_name_0, | |
4141 | 'pandas_type': 'int64', | |
4142 | 'numpy_type': 'int64', | |
4143 | 'metadata': None}], | |
4144 | 'pandas_version': '0.23.4'} | |
4145 | )}) | |
4146 | r2 = t2.to_pandas() | |
4147 | tm.assert_frame_equal(r2, e2) | |
4148 | ||
4149 | # Case 3: unnamed RangeIndex | |
4150 | e3 = pd.DataFrame({ | |
4151 | 'a': a_values | |
4152 | }, index=pd.RangeIndex(0, 8, step=2, name=None)) | |
4153 | t3 = pa.Table.from_arrays([a_arrow, rng_index_arrow], | |
4154 | names=['a', gen_name_0]) | |
4155 | t3 = t3.replace_schema_metadata({ | |
4156 | b'pandas': json.dumps( | |
4157 | {'index_columns': [gen_name_0], | |
4158 | 'column_indexes': [{'name': None, | |
4159 | 'field_name': None, | |
4160 | 'pandas_type': 'unicode', | |
4161 | 'numpy_type': 'object', | |
4162 | 'metadata': {'encoding': 'UTF-8'}}], | |
4163 | 'columns': [{'name': 'a', | |
4164 | 'field_name': 'a', | |
4165 | 'pandas_type': 'unicode', | |
4166 | 'numpy_type': 'object', | |
4167 | 'metadata': None}, | |
4168 | {'name': None, | |
4169 | 'field_name': gen_name_0, | |
4170 | 'pandas_type': 'int64', | |
4171 | 'numpy_type': 'int64', | |
4172 | 'metadata': None}], | |
4173 | 'pandas_version': '0.23.4'} | |
4174 | )}) | |
4175 | r3 = t3.to_pandas() | |
4176 | tm.assert_frame_equal(r3, e3) | |
4177 | ||
4178 | # Case 4: MultiIndex with named RangeIndex | |
4179 | e4 = pd.DataFrame({ | |
4180 | 'a': a_values | |
4181 | }, index=[pd.RangeIndex(0, 8, step=2, name='qux'), b_values]) | |
4182 | t4 = pa.Table.from_arrays([a_arrow, rng_index_arrow, b_arrow], | |
4183 | names=['a', 'qux', gen_name_1]) | |
4184 | t4 = t4.replace_schema_metadata({ | |
4185 | b'pandas': json.dumps( | |
4186 | {'index_columns': ['qux', gen_name_1], | |
4187 | 'column_indexes': [{'name': None, | |
4188 | 'field_name': None, | |
4189 | 'pandas_type': 'unicode', | |
4190 | 'numpy_type': 'object', | |
4191 | 'metadata': {'encoding': 'UTF-8'}}], | |
4192 | 'columns': [{'name': 'a', | |
4193 | 'field_name': 'a', | |
4194 | 'pandas_type': 'unicode', | |
4195 | 'numpy_type': 'object', | |
4196 | 'metadata': None}, | |
4197 | {'name': 'qux', | |
4198 | 'field_name': 'qux', | |
4199 | 'pandas_type': 'int64', | |
4200 | 'numpy_type': 'int64', | |
4201 | 'metadata': None}, | |
4202 | {'name': None, | |
4203 | 'field_name': gen_name_1, | |
4204 | 'pandas_type': 'unicode', | |
4205 | 'numpy_type': 'object', | |
4206 | 'metadata': None}], | |
4207 | 'pandas_version': '0.23.4'} | |
4208 | )}) | |
4209 | r4 = t4.to_pandas() | |
4210 | tm.assert_frame_equal(r4, e4) | |
4211 | ||
4212 | # Case 4: MultiIndex with unnamed RangeIndex | |
4213 | e5 = pd.DataFrame({ | |
4214 | 'a': a_values | |
4215 | }, index=[pd.RangeIndex(0, 8, step=2, name=None), b_values]) | |
4216 | t5 = pa.Table.from_arrays([a_arrow, rng_index_arrow, b_arrow], | |
4217 | names=['a', gen_name_0, gen_name_1]) | |
4218 | t5 = t5.replace_schema_metadata({ | |
4219 | b'pandas': json.dumps( | |
4220 | {'index_columns': [gen_name_0, gen_name_1], | |
4221 | 'column_indexes': [{'name': None, | |
4222 | 'field_name': None, | |
4223 | 'pandas_type': 'unicode', | |
4224 | 'numpy_type': 'object', | |
4225 | 'metadata': {'encoding': 'UTF-8'}}], | |
4226 | 'columns': [{'name': 'a', | |
4227 | 'field_name': 'a', | |
4228 | 'pandas_type': 'unicode', | |
4229 | 'numpy_type': 'object', | |
4230 | 'metadata': None}, | |
4231 | {'name': None, | |
4232 | 'field_name': gen_name_0, | |
4233 | 'pandas_type': 'int64', | |
4234 | 'numpy_type': 'int64', | |
4235 | 'metadata': None}, | |
4236 | {'name': None, | |
4237 | 'field_name': gen_name_1, | |
4238 | 'pandas_type': 'unicode', | |
4239 | 'numpy_type': 'object', | |
4240 | 'metadata': None}], | |
4241 | 'pandas_version': '0.23.4'} | |
4242 | )}) | |
4243 | r5 = t5.to_pandas() | |
4244 | tm.assert_frame_equal(r5, e5) | |
4245 | ||
4246 | ||
4247 | def test_metadata_compat_missing_field_name(): | |
4248 | # Combination of missing field name but with index column as metadata. | |
4249 | # This combo occurs in the latest versions of fastparquet (0.3.2), but not | |
4250 | # in pyarrow itself (since field_name was added in 0.8, index as metadata | |
4251 | # only added later) | |
4252 | ||
4253 | a_values = [1, 2, 3, 4] | |
4254 | b_values = ['a', 'b', 'c', 'd'] | |
4255 | a_arrow = pa.array(a_values, type='int64') | |
4256 | b_arrow = pa.array(b_values, type='utf8') | |
4257 | ||
4258 | expected = pd.DataFrame({ | |
4259 | 'a': a_values, | |
4260 | 'b': b_values, | |
4261 | }, index=pd.RangeIndex(0, 8, step=2, name='qux')) | |
4262 | table = pa.table({'a': a_arrow, 'b': b_arrow}) | |
4263 | ||
4264 | # metadata generated by fastparquet 0.3.2 with missing field_names | |
4265 | table = table.replace_schema_metadata({ | |
4266 | b'pandas': json.dumps({ | |
4267 | 'column_indexes': [ | |
4268 | {'field_name': None, | |
4269 | 'metadata': None, | |
4270 | 'name': None, | |
4271 | 'numpy_type': 'object', | |
4272 | 'pandas_type': 'mixed-integer'} | |
4273 | ], | |
4274 | 'columns': [ | |
4275 | {'metadata': None, | |
4276 | 'name': 'a', | |
4277 | 'numpy_type': 'int64', | |
4278 | 'pandas_type': 'int64'}, | |
4279 | {'metadata': None, | |
4280 | 'name': 'b', | |
4281 | 'numpy_type': 'object', | |
4282 | 'pandas_type': 'unicode'} | |
4283 | ], | |
4284 | 'index_columns': [ | |
4285 | {'kind': 'range', | |
4286 | 'name': 'qux', | |
4287 | 'start': 0, | |
4288 | 'step': 2, | |
4289 | 'stop': 8} | |
4290 | ], | |
4291 | 'pandas_version': '0.25.0'} | |
4292 | ||
4293 | )}) | |
4294 | result = table.to_pandas() | |
4295 | tm.assert_frame_equal(result, expected) | |
4296 | ||
4297 | ||
4298 | def test_metadata_index_name_not_json_serializable(): | |
4299 | name = np.int64(6) # not json serializable by default | |
4300 | table = pa.table(pd.DataFrame(index=pd.RangeIndex(0, 4, name=name))) | |
4301 | metadata = table.schema.pandas_metadata | |
4302 | assert metadata['index_columns'][0]['name'] == '6' | |
4303 | ||
4304 | ||
4305 | def test_metadata_index_name_is_json_serializable(): | |
4306 | name = 6 # json serializable by default | |
4307 | table = pa.table(pd.DataFrame(index=pd.RangeIndex(0, 4, name=name))) | |
4308 | metadata = table.schema.pandas_metadata | |
4309 | assert metadata['index_columns'][0]['name'] == 6 | |
4310 | ||
4311 | ||
4312 | def make_df_with_timestamps(): | |
4313 | # Some of the milliseconds timestamps deliberately don't fit in the range | |
4314 | # that is possible with nanosecond timestamps. | |
4315 | df = pd.DataFrame({ | |
4316 | 'dateTimeMs': [ | |
4317 | np.datetime64('0001-01-01 00:00', 'ms'), | |
4318 | np.datetime64('2012-05-02 12:35', 'ms'), | |
4319 | np.datetime64('2012-05-03 15:42', 'ms'), | |
4320 | np.datetime64('3000-05-03 15:42', 'ms'), | |
4321 | ], | |
4322 | 'dateTimeNs': [ | |
4323 | np.datetime64('1991-01-01 00:00', 'ns'), | |
4324 | np.datetime64('2012-05-02 12:35', 'ns'), | |
4325 | np.datetime64('2012-05-03 15:42', 'ns'), | |
4326 | np.datetime64('2050-05-03 15:42', 'ns'), | |
4327 | ], | |
4328 | }) | |
4329 | # Not part of what we're testing, just ensuring that the inputs are what we | |
4330 | # expect. | |
4331 | assert (df.dateTimeMs.dtype, df.dateTimeNs.dtype) == ( | |
4332 | # O == object, <M8[ns] == timestamp64[ns] | |
4333 | np.dtype("O"), np.dtype("<M8[ns]") | |
4334 | ) | |
4335 | return df | |
4336 | ||
4337 | ||
4338 | @pytest.mark.parquet | |
4339 | def test_timestamp_as_object_parquet(tempdir): | |
4340 | # Timestamps can be stored as Parquet and reloaded into Pandas with no loss | |
4341 | # of information if the timestamp_as_object option is True. | |
4342 | df = make_df_with_timestamps() | |
4343 | table = pa.Table.from_pandas(df) | |
4344 | filename = tempdir / "timestamps_from_pandas.parquet" | |
4345 | pq.write_table(table, filename, version="2.0") | |
4346 | result = pq.read_table(filename) | |
4347 | df2 = result.to_pandas(timestamp_as_object=True) | |
4348 | tm.assert_frame_equal(df, df2) | |
4349 | ||
4350 | ||
4351 | def test_timestamp_as_object_out_of_range(): | |
4352 | # Out of range timestamps can be converted Arrow and reloaded into Pandas | |
4353 | # with no loss of information if the timestamp_as_object option is True. | |
4354 | df = make_df_with_timestamps() | |
4355 | table = pa.Table.from_pandas(df) | |
4356 | df2 = table.to_pandas(timestamp_as_object=True) | |
4357 | tm.assert_frame_equal(df, df2) | |
4358 | ||
4359 | ||
4360 | @pytest.mark.parametrize("resolution", ["s", "ms", "us"]) | |
4361 | @pytest.mark.parametrize("tz", [None, "America/New_York"]) | |
4362 | # One datetime outside nanosecond range, one inside nanosecond range: | |
4363 | @pytest.mark.parametrize("dt", [datetime(1553, 1, 1), datetime(2020, 1, 1)]) | |
4364 | def test_timestamp_as_object_non_nanosecond(resolution, tz, dt): | |
4365 | # Timestamps can be converted Arrow and reloaded into Pandas with no loss | |
4366 | # of information if the timestamp_as_object option is True. | |
4367 | arr = pa.array([dt], type=pa.timestamp(resolution, tz=tz)) | |
4368 | table = pa.table({'a': arr}) | |
4369 | ||
4370 | for result in [ | |
4371 | arr.to_pandas(timestamp_as_object=True), | |
4372 | table.to_pandas(timestamp_as_object=True)['a'] | |
4373 | ]: | |
4374 | assert result.dtype == object | |
4375 | assert isinstance(result[0], datetime) | |
4376 | if tz: | |
4377 | assert result[0].tzinfo is not None | |
4378 | expected = result[0].tzinfo.fromutc(dt) | |
4379 | else: | |
4380 | assert result[0].tzinfo is None | |
4381 | expected = dt | |
4382 | assert result[0] == expected | |
4383 | ||
4384 | ||
4385 | def test_threaded_pandas_import(): | |
4386 | invoke_script("pandas_threaded_import.py") |