Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/pandas/core/interchange/from_dataframe.py: 12%

Shortcuts on this page

r m x   toggle line displays

j k   next/prev highlighted chunk

0   (zero) top of page

1   (one) first highlighted chunk

171 statements  

1from __future__ import annotations 

2 

3import ctypes 

4import re 

5from typing import Any 

6 

7import numpy as np 

8 

9from pandas.compat._optional import import_optional_dependency 

10 

11import pandas as pd 

12from pandas.core.interchange.dataframe_protocol import ( 

13 Buffer, 

14 Column, 

15 ColumnNullType, 

16 DataFrame as DataFrameXchg, 

17 DtypeKind, 

18) 

19from pandas.core.interchange.utils import ( 

20 ArrowCTypes, 

21 Endianness, 

22) 

23 

24_NP_DTYPES: dict[DtypeKind, dict[int, Any]] = { 

25 DtypeKind.INT: {8: np.int8, 16: np.int16, 32: np.int32, 64: np.int64}, 

26 DtypeKind.UINT: {8: np.uint8, 16: np.uint16, 32: np.uint32, 64: np.uint64}, 

27 DtypeKind.FLOAT: {32: np.float32, 64: np.float64}, 

28 DtypeKind.BOOL: {1: bool, 8: bool}, 

29} 

30 

31 

32def from_dataframe(df, allow_copy: bool = True) -> pd.DataFrame: 

33 """ 

34 Build a ``pd.DataFrame`` from any DataFrame supporting the interchange protocol. 

35 

36 Parameters 

37 ---------- 

38 df : DataFrameXchg 

39 Object supporting the interchange protocol, i.e. `__dataframe__` method. 

40 allow_copy : bool, default: True 

41 Whether to allow copying the memory to perform the conversion 

42 (if false then zero-copy approach is requested). 

43 

44 Returns 

45 ------- 

46 pd.DataFrame 

47 """ 

48 if isinstance(df, pd.DataFrame): 

49 return df 

50 

51 if not hasattr(df, "__dataframe__"): 

52 raise ValueError("`df` does not support __dataframe__") 

53 

54 return _from_dataframe(df.__dataframe__(allow_copy=allow_copy)) 

55 

56 

57def _from_dataframe(df: DataFrameXchg, allow_copy: bool = True): 

58 """ 

59 Build a ``pd.DataFrame`` from the DataFrame interchange object. 

60 

61 Parameters 

62 ---------- 

63 df : DataFrameXchg 

64 Object supporting the interchange protocol, i.e. `__dataframe__` method. 

65 allow_copy : bool, default: True 

66 Whether to allow copying the memory to perform the conversion 

67 (if false then zero-copy approach is requested). 

68 

69 Returns 

70 ------- 

71 pd.DataFrame 

72 """ 

73 pandas_dfs = [] 

74 for chunk in df.get_chunks(): 

75 pandas_df = protocol_df_chunk_to_pandas(chunk) 

76 pandas_dfs.append(pandas_df) 

77 

78 if not allow_copy and len(pandas_dfs) > 1: 

79 raise RuntimeError( 

80 "To join chunks a copy is required which is forbidden by allow_copy=False" 

81 ) 

82 if len(pandas_dfs) == 1: 

83 pandas_df = pandas_dfs[0] 

84 else: 

85 pandas_df = pd.concat(pandas_dfs, axis=0, ignore_index=True, copy=False) 

86 

87 index_obj = df.metadata.get("pandas.index", None) 

88 if index_obj is not None: 

89 pandas_df.index = index_obj 

90 

91 return pandas_df 

92 

93 

94def protocol_df_chunk_to_pandas(df: DataFrameXchg) -> pd.DataFrame: 

95 """ 

96 Convert interchange protocol chunk to ``pd.DataFrame``. 

97 

98 Parameters 

99 ---------- 

100 df : DataFrameXchg 

101 

102 Returns 

103 ------- 

104 pd.DataFrame 

105 """ 

106 # We need a dict of columns here, with each column being a NumPy array (at 

107 # least for now, deal with non-NumPy dtypes later). 

108 columns: dict[str, Any] = {} 

109 buffers = [] # hold on to buffers, keeps memory alive 

110 for name in df.column_names(): 

111 if not isinstance(name, str): 

112 raise ValueError(f"Column {name} is not a string") 

113 if name in columns: 

114 raise ValueError(f"Column {name} is not unique") 

115 col = df.get_column_by_name(name) 

116 dtype = col.dtype[0] 

117 if dtype in ( 

118 DtypeKind.INT, 

119 DtypeKind.UINT, 

120 DtypeKind.FLOAT, 

121 DtypeKind.BOOL, 

122 ): 

123 columns[name], buf = primitive_column_to_ndarray(col) 

124 elif dtype == DtypeKind.CATEGORICAL: 

125 columns[name], buf = categorical_column_to_series(col) 

126 elif dtype == DtypeKind.STRING: 

127 columns[name], buf = string_column_to_ndarray(col) 

128 elif dtype == DtypeKind.DATETIME: 

129 columns[name], buf = datetime_column_to_ndarray(col) 

130 else: 

131 raise NotImplementedError(f"Data type {dtype} not handled yet") 

132 

133 buffers.append(buf) 

134 

135 pandas_df = pd.DataFrame(columns) 

136 pandas_df.attrs["_INTERCHANGE_PROTOCOL_BUFFERS"] = buffers 

137 return pandas_df 

138 

139 

140def primitive_column_to_ndarray(col: Column) -> tuple[np.ndarray, Any]: 

141 """ 

142 Convert a column holding one of the primitive dtypes to a NumPy array. 

143 

144 A primitive type is one of: int, uint, float, bool. 

145 

146 Parameters 

147 ---------- 

148 col : Column 

149 

150 Returns 

151 ------- 

152 tuple 

153 Tuple of np.ndarray holding the data and the memory owner object 

154 that keeps the memory alive. 

155 """ 

156 buffers = col.get_buffers() 

157 

158 data_buff, data_dtype = buffers["data"] 

159 data = buffer_to_ndarray( 

160 data_buff, data_dtype, offset=col.offset, length=col.size() 

161 ) 

162 

163 data = set_nulls(data, col, buffers["validity"]) 

164 return data, buffers 

165 

166 

167def categorical_column_to_series(col: Column) -> tuple[pd.Series, Any]: 

168 """ 

169 Convert a column holding categorical data to a pandas Series. 

170 

171 Parameters 

172 ---------- 

173 col : Column 

174 

175 Returns 

176 ------- 

177 tuple 

178 Tuple of pd.Series holding the data and the memory owner object 

179 that keeps the memory alive. 

180 """ 

181 categorical = col.describe_categorical 

182 

183 if not categorical["is_dictionary"]: 

184 raise NotImplementedError("Non-dictionary categoricals not supported yet") 

185 

186 cat_column = categorical["categories"] 

187 if hasattr(cat_column, "_col"): 

188 # Item "Column" of "Optional[Column]" has no attribute "_col" 

189 # Item "None" of "Optional[Column]" has no attribute "_col" 

190 categories = np.array(cat_column._col) # type: ignore[union-attr] 

191 else: 

192 raise NotImplementedError( 

193 "Interchanging categorical columns isn't supported yet, and our " 

194 "fallback of using the `col._col` attribute (a ndarray) failed." 

195 ) 

196 buffers = col.get_buffers() 

197 

198 codes_buff, codes_dtype = buffers["data"] 

199 codes = buffer_to_ndarray( 

200 codes_buff, codes_dtype, offset=col.offset, length=col.size() 

201 ) 

202 

203 # Doing module in order to not get ``IndexError`` for 

204 # out-of-bounds sentinel values in `codes` 

205 if len(categories) > 0: 

206 values = categories[codes % len(categories)] 

207 else: 

208 values = codes 

209 

210 cat = pd.Categorical( 

211 values, categories=categories, ordered=categorical["is_ordered"] 

212 ) 

213 data = pd.Series(cat) 

214 

215 data = set_nulls(data, col, buffers["validity"]) 

216 return data, buffers 

217 

218 

219def string_column_to_ndarray(col: Column) -> tuple[np.ndarray, Any]: 

220 """ 

221 Convert a column holding string data to a NumPy array. 

222 

223 Parameters 

224 ---------- 

225 col : Column 

226 

227 Returns 

228 ------- 

229 tuple 

230 Tuple of np.ndarray holding the data and the memory owner object 

231 that keeps the memory alive. 

232 """ 

233 null_kind, sentinel_val = col.describe_null 

234 

235 if null_kind not in ( 

236 ColumnNullType.NON_NULLABLE, 

237 ColumnNullType.USE_BITMASK, 

238 ColumnNullType.USE_BYTEMASK, 

239 ): 

240 raise NotImplementedError( 

241 f"{null_kind} null kind is not yet supported for string columns." 

242 ) 

243 

244 buffers = col.get_buffers() 

245 

246 assert buffers["offsets"], "String buffers must contain offsets" 

247 # Retrieve the data buffer containing the UTF-8 code units 

248 data_buff, protocol_data_dtype = buffers["data"] 

249 # We're going to reinterpret the buffer as uint8, so make sure we can do it safely 

250 assert protocol_data_dtype[1] == 8 

251 assert protocol_data_dtype[2] in ( 

252 ArrowCTypes.STRING, 

253 ArrowCTypes.LARGE_STRING, 

254 ) # format_str == utf-8 

255 # Convert the buffers to NumPy arrays. In order to go from STRING to 

256 # an equivalent ndarray, we claim that the buffer is uint8 (i.e., a byte array) 

257 data_dtype = ( 

258 DtypeKind.UINT, 

259 8, 

260 ArrowCTypes.UINT8, 

261 Endianness.NATIVE, 

262 ) 

263 # Specify zero offset as we don't want to chunk the string data 

264 data = buffer_to_ndarray(data_buff, data_dtype, offset=0, length=data_buff.bufsize) 

265 

266 # Retrieve the offsets buffer containing the index offsets demarcating 

267 # the beginning and the ending of each string 

268 offset_buff, offset_dtype = buffers["offsets"] 

269 # Offsets buffer contains start-stop positions of strings in the data buffer, 

270 # meaning that it has more elements than in the data buffer, do `col.size() + 1` 

271 # here to pass a proper offsets buffer size 

272 offsets = buffer_to_ndarray( 

273 offset_buff, offset_dtype, offset=col.offset, length=col.size() + 1 

274 ) 

275 

276 null_pos = None 

277 if null_kind in (ColumnNullType.USE_BITMASK, ColumnNullType.USE_BYTEMASK): 

278 assert buffers["validity"], "Validity buffers cannot be empty for masks" 

279 valid_buff, valid_dtype = buffers["validity"] 

280 null_pos = buffer_to_ndarray( 

281 valid_buff, valid_dtype, offset=col.offset, length=col.size() 

282 ) 

283 if sentinel_val == 0: 

284 null_pos = ~null_pos 

285 

286 # Assemble the strings from the code units 

287 str_list: list[None | float | str] = [None] * col.size() 

288 for i in range(col.size()): 

289 # Check for missing values 

290 if null_pos is not None and null_pos[i]: 

291 str_list[i] = np.nan 

292 continue 

293 

294 # Extract a range of code units 

295 units = data[offsets[i] : offsets[i + 1]] 

296 

297 # Convert the list of code units to bytes 

298 str_bytes = bytes(units) 

299 

300 # Create the string 

301 string = str_bytes.decode(encoding="utf-8") 

302 

303 # Add to our list of strings 

304 str_list[i] = string 

305 

306 # Convert the string list to a NumPy array 

307 return np.asarray(str_list, dtype="object"), buffers 

308 

309 

310def parse_datetime_format_str(format_str, data): 

311 """Parse datetime `format_str` to interpret the `data`.""" 

312 # timestamp 'ts{unit}:tz' 

313 timestamp_meta = re.match(r"ts([smun]):(.*)", format_str) 

314 if timestamp_meta: 

315 unit, tz = timestamp_meta.group(1), timestamp_meta.group(2) 

316 if tz != "": 

317 raise NotImplementedError("Timezones are not supported yet") 

318 if unit != "s": 

319 # the format string describes only a first letter of the unit, so 

320 # add one extra letter to convert the unit to numpy-style: 

321 # 'm' -> 'ms', 'u' -> 'us', 'n' -> 'ns' 

322 unit += "s" 

323 data = data.astype(f"datetime64[{unit}]") 

324 return data 

325 

326 # date 'td{Days/Ms}' 

327 date_meta = re.match(r"td([Dm])", format_str) 

328 if date_meta: 

329 unit = date_meta.group(1) 

330 if unit == "D": 

331 # NumPy doesn't support DAY unit, so converting days to seconds 

332 # (converting to uint64 to avoid overflow) 

333 data = (data.astype(np.uint64) * (24 * 60 * 60)).astype("datetime64[s]") 

334 elif unit == "m": 

335 data = data.astype("datetime64[ms]") 

336 else: 

337 raise NotImplementedError(f"Date unit is not supported: {unit}") 

338 return data 

339 

340 raise NotImplementedError(f"DateTime kind is not supported: {format_str}") 

341 

342 

343def datetime_column_to_ndarray(col: Column) -> tuple[np.ndarray, Any]: 

344 """ 

345 Convert a column holding DateTime data to a NumPy array. 

346 

347 Parameters 

348 ---------- 

349 col : Column 

350 

351 Returns 

352 ------- 

353 tuple 

354 Tuple of np.ndarray holding the data and the memory owner object 

355 that keeps the memory alive. 

356 """ 

357 buffers = col.get_buffers() 

358 

359 _, _, format_str, _ = col.dtype 

360 dbuf, dtype = buffers["data"] 

361 # Consider dtype being `uint` to get number of units passed since the 01.01.1970 

362 data = buffer_to_ndarray( 

363 dbuf, 

364 ( 

365 DtypeKind.UINT, 

366 dtype[1], 

367 getattr(ArrowCTypes, f"UINT{dtype[1]}"), 

368 Endianness.NATIVE, 

369 ), 

370 offset=col.offset, 

371 length=col.size(), 

372 ) 

373 

374 data = parse_datetime_format_str(format_str, data) 

375 data = set_nulls(data, col, buffers["validity"]) 

376 return data, buffers 

377 

378 

379def buffer_to_ndarray( 

380 buffer: Buffer, 

381 dtype: tuple[DtypeKind, int, str, str], 

382 *, 

383 length: int, 

384 offset: int = 0, 

385) -> np.ndarray: 

386 """ 

387 Build a NumPy array from the passed buffer. 

388 

389 Parameters 

390 ---------- 

391 buffer : Buffer 

392 Buffer to build a NumPy array from. 

393 dtype : tuple 

394 Data type of the buffer conforming protocol dtypes format. 

395 offset : int, default: 0 

396 Number of elements to offset from the start of the buffer. 

397 length : int, optional 

398 If the buffer is a bit-mask, specifies a number of bits to read 

399 from the buffer. Has no effect otherwise. 

400 

401 Returns 

402 ------- 

403 np.ndarray 

404 

405 Notes 

406 ----- 

407 The returned array doesn't own the memory. The caller of this function is 

408 responsible for keeping the memory owner object alive as long as 

409 the returned NumPy array is being used. 

410 """ 

411 kind, bit_width, _, _ = dtype 

412 

413 column_dtype = _NP_DTYPES.get(kind, {}).get(bit_width, None) 

414 if column_dtype is None: 

415 raise NotImplementedError(f"Conversion for {dtype} is not yet supported.") 

416 

417 # TODO: No DLPack yet, so need to construct a new ndarray from the data pointer 

418 # and size in the buffer plus the dtype on the column. Use DLPack as NumPy supports 

419 # it since https://github.com/numpy/numpy/pull/19083 

420 ctypes_type = np.ctypeslib.as_ctypes_type(column_dtype) 

421 

422 if bit_width == 1: 

423 assert length is not None, "`length` must be specified for a bit-mask buffer." 

424 pa = import_optional_dependency("pyarrow") 

425 arr = pa.BooleanArray.from_buffers( 

426 pa.bool_(), 

427 length, 

428 [None, pa.foreign_buffer(buffer.ptr, length)], 

429 offset=offset, 

430 ) 

431 return np.asarray(arr) 

432 else: 

433 data_pointer = ctypes.cast( 

434 buffer.ptr + (offset * bit_width // 8), ctypes.POINTER(ctypes_type) 

435 ) 

436 return np.ctypeslib.as_array( 

437 data_pointer, 

438 shape=(length,), 

439 ) 

440 

441 

442def set_nulls( 

443 data: np.ndarray | pd.Series, 

444 col: Column, 

445 validity: tuple[Buffer, tuple[DtypeKind, int, str, str]] | None, 

446 allow_modify_inplace: bool = True, 

447): 

448 """ 

449 Set null values for the data according to the column null kind. 

450 

451 Parameters 

452 ---------- 

453 data : np.ndarray or pd.Series 

454 Data to set nulls in. 

455 col : Column 

456 Column object that describes the `data`. 

457 validity : tuple(Buffer, dtype) or None 

458 The return value of ``col.buffers()``. We do not access the ``col.buffers()`` 

459 here to not take the ownership of the memory of buffer objects. 

460 allow_modify_inplace : bool, default: True 

461 Whether to modify the `data` inplace when zero-copy is possible (True) or always 

462 modify a copy of the `data` (False). 

463 

464 Returns 

465 ------- 

466 np.ndarray or pd.Series 

467 Data with the nulls being set. 

468 """ 

469 null_kind, sentinel_val = col.describe_null 

470 null_pos = None 

471 

472 if null_kind == ColumnNullType.USE_SENTINEL: 

473 null_pos = pd.Series(data) == sentinel_val 

474 elif null_kind in (ColumnNullType.USE_BITMASK, ColumnNullType.USE_BYTEMASK): 

475 assert validity, "Expected to have a validity buffer for the mask" 

476 valid_buff, valid_dtype = validity 

477 null_pos = buffer_to_ndarray( 

478 valid_buff, valid_dtype, offset=col.offset, length=col.size() 

479 ) 

480 if sentinel_val == 0: 

481 null_pos = ~null_pos 

482 elif null_kind in (ColumnNullType.NON_NULLABLE, ColumnNullType.USE_NAN): 

483 pass 

484 else: 

485 raise NotImplementedError(f"Null kind {null_kind} is not yet supported.") 

486 

487 if null_pos is not None and np.any(null_pos): 

488 if not allow_modify_inplace: 

489 data = data.copy() 

490 try: 

491 data[null_pos] = None 

492 except TypeError: 

493 # TypeError happens if the `data` dtype appears to be non-nullable 

494 # in numpy notation (bool, int, uint). If this happens, 

495 # cast the `data` to nullable float dtype. 

496 data = data.astype(float) 

497 data[null_pos] = None 

498 

499 return data