torcharrow/velox_rt/dataframe_cpu.py (17 lines): - line 72: # TODO: using a dict for field type lookup - line 93: assert mask is None # TODO: remove mask parameter in _FullColumn - line 94: cols = data.values() # TODO: also allow data to be a single Velox RowColumn - line 105: # TODO this must be weakened (to deal with nulls, etc)... - line 324: # TODO: decide on nulls, here we assume all defined (mask = False) for new parent... - line 331: # TODO: decide on nulls, here we assume all defined (mask = False) for new parent... - line 1301: # TODO only flat columns supported... - line 1326: # TODO fix slow implementation by vectorization, - line 1362: # TODO Decide on tracing level: If we trace 'min' om a - line 1654: # TODO Add type translation. - line 1664: # TODO Add type translation - line 1693: # TODO: this actually puts the type annotations on the tuple wrong. - line 1711: # TODO decide on whether we nat to have arbitrarily nested wheres... - line 1944: # TODO implement - line 2079: # TODO extend that this works inside struct frames as well, - line 2141: # TODO should add reduce here as well... - line 2198: # TODO handle duplicate columns, if ever... csrc/velox/column.h (10 lines): - line 20: // TODO: Move uses of static variables into .cpp. Static variables are local to - line 46: // TODO: use row type instead of string - line 94: // TODO: Use Column once ConstantColumn is supported - line 161: // TODO: Decide on how we want to handle column-scalar type promotion. - line 234: // TODO: move this method as static... - line 246: // TODO: Model binary functions as UDF. - line 324: // TODO: deprecate this method and migrate to OperatorHandle::fromGenericUDF - line 345: // TODO: refactor into some type utility class - line 435: // TODO: return SimpleColumn instead? - line 551: // TODO: remove these method binding and migrate to the torcharrow/icolumn.py (9 lines): - line 227: # TODO use _column_copy, but for now this works... - line 240: # TODO: support non-primitive types - line 636: # TODO: check type annotations of inputs too in order to infer the input format - line 638: # TODO: if func is annotated, check whether its input parameter is IColumn when format="column" - line 1199: # TODO Add functionality for first and last - line 1224: # TODO: use pc.min once upgrade to later version of PyArrow - line 1242: # TODO: use pc.max once upgrade to later version of PyArrow - line 1559: # TODO: should we just move _py_arithmetic_op to INumericColumn since it only works for boolean/numeric types - line 1569: # TODO Use error handling to mke this more efficient.. csrc/velox/column.cpp (7 lines): - line 148: // TODO: Model binary functions as UDF. - line 155: // FIXME This is fragile as it assumes velox::TypeKind numbers numeric types - line 240: // TODO: This causes an extra type-based dispatch. - line 424: // TODO: This causes an extra type-based dispatch. - line 439: // TODO: This causes an extra type-based dispatch. - line 453: // TODO: This causes an extra type-based dispatch. - line 522: // TODO: Open source some part of utility codes in Koski (PyVelox?) csrc/velox/lib.cpp (7 lines): - line 127: // TODO: Move the Koksi binding of Velox type to OSS - line 610: // TODO: Move the Koksi binding of Velox type to OSS - line 648: // TODO: Move the Koksi binding of Velox type to OSS - line 677: // TODO: Move the Koksi binding of Velox type to OSS - line 715: "TypeKind", // TODO: Move the Koksi binding of Velox type to OSS - line 735: // TODO: Move the Koksi binding of Velox type to OSS - line 810: // TODO: we may only need to register UDFs that TorchArrow required? torcharrow/scope.py (6 lines): - line 91: # TODO: refactor these static methods out of scope.py - line 110: TODO: rename this method to _EmptyColumnBuilder to emphasize it returns a builder - line 139: # TODO: rename the dispatch key to be "_from_pylist" - line 170: # TODO: infer the type from the whole list - line 190: # TODO: More efficient interop, such as leveraging arrow - line 222: # TODO fix me torcharrow/velox_rt/map_column_cpu.py (6 lines): - line 69: # # TODO fix nullability - line 80: # TODO check that all non-masked items are legal numbers (i.e not nan) - line 167: # TODO: more efficient/straightfowrad interop - line 169: # FIXME: https://github.com/facebookresearch/torcharrow/issues/62 to_arrow doesn't work as expected for map - line 179: # TODO: should we propagate python list if both keys and vals are lists of strings? - line 190: # FIXME: https://github.com/facebookresearch/torcharrow/issues/62 torcharrow/_pytorch.py (6 lines): - line 122: # TODO: this implementation is very inefficient, we should wrap the column directly instead of round-tripping through python - line 149: # TODO: this implementation is very inefficient, we should wrap the column directly instead of round-tripping through python - line 174: # TODO: this implementation is very inefficient, we should wrap the column directly instead of round-tripping through python - line 190: # TODO: check that fields of named tuples match? - line 207: # TODO: this implementation is very inefficient, we should wrap the column directly instead of round-tripping through python - line 224: # TODO: this implementation is very inefficient, we should wrap the column directly instead of round-tripping through python torcharrow/dtypes.py (4 lines): - line 273: # TODO: perhaps this should be a private method - line 296: # TODO: this might cause name duplicates, do disambiguation - line 707: # # DESIGN BUG: TODO needs actually both sides for symmetric promotion rules ... - line 983: # TODO: we need a type annotation for Columns with statically accessible dtype torcharrow/velox_rt/numerical_column_cpu.py (4 lines): - line 35: # TODO: Deprecate _finalized since Velox Column doesn't have "Builder" mode - line 270: # TODO: Support native Velox execution for boolean ops - line 622: # TODO: round(-2.5) returns -2.0 in Numpy/PyTorch but returns -3.0 in Velox - line 868: # TODO: figure out zero copy from Velox vector torcharrow/idataframe.py (4 lines): - line 411: # TODO: check type annotations of inputs too in order to infer the input format - line 458: # TODO Make this abstract and add all the abstract methods here ... - line 459: # TODO Current short cut has 'everything', excpet for columns as a DataFrameDemo - line 460: # TODO Make GroupedDatFrame also an IGroupedDataframe to make it truly compositional torcharrow/velox_rt/string_column_cpu.py (4 lines): - line 59: # TODO refine this test - line 67: # TODO check that all non-masked items are strings - line 80: # TODO Add native kernel support - line 322: # TODO: calculate without materializing all the occurrences torcharrow/velox_rt/list_column_cpu.py (3 lines): - line 150: # TODO: more efficient/straightfowrad interop - line 171: # TODO: clarify int32 vs int64 - line 188: # TODO: pad_sequence also works for nest numeric list torcharrow/_interop.py (2 lines): - line 114: # TODO if not, adopt the type or? - line 139: # TODO Check why teh following assert isn't the case torcharrow/functional.py (2 lines): - line 48: # TODO: Support SizeProxy - line 67: # TODO: Perhaps this should be part of dispatch backend registration setup.py (2 lines): - line 87: # TODO: is there any better approach? - line 158: packages=find_packages() # TODO ??? torcharrow/velox_rt/functional.py (2 lines): - line 67: # TODO: automtically populate it - line 74: # TODO: implement this torcharrow/expression.py (1 line): - line 31: # TODO should we hijack this (or have yet another doc decorator) torcharrow/imap_column.py (1 line): - line 83: # TODO improve perf by looking at lists instead of first building a map torcharrow/istring_column.py (1 line): - line 191: # TODO: docstring csrc/velox/functions/functions.h (1 line): - line 30: // TODO: support type promotion in TorchArrow-Velox backend so registering less tutorial/tutorial.py (1 line): - line 119: # TODO: Fix this! torcharrow/velox_rt/column.py (1 line): - line 10: # TODO: Rename this class to IColumnVelox or IColumnCpu