def export_hdf5(dataset, path, column_names=None, byteorder="=", shuffle=False, selection=False, progress=None, virtual=True, sort=None, ascending=True):
"""
:param DatasetLocal dataset: dataset to export
:param str path: path for file
:param lis[str] column_names: list of column names to export or None for all columns
:param str byteorder: = for native, < for little endian and > for big endian
:param bool shuffle: export rows in random order
:param bool selection: export selection or not
:param progress: progress callback that gets a progress fraction as argument and should return True to continue,
or a default progress bar when progress=True
:param: bool virtual: When True, export virtual columns
:return:
"""
if selection:
if selection == True: # easier to work with the name
selection = "default"
# first open file using h5py api
with h5py.File(path, "w") as h5file_output:
h5table_output = h5file_output.require_group("/table")
h5table_output.attrs["type"] = "table"
h5columns_output = h5file_output.require_group("/table/columns")
# i1, i2 = dataset.current_slice
N = len(dataset) if not selection else dataset.selected_length(selection)
if N == 0:
raise ValueError("Cannot export empty table")
logger.debug("virtual=%r", virtual)
logger.debug("exporting %d rows to file %s" % (N, path))
# column_names = column_names or (dataset.get_column_names() + (list(dataset.virtual_columns.keys()) if virtual else []))
column_names = column_names or dataset.get_column_names(virtual=virtual, strings=True, alias=False)
logger.debug("exporting columns(hdf5): %r" % column_names)
sparse_groups = collections.defaultdict(list)
sparse_matrices = {} # alternative to a set of matrices, since they are not hashable
for column_name in list(column_names):
sparse_matrix = dataset._sparse_matrix(column_name)
if sparse_matrix is not None:
# sparse columns are stored differently
sparse_groups[id(sparse_matrix)].append(column_name)
sparse_matrices[id(sparse_matrix)] = sparse_matrix
continue
dtype = dataset.data_type(column_name)
shape = (N, ) + dataset._shape_of(column_name)[1:]
h5column_output = h5columns_output.require_group(column_name)
if vaex.array_types.is_string_type(dtype):
# TODO: if no selection or filter, we could do this
# if isinstance(column, ColumnStringArrow):
# data_shape = column.bytes.shape
# indices_shape = column.indices.shape
# else:
byte_length = dataset[column_name].str.byte_length().sum(selection=selection)
if byte_length > max_int32:
dtype_indices = 'i8'
else:
dtype_indices = 'i4'
data_shape = (byte_length, )
indices_shape = (N+1, )
array = h5column_output.require_dataset('data', shape=data_shape, dtype='S1')
if byte_length > 0:
array[0] = array[0] # make sure the array really exists
index_array = h5column_output.require_dataset('indices', shape=indices_shape, dtype=dtype_indices)
index_array[0] = index_array[0] # make sure the array really exists
null_value_count = N - dataset.count(column_name, selection=selection)
if null_value_count > 0:
null_shape = ((N + 7) // 8, ) # TODO: arrow requires padding right?
null_bitmap_array = h5column_output.require_dataset('null_bitmap', shape=null_shape, dtype='u1')
null_bitmap_array[0] = null_bitmap_array[0] # make sure the array really exists
array.attrs["dtype"] = 'str'
# TODO: masked support ala arrow?
else:
if dtype.kind in 'mM':
array = h5column_output.require_dataset('data', shape=shape, dtype=np.int64)
array.attrs["dtype"] = dtype.name
elif dtype.kind == 'U':
# numpy uses utf32 for unicode
char_length = dtype.itemsize // 4
shape = (N, char_length)
array = h5column_output.require_dataset('data', shape=shape, dtype=np.uint8)
array.attrs["dtype"] = 'utf32'
array.attrs["dlength"] = char_length
else:
try:
array = h5column_output.require_dataset('data', shape=shape, dtype=dtype.newbyteorder(byteorder))
except:
logging.exception("error creating dataset for %r, with type %r " % (column_name, dtype))
del h5columns_output[column_name]
column_names.remove(column_name)
array[0] = array[0] # make sure the array really exists
data = dataset.evaluate(column_name, 0, 1, parallel=False)
if np.ma.isMaskedArray(data):
mask = h5column_output.require_dataset('mask', shape=shape, dtype=np.bool)
mask[0] = mask[0] # make sure the array really exists
random_index_name = None
column_order = list(column_names) # copy
if shuffle:
random_index_name = "random_index"
while random_index_name in dataset.get_column_names():
random_index_name += "_new"
shuffle_array = h5columns_output.require_dataset(random_index_name + "/data", shape=(N,), dtype=byteorder + "i8")
shuffle_array[0] = shuffle_array[0]
column_order.append(random_index_name) # last item
h5columns_output.attrs["column_order"] = ",".join(column_order) # keep track or the ordering of columns
sparse_index = 0
for sparse_matrix in sparse_matrices.values():
columns = sorted(sparse_groups[id(sparse_matrix)], key=lambda col: dataset.columns[col].column_index)
name = "sparse" + str(sparse_index)
sparse_index += 1
# TODO: slice columns
# sparse_matrix = sparse_matrix[:,]
sparse_group = h5columns_output.require_group(name)
sparse_group.attrs['type'] = 'csr_matrix'
ar = sparse_group.require_dataset('data', shape=(len(sparse_matrix.data), ), dtype=sparse_matrix.dtype)
ar[0] = ar[0]
ar = sparse_group.require_dataset('indptr', shape=(len(sparse_matrix.indptr), ), dtype=sparse_matrix.indptr.dtype)
ar[0] = ar[0]
ar = sparse_group.require_dataset('indices', shape=(len(sparse_matrix.indices), ), dtype=sparse_matrix.indices.dtype)
ar[0] = ar[0]
for i, column_name in enumerate(columns):
h5column = sparse_group.require_group(column_name)
h5column.attrs['column_index'] = i
# after this the file is closed,, and reopen it using out class
dataset_output = vaex.hdf5.dataset.Hdf5MemoryMapped(path, write=True)
column_names = vaex.export._export(dataset_input=dataset, dataset_output=dataset_output, path=path, random_index_column=random_index_name,
column_names=column_names, selection=selection, shuffle=shuffle, byteorder=byteorder,
progress=progress, sort=sort, ascending=ascending)
import getpass
import datetime
user = getpass.getuser()
date = str(datetime.datetime.now())
source = dataset.path
description = "file exported by vaex, by user %s, on date %s, from source %s" % (user, date, source)
if dataset.description:
description += "previous description:\n" + dataset.description
dataset_output.copy_metadata(dataset)
dataset_output.description = description
logger.debug("writing meta information")
dataset_output.write_meta()
dataset_output.close_files()
return
def export_hdf5(dataset, path, column_names=None, byteorder="=", shuffle=False, selection=False, progress=None, virtual=True, sort=None, ascending=True):
"""
:param DatasetLocal dataset: dataset to export
:param str path: path for file
:param lis[str] column_names: list of column names to export or None for all columns
:param str byteorder: = for native, < for little endian and > for big endian
:param bool shuffle: export rows in random order
:param bool selection: export selection or not
:param progress: progress callback that gets a progress fraction as argument and should return True to continue,
or a default progress bar when progress=True
:param: bool virtual: When True, export virtual columns
:return:
"""
if selection:
if selection == True: # easier to work with the name
selection = "default"
# first open file using h5py api
with h5py.File(path, "w") as h5file_output:
h5table_output = h5file_output.require_group("/table")
h5table_output.attrs["type"] = "table"
h5columns_output = h5file_output.require_group("/table/columns")
# i1, i2 = dataset.current_slice
N = len(dataset) if not selection else dataset.selected_length(selection)
if N == 0:
raise ValueError("Cannot export empty table")
logger.debug("virtual=%r", virtual)
logger.debug("exporting %d rows to file %s" % (N, path))
# column_names = column_names or (dataset.get_column_names() + (list(dataset.virtual_columns.keys()) if virtual else []))
column_names = column_names or dataset.get_column_names(virtual=virtual, strings=True)
logger.debug("exporting columns(hdf5): %r" % column_names)
sparse_groups = collections.defaultdict(list)
sparse_matrices = {} # alternative to a set of matrices, since they are not hashable
for column_name in list(column_names):
sparse_matrix = dataset._sparse_matrix(column_name)
if sparse_matrix is not None:
# sparse columns are stored differently
sparse_groups[id(sparse_matrix)].append(column_name)
sparse_matrices[id(sparse_matrix)] = sparse_matrix
continue
dtype = dataset.dtype(column_name)
if column_name in dataset.get_column_names(virtual=False):
column = dataset.columns[column_name]
shape = (N,) + column.shape[1:]
else:
shape = (N,)
h5column_output = h5columns_output.require_group(column_name)
if dtype == str_type:
# TODO: if no selection or filter, we could do this
# if isinstance(column, ColumnStringArrow):
# data_shape = column.bytes.shape
# indices_shape = column.indices.shape
# else:
byte_length = dataset[column_name].str.byte_length().sum(selection=selection)
if byte_length > max_int32:
dtype_indices = 'i8'
else:
dtype_indices = 'i4'
data_shape = (byte_length, )
indices_shape = (N+1, )
array = h5column_output.require_dataset('data', shape=data_shape, dtype='S1')
array[0] = array[0] # make sure the array really exists
index_array = h5column_output.require_dataset('indices', shape=indices_shape, dtype=dtype_indices)
index_array[0] = index_array[0] # make sure the array really exists
null_value_count = N - dataset.count(column_name, selection=selection)
if null_value_count > 0:
null_shape = ((N + 7) // 8, ) # TODO: arrow requires padding right?
null_bitmap_array = h5column_output.require_dataset('null_bitmap', shape=null_shape, dtype='u1')
null_bitmap_array[0] = null_bitmap_array[0] # make sure the array really exists
array.attrs["dtype"] = 'str'
# TODO: masked support ala arrow?
else:
if dtype.kind in 'mM':
array = h5column_output.require_dataset('data', shape=shape, dtype=np.int64)
array.attrs["dtype"] = dtype.name
elif dtype.kind == 'U':
# numpy uses utf32 for unicode
char_length = dtype.itemsize // 4
shape = (N, char_length)
array = h5column_output.require_dataset('data', shape=shape, dtype=np.uint8)
array.attrs["dtype"] = 'utf32'
array.attrs["dlength"] = char_length
else:
try:
array = h5column_output.require_dataset('data', shape=shape, dtype=dtype.newbyteorder(byteorder))
except:
logging.exception("error creating dataset for %r, with type %r " % (column_name, dtype))
del h5columns_output[column_name]
column_names.remove(column_name)
array[0] = array[0] # make sure the array really exists
data = dataset.evaluate(column_name, 0, 1)
if np.ma.isMaskedArray(data):
mask = h5column_output.require_dataset('mask', shape=shape, dtype=np.bool)
mask[0] = mask[0] # make sure the array really exists
random_index_name = None
column_order = list(column_names) # copy
if shuffle:
random_index_name = "random_index"
while random_index_name in dataset.get_column_names():
random_index_name += "_new"
shuffle_array = h5columns_output.require_dataset(random_index_name + "/data", shape=(N,), dtype=byteorder + "i8")
shuffle_array[0] = shuffle_array[0]
column_order.append(random_index_name) # last item
h5columns_output.attrs["column_order"] = ",".join(column_order) # keep track or the ordering of columns
sparse_index = 0
for sparse_matrix in sparse_matrices.values():
columns = sorted(sparse_groups[id(sparse_matrix)], key=lambda col: dataset.columns[col].column_index)
name = "sparse" + str(sparse_index)
sparse_index += 1
# TODO: slice columns
# sparse_matrix = sparse_matrix[:,]
sparse_group = h5columns_output.require_group(name)
sparse_group.attrs['type'] = 'csr_matrix'
ar = sparse_group.require_dataset('data', shape=(len(sparse_matrix.data), ), dtype=sparse_matrix.dtype)
ar[0] = ar[0]
ar = sparse_group.require_dataset('indptr', shape=(len(sparse_matrix.indptr), ), dtype=sparse_matrix.indptr.dtype)
ar[0] = ar[0]
ar = sparse_group.require_dataset('indices', shape=(len(sparse_matrix.indices), ), dtype=sparse_matrix.indices.dtype)
ar[0] = ar[0]
for i, column_name in enumerate(columns):
h5column = sparse_group.require_group(column_name)
h5column.attrs['column_index'] = i
# after this the file is closed,, and reopen it using out class
dataset_output = vaex.hdf5.dataset.Hdf5MemoryMapped(path, write=True)
column_names = vaex.export._export(dataset_input=dataset, dataset_output=dataset_output, path=path, random_index_column=random_index_name,
column_names=column_names, selection=selection, shuffle=shuffle, byteorder=byteorder,
progress=progress, sort=sort, ascending=ascending)
import getpass
import datetime
user = getpass.getuser()
date = str(datetime.datetime.now())
source = dataset.path
description = "file exported by vaex, by user %s, on date %s, from source %s" % (user, date, source)
if dataset.description:
description += "previous description:\n" + dataset.description
dataset_output.copy_metadata(dataset)
dataset_output.description = description
logger.debug("writing meta information")
dataset_output.write_meta()
dataset_output.close_files()
return