def array_to_meta_thrift_struct(array, name, format):
type = array.dtype.kind
slice = name
l = len(array.shape)
# initial load, compute slice
if format == '%':
if l > 2:
slice += '[0]' * (l - 2)
for r in range(l - 2):
array = array[0]
if type == 'f':
format = '.5f'
elif type == 'i' or type == 'u':
format = 'd'
else:
format = 's'
else:
format = format.replace('%', '')
l = len(array.shape)
reslice = ""
if l > 2:
raise ExceedingArrayDimensionsException
elif l == 1:
# special case with 1D arrays arr[i, :] - row, but arr[:, i] - column with equal shape and ndim
# http://stackoverflow.com/questions/16837946/numpy-a-2-rows-1-column-file-loadtxt-returns-1row-2-columns
# explanation: http://stackoverflow.com/questions/15165170/how-do-i-maintain-row-column-orientation-of-vectors-in-numpy?rq=1
# we use kind of a hack - get information about memory from C_CONTIGUOUS
is_row = array.flags['C_CONTIGUOUS']
if is_row:
rows = 1
cols = len(array)
if cols < len(array):
reslice = '[0:%s]' % (cols)
array = array[0:cols]
else:
cols = 1
rows = len(array)
if rows < len(array):
reslice = '[0:%s]' % (rows)
array = array[0:rows]
elif l == 2:
rows = array.shape[-2]
cols = array.shape[-1]
if cols < array.shape[-1] or rows < array.shape[-2]:
reslice = '[0:%s, 0:%s]' % (rows, cols)
array = array[0:rows, 0:cols]
# avoid slice duplication
if not slice.endswith(reslice):
slice += reslice
bounds = (0, 0)
if type in NUMPY_NUMERIC_TYPES and array.size != 0:
bounds = (array.min(), array.max())
array_chunk = GetArrayResponse()
array_chunk.slice = slice
array_chunk.rows = rows
array_chunk.cols = cols
array_chunk.format = "%" + format
array_chunk.type = type
array_chunk.max = "%s" % bounds[1]
array_chunk.min = "%s" % bounds[0]
return array, array_chunk, rows, cols, format
def dataframe_to_thrift_struct(df, name, roffset, coffset, rows, cols, format):
"""
:type df: pandas.core.frame.DataFrame
:type name: str
:type coffset: int
:type roffset: int
:type rows: int
:type cols: int
:type format: str
"""
original_df = df
dim = len(df.axes)
num_rows = df.shape[0]
num_cols = df.shape[1] if dim > 1 else 1
array_chunk = GetArrayResponse()
array_chunk.slice = name
array_chunk.rows = num_rows
array_chunk.cols = num_cols
array_chunk.type = ""
array_chunk.max = "0"
array_chunk.min = "0"
format = format.replace("%", "")
if not format:
if num_rows > 0 and num_cols == 1: # series or data frame with one column
try:
kind = df.dtype.kind
except AttributeError:
try:
kind = df.dtypes[0].kind
except (IndexError, KeyError):
kind = "O"
format = array_default_format(kind)
else:
format = array_default_format(DEFAULT_DF_FORMAT)
array_chunk.format = "%" + format
if (rows, cols) == (-1, -1):
rows, cols = num_rows, num_cols
elif (rows, cols) == (0, 0):
# return header only
r = min(num_rows, DATAFRAME_HEADER_LOAD_MAX_SIZE)
c = min(num_cols, DATAFRAME_HEADER_LOAD_MAX_SIZE)
array_chunk.headers = header_data_to_thrift_struct(r, c, [""] * num_cols, [(0, 0)] * num_cols, lambda x: DEFAULT_DF_FORMAT, original_df, dim)
array_chunk.data = array_data_to_thrift_struct(rows, cols, None, format)
return array_chunk
rows = min(rows, MAXIMUM_ARRAY_SIZE)
cols = min(cols, MAXIMUM_ARRAY_SIZE, num_cols)
# need to precompute column bounds here before slicing!
col_bounds = [None] * cols
dtypes = [None] * cols
if dim > 1:
for col in range(cols):
dtype = df.dtypes.iloc[coffset + col].kind
dtypes[col] = dtype
if dtype in NUMPY_NUMERIC_TYPES and df.size != 0:
cvalues = df.iloc[:, coffset + col]
bounds = (cvalues.min(), cvalues.max())
else:
bounds = (0, 0)
col_bounds[col] = bounds
else:
dtype = df.dtype.kind
dtypes[0] = dtype
col_bounds[0] = (df.min(), df.max()) if dtype in NUMPY_NUMERIC_TYPES and df.size != 0 else (0, 0)
df = df.iloc[roffset: roffset + rows, coffset: coffset + cols] if dim > 1 else df.iloc[roffset: roffset + rows]
rows = df.shape[0]
cols = df.shape[1] if dim > 1 else 1
def col_to_format(c):
return get_column_formatter_by_type(format, dtypes[c])
iat = df.iat if dim == 1 or len(df.columns.unique()) == len(df.columns) else df.iloc
def formatted_row_elements(row):
return get_formatted_row_elements(row, iat, dim, cols, format, dtypes)
array_chunk.headers = header_data_to_thrift_struct(rows, cols, dtypes, col_bounds, col_to_format, df, dim)
array_chunk.data = array_data_to_thrift_struct(rows, cols, formatted_row_elements, format)
return array_chunk
def dataframe_to_thrift_struct(df, name, roffset, coffset, rows, cols, format):
"""
:type df: pandas.core.frame.DataFrame
:type name: str
:type coffset: int
:type roffset: int
:type rows: int
:type cols: int
:type format: str
"""
dim = len(df.axes)
num_rows = df.shape[0]
num_cols = df.shape[1] if dim > 1 else 1
array_chunk = GetArrayResponse()
array_chunk.slice = name
array_chunk.rows = num_rows
array_chunk.cols = num_cols
array_chunk.type = ""
array_chunk.max = "0"
array_chunk.min = "0"
format = format.replace("%", "")
if not format:
if num_rows > 0 and num_cols == 1: # series or data frame with one column
try:
kind = df.dtype.kind
except AttributeError:
try:
kind = df.dtypes[0].kind
except IndexError:
kind = "O"
format = array_default_format(kind)
else:
format = array_default_format("f")
array_chunk.format = "%" + format
if (rows, cols) == (-1, -1):
rows, cols = num_rows, num_cols
rows = min(rows, MAXIMUM_ARRAY_SIZE)
cols = min(cols, MAXIMUM_ARRAY_SIZE, num_cols)
# need to precompute column bounds here before slicing!
col_bounds = [None] * cols
dtypes = [None] * cols
if dim > 1:
for col in range(cols):
dtype = df.dtypes.iloc[coffset + col].kind
dtypes[col] = dtype
if dtype in NUMPY_NUMERIC_TYPES:
cvalues = df.iloc[:, coffset + col]
bounds = (cvalues.min(), cvalues.max())
else:
bounds = (0, 0)
col_bounds[col] = bounds
else:
dtype = df.dtype.kind
dtypes[0] = dtype
col_bounds[0] = (df.min(), df.max()) if dtype in NUMPY_NUMERIC_TYPES else (0, 0)
df = df.iloc[roffset: roffset + rows, coffset: coffset + cols] if dim > 1 else df.iloc[roffset: roffset + rows]
rows = df.shape[0]
cols = df.shape[1] if dim > 1 else 1
def col_to_format(c):
return format if dtypes[c] in NUMPY_NUMERIC_TYPES and format else array_default_format(dtypes[c])
iat = df.iat if dim == 1 or len(df.columns.unique()) == len(df.columns) else df.iloc
array_chunk.headers = header_data_to_thrift_struct(rows, cols, dtypes, col_bounds, col_to_format, df, dim)
array_chunk.data = array_data_to_thrift_struct(rows, cols,
lambda r: (("%" + col_to_format(c)) % (iat[r, c] if dim > 1 else iat[r])
for c in range(cols)), format)
return array_chunk