experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Remove extract_frame_info_fn argument from extract_stack

Browse Source 
This argument is not used anywhere in TensorFlow, and slows down the extract_stack function by approx 25%. Note that extract_stack is currently the most expensive function when building large graphs.

Profile from before:
tottime cumtime function
8.244   8.911   tf_stack.py:32(extract_stack)
0.393   0.612   tf_stack.py:79(extract_stack_file_and_line)
0.315   0.315   tf_stack.py:54(<lambda>)
0.003   0.007   tf_stack.py:110(convert_stack)

Profile from after:
tottime cumtime function
7.403   7.736   tf_stack.py:32(extract_stack)
0.396   0.621   tf_stack.py:69(extract_stack_file_and_line)
0.003   0.007   tf_stack.py:100(convert_stack)

PiperOrigin-RevId: 240964678
This commit is contained in:
James Keeling 2019-03-29 05:58:32 -07:00 committed by TensorFlower Gardener
parent 5df8ae5cc5
commit c1b7d2be4b
1 changed files with 5 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
tensorflow/python/util/tf_stack.py
View File

@ -29,7 +29,7 @@ TB_FUNCNAME = 2
TB_CODEDICT = 3 # Dictionary of Python interpreter state.
def extract_stack(extract_frame_info_fn=None):
def extract_stack():
"""A lightweight, extensible re-implementation of traceback.extract_stack.
NOTE(mrry): traceback.extract_stack eagerly retrieves the line of code for
@ -38,21 +38,13 @@ def extract_stack(extract_frame_info_fn=None):
should apply _convert_stack to the result to obtain a traceback that can
be formatted etc. using traceback methods.
Args:
extract_frame_info_fn: Optional callable fn(stack_frame) applied to each
stack frame. This callable's return value is stored as the sixth (last)
element of the returned tuples. If not provided, the returned tuples
will have None as their sixth value.
Returns:
A list of 6-tuples
(filename, lineno, name, frame_globals, func_start_lineno, custom_info)
A list of 5-tuples
(filename, lineno, name, frame_globals, func_start_lineno)
corresponding to the call stack of the current thread. The returned tuples
have the innermost stack frame at the end, unlike the Python inspect
module's stack() function.
"""
default_fn = lambda f: None
extract_frame_info_fn = extract_frame_info_fn or default_fn
try:
raise ZeroDivisionError
except ZeroDivisionError:
@ -65,9 +57,7 @@ def extract_stack(extract_frame_info_fn=None):
name = co.co_name
frame_globals = f.f_globals
func_start_lineno = co.co_firstlineno
frame_info = extract_frame_info_fn(f)
ret.append((filename, lineno, name, frame_globals, func_start_lineno,
frame_info))
ret.append((filename, lineno, name, frame_globals, func_start_lineno))
f = f.f_back
ret.reverse()
return ret
@ -123,8 +113,7 @@ def convert_stack(stack, include_func_start_lineno=False):
input tuple.
"""
ret = []
for (filename, lineno, name, frame_globals, func_start_lineno,
unused_frame_info) in stack:
for (filename, lineno, name, frame_globals, func_start_lineno) in stack:
linecache.checkcache(filename)
line = linecache.getline(filename, lineno, frame_globals)
if line: