pyctb 0.0.6

Python Customized Tracebacks

Python Customized Tracebacks

Installation

pip install pyctb 

Synopsis

import pyctb

# add a pre-existing group of custom rendering functions
pyctb.add_group('torch')

# or add your own:
# pyctb.add(your_cls, your_render_func)
# pyctb.add(your_cls2, your_render_func2)
# ...

# for each frame in the traceback:
# display 1 piece of code context before the current executing piece
# display 0 pieces of code context after the current executing piece
pyctb.config(before_ctx=1, after_ctx=0)

# activate the custom tracebacks
pyctb.on()

# your program code below.  any exception will generate a customized traceback
...

# exceptions will now generate the default traceback as before
pyctb.off()

Introduction

pyctb is tool for customizing the information displayed in a Python traceback. It is useful to quickly see the values of arguments or operands used at the point where your program throws an exception, without the need for a debugger. For each traceback frame, pyctb adds extra lines of the form var = render_func(value) for each variable in the currently executing function of that frame, using a render_func registered with that class. By default, object is registered to use repr. The user may register custom functions for rendering specific types of their choice. Such specific rendering functions are needed for types that have too large of an output with just repr.

pyctb relies on the excellent stack_data repo from Alex Hall for inspecting the frame information variable names and values.

You may register a new rendering function using:

pyctb.add(cls, func)

Objects are rendered using the most specific (in method-resolution order) rendering function registered. For example:

pyctb.add(Animal, render_animal)
pyctb.add(Cat, render_cat)
pyctb.add(Tabby, render_tabby)

Then, a Tabby will be rendered with render_tabby. A Persian will use render_cat, and Dog will use render_animal, and so forth. By default, object is registered with repr.

Tests

# compare default and customized tracebacks using the `torch` group of
# renderers.  In each frame of the traceback, <before_ctx> and <after_ctx>
# specify the number of pieces of code context to be displayed before and after
# each currently executing piece.  Using 0 for both <before_ctx> and
# <after_ctx> is equivalent to the default behavior.
python -m tests.torch_test <before_ctx> <after_ctx>

# same but with tensorflow tests.
python -m tests.tf_test <before_ctx> <after_ctx>

The tests.torch_test runs two functions called binop_test and matmul_test. Each function raises an exception, and it is run twice. First with pyctb turned off, and then with it turned on. The default and custom tracebacks are shown as output.

Notice that the custom traceback provides additional name = value lines for all arguments to the function or operand that raised the exception. In this case, using the torch group, the operands were torch.Tensor, and the custom rendering function displays them in a format shape:dtype:device.

Additionally, you can see that there is one piece of leading code context (if it exists in that frame) shown before the currently executing piece.

# Run the torch_tests using 1 piece of leading context and non trailing context
(pytorch181) henry@henry-gs65:pyctb$ python -m tests.torch_test 1 0
========= Default traceback for binop_test =========
Traceback (most recent call last):
  File "/home/henry/ai/projects/pyctb/tests/torch_test.py", line 25, in main
    test()
  File "/home/henry/ai/projects/pyctb/tests/torch_test.py", line 16, in binop_test
    return a / b
RuntimeError: The size of tensor a (5) must match the size of tensor b (6) at non-singleton dimension 1


========= Custom traceback for binop_test =========
1 pieces of leading context
0 pieces of trailing context
Custom Traceback (most recent call last):
  File "/home/henry/ai/projects/pyctb/tests/torch_test.py", line 34, in main
    try:
-->     test()
         test = <function binop_test at 0x7f86bbfeda60>
  File "/home/henry/ai/projects/pyctb/tests/torch_test.py", line 16, in binop_test
    b = torch.randn(3,6,7)
--> return a / b
         a = [3,5,7]:float32:cpu
         b = [3,6,7]:float32:cpu
RuntimeError: The size of tensor a (5) must match the size of tensor b (6) at non-singleton dimension 1


========= Default traceback for matmul_test =========
Traceback (most recent call last):
  File "/home/henry/ai/projects/pyctb/tests/torch_test.py", line 25, in main
    test()
  File "/home/henry/ai/projects/pyctb/tests/torch_test.py", line 9, in matmul_test
    d = torch.matmul(a, b)
RuntimeError: Expected batch2_sizes[0] == bs && batch2_sizes[1] == contraction_size to be true, but got false.  (Could this error message be improved?  If so, please report an enhancement request to PyTorch.)


========= Custom traceback for matmul_test =========
1 pieces of leading context
0 pieces of trailing context
Custom Traceback (most recent call last):
  File "/home/henry/ai/projects/pyctb/tests/torch_test.py", line 34, in main
    try:
-->     test()
         test = <function matmul_test at 0x7f87e7ce8c10>
  File "/home/henry/ai/projects/pyctb/tests/torch_test.py", line 9, in matmul_test
    c = b * 5
--> d = torch.matmul(a, b)
         a = [3,5,7]:float32:cpu
         b = [3,5,7]:float32:cpu
RuntimeError: Expected batch2_sizes[0] == bs && batch2_sizes[1] == contraction_size to be true, but got false.  (Could this error message be improved?  If so, please report an enhancement request to PyTorch.)