Open pbsds opened 2 years ago
cmckain
commented
2 years ago I have implemented a similar feature in my own fork of Rich (my own requirements are probably not general enough for direct inclusion into Rich). If anyone needs inspiration, cmckain@rich and take note of locals_to_ignore and clean_values.
willmcgugan
commented
2 years ago How about?
locals_suppress=<callable>
locals_max_length=<optional int
actforjason
commented
2 years ago
locals_suppress=<callable>locals_max_length=<optional int
rich.traceback.install doesn't have these parameters，how do it？
femtomc
commented
2 years ago Bumping this issue -- running into similar problems where locals prints to egregiously large numpy arrays.
Really like having the locals, but having to turn them off when working with arrays is not the best.
Have you checked the issues for a similar suggestions?
318 and #1378 is related
How would you improve Rich?
Showing the local variables in Rich tracebacks is incredibly helpful, but this does at times become too verbose prompting me to disable it. Especially when dealing with types such as
torch.nn.Module, which feature very long__repr__and__str__outputs (example of a single frame featured below). My feature request is twofold: the ability to filter certain local variables entierly, and the ability to suppress/truncate the length of others.Variable names that start with a double underscore prefix (e.g.
__init__) is a prime target for filtering. I imagine parametrizing such behavior inrich.traceback.installwith an optional lambda function (e.g.locals_filterer=lambda x: not x.startswith("__")).For suppressing/truncating the length, one could either fallback to the default
object.__repr__format, or simply clip all__repr__() output tonlines with trailing ellipsis (as dictionaries already seem to do). For deciding which locals to suppress, if not all, one could either provide a list of qualified class names (type(obj).__qualname__) and types (kinda like the existingsuppressargument for modules), or provide a lambda which determines if it should be supressed (e.g.locals_suppress=lambda x: any(isinstance(x, i) for i in list_of_types_to_suppress))Example output (note the `module = MyMachineLearningModel`):
``` │ /home/pbsds/ntnu/xxxxxxxxxxxxxxxxxxxxxxxxxx/yyyyyy/cli.py:332 in log_training_setup │ │ │ │ 329 │ │ if logger is not None: │ │ 330 │ │ │ print(f"{logger.__class__.__qualname__} hparams:") │ │ 331 │ │ │ print_dict_twocolumn(config.logging) │ │ ❱ 332 │ │ │ logging.log_config(logger, logger = {"_class": logger.__class__.__name__} | │ │ 333 │ │ │ │ │ 334 │ │ │ # host info │ │ 335 │ │ │ def cmd(cmd: Union[str, list[str]]) -> str: │ │ │ │ ╭───────────────────────────────────────────────────────── locals ─────────────────────────────────────────────────────────╮ │ │ │ args = Namespace(mode='module', data_dir=What problem does it solve for you?
When running my program on a supercluster, which often feature a queue time of several hours, it sucks for it to immediately crash without me having set
show_locals=Trueinrich.traceback.installsimply due to it being too verbose. Having it be enabled all the time, without it being too noisy, would greatly reduce development friction and put a smile on my face.