Skip to main content

Personal collection of common utilities

Project description

shinyutils

Various utilities for common tasks. :sparkles: :sparkles: :sparkles:

Setup

Install with pip. Additional features can be enabled with the [<feature>] syntax shown below. Available optional features are:

  • color: color support for logging and argument parsing
  • plotting: support for matplotlib and seaborn
pip install shinyutils  # basic install
pip install "shinyutils[color]"  # install with color support
pip install "shinyutils[color,plotting]"  # install with color and plotting support

Components

subcls

Utility functions for dealing with subclasses.

Functions

  • get_subclasses(cls): returns a list of all the subclasses of cls.
  • get_subclass_names(cls): returns a list of names of all subclasses of cls.
  • get_subclass_from_name(base_cls, cls_name): return the subclass of base_cls named cls_name.

argp

Utilities for argument parsing.

LazyHelpFormatter

HelpFormatter with sane defaults, and colors (courtesy of crayons)! To use, simply pass formatter_class=LazyHelpFormatter when creating ArgumentParser instances.

arg_parser = ArgumentParser(formatter_class=LazyHelpFormatter)
sub_parsers = arg_parser.add_subparsers(dest="cmd")
sub_parsers.required = True
# `formatter_class` needs to be set for sub parsers as well.
cmd1_parser = sub_parsers.add_parser("cmd1", formatter_class=LazyHelpFormatter)

CommaSeparatedInts

ArgumentParser type representing a list of int values. Accepts a string of comma separated values, e.g., '1,2,3'.

InputFileType

FileType restricted to input files, (with '-' for stdin). Returns a file object.

OutputFileType

FileType restricted to output files (with '-' for stdout). The file's parent directories are created if needed. Returns a file object.

InputDirectoryType

ArgumentParser type representing a directory. Returns a Path object.

OutputDirectoryType

ArgumentParser type representing an output directory. The directory is created if it doesn't exist. Returns a Path object.

ClassType

ArgumentParser type representing sub-classes of a given base class. The returned value is a class.

class Base:
    pass

class A(Base):
    pass

class B(Base):
    pass

arg_parser.add_argument("--cls", type=ClassType(Base), default=A)

KeyValuePairsType

ArgumentParser type representing mappings. Accepts inputs of the form str=val,[...] where val is int/float/str. Returns a dict.

shiny_arg_parser

ArgumentParser object with LazyHelpFormatter, and arguments from sub-modules.

logng

Utilities for logging.

build_log_argp

Creates an argument group with logging arguments.

>>> arg_parser = ArgumentParser()
>>> _ = build_log_argp(arg_parser)  # returns the parser
>>> arg_parser.print_help()
usage: -c [-h] [--log-level {DEBUG,INFO,WARNING,ERROR,CRITICAL}]

optional arguments:
  -h, --help            show this help message and exit
  --log-level {DEBUG,INFO,WARNING,ERROR,CRITICAL}

This function is called on shiny_arg_parser when shinyutils is imported.

conf_logging

Configures global logging using arguments returned by ArgumentParser.parse_args. log_level can be over-ridden with the keyword argument. Colors (enabled by default if rich is installed) can be toggled.

args = arg_parser.parse_args()
conf_logging(args)
conf_logging(args, log_level="INFO")  # override `log_level`
conf_logging(use_colors=False)  # disable colors

When imported, shinyutils calls conf_logging without any arguments.

matwrap

Wrapper around matplotlib and seaborn.

MatWrap

from shinyutils.matwrap import MatWrap as mw  # do not import `matplotlib`, `seaborn`

mw.configure()  # this should be called before importing any packages that import matplotlib

fig = mw.plt().figure()
ax = fig.add_subplot(111)  # `ax` can be used normally now

# Use class methods in `MatWrap` to access `matplotlib`/`seaborn` functions.
mw.mpl()  # returns `matplotlib` module
mw.plt()  # returns `matplotlib.pyplot` module
mw.sns()  # returns `seaborn` module

Use mw.configure to configure plots. Arguments (defaults in bold) are:

  • context: seaborn context (paper/poster/talk/notebook)
  • style: seaborn style (white/whitegrid/dark/darkgrid/ticks)
  • font: any font available to fontspec (default Latin Modern Roman)
  • latex_pkgs: additional latex packages to be included before defaults
  • **rc_extra: matplotlib rc parameters to override defaults mw.configure() is called when shinyutils.matwrap is imported.

Use add_parser_config_args to add matwrap config options to an argument parser.

>>> arg_parser = ArgumentParser()
>>> _ = mw.add_parser_config_args(arg_parser, group_title="plotting options")  # returns the parser group
>>> arg_parser.print_help()
usage: -c [-h] [--plotting-context {paper,notebook,talk,poster}]
          [--plotting-style {white,dark,whitegrid,darkgrid,ticks}]
          [--plotting-font PLOTTING_FONT]
          [--plotting-latex-pkgs PLOTTING_LATEX_PKGS [PLOTTING_LATEX_PKGS ...]]
          [--plotting-rc-extra PLOTTING_RC_EXTRA]

optional arguments:
  -h, --help            show this help message and exit

plotting options:
  --plotting-context {paper,notebook,talk,poster}
  --plotting-style {white,dark,whitegrid,darkgrid,ticks}
  --plotting-font PLOTTING_FONT
  --plotting-latex-pkgs PLOTTING_LATEX_PKGS [PLOTTING_LATEX_PKGS ...]
  --plotting-rc-extra PLOTTING_RC_EXTRA

group_title is optional, and if omitted, matwrap options will not be put in a separate group. When shinyutils.matwrap is imported, this function is called on shiny_arg_parser.

Plot

Plot is a wrapper around a single matplotlib plot, designed to be used as a context manager.

from shinyutils.matwrap import Plot

with Plot(save_file, title, sizexy, labelxy, logxy) as ax:
  ...

Only the save_file argument is mandatory. When entering the context, Plot returns the plot axes, and when leaving, the plot is saved to the provided path.

pt

Utilities for pytorch.

PTOpt

Wrapper around pytorch optimizer and learning rate scheduler.

from shinyutils.pt import PTOpt
opt = PTOpt(
        weights,  # iterable of parameters to update
        optim_cls,  # subclass of torch.optim.Optimizer
        optim_params,  # arguments to initialize optim_cls
        lr_sched_cls,  # subclass of torch.optim.lr_scheduler._LRScheduler
        lr_sched_params,
)
...
opt.zero_grad()
loss.backward()
opt.step()

lr_sched_ arguments are optional, and control the learning rate schedule. The class can also be used with argument parsers.

>>> arg_parser = ArgumentParser(formatter_class=LazyHelpFormatter)
>>> PTOpt.add_parser_args(
        arg_parser,
        arg_prefix="test",  # all options will be prefixed with "test-"
        group_title="pt test",  # if None, separate group will not be created
        default_optim_cls=Adam,
        default_optim_params=None,  # if None, default is an empty dict
        add_lr_decay=True,
    )
>>> arg_parser.print_help()
options:
  -h, --help                            show this help message and exit (optional)

pt test:
  --test-optim-cls cls                  ({Adadelta / Adagrad / Adam / AdamW / SparseAdam /
                                          Adamax / ASGD / SGD / Rprop / RMSprop / LBFGS}
                                          default: Adam)
  --test-optim-params str=val[,...]     (default: {})
  --test-lr-sched-cls cls               ({LambdaLR / MultiplicativeLR / StepLR / MultiStepLR /
                                        ExponentialLR / CosineAnnealingLR / CyclicLR /
                                        CosineAnnealingWarmRestarts / OneCycleLR} optional)
  --test-lr-sched-params str=val[,...]  (default: {})
>>> args = arg_parser.parse_args(...)
>>> opt = PTOpt.from_args(weights, args, arg_prefix="test")

PTOpt can also add help options to argument parsers to display signatures for optimizer and learning rate schedule classes.

>>> arg_parser = ArgumentParser()
>>> PTOpt.add_help(arg_parser)
>>> arg_parser.print_help()
usage: -c [-h] [--explain-optimizer EXPLAIN_OPTIMIZER]
          [--explain-lr-sched EXPLAIN_LR_SCHED]

optional arguments:
  -h, --help            show this help message and exit

pytorch help:
  --explain-optimizer EXPLAIN_OPTIMIZER
                        describe arguments of a torch optimizer
  --explain-lr-sched EXPLAIN_LR_SCHED
                        describe arguments of a torch lr scheduler

>>> arg_parser.parse_args(["--explain-optimizer", "Adam"])
Adam(params, lr=0.001, betas=(0.9, 0.999), eps=1e-08, weight_decay=0, amsgrad=False)
...

The help options are added to shiny_arg_parser when shinyutils.pt is imported.

FCNet

FCNet is a template class for fully connected networks.

from shinyutils.pt import FCNet
net = FCNet(
        in_dim,  # input dimension
        out_dim,  # output dimension
        hidden_sizes,  # list of hidden layer sizes
        hidden_act,  # hidden layer activations (default relu)
        out_act,  # output layer activation (default None)
)

Like PTOpt, this class also supports construction through command line arguments.

>>> arg_parser = ArgumentParser(formatter_class=LazyHelpFormatter)
>>> FCNet.add_parser_args(
        arg_parser,
        arg_prefix="test",
        group_title="fcnet",
        default_indim=None,  # None means the option is mandatory
        default_outdim=1,
        default_hidden_sizes=None,
        default_hidden_act=F.relu,
        default_out_act=None,  # here, None means no output activation
)
>>> arg_parser.print_help()
options:
  -h, --help                           show this help message and exit (optional)

fcnet:
  --test-fcnet-indim int               (required)
  --test-fcnet-outdim int              (default: 1)
  --test-fcnet-hidden-sizes int,[...]  (required)
  --test-fcnet-hidden-act func         (default: relu)
  --test-fcnet-out-act func            (optional)
>>> args = arg_parser.parse_args(...)
>>> net = FCNet.from_args(args, arg_prefix="test")

NNTrainer

This class trains a model on a dataset, and accepts multiple dataset "formats".

from shinyutils.pt import *
nn_trainer = NNTrainer(
    batch_size,  # only mandatory argument
    data_load_workers,  # default 0
    shuffle,  # default True
    pin_memory,  # default True
    drop_last,  # default True
    device,  # default cuda if available else cpu
)
nn_trainer.set_dataset(
    dataset,  # can be a torch Dataset, a tuple of torch Tensors, or a tuple of numpy arrays
)
model = FCNet(...)
opt = PTOpt(...)
loss_fn = torch.nn.functional.mse_loss
nn_trainer.train(model, opt, loss_fn, iters)

SetTBWriterAction

argparse action to create a tensorboard summary writer. The writer is stored in the tb_writer attribute of the argument namespace; this can be overridden by setting SetTBWriterAction.attr. The usage is shown below with the tensorboard option that is added to shiny_arg_parser on importing the module.

shiny_arg_parser.add_argument(
    "--tb-dir",
    type=OutputDirectoryType(),
    help="tensorboard log directory",
    default=None,
    action=SetTBWriterAction,
)
shiny_arg_parser.set_defaults(**{SetTBWriterAction.attr: Mock(SummaryWriter)})

shiny_arg_parser.tb_writer will contain a SummaryWriter like object. If no log directory is provided through the command line, this object will be a dummy. So tensorboard functions can be called on the writer without extra checks.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

shinyutils-5.0.2.tar.gz (22.9 kB view hashes)

Uploaded Source

Built Distribution

shinyutils-5.0.2-py3-none-any.whl (19.0 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page