trainer-py 0.2.0

Lightweight context-based library for benchmarking

Trainer ⏱

Lightweight benchmarking package with zero dependencies

Installation

pip install trainer-py

Usage

from trainer import Trainer

t = Trainer().round(2)

with t('metric1'):
    time.sleep(0.5)

with t('metric2'):
    time.sleep(0.3)

m = t.add_total('total').metrics
print(m)

output:

{
    'metric1': {'start': 1656844808.09, 'end': 1656844808.59, 'interval': 0.5},
    'metric2': {'start': 1656844808.59, 'end': 1656844808.89, 'interval': 0.3},
    'total': {'start': 1656844808.09, 'end': 1656844808.89, 'interval': 0.8}
}

for more, see examples

Features

Contexts

• using Python contexts, you can indent specific parts of the code you want to benchmark
• each measured code part produces a 'metric' dictionary with name (key), start (epoch), end (epoch), interval (epoch)
• all metrics can be retrieved at any point by calling trainer.metrics

Total Execution Time

• you can add the total execution time by executing trainer.add_total() at the end
• by default, the total time considers the first metric's start time as its start, but it is possible to prematurely start measurement by calling trainer.start_measuring() instead

Rounding

• all epoch timestamps can be rounded to any number of decimals using trainer.round(<int>)
• all epoch timestamps can be rounded to full seconds (making them integers instead of floats) by executing trainer.round() or trainer.round(0)

Turning off Trainer

• to turn off trainer, the DummyTrainer class is provided. This class can be switched on, for example, based on a value of an environment variable

    trainer = Trainer() if ENV_VAR else DummyTrainer()

• calling .metrics at the end will always return in an empty dictionary which is falsy in Python. Your following code that uses the metrics dictionary can first do a check for this falsy value

go to examples