rich-logger

Table logger using Rich, aimed at Pytorch Lightning logging

Features

• display your training logs with pretty rich tables
• describe your fields with goal ("higher_is_better" or "lower_is_better"), format and name
• a field descriptor can be matched with any regex
• a field name can be computed as a regex substitution
• works in Jupyter notebooks as well as in a command line
• integrates easily with Pytorch Lightning

Demo

from rich_logger import RichTablePrinter
import time
import random
from tqdm import trange

logger_fields = {
    "step": {},
    "(.*)_precision": {
        "goal": "higher_is_better",
        "format": "{:.4f}",
        "name": r"\1_p",
    },
    "(.*)_recall": {
        "goal": "higher_is_better",
        "format": "{:.4f}",
        "name": r"\1_r",
    },
    "duration": {"format": "{:.1f}", "name": "dur(s)"},
}


def optimization():
    printer = RichTablePrinter(key="step", fields=logger_fields)
    printer.hijack_tqdm()

    t = time.time()
    for i in trange(10):
        time.sleep(random.random() / 3)
        printer.log(
            {
                "step": i,
                "task_precision": i / 10.0 if i < 5 else 0.5 - (i - 5) / 10.0,
            }
        )
        time.sleep(random.random() / 3)
        printer.log(
            {
                "step": i,
                "task_recall": 0.0 if i < 3 else (i - 3) / 10.0,
                "duration": time.time() - t,
            }
        )
        printer.log({"test": i})
        t = time.time()
        for j in trange(5):
            time.sleep(random.random() / 10)

    printer.finalize()


optimization()

Use it with PytorchLightning

from rich_logger import RichTableLogger

trainer = pl.Trainer(..., logger=[RichTableLogger(key="epoch", fields=logger_fields)])