Long decimals for rounded values in logs

[jedick]

I wonder if there's a way to round or truncate values in the test logs, specifically the results printed on screen. The logs have many decimal digits for rounded values. Here's a small example:

import pytorch_lightning as pl
from pytorch_lightning.loggers import CSVLogger

class MyModel(pl.LightningModule):
    def __init__(self):
        super().__init__()

    def test_step(self, batch, batch_idx):
        pass

    def on_test_epoch_end(self):
        strs = ["80.000", "80.100", "80.125", "80.200", "80.500", "80.625"]
        for str in strs:
            self.log(str, float(str))

model = MyModel()
logger = CSVLogger("logs")
trainer = pl.Trainer(logger=logger)
trainer.test(model, dataloaders = [0])

The screen output looks like this, and the same values are in the CSV file.

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃        Test metric        ┃       DataLoader 0        ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│          80.000           │           80.0            │
│          80.100           │     80.0999984741211      │
│          80.125           │          80.125           │
│          80.200           │     80.19999694824219     │
│          80.500           │           80.5            │
│          80.625           │          80.625           │
└───────────────────────────┴───────────────────────────┘

This is related to floating-point representation. Decimals that are sums of negative powers of two (1/2 + 1/8 = 0.625) are represented exactly, but 80.1 is represented as 80.0999984741211.

I could just save the unrounded values to CSV and format them myself, but I would like to have cleaner output here as well. I think nice formatting of metrics helps with interpreting results of experiments and sharing with colleagues.

There is a related open discussion but with a different problem (undesired rounding to the nearest integer): #19496