Strange visualization of the log during the test

[utrobinmv]

🐛 Bug

Hello,

in module 'test_epoch_end' I calculate the final metric

score = self.metric(valid_preds, valid_labels)

print(score)

self.log("Final_metric", score)

logging_info('Test end Ok!')

but after getting the output, I found that the accuracy of the data written to the log differs from the original value

Result

console output:

0.8403041825095058
Test end Ok!

DATALOADER:0 TEST RESULTS
{'Final_metric': 0.8403041958808899}

score 0.8403041825095058 != 0.8403041958808899

I think I can draw the wrong conclusion from such an error in the results?

pytorch-lightning==1.4.1
torch==1.9.0

[tchaton]

Dear @utrobinmv,

Would you mind opening providing a reproducible script using the boring_model: https://github.com/PyTorchLightning/pytorch-lightning/blob/master/pl_examples/bug_report_model.py

It is unlikely we would be able to help otherwise.

Best,
T.C

[qqueing]

this is similar : #8887

[SkafteNicki]

This is due to the precision in float32 vs float64:

print(torch.tensor([0.8403041825095058], dtype=torch.float32).item())
# 0.8403041958808899

print(torch.tensor([0.8403041825095058], dtype=torch.float64).item())
# 0.8403041825095058

[tchaton]

I believe @SkafteNicki provided an answer and unfortunately, there is nothing we can do unless recommending you to use Trainer(precision=64).

Closing this issue.