Unpin pytorch-lightning<1.2

[akihironitta]

pytorch-lightning<1.2 was added in #549 to avoid the following errors.

Description of errors with pytorch-lighting>=1.2

On the current master, the following tests:

test_logistic_regression_model
test_cli_run_log_regression[--max_epochs 1 --max_steps 2]
test_cli_run_self_supervised_simclr[--data_dir /home/runner/work/pytorch-lightning-bolts/pytorch-lightning-bolts/datasets --gpus 0 --fp32 --max_epochs 1 --max_steps 3 --fast_dev_run 1 --batch_size 2 --online_ft]
test_cli_run_self_supervised_byol[ --data_dir /home/runner/work/pytorch-lightning-bolts/pytorch-lightning-bolts/datasets --max_epochs 1 --max_steps 3 --fast_dev_run 1 --batch_size 2 --num_workers 0 --online_ft]
test_cli_run_self_supervised_simsiam[ --dataset cifar10 --data_dir /home/runner/work/pytorch-lightning-bolts/pytorch-lightning-bolts/datasets --max_epochs 1 --max_steps 3 --fast_dev_run 1 --batch_size 2 --num_workers 0 --online_ft --gpus 0 --fp32]

fail with the error:

ValueError: The `preds` should be probabilities, but values were detected outside of [0,1] range.

(Found while fixing ssl tests in #541.)

[Borda]

Yeah, we shall get compatibility with PL 1.2

[mariaculman18]

I got the same error, to solve it:

pip uninstall pytorch-lightning
pip install pytorch-lightning==1.1.8

[Borda]

@ananyahjha93 any update here?

[kaushikb11]

@ananyahjha93 Could you look into it? Would like to try new Lightning features with models in Bolts.

[joseporiolayats]

Hi, still not working with main and nightly version. It works with 1.1.8

[summelon]

Any solution with v1.2.6?
Indeed it works on v1.1.8.
While some features I want lose in v1.1.8

[mariaculman18]

Still not working for v1.3.1

I wanted to use CombinedLoader feature, but had to go back to v1.1.8.

[summelon]

Still not working for v1.3.1

I wanted to use CombinedLoader feature, but had to go back to v1.1.8.

acc_metric = torchmetrics.Accuracy()
...
acc_now = acc_metric(preds.argmax(dim=1), labels)

For now, this works well for me.

[mariaculman18]

Still not working for v1.3.1
I wanted to use CombinedLoader feature, but had to go back to v1.1.8.

acc_metric = torchmetrics.Accuracy()
...
acc_now = acc_metric(preds.argmax(dim=1), labels)

For now, this works well for me.

Thanks! Now working on the latest version :D

[talhaanwarch]

Still not working for v1.3.1
I wanted to use CombinedLoader feature, but had to go back to v1.1.8.

acc_metric = torchmetrics.Accuracy()
...
acc_now = acc_metric(preds.argmax(dim=1), labels)

For now, this works well for me.

but this would not work for multilabel F1 score

[Borda]

Mind update to the latest TM, shall be fixed and all logits

[yeus]

Mind update to the latest TM, shall be fixed and all logits

almost. Version 0.4.1 still has the issue with Lightning-AI/torchmetrics#354:

ValueError: The highest label in preds should be smaller than num_classes.

although that issue seems to be fixed in master...

[Borda]

although that issue seems to be fixed in master...

new TM release is coming next week :]