goseesomething comments

Results 3 comments of


                                            goseesomething

The provided lr scheduler `LambdaLR` doesn't follow PyTorch's LRScheduler API. You should override the `LightningModule.lr_scheduler_step` hook with your own logic if you are using a custom LR scheduler.

The fix is here. https://github.com/Lightning-AI/lightning/issues/15912 ![image](https://github.com/clovaai/donut/assets/40739576/f57382fc-4158-4ed3-b288-8f958a335214)

CUDA error: device-side assert triggered

try edit learning rate params. It's because the dataset is difference and the params is not the best conditions, that's why the calculation between mask and pred coming out NAN...

using ICDAR2015 dataset train from raw and the model doesn't converge? ICDAR2015 EPO1200 Batch_size 16 不收敛

> > Does anyone face the problem? > > 有人遇到這問題嗎? > > 是可以收敛的，你检查一下你数据集的问题，我换了backbone从头训练，大概600多epochs的时候收敛了感謝回覆~ 後來找到原因了SGD是比較不穩定～的確有時候會不收斂然後也感謝通知是600多epo左右～我下次再訓會留意這個數字！然後之前訓沒有找到對的conda list 後來跟著下面凡Issue內容重整conda list環境就訓得起來了： https://github.com/MhLiao/DB/issues/323 這邊的conda list