pytorch - train and validation metrics and loss using torchvision - Stack Overflow

admin管理员组

文章数量:1287523

I am new to pytorch and torchvision.

Apparently, the __call__ method of a model returns different things if they are on train mode or not. The first case returns the loss, the second returns the outputs.

This is quite bothersome for two reasons:

1. I need to do very strange things to compute the validation loss to plot the history as one will do with TF,
   • either I have a big error if I have layers like dropout or batch norm
   • or using a function of 100 lines of code.

source: .

2. Also, if I want to use torchmetrics with the train data, I need to change to eval and send again the train_dataloader. Seems like a waste of computation time.

Is this indeed the case? Otherwise, how can I do this two things using best practices? My code looks quite messy with all this.

I am new to pytorch and torchvision.

Apparently, the __call__ method of a model returns different things if they are on train mode or not. The first case returns the loss, the second returns the outputs.

This is quite bothersome for two reasons:

1. I need to do very strange things to compute the validation loss to plot the history as one will do with TF,
   • either I have a big error if I have layers like dropout or batch norm
   • or using a function of 100 lines of code.

source: https://stackoverflow/a/71315672/5931672.

2. Also, if I want to use torchmetrics with the train data, I need to change to eval and send again the train_dataloader. Seems like a waste of computation time.

Is this indeed the case? Otherwise, how can I do this two things using best practices? My code looks quite messy with all this.

• pytorch
• torchvision

Share Improve this question Follow asked Feb 24 at 10:17 J Agustin BarrachinaJ Agustin Barrachina 4,10033 gold badges4242 silver badges6767 bronze badges 1

• What version of torchvision are you using? Looking at the code for rcnn for example it seems to return losses so long as targets are passed – Karl Commented Feb 24 at 19:23

Add a comment  | 

1 Answer 1

Sorted by: Reset to default 0

Ok, apparently I'm not the only one to think this is not good, and many others had the same issue:

• Add an option for Dection Models to return cost in eval mode as well (validation, test loss) · Issue #1574
• [models] Suggestion of GeneralizedRCNN forward output · Issue #1775
• Calculate Training Accuracy on resnet152 · Issue #2578

本文标签： pytorchtrain and validation metrics and loss using torchvisionStack Overflow