PyTorch CNN gives different results

[chandra_sutrisno]

I am trying to make a face recognition model with PyTorch. My model performs well with both loss scores for training and validation close to 0. The problem is when I tested it using same input, it gives me different results like below:

and below

The left side is the true label and the right one is the prediction output. I set random seed and ran

model.eval()

before feeding the input to the model.

Does anyone know how to solve this?

Please advise

[Nikronic (Nikan Doosti)]

Hi,

What is the problem? It seems it just doing fine. Could you explain what is your expectation and what is the difference of these values from your expectations?
As I see, it has only one error which is normal as you testing the model and not training and in any model that has not overfitted, test loss/ test accuracy is higher/lower than train mode.

Bests

[chandra_sutrisno]

Hi @Nikronic,

I missed some words, sorry for that. I meant i tested it using the same input and it gave me different results.

[Nikronic (Nikan Doosti)]

So, what I understand is that if you pass input x to model in case of model.train() everything is ok but for the same input x, model.eval() is giving worse result. model.eval() only affect layers such as batchnorm and dropout, so I guess you have such layers in your model definition. I think this post in that case can help alot as this is a tricky challenge sometimes:

Model.eval() gives incorrect loss for model with batchnorm layers

The testing code which is called every 100 training iters is as following: def test_model(model): psnr_test_avg = 0 loss_test_avg = 0 model.eval() for iteration, test_data in enumerate(dataloader_test, 0): label_test = test_data['label_patch'] residue_test = test_data['residue_patch'] stacked_test = test_data['stacked_patch'] microshift_test = test_data['train_patch'] inputs_test, residues_test, labels_test, microshifts_test = Variabl…

There are few tricks to solve issue in the aforementioned thread. I hope it helps.

Bests

[chandra_sutrisno]

Hi @Nikronic

Thank you. I will check later.

[Dicko]

I am hoping somebody can help me because I am banging by head against a brick wall
I am using the RetinaNet notebook (link posted in my post above), after running different experiments my results were getting worse and no where near what I got the first time round so I decided to do a check.
I looked at the model where I got good reults, an mAP of 0.36
So I took tha same model, same notebook, same data, changed NOTHING and my results are terrible mAP 0.15. Has anyone come across this before ? I haven’t changed a thing but I know that something must be different and I can’t seem to find out what it is!.
Thanks