Test with single image gives different result in each runtime

[gihunsong]

Hi all,

I pretrained myself using EfficientNet B0 with 5 classes, and..

Now I have been loaded using below code:

model = EfficientNet.from_pretrained('efficientnet-b0', weight_path = './weights_path/model.pt', num_classess=5) and i'm using model.eval(), and with torch.no_grad() also.

In a single test phase, i tried two times inference using below code:

outputs = model(inputs) print(outputs) outputs = model(inputs) print(outputs) both outputs printing gives same result but,

If i run again, it gives different result.

Maybe i'm missing something in weight loading part but I couldn't get guess easily.

Please help me, regards.

[gihunsong]

running on cpu and gpu also gives same situation..

[gihunsong]

It seems like, setting seeds give same result for each time, but i'm trying to use same input (single image) so.. it's not understandable..

[rkakash59]

@gihunsong Even I am facing the same issue.

[rkakash59]

@gihunsong you can try : model = EfficientNet.from_name('efficientnet-b0', num_classes=5) model.load_state_dict(weights_path)

[gihunsong]

@gihunsong you can try : model = EfficientNet.from_name('efficientnet-b0', num_classes=5) model.load_state_dict(weights_path)

@rkakash59 do you mean that you solved the problem that i mentiond using your suggestion?

[rkakash59]

@gihunsong that was a suggestion.

I missed model.eval() during inference so l was getting inconsistent results. Now I am getting consistent results.

[gihunsong]

@gihunsong that was a suggestion.

I missed model.eval() during inference so l was getting inconsistent results. Now I am getting consistent results.

@rkakash59 Thanks for your reply.

Could you share your a .py file that you have been using model.eval() part? It'll be helpful for me..

Because I am also using model.eval() in inference but it gives different result every time when i load model and do inference.

[gihunsong]

@rkakash59 I'm still struggling in this issue.

you suggested the below code: model = EfficientNet.from_name('efficientnet-b0', num_classes=5) model.load_state_dict(weights_path)

could you share your code more in detail upper part of your suggested code?

[ndcuong91]

So weird that no one can explain this issue. My code worked normally with 2 classes but It had problem with 5 classes