Inconsistencies in ADE20K model inference output for the same input image.

[MartaFdez97]

Hello,
I am doing inference with the ade20k model and I am seeing that for the same input, I get a different output. I have saved the data and model variables to be always the same and yet I still have variability. Why does this happen? Also, I have removed the rescale.
My code:

   import pickle

    with open('data.pkl', 'rb') as f:
        data = pickle.load(f)
    with open('model.pkl', 'rb') as f:
        model = pickle.load(f)
    
    # forward the model
    with torch.no_grad():
        result = model(return_loss=False, rescale=False, **data)
    return result

[uyzhang]

There is a random reasoning part in ham_head, which may fluctuate slightly, but it exists.

[MartaFdez97]

It fluctuates even if the weights are the fixed? I do not get how can It be possible

[uyzhang]

Yes, you can know the detailed Implementation with the code