Decreasing validation accuracy in Quantization-aware training?

I see a decreasing validation accuracy in this tutorial:https://pytorch.org/tutorials/advanced/static_quantization_tutorial.html#quantization-aware-training

in the article, validation accuracy decreases from
Epoch 0 :Evaluation accuracy on 300 images, 77.67
to
Epoch 7 :Evaluation accuracy on 300 images, 75.67

Is it continuously decreasing or just for one Epoch? Does it recover?

hi @icyhearts, this is expected for the tutorial as written, as it is using a toy dataset for speed (a tiny subsample of ImageNet, without enough images in each class). You would need to run it on a real dataset to get better numerics on the validation set.