Single model vs multiple model shows different Accuracy

Hi Guys,

I’m currently executing several kinds of models at the same time as multiple processes.

and I’m using same Imagenet 1k dataset which inside of my data folder.

the problem is when I do validate the single(mobilenetv2) model and when I do validate with multiple(mobiletv2 + other dnns)models, the accuracy shows slightly different.

I think multiple dataloader looking at the same data folder might be the problem but not sure it is.

Do you know what makes this problem?

I use PyTorch 1.4.0 on NVIDIA Jetson Xavier
I got the data sets and dataloader code from the below page
one thing different is that i set num_worker as a 0.


How reproducible is this behavior, i.e. are you always seeing this issue using e.g. 10 different runs with different seeds?