Slight differences in softmax outputs depending on batch size

[mikkoim]

How to reproduce:

Run 03_predict.py with different batch sizes and compare the outputs.

Probably due to something in the timm-models?