Changing batch_size cause loss different

[YingdiZhang]

When I fine tune VGG in Stanford Dog Breed Dataset,I found different batch_size cause loss different.
for example, when I set batch_size = 1,the loss plot looks like this:

but when I set batch_size = 32, the loss plot seems normal and looks like this:

what is the reason of this phenomena?

[ptrblck]

Is it really the loss or do you mean accuracy?
If it’s the loss, both plots look fishy, since the loss should go down.