First, fastai has a concept of a test_dl, which no matter what your dataloader was made of, will accept items and preprocess it similar to your validation dataloader. So, we can do something like so:
dl = learn.dls.test_dl([f for f in frames_dict.values()])
preds = learn.get_preds(dl=dl)
preds = preds.argmax(dim=-1)
names = [learn.dls.vocab[i] for i in preds]
I can’t 100% be sure of what you’re doing, but this is a start.