Hi,
I am working with the Planet dataset. It does not fit my GPU’s RAM (6gb - 1660ti), so I am trying to use the float16 format by using the following code:
learn = cnn_learner(data, arch, metrics=[acc_02, f_score]).to_fp16()
data = (src.transform(tfms, size=256)
.databunch(bs = 20).normalize(imagenet_stats))
learn.data.valid_dl.add_tfm(to_half)
However I did not notice any significant improvements on memory used or training speed. How to verify that both the model and the dataset are converted to float16?