sampling

The sampling config group determines some behaviour of the dataloader within MultiTaskDataModule as well as the underlying TaskDataset.

full

sample_num

default: 0

• number of samples per epoch

• if 0 will use len(dataset) samples (even if balanced sampling is active)

balanced

default: false

• if true will try to sample equally from each (target) class

• if false samples randomly over the split

• unbalanced sampling might activate weights in loss criterion, see loss.auto_activate_weighing

batch_size

default: 300

• number of samples used in one forward+backward pass

• batch size will be overwritten if you activate tune.bs to automatically tune the batch size

drop_last

default: false

• whether the final (incomplete) batch will be dropped at the end of the epoch

• directly passed to DataLoader

enable_caching

default: false

• if true activates a caching mechanism which trades in RAM usage for less disk access

• only works for already preprocessed datasets!

cache_max_size

default: 10000

• sets a max_size of cache (in terms of images), cache will be disabled if more images are in the datasets loaded

• avoid exploding RAM by keeping low or downsize images during preprocess

• from experience increasing above 10000 does not yield any more benefits, but this may be very case dependent

• for full optimization consider experimenting with num_workers entry of the config