Derk Mus

The tool for GPU utilization is Grafana. This uses the existing samplers, no need for reimplementation. How this custom queue works is: 1. Select a random subject (in each worker)...

> Nice. Using a map-style dataset would ensure that patches are extracted from each subject once, and would let us define an epoch, right? Yes that is right > I'm...

I will test and work it out a bit more and then submit a PR

To come back to this issue. The implementation now looks like this ```python class BatchedPatchesDataset(IterableDataset): def __init__(self, subjects_datasets, weights, sampler, samples_per_volume): self.subjects_datasets = subjects_datasets self.weights = weights self.sampler = sampler...

1. Yes, this can be changed to accept both one dataset and a list of datasets with corresponding weights. 2. This would be difficult I think (and for more complex...

Implementation is updated in the meantime to use the Torch DataPipes (see https://github.com/pytorch/data): ```python from itertools import islice import torchio as tio from torch.utils.data import DataLoader from torch.utils.data import IterDataPipe...

@fepegar changes look good. only minor comment is access to the private attribute _loaded outside the class, but not really part of this merge request

yes, in the test. If python would have real support for private/protected attribute it is not even possible to access it in the test. See also comment here for example:...

And any new estimate? :)

I have a question about this test in test_modeling_common.py ```python def test_forward_signature(self): config, _ = self.model_tester.prepare_config_and_inputs_for_common() for model_class in self.all_model_classes: model = model_class(config) signature = inspect.signature(model.forward) # signature.parameters is an...