Skip to content

ECVL 0.4.0a

Pre-release
Pre-release
Compare
Choose a tag to compare
@MicheleCancilla MicheleCancilla released this 09 Jun 14:21
· 58 commits to master since this release
v0.4.0a
c4fdc36
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.

This pre-release implement a parallel batch loader that instantiates several producers (workers) and allows to load images while running the training code. The loading + augmentation function used by the producers, being virtual, can be easily reimplemented to modify its behavior as desired.
Main changes:

  • Splits can now have names different from training, validation, test and they can count more than three.
  • DatasetAugmentations must be as long as the number of splits that are in the YAML file.
  • In the DLDataset constructor the user can specify whether to automatically drop the last samples of the split if they are not enough to compose a batch, set the number of workers and the ratio that will be used to calculate the maximum dimension of the queue.
  • ResetCurrentBatch became ResetBatch. It can be specified for which split you want to reset the batch counter and if you want to shuffle the batch.
  • ColorType for loaded images in DLDataset is now RGB by default (instead of BGR).
Assets 2
Loading