All notable changes to this project will be documented in this file.
[0.1.3] - 2018-10-22
identify.get_apozfails to identify all generators and resulted high memory usage (possibly an infinite loop) when calculating layer outputs
identify.get_apozcalculates incorrect apoz when layers are re-used
[0.1.2] - 2018-10-04
- Bug when pruning Conv layers with smaller input than filter shape (identified in a VGG-like architecture).
[0.1.1] - 2018-08-07
- Updated to work with keras >= 2.2.0
- No longer triggers numpy FutureWarning: "Using a non-tuple sequence for multidimensional indexing is deprecated".
- Updated flowers example to work with latest keras and tensorflow. It probably won't work with old versions now.
0.1.0 - 2018-05-15
- Deleting all neurons in a layer now removes the whole branch. If there is only one branch, this will cause an error to be raised.
- Enabled resuming pruning from the last checkpoint in
inception_flowers_pruneexample. This is now the default behaviour.
- Added a changelog!
- Updated to work with Keras >= 2.1.3.
- Massively reduced test time.
- Updated tox.ini to test many combinations of keras and tensorflow versions.
- Support for Keras < 2.1.3 will be removed in a future release.
- Shared layers are no longer broken when
delete_channelsis applied to upstream layers·
- Added memory cleanup to
inception_flowers_pruneexample; prevents memory leak.