v1.0.0-alpha Release
syncrostone
released this
13 Jul 03:58
·
24 commits
to master
since this release
This is the alpha pre-release of the stable v1.0.0 release. Below is a summary of changes and updates from the previous release, v0.4.
New Features
- Exhaustive documentation now available!
- Now compatible with both TF1 and TF2 (still uses tensorflow graphs).
- Add fixed weights feature-- can choose not to train certain weights in W_rec, W_in, W_out.
- Revamps Simulation to make it relevant and functional.
- Adds new MatchToCategory task.
- Add support for custom loss and regularization functions.
Other New Features
- Added options for constant uniform, constant gaussian, glorot gaussian, glorot uniform initializations.
- Automatically build and initialize in get_weights, save, test, train, train_curric. No model.build() calls needed from now on.
- Added train_curric function.
Breaking Changes
- No long supports TF < 1.13.1
- Refactors RDM Task--> PerceptualDiscrimination & changes some functionality.
- Refactors Romo Task--> DelayedDiscrimination & changes some functionality.
- Refactors Curriculum's self.stopTraining --> self. stop_training. This change should not affect most users since it is a variable used internally only by curriculum.py and rnn.py
- Removes sussillo regularizer because it was specific to the transfer-function used. Can be defined and used via the custom regularizer.
- Remove get_initial_weights function. Use get_weights instead.
Other changes (may affect results)
- Default random initialization changed to glorot gaussian from constant uniform. This may affect results.
- Unexpected behavior in Dale's fixed and Dale's setup moved into WeightInitializer. This should not affect most users but may affect results.
- Default training iters changed from 10000 to 50000
Other Changes (should not affect users)
- BasicScan internal implementation slightly change. Should not affect users unless users inherited BasicScan to make a custom Model.
- Use get_effective to get the weight matrices to regularize. Should not affect users.
Bug Fixes
- Bug in binary cross entropy loss function fixed.