weekly daily-dev to dev merge #90
Conversation
regular dev to master merge
- added stratified sampling - added latin hypercube sampling with sudoku constraint - added latin hypercub e - commented several functions - moved shuffle as an option in sample_reducer - added 'random_method' parameter to Scan() - cleaned up and modified README.md - added latin hypercube sampling with sudoku constraint submodule - added the new features to tests
# the new Randomizer() supports: - Quantum randomness (vacuum based) - Ambient Sound based randomness - Sobol sequences - Halton sequences - Latin hypercube - Improved Latin hypercube - Latin hypercube with a Sudoku-style constraint - Uniform Mersenne - Cryptographically sound uniform - Reduced testing script rounds to 5 from 10 - FIX: a bug in sample_reduce that prevented stratification from working - TODO: need to think how stratification works with various randomizers and if it makes sense to have a separate option or random_method as it is now Also updated the README.md
|
Hello @mikkokotila! Thanks for updating the PR.
Comment last updated on October 01, 2018 at 21:17 Hours UTC |
Pull Request Test Coverage Report for Build 234
💛 - Coveralls |
fix a missing SSL cert problem for quantum random number API call
regular master >> daily-dev
- improved / added docstrings - pepified - added Randomizer as a stand-alone package - turned "clear_tf_session" True by default in Scan() - added tf_config function to /utils as a new home for tf resource management functionality
This is now moved into its own package in a separate repo and installs with 'pip install chances'.
TrainingLog adds a callback function with live view of the training process against any metric.
- pepification - moved 'model' to the first parameters in Scan() - made 'dataset_name' and 'experiment_no' optional - 'dataset_name' becomes timestamp with second accuracy if None - added 'print_params' option which prints the params every round
- created custom exceptions / error handling for the most common (and annoying) cases - added an experimental best model saving - created exceptions.py in /utils - fixed an issue in reporting.py with table() where both metric and sort_by needed to be separately stated unless metric was val_acc
- creates two new objects 'saved_weights' and 'saved_models' - stores the model (as json) and the model weights for each round - added new class Predict() in /utils/predict.py
- added kfold splitting to /utils/validation_split.py - added f1-score based model validation
See individual commits for more.