NEWS
automl 1.3.2 (2020-01-16)
- one step forward, one step back!
MAPE cost function is back and validated
RMSE cost function removed (derivative
challenge + no added value vs MSE imo)
- bug fix: 'chkgradevery' is now available
automl 1.3.1 (2020-01-09)
- direct ‘overfitting’ & ‘regularization’
modes for auto_runtype
- batch normalization set to 0 by default
(not good for regression targets)
- bug fix: now exits if no gain when no cros-
s validation sample
- bug fix: testgainunder now kept as is
- MAPE cost function removed (probably an er-
ror to be fixed in derivative: todo list)
automl 1.3.0 (2019-12-15)
- rmse & mape cost function added for regres-
sion
automl 1.2.9
- documentation update (typo correction and
more detailed explanations concerning
modexec and autopar)
thanks to Alexey
automl 1.2.8 (2019-03-16)
- time limit for sub modelizations in automl_
train function, to avoid waiting too long
for a specific particle to finish its mode-
lization
automl 1.2.7 (2019-01-27)
- lightening models produced by removing
unnecessary data (Z, A layers, etc...)
automl 1.2.6 (2018-12-08)
- just a proper Authors@R field 4 CRAN :-)
automl 1.2.5
- automl_train bug fix when particle produces
no exploitable model (continued work)
automl 1.2.4
- automl_train bug fix when particle produces
no exploitable model
automl 1.2.3
- dropout bug fix (now random at each mini
batch and reproductible)
- bug fix: continue training on auto trained model
with manual function
automl 1.2.2
- pkgdown site with vignettes (Rmd file added)
automl 1.2.1
- pkgdown site (nb: Rnw vignettes are not included)
automl 1.2.0 (2018-11-23)
- New param 'mdlref' for automl_train_manual:
to start training from saved model (shape,
weights...) for fine tuning
- New param 'mdlref' for automl_train:
to start training with loaded hpar and autopar
(not the model)
- New 'auto_runtype' autopar for automl_train
to run automatically the 2 steps below;
1: overfitting, goal: performance
2: regularization, goal: generalization
- dropout bug fix
- lambda bug fix
automl 1.1.0
- same train/test sampling for each particle with automl_train
- stick to a format in variables naming
automl 1.0.9
automl 1.0.8 (2018-11-08)
- New automatic hyperparameters adjustments below:
'auto_psovelocitymaxratio' autopar PSO velocity max ratio
'auto_layer' autopar layer shape (layers number, nodes
number per layer, activation types and dropout ratios)
- 'auto_lambda' bug fix
automl 1.0.7
automl 1.0.6 (2018-10-21)
- vignette howto_automl: why, how and basic howto
- 'auto_lambda' autopar regularization hyperparameter
- seed bug fix for reproductibility
- display enhancement in NN structure display
automl 1.0.5 (2018-09-13)
- first public release on CRAN
- There's so much to do; transfert learning, CNN, RNN ...
feel free to join