-
Notifications
You must be signed in to change notification settings - Fork 12
/
NEWS
executable file
·121 lines (103 loc) · 3.56 KB
/
NEWS
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
automl 1.3.2
============================================
* 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
============================================
* 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
============================================
* 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
============================================
* 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
============================================
* lightening models produced by removing
unnecessary data (Z, A layers, etc...)
automl 1.2.6
============================================
* 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
============================================
* 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
============================================
* testcvsize = 0 bug fix
automl 1.0.8
============================================
* 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
============================================
* vignette completion
automl 1.0.6
============================================
* 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
============================================
* first public release on CRAN
* There's so much to do; transfert learning, CNN, RNN ...
feel free to join