1. corrected the parameter list for the Delay command in NDL.
2. added the options for the rmprop gradient update type.
3. added the ConvertDBN command.
This commit is contained in:
Dong Yu 2015-01-04 23:54:16 -08:00
Родитель e773dc6264
Коммит 6bd6dbeb47
5 изменённых файлов: 14334 добавлений и 14099 удалений

1
.gitignore поставляемый
Просмотреть файл

@ -159,3 +159,4 @@ $RECYCLE.BIN/
*.lyx~
*.bak
*.lyx#

Разница между файлами не показана из-за своего большого размера Загрузить разницу

Просмотреть файл

@ -115,8 +115,8 @@ Jie Gao, Avner May, Baolin Peng, Andreas Stolcke, Malcolm Slaney
\end_layout
\begin_layout Date
MSR-TR-2014-112 (DRAFT v0.3: Dec.
11, 2014)
MSR-TR-2014-112 (DRAFT v0.4: Jan.
4, 2015)
\end_layout
\begin_layout Standard

Просмотреть файл

@ -3265,7 +3265,7 @@ status open
\begin_layout Plain Layout
Delay(m, [delayTime=1, defaultPastValue=0.1])
Delay(rows, [cols], m, [delayTime=1, defaultPastValue=0.1])
\end_layout
\end_inset
@ -3273,6 +3273,18 @@ Delay(m, [delayTime=1, defaultPastValue=0.1])
\end_layout
\begin_layout Itemize
rows - the number of rows in the delay node (and in the input matrix).
This parameter is needed because under some loopy conditions the dimensions
cannot be automatically inferred from the input matrix.
\end_layout
\begin_layout Itemize
cols - the number of columns in the delay node (and in the input matrix).
This parameter is optional since it will be set based on the minibatch
size during training and testing.
\end_layout
\begin_layout Itemize
m - input matrix to be delayed.
Each column is a sample.

Просмотреть файл

@ -1789,9 +1789,113 @@ gradUpdateType
: gradient update type.
Valid values are None (default, no special treatment to the gradient),
AdaGrad, and RmsProp.
When gradUpdateType equals to RmsProp, you can control the behavior of
the gradient update using following parameters:
\end_layout
\begin_deeper
\begin_layout Itemize
\emph on
rms_wgt_inc
\emph default
\begin_inset Index idx
status open
\begin_layout Plain Layout
rms_wgt_inc
\end_layout
\end_inset
: multiplicative increment of the learning rate scale.
Default is 1.2.
\end_layout
\begin_layout Itemize
\emph on
rms_wgt_dec
\emph default
\begin_inset Index idx
status open
\begin_layout Plain Layout
rms_wgt_dec
\end_layout
\end_inset
: multiplicative decrement of the learning rate scale.
Default is 0.75.
\end_layout
\begin_layout Itemize
\emph on
rms_wgt_max
\emph default
\begin_inset Index idx
status open
\begin_layout Plain Layout
rms_wgt_max
\end_layout
\end_inset
: maximum learning rate scale allowed.
A value closer to 1 makes the learning rate adjustment more stable but
slower.
Default is 10.
\end_layout
\begin_layout Itemize
\emph on
rms_wgt_min
\emph default
\begin_inset Index idx
status open
\begin_layout Plain Layout
rms_wgt_min
\end_layout
\end_inset
: minimum learning rate scale allowed.
A value closer to 1 makes the learning rate adjustment more stable but
slower.
Default is 0.1.
\end_layout
\begin_layout Itemize
\emph on
rms_gamma
\emph default
\begin_inset Index idx
status open
\begin_layout Plain Layout
rms_gamma
\end_layout
\end_inset
: smoothing factor used to estimate the moving average of the variance.
The smaller the value, the quicker it forgets the past information.
Default is 0.99.
\end_layout
\end_deeper
\begin_layout Itemize
\emph on
@ -4379,6 +4483,60 @@ minibatchSize
– the minibatch size to use when creating the label mapping file.
\end_layout
\begin_layout Section
ConvertDBN Command
\begin_inset Index idx
status open
\begin_layout Plain Layout
ConvertDBN Command
\end_layout
\end_inset
\end_layout
\begin_layout Standard
This command is used to convert a model generated by Microsoft's dbn.exe
tool to a CNTK model.
This command is useful when you want to compare the performance of these
two tools (dbn.exe only supports simple fully connected deep neural networks),
port existing models trained with dbn.exe to CNTK, or if you want to use
the RBM pre-training which is available in dbn.exe but not in CNTK right
now.
The related parameters are
\end_layout
\begin_layout Itemize
modelPath
\begin_inset Index idx
status open
\begin_layout Plain Layout
modelPath
\end_layout
\end_inset
– the full path of the generated CNTK model.
\end_layout
\begin_layout Itemize
dbnModelPath
\begin_inset Index idx
status open
\begin_layout Plain Layout
dbnModelPath
\end_layout
\end_inset
– the full path of the model to be converted.
\end_layout
\begin_layout Section
Additional Top-Level Configurations
\end_layout