These examples demonstrate several ways to train and evaluate acoustic models using CNTK.
Below is a brief description of the examples.
** Note that these examples are designed to demonstrate functionality of CNTK. The particular parameter settings do not necessarily result in state of the art performance. **
To Use:
=======
Modify the following files:
* globals.config in "configs" to reflect your current experimental setup)
* modify "DeviceNumber" in globals.config to specify CPU (<0) or GPU (>=0)
* all SCP files (lists of files) in "lib/scp" to point to your feature files
Run the command line with both globals.config and the desired config, separated by a +
* for example: cntk configFile=globals.config+rnnlu.config
* note that full paths to config files need to be provided if you are not inside the config directory
* for example
* C:\dev\cntk5\x64\release\CNTK.exe configFile=C:\dev\cntk5\ExampleSetups\SLU\globals.config+C:\dev\cntk5\ExampleSetups\SLU\rnnlu.config
Scoring
* ./score.sh
* however, need to supply feature and lable files, which are not included in this
* distribution due to copy right issues.
Path Definitions:
=================
* globals.config [defines paths to feature and label files and experiments]
Check training loss
==========================
$ grep Finish log_LSTM_LSTMTest.log
Finished Epoch[1]: [Training Set] Train Loss Per Sample = 0.62975813 EvalErr Per Sample = 0.62975813 Ave Learn Rate Per Sample = 0.1000000015 Epoch Time=5250.689
Finished Epoch[1]: [Validation Set] Train Loss Per Sample = 0.2035009 EvalErr Per Sample = 0.2035009
------ code changed and the following need to be verified ----
------ May 29 2015
--------------------------------------------------------------
Network Training Examples:
==========================
* rnnlu.config
# iter 10, learning rate 0.1
accuracy: 98.01%; precision: 93.75%; recall: 94.04%; FB1: 93.89
# iter 20, learning rate 0.1
accuracy: 98.04%; precision: 94.05%; recall: 94.15%; FB1: 94.10
# iter 30, leraning rate 0.1
accuracy: 98.03%; precision: 94.05%; recall: 94.15%; FB1: 94.10
802b2194b5