search

microsoft / CNTK

Microsoft Cognitive Toolkit (CNTK), an open source deep-learning toolkit
https://docs.microsoft.com/cognitive-toolkit/
Other
17.51k stars 4.28k forks source link

Excluding a training sample having specific target label value in DNN training #2369

Open bansungmin opened 7 years ago

bansungmin commented 7 years ago

Using CNTK brainScript I am training feedforward DNN having 1 output node with mean square error(MSE) criterion. "SquareError" block is used in NDL. When calculating stochastic gradient descent, I would like to exclude a training sample in DNN training if it has a specific target label value. It seems that this function is not provided in CNTK, so I want to edit CNTK code. Which source file do I have to check? I am not good at C++ programing, so gentle explanation may be requried. Thank you.

sayanpa commented 7 years ago

We encourage you to using the python API's as they are easier and more up-to-date. BrainScript and NDL are supported in legacy mode. Adding new functionality via BrainScript if you are not familiar with the C++ is something I would not recommend.

bansungmin commented 7 years ago

Thank you for your reply. I am sorry, but I am familiar with the C++ than python API's environment. If you guide me which source code is related to my question, I will choose which method is right for me. In python API's, how training sample having specific target label value can be exluded in DNN training in detail? Is this possible on script lebel?

cha-zhang commented 7 years ago

This should be done by the reader. That is, you don't need to touch CNTK core C++ code.