Function: RNNForward -> Feature Not Implemented.

[ChangLiangLiuu]

I am doing evaluation for a BLSTM model using CPU. I am using the C++ CNTK API for the evaluation. However I got this error Inside File: Source/Math/Matrix.cpp Line: 4423 Function: RNNForward -> Feature Not Implemented.

Does the Evaluation dll support BLSTM evaluation on CPU?

[guoli-ye]

BLSTM is only supported by GPU decoding in my experience.

-GUoli

From: Changliang Liu [mailto:[email protected]] Sent: Friday, December 22, 2017 10:06 AM To: Microsoft/CNTK [email protected] Cc: Subscribed [email protected] Subject: [Microsoft/CNTK] Function: RNNForward -> Feature Not Implemented. (#2798)

I am doing evaluation for a BLSTM model using CPU. I am using the C++ CNTK API for the evaluation. However I got this error Inside File: Source/Math/Matrix.cpp Line: 4423 Function: RNNForward -> Feature Not Implemented.

Does the Evaluation dll support BLSTM evaluation on CPU?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FMicrosoft%2FCNTK%2Fissues%2F2798&data=02%7C01%7Cguoye%40microsoft.com%7Ca8eb84536df143e6c2d208d54966b107%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636495627762211877&sdata=8bAcPuxDXbXvVoVHUW%2BFxo7mJ%2BCHpSRBhOvraifBmcw%3D&reserved=0, or mute the threadhttps://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAMmbcaQBjfEjEUAG3n_poy26ljs8lu-1ks5tC--VgaJpZM4RLSkf&data=02%7C01%7Cguoye%40microsoft.com%7Ca8eb84536df143e6c2d208d54966b107%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636495627762211877&sdata=rpzSPns3Hd8NY27bxcfAsy5WzJAfV1Rw%2B0pYzzvz7qc%3D&reserved=0.

[ChangLiangLiuu]

Is there any plan to implement this ?

[LLLLuo]

I also meet the same problem when I doing evaluation for a BLSTM model using CPU, did you solve this problem?

[timolohrenz]

I also use the CNTK C++ API to evaluate a BLSTM model. In my case it doesn't even throw some error, but gives me obviously wrong model outputs.

Please add CPU support for evaluating BLSTMs.

[j-a-lewis]

In a post in another thread I describe implementing the alternative code suggested for a CPU-only machine in the CNTK module throwing the exception. Alternative code seems to give a reasonable BLSTM acoustic modeling fit in about 10 epochs vs. the ~100 epochs required with a DNN feedforward network for the same data. https://github.com/Microsoft/CNTK/issues/1380#issuecomment-582265598