Microsoft/CNTK

Microsoft Cognitive Toolkit (CNTK), an open source deep-learning toolkit http://cntk.ai

Kerry-Cho

@kerry-Cho

Is it normal?

delzac

@delzac

It shouldn't, did any error appear?

Kerry-Cho

@kerry-Cho

yes

no error just stop

delzac

@delzac

You need to do some debugging now i guess.

Kerry-Cho

@kerry-Cho

Okay. Thank you.
Can I debug and ask again?

delzac

@delzac

sure :)

Kerry-Cho

@kerry-Cho

@delzac Is there a Unet example in cntkx?

Kerry-Cho

@kerry-Cho

Found!

Kerry-Cho

@kerry-Cho

hi. how to used copy the Weight value of the PretrainedModel?

delzac

@delzac

@kerryJo you can take a look at this https://github.com/Microsoft/CNTK/blob/master/Examples/Image/TransferLearning/TransferLearning.py

Typically, once you have a reference to the parameter object (W) you can access the attribute (W.values) and you can get the weights in numpy

Kerry-Cho

@kerry-Cho

@delzac Thanks bro. Fina

I finally found a good example of CNTK image segmentation.
I do not know why I do not include this in the tutorial.
Many people do not know if good examples exist.
Did not I know it alone?

https://notebooks.azure.com/Microsoft-Learning/projects/dev290x

Gregor

@Gregor959

Hi, I understand that 2.7 is the last release.. What does that mean. As far as i understand it works in C# for both training deep neural networks and recalling from them. However the documentation and examples in C# seem incomplete. Furthermore since it is the last release no further major release will be made . Is that a correct summary of the current and final state of affairs?

delzac

@delzac

@Gregor959 Python APIs and related work are complete. C# is more patchy, but still better than anything out there (compared to other frameworks)

Gregor

@Gregor959

@delzac .. there are many deeplearning libraries ( Tensorflow, keras) using python and also more popular ones and ones that are still getting supported with future development. As far as C# and F# are concerned this was the only officially supported library on .net . The .net community has been waiting 5 years for microsoft to get the deep learning on .net . We have existing projects on .net that are screaming for deeplearning... Will the c# interface and examples be completed under 2.7 ? How is the F# looking: 100% complete as pythons's , including the documentation?

delzac

@delzac

@Gregor959 For more C# examples, you can check out https://github.com/anastasios-stamoulis/deep-learning-with-csharp-and-cntk

willprgm

@willprgm

@Gregor959 I'm also wondering about the "last main release" notice for version 2.7. Does anybody know if MS has abandoned CNTK? I've spent months developing a project using CNTK and now I feel like I should start over with a different AI toolkit. So far, I've had no luck trying to find out from MS. At this point, I think it would be a waste of time to put any time on learning any thing new for CNTK.

willprgm

@willprgm

Does anyone know anything about the fate of CNTK? Has MS abandoned it?

delzac

@delzac

@willprgm you can check out Microsoft/CNTK#3646 on the way forward

maxima120

@maxima120

I have joined to ask the same question.. @willprgm

"The CNTK 2.7 release has full support for ONNX 1.4.1, and we encourage those seeking to operationalize their CNTK models to take advantage of ONNX and the ONNX Runtime." - I fail to understand relation between CNTK future and ONNX. I thought ONNX is just model export/import format..

maxima120

@maxima120

Amazon should take over. Anyone knows Bezos phone number?

duxianghong

@duxianghong

in the future, MS will publish the CNTK 3....?

so the CNTK 2.7 is the last main release. right?

delzac

@delzac

Yes cntk2.7 will be the last main release. More details pleases check out here #3646

Basheer Radman

@Basheer94517854_twitter

Hello

Emad Barsoum

@ebarsoumMS

Yes, 2.7 is the latest release.

TchuBotin

@TchuBotin

Hi! I'm training my own data with complex objects:

python 3.5
CNTK 2.6
no GPUs
epoch=20
C.CNTK.TRAIN_E2E = True C.CNTK.E2E_MAX_EPOCHS = 20
C.NUM_ROI_PROPOSALS = 2000 C.DATA.NUM_TRAIN_IMAGES = 39
C.DATA.NUM_TEST_IMAGES = 7 C.DATA.PROPOSAL_LAYER_SCALES = [4, 8, 12]
...
So, I'm getting such outputs, is it normal or does it mea a error?
PROGRESS: 0.00%
CUDA driver version is insufficient for CUDA runtime version
CUDA driver version is insufficient for CUDA runtime version
CUDA driver version is insufficient for CUDA runtime version
CUDA driver version is insufficient for CUDA runtime version
CUDA driver version is insufficient for CUDA runtime version
CUDA driver version is insufficient for CUDA runtime version
CUDA driver version is insufficient for CUDA runtime version
CUDA driver version is insufficient for CUDA runtime version
...
Thank you for some help.

delzac

@delzac

@TchuBotin Sounds like you need to update your GPU driver

TchuBotin

@TchuBotin

Thx @delzac, but I don't have GPU only CPU:

Selected CPU as the process wide default device.
Using base model: AlexNet
lr_per_sample: [0.001, 0.001, 0.001, 0.001, 0.001, 0.001, 0.001, 0.001, 0.001, 0.001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 0.0001, 1e-05]
Training model for 7 epochs.

Training 57513152 parameters in 27 parameter tensors.

Build info:

            Built time: Sep 14 2018 01:23:48
            Last modified date: Fri Sep 14 01:16:52 2018
            Build type: Release
            Build target: CPU-only
            With ASGD: yes
            Math lib: mkl
            Build Branch: HEAD
            Build SHA1: 7c1b0fadb64f83a41739020b24fb2d4950015229
            MPI distribution: Microsoft MPI
            MPI version: 7.0.12437.6

Perhaps, my problem is that the first time I'm trying CNTK.

Mukundh Bhushan

@MukundhBhushan

Hi guys, I created a python package a couple of days back which helps generate ML code based a yml like file called "mlm". This package as of now supports keras and a little of tf2. I would really appreciate if you guys could help out with the CNTK part as well.

This is the git hub repo: https://github.com/NebutechOpenSource/MLGen

Thank you

Michael Coen

@mhcoen

Hi all. We've been using CNTK with Cuda Toolkit v10.0 and cuDNN v7.5.1. I just tried CNTK with the latest releases of the CUDA Toolkit and cuDNN library for Cuda 10.2. CNTK works fine with these but I was expecting I might see a performance improvement during training given some of NVidia's claims. However, benchmark scores during training are unchanged. Any thoughts here? Was my expectation unreasonable? Just wondering what I'm not getting. (The testing was done on a GTX 1080. I have a newer (but smaller) Quadro T2000 in my laptop that I haven't upgraded for CUDA 10.2 yet. Any reason to expect I'd see an improvement with a different GPU?)

Many thanks.

ambarchakrabarty-DnA

@ambarchakrabarty-DnA

Hi I am looking for an AI solution that can automatically read data from excel, add conditional logic to excel data currently being added to manually to excel and then finally load data into MS Access . Can I use CNTK to deliver this solution ?

delzac

@delzac

@ambarchakrabarty-DnA No cntk is a deep learning framework and doesn't deal with excel. You might want to try PyXLL

ambarchakrabarty-DnA

@ambarchakrabarty-DnA

@delzac many thanks for getting back . Sure, I will have a look. Thanks !

JamesLear92

@JamesLear92

Hi, can someone help me understand why CNTK is not getting further releases? There is nothing else that can compete when it comes to C# development.
What should C# developers use instead? Tensorflow.Net is not as C# friendly, and won't work at all with a GTX 760