search

zhongkaifu / RNNSharp

RNNSharp is a toolkit of deep recurrent neural network which is widely used for many different kinds of tasks, such as sequence labeling, sequence-to-sequence and so on. It's written by C# language and based on .NET framework 4.6 or above versions. RNNSharp supports many different types of networks, such as forward and bi-directional network, sequence-to-sequence network, and different types of layers, such as LSTM, Softmax, sampled Softmax and others.
BSD 3-Clause "New" or "Revised" License
285 stars 92 forks source link

Example file #10

Closed LiangYuzhi closed 7 years ago

LiangYuzhi commented 8 years ago

Is there any runnable example? I've change the code a little bit and tried to run it, but the error rate is unreasonable high

zhongkaifu commented 8 years ago

You can download demo package from release section. It contains runnable examples.

What's command line did you run ? What's change did you make ? And what's the error rate before your changes ?

LiangYuzhi commented 8 years ago

Thank you for your timely reply:)

The description of the demo package is 'To get the demo package, please access [DOWNLOADS] page and download the package', what is the download link?

Basically, I have a set of character embedding generated by word2vec, and let it as input of RNNSharp to do some Chinese segment task. It's similar to the task in this paper: http://arxiv.org/pdf/1404.4714.pdf

I've change the code as taking a matrix as input (though this is not necessary for this task, but I have following tasks has to take matrix rather than load word embedding as input). By setting breakpoint, I can see the data is loaded properly: image

I used 4 tags 'S/B/M/E' as output, But the segmentation result is always BEBEBESBE/BEBEBEBE.

For training, I used -layersize 50 -crf 1

bratao commented 8 years ago

LiangYuzhi, could you try without a crf?

Here, I couldn't find a input that crf works. I have the impression that or it is not working well, or it expect an specific input that I haven't figured out. On Mar 10, 2016 2:04 AM, "LiangYuzhi" notifications@github.com wrote:

Thank you for your timely reply:)

The description of the demo package is 'To get the demo package, please access [DOWNLOADS] page and download the package', what is the download link?

Basically, I have a set of character embedding generated by word2vec, and let it as input of RNNSharp to do some Chinese segment task. It's similar to the task in this paper: http://arxiv.org/pdf/1404.4714.pdf

I've change the code as taking a matrix as input (though this is not necessary for this task, but I have following tasks has to take matrix rather than load word embedding as input). By setting breakpoint, I can see the data is loaded properly: [image: image] https://cloud.githubusercontent.com/assets/17334812/13659059/4c5e98d6-e6b8-11e5-95df-6eda24bdd70b.png

I used 4 tags 'S/B/M/E' as output, But the segmentation result is always BEBEBESBE/BEBEBEBE.

For training, I used -layersize 50 -crf 1

— Reply to this email directly or view it on GitHub https://github.com/zhongkaifu/RNNSharp/issues/10#issuecomment-194667366.

zhongkaifu commented 8 years ago

@LiangYuzhi Sorry that I didn't update readme file in time, you can get demo package from "release" section. It contains some demo as your example.

Which modeltype did you use ? SimpleRNN or LSTM ? Which your entire command line ? Could you please show some information from log ?

bzaar commented 8 years ago

@zhongkaifu I couldn't find a "release" section in this repository. +1 for adding an example file.

zhongkaifu commented 8 years ago

Hi bzaar, You could get demo and example there: https://github.com/zhongkaifu/RNNSharp/releases

LiangYuzhi commented 7 years ago

The error rate problem might be caused by my training set, this issue can be closed. Thanks