[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-347791316 @szha May I ask you another question? I just noticed the symbol API SoftmaxOutput function has two parameters "ignore_label=_Null" and "use_ignore=_Null" which I think may very helpful when I training models, I didn't however find these two parameters in the Gluon Loss API "SoftmaxCrossEntropyLoss" function. So How can I do the same thing as the symbol API does in gluon? This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345927179 These words as separated tokens, I think, will bring some noisy word infos into hidden vector produced by LSTM structure instead of padding tokens that bring some meaningless infos. Does this form will improve the performance? I don't have any idea about it. Anyway, I will try it. But I'm still confused its theory or mechanism behind this trick. Do you have any idea about it? I still noticed that the last five padding tokens. Is this padding token still in the vocabulary set? say ``` "word2idx={'':0, '':1}" ``` This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345927179 These words as separated tokens, I think, will bring some noisy word infos into hidden vector produced by LSTM structure instead of padding tokens that bring some meaningless infos. Does this form will improve the performance? I don't have any idea about it. Anyway, I will try it. But I'm still confused its theory or mechanism behind this trick. Do you have any idea about it? I still noticed that the last five padding tokens. Is this padding token still in the vocabulary set? say ``` "word2idx={'':0, '':1}" ``` This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345927179 These words as separated tokens, I think, will bring some noisy word infos into hidden vector produced by LSTM structure instead of padding tokens that bring some meaningless infos. Does this form will improve the performance? I don't have any idea about it. Anyway, I will try it. But I'm still confused its theory or mechanism behind this trick. Do you have any idea about it? I still noticed that the last five padding tokens. Is this padding token still in the vocabulary set? say "word2idx={'':0, '':1}" This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345927179 These words as separated tokens, I think, will bring some noisy word infos into hidden vector produced by LSTM structure instead of padding tokens that bring some meaningless infos. Does this form will improve the performance? I don't have any idea about it. Anyway, I will try it. But I'm still confused its theory or mechanism behind this trick. Do you have any idea about it? I still noticed that the last five padding tokens. Is this padding token still in the vocabulary set? say "word2idx={'':0, '':1}" This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345927179 These words as separated tokens, I think, will bring some noisy word infos into hidden vector produced by LSTM structure instead of padding tokens that bring some meaningless infos. Does this form will improve the performance? I don't have any idea about it. Anyway, I will try it. But I'm still confused its theory or mechanism behind this trick. Do you have any idea about it? I still noticed that the last five padding tokens. Is this padding token still in the vocabulary set? say "word2idx={'':0, '':1}" This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345920699 Let me pose an example, say I have a question like "What is found on the floor in image1327 ?" and bptt=20, then I should turn it into ``` "What is found on the floor in image1327 ? tokens and put only one ". After that, padding with "" it looks like below "What is found on the floor in image1327 ? tokens and put only one <#><#><#><#><#><#>". ``` Am I right? This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345920699 Let me pose an example, say I have a question like "What is found on the floor in image1327 ?" and bptt=20, then I should turn it into "What is found on the floor in image1327 ? tokens and put only one ". After that, padding with "" it looks like below "What is found on the floor in image1327 ? tokens and put only one <#><#><#><#><#><#>". Am I right? This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345920699 Let me pose an example, say I have a question like "What is found on the floor in image1327 ?" and bptt=20, then I should turn it into "What is found on the floor in image1327 ? tokens and put only one ". After that, padding with "" it looks like below "What is found on the floor in image1327 ? tokens and put only one ". Am I right? This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345920699 Let me pose an example, say I have a question like "What is found on the floor in image1327 ?" and bptt=20, then I should turn it into "What is found on the floor in image1327 ? tokens and put only one ". After that, padding with "" it looks like below "What is found on the floor in image1327 ? tokens and put only one ". Am I right? This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345877657 How to do the "separate token"? This really is my first time hearing about this trick. Could you explain more detail over how to do separate token ""? I really need to solve this problem which I have struggled for several weeks!! Thank you so much! This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345619597 I see. Thank you for helping me a lot. My model can overfit this data set. I had record my training loss, it decreases all the time during training process goes, the validation loss however increases after dozens of epochs. I think this is the clue of overfitting. Also, my image feature matches the questions. For second way for padding, do you mean I should use bucketing to try it again? I do not clearly understand the first way you said " use a separate token for padding". What's that mean? This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345613723 I still wander whether the way I pad end character is correct ? This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345611895 I just examine the data batch I provide to the model. one of batches looks like below and the next one is its corresponding text version. I can assure you that my data pipeline is working correctly. The data set I use to train the model is the same as used in this paper, and I just want to reproduce this experiment result see whether it really work. I think it is not overfitting cause the training loss and validations loss are both high. I think it's still underfitting. Can you just read this paper to check if my model structure is correct? Cause it is not too complicated and won't take you too much time. maybe about ten minutes at most. ![image](https://user-images.githubusercontent.com/13534043/33006930-47302874-ce07-11e7-95a0-f63d539f30b9.png) ![image](https://user-images.githubusercontent.com/13534043/33006123-19511100-ce04-11e7-9867-7e856d59255c.png) This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345611895 I just examine the data batch I provide to the model. one of batches looks like below and the next one is its corresponding text version. I can assure you that my data pipeline is working correctly. The data set I use to train the model is the same as used in this paper, and I just want to reproduce this experiment result see whether it really work. I think it is not overfitting cause the training loss and validations loss are both high. I think it's still underfitting. Can you just read this paper to check if my model structure is correct? Cause it is not too complicated and won't take you too much time. maybe about ten minutes at most. ![image](https://user-images.githubusercontent.com/13534043/33006782-a80a5e0e-ce06-11e7-8af3-e6cf6934d987.png) ![image](https://user-images.githubusercontent.com/13534043/33006123-19511100-ce04-11e7-9867-7e856d59255c.png) This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345598672 @szha I do not understand the meaning of "isolate the problem". Can you explain it? And Is it OK for you that I have your Wechat or email? This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
[GitHub] liuzhi136 commented on issue #8720: Implementation Help!!!
liuzhi136 commented on issue #8720: Implementation Help!!! URL: https://github.com/apache/incubator-mxnet/issues/8720#issuecomment-345585456 @mli @piiswrong @szha This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services