05:51 < jeffin143> error while loading shared libraries: libboost_program_options.so.1.69.0: cannot open shared object file: No such file or directory

05:51 < jeffin143> Does anyone know why this error ?

07:34 < lozhnikov> jeffin143: Perhaps boost isn't installed or has been updated. Try to install boost or rebuild the project from scratch.

07:49 < jenkins-mlpack2> Project docker mlpack nightly build build #417: STILL UNSTABLE in 3 hr 35 min: http://ci.mlpack.org/job/docker%20mlpack%20nightly%20build/417/

08:37 < jeffin143> lozhnikov : I tried building project from scratch a lot of times, still it throws the error, I have written the word2vec class, but just needs some testing

08:38 < jeffin143> but i cannot do it since i am unable to rectify this issue

08:38 < jeffin143> i will download the mlpack again and then run again

08:39 < zoq> jeffin143: Can you make sure libboost_program_options.so.1.69.0 does exist?

08:44 < jeffin143> zoq : i guess the issue is, while running cmake the mlpack detects my boost version as 1.69

08:45 < jeffin143> but when i runs it in c++ , as

08:45 < jeffin143> std::cout << "Using Boost " << BOOST_VERSION / 100000 << "." // major version << BOOST_VERSION / 100 % 1000 << "." // minor version << BOOST_VERSION % 100 // patch level << std::endl;

08:45 < jeffin143> it shows 1.65.1

08:45 < jeffin143> for some odd reasons

08:47 < zoq> jeffin143: can you delete the build folder an rerun the cmake step?

08:48 < jeffin143> dpkg -s libboost-dev | grep 'Version'

08:48 < jeffin143> this gives me my version as 1.65.1

08:48 < jeffin143> but i don't know how cmake gives me as 1.69

08:49 < zoq> jeffin143: Maybe you have both version installed?

08:49 < zoq> jeffin143: I guess you could just remove the boost package and install it again?

08:50 < jeffin143> zoq : https://pastebin.com/dgedKCH9 : line 29

08:51 < zoq> jeffin143: Are the boost libs in /usr/lib/?

08:52 < jeffin143> there is not boost directory in /usr/lib

08:52 < jeffin143> it only has libboost_program_options.so.1.67.0

08:53 < jeffin143> i mean, it has some more libs

08:53 < zoq> jeffin143: Strange, let's remove those boost packages and install the latest version?

08:53 < jeffin143> yeah most probably my boost has been messed up badly

08:53 < jeffin143> I wasted my whole day compiling the tar file of 1.69 and then getting these errors

08:54 < jeffin143> I will get back to you in a while

10:34 < jeffin143> Do we have softmax function ?

10:43 < jeffin143> suppose I have the following model : https://ibb.co/rvXwtmV

10:43 < jeffin143> one input layer, one hidden layer and one output layer

10:44 < jeffin143> ouput has softmax as activation

10:44 < jeffin143> is the following model correct : https://pastebin.com/yYBUUrtc

10:50 < zoq> jeffin143: Yes, looks correct, I think you intentionally left another activation function after the first linear layer?

11:07 < AndroUser> Zoq: for some reasons , the output layer is not having size of 1*V , but having size 1*1 , I mean the layer has only 1 node as output

11:07 < AndroUser> Whereas I want V nodes,

11:12 < zoq> jeffin143: Hm, will check later, for now you could use the Sigmoid function instead.

11:21 < jeffin143> Yeah* , I will go with sigmoid

12:24 < jeffin143> Huh , finally resolved the issue *

12:24 < jeffin143> Install the rpm package

12:26 < jeffin143> Lozhnikov : I am almost done with the implementation, raw implementation it is not at all optimised and also not so good, I will make an intial commit once I am done with the testing of the class ,

12:27 < jeffin143> I haven't gone through the exact code of Google implementation of word2vec , since I am not able to understand the code structure, so I went through some implementations of it , and found out that they are using softmax , but we have logsoftmax , so I will go with it.

13:15 < Yashwants19> Hi rcurtin: While working on Go-bindings I am facing some issues.

13:16 < Yashwants19> Go doesn't provide a easy to load or save the CSV

13:21 < Yashwants19> It is very difficult when we convert those CSV into mat.Dense because go take CSV string array.

13:22 < Yashwants19> So should we develop something for Loading and saving of csv in our desired format

13:22 < Yashwants19> And When I was working with preprocess_split It take labels as mat.Dense and split them according. but other doesn't take labels as mat.Dense they take labels as mat.VecDense.

13:23 < Yashwants19> Here can we convert all our go-bindings functionality in mat.Dense rather than mat.VecDense.

13:40 < favre49> Another update - Seems like my laptop's problem isn't that trivial, and the repair technician here can't fix it. I either have to go to Delhi to see if a certified service center can repair it, or buy a new one

13:43 < favre49> Considering the poor build quality of my current laptop (heard a lot of complaints by another guy who has the same one) I may buy a new one, not sure motherboard repair is worth the investment, Any suggestions?

14:26 < Toshal> zoq: I need your help to decide the adjustment. There are two ways to make the adjustments in the visitors. One is to add one more overload of the `()` operator as mentioned in the link. And the other is to use SFINAE for overloading the `()` Operator. What do you think will be the best?

14:27 < Toshal> I personally feel SFINAE as in future more such type of addition may be required.

14:35 < lozhnikov> jeffin143: What do you mean by logsoftmax? I am not sure it's correct, it depends on the loss function.

14:38 < lozhnikov> jeffin143: We are not obliged to write exactly the same as the original google implementation. But I think we must follow the papers. Did you read the papers I sent you? Does your implementation correspond to the ideas proposed in the papers?

14:42 < lozhnikov> jeffin143: As far as I know mlpack doesn't provide the log softmax activation function.

15:14 < rcurtin> zoq: I guess I could just do it differently for future tags. Probably nobody is going to go download old versions :)

15:15 < rcurtin> Yashwants19: hmmm, good find. I'm quite surprised Go has no CSV reading or loading functionality into matrices

15:16 < rcurtin> I'm not opposed to providing a converter from what Go's CSV utilities already provide but I would prefer to avoid writing and maintaining a Go CSV reader :)

15:17 < rcurtin> for the mat.Dense and mat.VecDense issue, what we have done in other languages is to make the function acceot either type (if possible) and then cast to the necessary type

15:17 < rcurtin> *accept (not acceot, sorry---awful phone keyboard :))

15:18 < rcurtin> so for instance a VecDense is just a Dense matrix with one dimension of size 1

15:18 < rcurtin> and also, a Dense matrix that has one dimension of size 1 is a VecDense

15:18 < rcurtin> this typically requires runtime checks on the matrix/vector dimensions and an exception to be thrown if something is wrong

15:26 < Yashwants19> Cool I will work on casting. For now I will skip the csv part :)

15:26 < Yashwants19> Thank You.

15:47 < jeffin143> lozhnikov : yes I went through the paper

15:47 < jeffin143> So they are two methods too implement cbow, skip gram

15:48 < jeffin143> They only difference is suppose there is X and y , then for cbow X is input and y is output

15:48 < jeffin143> Where for skip gram y is input and X is output

15:48 < rcurtin> Yashwants19: sounds good, thanks for checking it out :)

15:48 < jeffin143> Both follows a single hidden layer neural networks

15:50 < jeffin143> One input one output and out hidden layer, the output layer has softmax function and calculate the error and then backward propagation ,simple to a single neural networks, the intermediate weights are the vector ( embeding vectors)

15:50 < jeffin143> Also : https://github.com/mlpack/mlpack/blob/master/src/mlpack/methods/ann/layer/log_softmax.hpp

16:33 < lozhnikov> jeffin143: Yes, I see. I didn't look through the layers.

16:37 < lozhnikov> jeffin143: Okay, I glanced over only one of the papers. I am going to glance over the other papers and over the original implementation. Then I'll look through your code.

17:10 < zoq> favre49: Maybe you can go with a similar setup as Ryan uses, Phone (ssh + screen/tmux + vim), not sure this is something for everyone but I like the idea :)

17:10 < zoq> favre49: Honestly, not familiar enough with notebooks, that I could recommend something, I do have an xps and a macbook and both are okay; don't really like the touchpad on the xps, but I do like the keyboard, on the macbook the touchpad is good but the keyboard needs some time to get used to it.

17:10 < zoq> Toshal: hmm, you have to adapt SFINAE once you add a new type, for the first idea we just have to add another overload. Do you think the second idea is faster?

17:11 < zoq> rcurtin: If the modifcation of old tags is simple I guess we could do it anyway.

17:25 < Toshal> zoq: No I don't know whether the second idea is faster.

17:26 < Toshal> I will stick to the overload then.

17:26 < zoq> Toshal: I think the optimizer will optimize out the extra overload, so we might end up with what we get using SFINAE.

17:26 < zoq> Toshal: If you already started, we can go with SFINAE.

18:08 < abernauer> rcurtin: I ran R through gdb loaded the shared object. Set some breakpoints and got a better stack trace.

21:45 < abernauer> rcurtin: Wanted to give you some more insight on that stack trace. I printed the variable name, after passing "Principal Components Analysis" had the value "\220" which in octal is that question mark symbol in ascii.

22:19 < rcurtin> abernauer: sounds good, looks like you've identified the issue---it seems to me like strings are not being passed from R to C++ correctly

22:19 < rcurtin> (however I am not sure the details of why that is or how it can be fixed; my guess is maybe an error in how the types are being passed)

23:26 < rcurtin> zoq: yeah, I'll see how easy changing tags are when I have some time