ChanServ changed the topic of #mlpack to: "mlpack: a fast, flexible machine learning library :: We don't always respond instantly, but we will respond; please be patient :: Logs at http://www.mlpack.org/irc/
< R-Aravind[m]> <anjishnu[m] "I haven't looked too closely but"> Thank you!
vansika__ has quit [Read error: Connection reset by peer]
vansika__ has joined #mlpack
ImQ009 has joined #mlpack
ggalan87[m] has left #mlpack []
qian[m] has left #mlpack []
< nishantkr18[m]> zoq: regarding https://github.com/mlpack/examples/pull/111 ; I tried uploading the xml model in the same directory as my jupyter notebook and running `data::Load("./50qNetwork.xml", "episode", qNetwork);`. It doesn't seem to work though
< zoq> nishantkr18[m]: Did you get any error?
< nishantkr18[m]> Yes, I'll rerun and paste the error now
< zoq> thanks
< nishantkr18[m]> Oh. Not a problem :)
< zoq> But, the trained model dosn't produce the correct results?
< nishantkr18[m]> Currently it won't because the parametrs get reset during initialization. https://github.com/mlpack/mlpack/pull/2578 should solve the problem though
< zoq> Right
< zoq> If someone has a minute, please take a look at #2578, we just need another approval :)
< nishantkr18[m]> zoq: Here's the error: https://pastebin.ubuntu.com/p/Yjb9nfPbcH/
< zoq> nishantkr18[m]: Hm, did you run any cell twice? Or was that in a single run?
< nishantkr18[m]> zoq: A single Restart and runall. No dublicate cells
< nishantkr18[m]> *duplicate
< zoq> nishantkr18[m]: Okay, let's see if I can reproduce the issue.
< nishantkr18[m]> Shall I commit the updated notebook with the q and policy networks?
< nishantkr18[m]> For u to test?
< zoq> nishantkr18[m]: Ohh, yeah, that would be great.
< zoq> nishantkr18[m]: Can you upload the file as well?
< nishantkr18[m]> Yes, u mean the networks in the form of xml file right?
< zoq> right
< nishantkr18[m]> sure, just a min
< zoq> thanks
< nishantkr18[m]> done :)
< zoq> nishantkr18[m]: I guess, you already tested the same code local, and it works?
< nishantkr18[m]> Yes
< zoq> nishantkr18[m]: Hm, takes a really long time to load.
< nishantkr18[m]> Yes, and at the end the error appears
< zoq> nishantkr18[m]: Do you have the chance to save the models as .bin instead of .xml?
< nishantkr18[m]> Ok, I'll try that
< nishantkr18[m]> The same thing is happening again with bin file :(
< zoq> nishantkr18[m]: Okay, let's see if we can get some more information.
< nishantkr18[m]> Does any other notebook exist in examples repo that deals with loading/saving a model?
< zoq> nishantkr18[m]: Testing the save/load part, but this also takes a really long time, so I guess it will fail as well.
< nishantkr18[m]> Hmm..
< nishantkr18[m]> BTW, I was able to get a decent score after ~600 episodes of training SAC on bipedal on my laptop (which took around 1.5 hrs)
< nishantkr18[m]> After 600 episodes, the performance dropped..
< zoq> nishantkr18[m]: Hm
< zoq> nishantkr18[m]: same error in the other notebook
< nishantkr18[m]> <zoq "nishantkr18: same error in the o"> oh..
< zoq> nishantkr18[m]: I guess, if we stop early, we can work with the pretrained model.
< nishantkr18[m]> Yeah I had managed to save the model every 10 episodes, so I have the best model saved with me
< zoq> Nice
< nishantkr18[m]> But I think if we add one more layer, It should perform much better
< nishantkr18[m]> Right Now, the best agent is barely capable to walk
< nishantkr18[m]> So the simulated run of the best agent is not that satisfying 😅
< zoq> You are currently training another one?
< nishantkr18[m]> Nope, That's the issue. U see, adding one more layer takes a lot of time to train.
< zoq> nishantkr18[m]: Yeah, if you like I can train it on a remote machine.
< nishantkr18[m]> I tired training in parts, but since the replay buffer gets reseted after every restart of training, the performance drops if I restart training
< nishantkr18[m]> <zoq "nishantkr18: Yeah, if you like I"> Yeah, that would be really great!
< zoq> nishantkr18[m]: No problem, let's wait for #2578 and start training?
< nishantkr18[m]> zoq: yeah sure. btw, here's the snippet I use for training: https://gist.github.com/nishantkr18/3791afb101a963e3af4fc53450d25450
< zoq> Wonder if #2458 might work with cling.
< zoq> nishantkr18[m]: I guess a workaround would be to set the weights and parameter manually.
< shrit[m]> <rcurtin "for any other serialization issu"> @rcurtin What is the virtually identical strategy?
< rcurtin> shrit[m]: the recursion to set 'dataset' in the patch I sent
< shrit[m]> rcurtin: I have just seen your email of Friday, It was classified in the spam in my box sorry
< shrit[m]> OK I see.
< chopper_inbound[> Hello everyone! here is my weekly update blog https://mrityunjay-tripathi.github.io/gsoc-with-mlpack/coding_period/week11_and_12.html. Please have a look :)
< shrit[m]> @rcurtin Tests related to binary tree are passing fine, I am applying the same modification on other trees, and see the final results.
< rcurtin> shrit[m]: classified as spam?? :( do you have any info on why?
< rcurtin> I thought I had DKIM and DMARC working right
< shrit[m]> I do not know why, This is the reason, I did not answer you sorry, I do not check the spam always.
< shrit[m]> I think it is related to outlook, even for me this happen and I think I have DKIM and DMARC set correctly I think, at the same time, I do not believe that outlook implement the email RFC correctly.
< shrit[m]> If you ping me with a mail on gmail. I can verify if it goes to spam
< shrit[m]> Did you changed the IP for you email?
< rcurtin> yeah, when mlpack.org went out two weeks ago, it came back up with a new IP
< rcurtin> actually I am not sure what a gmail address for you is, but it seems like most other people with gmail don't have an issue receiving emails from me
< shrit[m]> I know gmail is usually fine, it is only outlook that is usually doing this.
< rcurtin> I'll have to see if I can figure out any "tricks" to get outlook to receive my email
< rcurtin> in the mean time if I email you I'll use the shrit.me account :)
< shrit[m]> Perfect, you can contact the outlook spam services they will unlist the IP from the spam
< rcurtin> done... we'll see what happens :)
< shrit[m]> :+1
PratikPriyadarsa has joined #mlpack
< PratikPriyadarsa> Hello, I am willing to contribute to mlpack, can anyone please guide me to start?
< zoq> PratikPriyadarsa: Hello, checkout https://www.mlpack.org/community.html#getting-involved for a starting point.
gtank____ has joined #mlpack
gtank___ has quit [Ping timeout: 240 seconds]
gtank____ is now known as gtank___
carloscarlos[m] has joined #mlpack
TanayMehtaGitter has joined #mlpack
nishantkr18[m] has joined #mlpack
himanshu_pathak[ has joined #mlpack
chopper_inbound[ has joined #mlpack
srinivasyadav227 has joined #mlpack
jacob-earleGitte has joined #mlpack
khimrajGitter[m] has joined #mlpack
AbdullahKhilji[m has joined #mlpack
sailor[m] has joined #mlpack
ArunavShandeelya has joined #mlpack
PulkitgeraGitter has joined #mlpack
yashwants19[m] has joined #mlpack
yashwants19[m] has quit [Changing host]
yashwants19[m] has joined #mlpack
PratikPriyadarsa has joined #mlpack
< kartikdutt18[m]> Hey sakshamb189 , KimSangYeon-DGU , Here is the link to my [GSoC Work Report](https://github.com/kartikdutt18/GSoC-Work-Report). Kindly let me know what you think.
< shrit[m]> rcurtin are you here?
ImQ009 has quit [Quit: Leaving]
< rcurtin> shrit[m]: let me take a quick look
< rcurtin> shrit[m]: I see this in the "install build dependencies":
< rcurtin> Get:17 http://azure.archive.ubuntu.com/ubuntu xenial/universe amd64 libcereal-dev amd64 1.1.2-4 [163 kB]
< rcurtin> is 1.1.2-4 too old?
< rcurtin> maybe you might need to find a PPA to install a newer version
< rcurtin> I can see at least that 1.1.2 is from 2015
< rcurtin> so, the version must be too old
< rcurtin> judging by https://github.com/USCiLab/cereal/issues/360, it looks like the minimum version of cereal is v1.2.2 (I guess we should codify that in CMake)
< shrit[m]> Exactly, but even if we did this modification, can we count on the PPA to install the specified version?
< rcurtin> if you can find a PPA that has a new enough version, then yes, presumably
< rcurtin> alternately you could build cereal from source and install it in the build rules, but that is just a little bit uglier :)
< shrit[m]> rcurtin what do you think is the best for the user?
< rcurtin> the user doesn't matter here, these are just the CI rules
< shrit[m]> Yes, I see, we just have to specify the version in FindCereal
< rcurtin> I think it's fine to require 1.2.2 or newer; that build node is xenial, which is Ubuntu 16.04... that's ancient
< shrit[m]> The modification I made for RectangleTree is not working, I am getting the same memory error
< rcurtin> I am looking through the code. it seems like the storage of `dataset` is different in `RectangleTree`
< rcurtin> take a look at the constructors, and how `ownsDataset` is used throughout the class
< rcurtin> you might want to trace through the tree's constructor to understand how the structure is built
< shrit[m]> Ok
< rcurtin> I am in a meeting now, so I can't verify this, but I think maybe that each RectangleTree may hold its own dataset; however, I am not sure
< rcurtin> although, it seems like based on the copy constructor that that is not true
< rcurtin> can you maybe print the memory error to start debugging it in the same way that the BinarySpaceTree error was debugged?
< jeffin143[m]> Can some one merge 2579
< shrit[m]> I am debugging the same way, since it was the same error, I will verify if the dataset returned by the tree.Dataset() is NULL
< rcurtin> ok, great
< shrit[m]> I am waiting to finish compilation
< rcurtin> jeffin143[m]: done :)
< jeffin143[m]> Thanks :(
< jeffin143[m]> :)*
< rcurtin> :)
< himanshu_pathak[> Hey rcurtin: Can you answer this:)
< shrit[m]> This error is caused by the serialization process of the RectangleTree, more specifically in the save function. I will try to get more details tomorrow.
< rcurtin> shrit[m]: no problem, I will at least now read through the sources to understand what the situation is with the ownsDataset and dataset members
< rcurtin> shrit[m]: the `dataset` member in `RectangleTree` is, as far as I can tell, just like the one for `BinarySpaceTree`: every node in the tree has the same `dataset`, but the root "owns" it
< rcurtin> and thus, like `BinarySpaceTree`, during serialization, only the root should serialize `dataset`, and when loading, after the root has deserialized its children, it should do a traversal to set all of its descendants' `dataset` pointer to its `dataset` pointer
< rcurtin> sorry for the confusion in what I wrote earlier... it is hard to keep focus between a meeting and also helping debug, so I confused myself :)
< rcurtin> zoq: did you want to review #2325 (one-hot encoding binding)? there is a pending review for it, but if you are busy, no worries---it's approved, but I don't want to merge if you wanted to take a look :)
< zoq> rcurtin: I'll take a look.