mlpack IRC logs, 2018-06-17

Logs for the day 2018-06-17 (starts at 0:00 UTC) are shown below.

June 2018
--- Log opened Sun Jun 17 00:00:07 2018
04:52 -!- wenhao [731bce35@gateway/web/freenode/ip.] has joined #mlpack
05:08 -!- vivekp [~vivek@unaffiliated/vivekp] has quit [Read error: Connection reset by peer]
05:12 -!- vivekp [~vivek@unaffiliated/vivekp] has joined #mlpack
06:58 -!- ImQ009 [~ImQ009@unaffiliated/imq009] has joined #mlpack
07:07 -!- wenhao [731bce35@gateway/web/freenode/ip.] has quit [Ping timeout: 260 seconds]
07:36 -!- manish7294 [8ba7aca8@gateway/web/freenode/ip.] has joined #mlpack
07:41 < manish7294> zoq: Things are working now but somehow accuracy prediction doesn't seems to be valid (for instance shogun's knn is predicting 100% accuracy on iris, wine and 20% on some other datasets). Can you please take a look at the code?
08:28 -!- Guest63658 is now known as anirudhm
08:28 -!- anirudhm [sid227710@gateway/web/] has left #mlpack []
09:02 -!- manish7294 [8ba7aca8@gateway/web/freenode/ip.] has quit [Ping timeout: 260 seconds]
09:12 < ShikharJ> zoq: I don't seem to get the issue with the line `outputTemp.slice(outMap) += bias(outMap);` inside the convolution layer?
09:21 < zoq> ShikharJ: bias is of size outSize but outMap will iterate over outSize * batchSize (outMap < outSize * batchSize).
09:22 < jenkins-mlpack> Project docker mlpack nightly build build #351: FAILURE in 1 day 2 hr:
09:24 < zoq> manish7294: I'll take a look at the code later today.
09:26 < ShikharJ> zoq: So can I just change the bias statement to `bias(outMap)`
09:27 < ShikharJ> zoq: Ignore the above, I guess I'm still not able to understand the problem, isn't bias called at every iteration on the output slice?
09:27 < ShikharJ> zoq: How would that cause a problem?
09:33 < zoq> ShikharJ: The bias has to be added for each output slice right, but we have one bias term per output so let's say the outSize is 3, the bias is of size 3 e.g. [0.1 0.5 0.3]. If I use a batchSize = 2 the for loop is iterating over 0, 1, 2, 3, 4, 5 (index), since we use the condition outMap < outSize * batchSize.
09:37 < ShikharJ> zoq: Do you think I can simply use a `bias(outMap % outSize);` for correcting this?
09:38 < zoq> ShikharJ: I think that should work, yes
09:38 < ShikharJ> zoq: Thanks for explaining the issue, I really appreciate it. Your debugging skills are really good :)
09:39 < zoq> ShikharJ: Happy to help :)
09:40 < zoq> ShikharJ: There is another issue which we have to solve, but let's fix this one first.
09:42 < ShikharJ> zoq: You mean with the pooling layers? I'm already finding a fix for that.
09:43 < zoq> ShikharJ: Good, there is another batch related issue inside the GAN class, some submatrix mismatch.
12:25 < ShikharJ> zoq: I think that must be because Gradients has not been modified as of yet. I'll let you know if I get stuck with anything.
12:36 -!- witness_ [uid10044@gateway/web/] has quit [Quit: Connection closed for inactivity]
12:42 < ShikharJ> zoq: The BatchSupport PR has also been debugged now. Thanks a lot for your help, wouldn't have been possible without it :)
13:18 < jenkins-mlpack> Project docker mlpack nightly build build #352: NOW UNSTABLE in 3 hr 56 min:
13:29 < ShikharJ> zoq: Now with BatchSupport and DCGAN ready, we just need to focus on the GANOptimizer class, and the we're good to go with the rest of the implementations.
15:50 < zoq> ShikharJ: Agreed, the batch support should give us some speedups.
15:53 < zoq> manish7294: checked the wine and iris dataset and the results are reasonable, can you comment on the dataset that might produce some strange results?
16:14 -!- manish7294 [8ba79986@gateway/web/freenode/ip.] has joined #mlpack
16:15 < manish7294> zoq: Can you try balance_scale and letters?
16:15 < manish7294> and did you got 100% on iris and wine?
16:41 -!- manish7294 [8ba79986@gateway/web/freenode/ip.] has quit [Ping timeout: 260 seconds]
18:49 -!- witness_ [uid10044@gateway/web/] has joined #mlpack
19:00 -!- vivekp [~vivek@unaffiliated/vivekp] has quit [Ping timeout: 248 seconds]
19:15 < zoq> manish7294: The letters failed on my end, have to take a closer look into the issue, maybe it's already fixed? And yes I got 100% on wine and iris, but they are simple datasets so I'm not really surprised, do you think the results are not reasonable?
19:16 < ShikharJ> zoq: Could you review the BatchSupport PR when you get some time?
19:16 < zoq> ShikharJ: yes
19:26 < ShikharJ> zoq: Thanks for reviewing on such a short notice :)
19:26 < zoq> ShikharJ: Hope this solves the issue, MSVC is somewhat picky :)
19:29 < ShikharJ> zoq: Yeah, hopefully that was the issue that we were facing.
20:09 -!- ImQ009 [~ImQ009@unaffiliated/imq009] has quit [Quit: Leaving]
20:54 -!- witness_ [uid10044@gateway/web/] has quit [Quit: Connection closed for inactivity]
20:55 < ShikharJ> zoq: That was precisely the issue with MSVC.
22:03 -!- sumedhghaisas_ [5c082148@gateway/web/freenode/ip.] has quit [Ping timeout: 260 seconds]
22:52 < zoq> ShikharJ: Great, I'll take a second look at the code tomorrow, so that it can be merged in the next days.
22:53 < zoq> manish7294: With the letter dataset I get the following error message: 'libraries/shogun/src/shogun/metric/LMNNImpl.cpp line 69: The initial transform must be a square matrix of size equal to the number of features'.
--- Log closed Mon Jun 18 00:00:08 2018