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/
psuresh has quit [Remote host closed the connection]
< rcurtin>
kartikdutt18Gitt: nice timing, you posted the comment on models#3 literally just as I was tabbing over to it to take a look at it :-D
< kartikdutt18Gitt>
Yes,I just read your comment on copy constructors so I thought you could take a look at it.
< rcurtin>
:)
< rcurtin>
my company uses a really different style for their Julia code... it's really hard to switch back to the mlpack style guide for mlpack reviews :)
< rcurtin>
I keep wanting to make comments like "this doesn't follow BlueStyle" then I go "wait a second..."
< rcurtin>
also it's 4 space indents at my company, 2 space for mlpack... I think that's the one that throws me off the most
< kartikdutt18Gitt>
Ohh, that must get confusing at times.
< rcurtin>
by the time I'm done reviewing this PR and go back to company work, everything there will look wrong :-D
< kartikdutt18Gitt>
xD
< rcurtin>
okay, there we go, hopefully the comments are helpful. awesome effort, I'm really looking forward to the "new" models repository! :)
< kartikdutt18Gitt>
Great, Thanks a lot, hopefully it will be ready soon. Just one more thing that I left a comment for too, Could we add azure and style checks in the repo as well?
< zoq>
Yes, I can create the jobs later.
< kartikdutt18Gitt>
Great, Thanks a lot.
< kartikdutt18Gitt>
Hey @rcurtin, about your comment on storing LeNet as FFN, I might keep it as it because when I add more models they might become part of some other model so storing as sequential makes sense. What do you think?
< rcurtin>
kartikdutt18Gitt: yeah, I see what you mean, but that should be pretty easy, maybe you could make an AsSequential() member for LeNet5 that could return a new Sequential layer that held the LeNet?
< rcurtin>
I guess, I would expect that if people want to use the LeNet class, they're going to think it's turn-key, and they can, e.g., Train() and Predict() with it
< kartikdutt18Gitt>
Ahh, I can do that. Will make the changes. Thanks a lot.
< rcurtin>
the right design really depends on the usage---I guess that I had figured that people would be trying to use the LeNet class directly; let me know if you were thinking otherwise
< kartikdutt18Gitt>
I guess we could support both, that makes more sense than supporting only one (sequential or FFN) and having a FFN is easier for anyone who uses the model so I think FFN might be better.
< Param-29Gitter[m>
Hello guys it's 1st Thursday are we going to have a video meet up? @rcurtin ?
< rcurtin>
oh, crap!
< rcurtin>
I forgot to send an email
< rcurtin>
thank you for reminding me :)
< rcurtin>
I felt like I was forgetting something today...
< Param-29Gitter[m>
I have a reminder set up for every Thursday :)
< rcurtin>
I think, maybe I should figure out how to add a calendar invite to those reminder emails
< Param-29Gitter[m>
Sounds interesting... Never tried that.
< chopper_inbound4>
hey Lakshya Ojha (Gitter) my internet is somewhat unstable....I was saying that I am currently working on the Multihead Attention layer, but I do need help for the word embeddings (there are multiple)...so it would be great if you can help :) jeffin143 is also working on same thing I guess?
< jeffin143[m]>
chopper_inbound (@chopper_inbound:matrix.org): can I know specifically on which algorithm you are working
< jeffin143[m]>
chopper_inbound (@chopper_inbound:matrix.org): I have close eye on that :)
ImQ009 has quit [Quit: Leaving]
< chopper_inbound4>
Positional Embedding
< chopper_inbound4>
but currently I am more focussed towards multihead attention....
< chopper_inbound4>
I am open to any help in word embeddings......
< LakshyaOjhaGitte>
chopper_inbound you mrityunjay right?
< chopper_inbound4>
yeah :)
< LakshyaOjhaGitte>
yeadh
< LakshyaOjhaGitte>
Well I did have a talk with sreenik
< LakshyaOjhaGitte>
He said maybe it is a good idea to talk this with zoq about the transformer layer
< LakshyaOjhaGitte>
I am also not sure about how we should proceed.
< AnjishnuGitter[m>
Hi Mrituyunjay, I wanted to know about the status of the multi-headed attention module, which is listed as a part of the transformer task list. You have an open issue or maybe an PR about this right? Could you probably tell me about the current status of the work so far? And then as @ojhalakshya suggests, we can talk it out with zoq
< LakshyaOjhaGitte>
But yes I am interested to help with the same, lets where we go from here.
< LakshyaOjhaGitte>
see*
< LakshyaOjhaGitte>
Maybe we can talk about this tomorrow, seems pretty late for today :-D
< chopper_inbound4>
though I have made it through the scaled-dot-product attention but parallelization of these inputs is yet to be done....
< chopper_inbound4>
okk :)
< AnjishnuGitter[m>
Okay! That’s great. Let us know if you would need any help with any of the stuff. Goodnight for now i guess :)
< NishantKumarGitt>
@zoq We haven't used any continuous space env for testing any algo right?
< zoq>
NishantKumarGitt: Right, because we don't have a method that works with continuous envs, the PPO PR will change that.
< NishantKumarGitt>
hmmm.. also, I noticed that the current implementation of Pendulum env is not similar to gym's implementation wrt rewards, am I right?
< zoq>
NishantKumarGitt: Hm, maybe something changed, but I think it uses the same rewards.
< NishantKumarGitt>
@zoq oh sorry, I meant return.. Mlpack's implementation returns around -35 for a random agent playing 1 episode, Whereas gym's impl returns around -1000
< zoq>
NishantKumarGitt: Same, initial state and action?
< NishantKumarGitt>
well, I havent checked about the same initial state and action, but the random agent has played 100 episodes in both implementations, and gives the same -35 and -1000 respectively
< zoq>
NishantKumarGitt: Not sure I get the issue, if the random agent performs a random action, it's not necessary returning the same reward, maybe I missed something?
< zoq>
NishantKumarGitt: Don't mind to adapt the gym calculation.
< NishantKumarGitt>
> `zoq on Freenode` Nishant Kumar (Gitter): Not sure I get the issue, if the random agent performs a random action, it's not necessary returning the same reward, maybe I missed something?
< NishantKumarGitt>
Oh no, I was talking about return.. Am sorry i started with rewards but meant return.. Here the return after a finite number of steps(lets say 200) should be around equal for random agents right?
< NishantKumarGitt>
> `zoq on Freenode` Nishant Kumar (Gitter): Don't mind to adapt the gym calculation.
< NishantKumarGitt>
Actually I was going through the PPO PR implementation, and it seems that -136 is used as a reward threshold.. But any random agent is able to generate -35.. Thats what I get after running a PPO agent with no training provided, for 1000 episodes.. 😅
< zoq>
NishantKumarGitt: Not necessarily, you can set a random seed for each run to see different results.
< NishantKumarGitt>
> `zoq on Freenode` Nishant Kumar (Gitter): Not necessarily, you can set a random seed for each run to see different results.
< NishantKumarGitt>
All right, I'll try that :)
< zoq>
NishantKumarGitt: Good point, my first guess is we would have to update the PPO reward.
< NishantKumarGitt>
Yeah, same here.. I m also not so sure about using randomReplay for PPO.. havent seen it in other implementations