06:20 < jarvis_> Hello developers, I am a second year computer science student at BITS Pilani, India. Here I have been a member of the Embedded Systems and Robotics lab since my freshman year itself. I work on robotics and artificial intelligence (specifically computer vision, deep learning and reinforcement learning). I have a strong foundation in deep learning and reinforcement learning and have decent experience in C++ programming.

06:20 < jarvis_> I am also a student at the Udacity self-driving car nanodegree, and as part of the programme I have implemented various deep learning architectures with established personnel in machine learning (PhD, scientists, professionals, etc).

06:20 < jarvis_> I noticed that there was a project previously which implemented the GoogLeNet architecture. I was wondering what is the interest in deploying a similar computer vision deep learning architecture this year - something like ResNet, R-CNN, DenseNet, etc. I have done similar projects before, have and continue to review the literature on this topic and am interested to contribute if a mentor is interested.

06:22 < jarvis_> If deploying such an architecture is not high on priority, I am also interested in the reinforcement learning applied on games. I have worked with deep neural networks as a student contributor to Udacity's real open source self-driving car and have experience with reinforcement learning due to my position as an undergrad robotics researcher at the robotics lab in my college. I am equally excited about this project and can keep up wi

06:22 < jarvis_> Kindly let me know your views

06:23 < jarvis_> Thanks, Rohan Saxena

15:18 < zoq> jarvis_: Hello and welcome, other interesting ideas are definitely suitable, we just have to make sure someone is able to mentor the project and it's enough work. As for the GoogLeNet project, there was the implementation of a method to localize objects and the network implementation itself.

15:18 < zoq> jarvis_: I think implementing a single model e.g. ResNet isn't enough work, but if you like to extend your project idea by one or two recent models it would make a great project. Let us know what you think.

18:01 < mikeling> rcutin: hi does my comment make sense to you? :)

19:32 < jarvis_> zoq: I'm glad that you are interested in my idea. I just gave an example of ResNet, I would love to implement more models! How about I implement ResNet (one of the best CNN architectures), VGGNet and Spatial Transformer Networks (https://arxiv.org/pdf/1506.02025.pdf)? Let me know your suggestions.

19:56 < darkknight_> '

19:57 < darkknight_> s

20:18 < kris> zoq: in gym_tcp_ip i wanted to add new q learning example

20:20 < kris> so i have made a entry to CMakeLists.txt on line 50 in set and added QLearning.cpp what other changes should i make

20:28 < kris> should i make all the hpp files as header files as include_directories(include) and the generate cmake file.

20:28 < kris>

20:29 < kris> as the error i am presently getting is of multiple definition of gym::*

20:29 < zoq> kris: You just wanted a simple solution to build your q learning example code right?

20:30 < kris> zoq: Yes

20:30 < kris> actually i am reading a tutorial on cmake so i came up with solution

20:37 < zoq> That is what I did: https://gist.github.com/zoq/6bb9c24f38d019ec55ed9acfc0a02cd4

20:39 < zoq> I added line 72-74, 79 and 82

20:42 < kris> zoq: Also why in example cpp env.done = 1 only after 15 steps in one episode. I think open ai cartpole says that we have to stay upright for at least 100 steps in one episode.

20:42 < kris> @zoq thanks i will take look and see that works

20:49 < zoq> kris: Note env.action_space.sample(); returns actions sampled from a simple agent, so it's possible that the agent isn't able to solve the environment. In case of the cartpole environment once you're in a position where you can't get back you failed.

20:50 < zoq> kris: Once you failed you can restart the environment and try again; jsut use env.reset();

20:55 < kris> but in the rendering video the cart does not go off screen

20:56 < kris> zoq: but in the rendering video the cart does not go off screen

20:57 < zoq> kris: But I guess it's leaning to the left or right side?

20:57 < zoq> kris: Also, I would probably start with a simple method e.g. Policy Gradients.

22:09 < zoq> jarvis_: I really like Spatial Transformer Networks, the ability to perform geometric transformations on the input is really clever.

22:09 < zoq> jarvis_: ResNet is also an interesting candidate, note that you have to implement some extra parts e.g. batch norm and we should also take a look at the conv layer, there should be some ways to make the layer faster.

22:10 < zoq> jarvis_: About VGGNet, do you think it's reasonable to implement the model since we already have the GoogLeNet and if I remember right VGGNet is way more expensive when it comes to the number of operations/runtime? I could be wrong, but maybe Inception V4 might be a better candidate. What do you think?

22:12 < chvsp> Hi. I have an older fork of the repo which has bias_layer.hpp file in it. In the current mlpack master, I am not able to find that file, hence my build is failing. It would be great if you could help me out. Thanks.

22:13 < rcurtin> chvsp: I don't understand, wouldn't it be better to simply upgrade the repo and use the latest code?

22:13 < rcurtin> otherwise I imagine you could find bias_layer.hpp by finding an older revision on github and downloading the file

22:13 < rcurtin> I can help you through that process if you like

22:16 < zoq> chvsp: Also, there is no extra bias layer anymore, the linear layer combines the former linear layer + bias layer.

22:19 < chvsp> @zoq Ok got it. I was actually working on the PR for Kathir Subavathi Init issue

22:31 < zoq> chvsp: I see, I think the easiest solution is to close the PR and as rcurtin already said open a new PR based on master. The transition should be straightforward.

22:32 < chvsp> I am trying to port the code from the refactored FFN test

22:32 < chvsp> Yes will do that

22:34 < zoq> Okay sounds good, whatever works best for you, if you like to work in the current PR, it's fine with me. If you need any help let us know.

22:46 < chvsp> Thanks a lot! I am currently working on it, will let you know if I face any issues. :)