reweight-gpt VS jetson-containers

Instead of self attention, I tried to generate the self-attention matrix directly using lateral connections among the inputs. The method is like LSTM but it gates all the past inputs using separate gates for each input (it can be parallelized). It's very easy to implement the method into the current GPT architectures. You just remove the attention part and replace it with learnable weights. Her is a working implementation (around100 lines!): Code: https://github.com/hunar4321/reweight-gpt In my experience, it learns very well and it can super-pass the self-attention mechanism if the number of the parameters are matched. (I tested it on small datasets for next character prediction. I haven't systematically compared these two methods yet).

https://github.com/dusty-nv/jetson-containers This one might be helpful

That's the worst solution because you lose CUDA support, if you don't use the NANO's GPU you can use a Rpi4 instead that has a more powerful CPU. Docker is the solution instead, there are ready images for Jetson: https://github.com/dusty-nv/jetson-containers

Finally, if you don't want the full ros2 desktop on the nano (you may struggle with memory anyway) then jetson containers can run foxy etc with your existing jetpack version. https://github.com/dusty-nv/jetson-containers

I had to modify this dockerfile to get pangolin to work, but now it is the ORB_SLAM2_CODA portion that I cannot figure out.

Have you tried looking at some jetson nano docker images? https://github.com/dusty-nv/jetson-containers

Yeah, Nvidia has been very slow to release a 20.04 version, but they have provided a lot of Docker containers just for the Jetson for people to use. Here is the Nvidia dev that works on the project.

There are also pip wheels provided. You can easily install them into your system, even without using containers. Also, look at the l4t-ml Dockerfile, which merges some stuff from several containers, using Docker multi-stage builds.