Because people come for Yann, but then stay for me? 😜 Haha, just joking. I think that Yann's name is indeed a major attractor to this channel. Hopefully, people will see there's much more that we are offering.
@@alfcnz haha this is definitely true XD Yann's lectures are great but I also really enjoy how fun your lectures are. I think the other contributing factor is the level of expertise of the students, lots seem to be post-grad so these practical exercises are maybe not so useful for them? On the other hand, I have done 2 undergrad courses (one math, one software) and have been working as software engineer for 3 years so it suits me better to take the entire course without skipping lectures.
Yeah, PhD students told me I'm too slow, but I'm teaching masters as well. They are fine with the speed, so I'm keeping the pace. We get students with any kind of background, from the one with zero programming experience to those who are rusty in math. And then yes, we get undergrads as well. So, you know, the blanket's always gonna be short 🤷🏼♂️
The speed is fine with me too. Listener always have the option to set the speed of the video to match their preference. Yann's name attracted me to this course at first. Then I found out about Alf. He is a knowledgeable, funny and kind professor who takes the time to answer his students' questions. I find practicum is an important part of understanding Yann's high level lectures. Thanks for your contribution and hard work, Alf!
Hey Alfredo! Did anyone submit a working complete python code on the truck-backer-upper (including the controller)? It would be great if you can share it.
Hi Alfredo, I'm trying to code the controller model. How should we define the end of an episode during the trainning loop. Should I reimplement the is_jackknifed and is_offscreen methods but from the output of the emulator ? Set a max episode lenght ? Thanks
For the forward pass the author of the paper is using the actual truck dynamics, with its `is_jackknifed()` and `is_offscreen()` methods. Then, for the backward pass, he's using the emulator. So, the emulator is solely used to backpropagate the gradients through the model(s).
I have two questions: when the emulator is trained, you use a random steering signal and the current position state, what is the reason that you do not start with a random position as well. second why are the model sizes 25 and 45?
Cause in the first step inputs/outputs were recorded simultaneously (as random trajectories/paths), so that if you use inputs/outputs one by one as they were recorded you will get just one full trajectory by full trajectory, which will lead to memorising and less generalisation 😉 I assume these specific sizes were dictated by 90s computing capacities 😁, ofc the more neurons the better* *not every time
The initial position and the control are indeed random. Why not using random position after the first one? Well, it actually is random, but from a distribution more similar to the real one (instead of using an uninformative uniform prior). Yes, those numbers are from the actual paper.
