Hi! I've just started digging into RL and cannot understand one point: You always show at the end the "last transition" path. I can understand why the algorithm always finds the correct and optimal way when we use the policy random + known value, but why when we are using only random policy the algorithm still get to the final point (I mean even when we are using MC and TD principles). I see, that the algorithm can build the Look-up table, and when we are using MC/TD it makes this much quicker, but why does it start always getting the max value after some amount of interatons?

Thanks a lot for this perfect explanation!

Hi, I'm using a Witty Pi4 L3V7 as a power management board for my project with a Raspberry Pi Zero 2W. I use two 10Ah 3.7V Li-Po batteries in parallel, and it works great. The board has an integrated RTC, and with the software, you can schedule the Raspberry Pi's power-up and shutdown. Moreover, the board's software allows you to monitor energy consumption in real time, including battery voltage, current drawn from the battery pack, and more.

Really interesting Luke, great work. Going through your channel is a treasure trove. Love your unique offering of short to medium practical ai applications.

So it was straightforward interface with jetson nano or you needed to make some alterations to camera module ?

So it's interesting so the Nvidia computer can actually train models. And the raspberry pi with the AI hat is much better at just running pre-trained models which as you say is probably what most robotics or self-driving car or other kind of developers really want. They probably don't want to retrain a bunch of models or train a bunch of models for scratch they want to be able to run this stuff that the experts have already trained and just quickly slap it on to whatever machinery they're using and make it work. And I do think you're right that that's probably what most hardware developers want. It will help however be fun to see what develops with the super in terms of online people developing open models and training them and how they train them because that would also be fun to get into but can't as easily directly lead to products. Or even your own home products.

Thank you for explaining the differences so clearly and succinctly and not patting your video like everybody else does these days. That earned a subscription.

Hi, I am trying to do sequence to vector prediction where a fixed length time series is used to predict some set of parameters. (Imagine having ECG graph of patient as time series and then trying to predict patient's BP , pulse rate and other cardiac parameters). What's most suitable for such task ? I was trying to use self-attention mechanism with no success.

I have data samples, where each sample is represented by 200 x 6 (time steps x features) numeric time series, while the label for each sample is 4 x 1 numeric vector. I tried to train multitask self-attention network (like transformer), but all it gives me is (roughly) an average 4 x 1 vector of all the training samples. Can you please make video where we train transformer (or LSTM with attention ) for (numeric time) series to (numeric) vector prediction model ?

Easily the single best video I have watched yet, over many years, that clearly and intelligently breaks down the basic development, testing and deployment of an edge AI system. BRILLIANT! Thank you so much 🎉

Have you ever successfully convert .onnx to .hef file?

😮

Parrafin based oil

Fill with transformer oil you find in powerline transformers its perfect for insultation and heat distributor

Just put a heat sink the size of your fist if you have that much room

I agree with the conclusion but the video didn’t cover WHY the Pi +AI HAT is the way to go. I presume the Pi hat has an NPU that does small bit vector functions as well as the Nano doing large ones -is that the trick?

I'd imagine the black door and window don't help the cooling much. My thought was to use a reflective face or light color with a "peephole" for the camera. I'm uncertain whether or not having a reflective surface would cause issues with the camera lense. However, I've no clue what I'm talking about in regards to your project. Cool to see what you're working on! Thanks for sharing!

Nice (size) comparison. I'm curious for the result.

Namo? Lol

Thanks, saved me some money buying a jetson without thinking about the use case. I just got the Pi5 and AI hat and just using it for inference. Have you found a way to connect an SSD instead of the SD card? I hear inference will be better on an SSD

Could you please send me a link to the raspi battery supply you're using? Loved the video

The Pi is a berry flexible platform. As a phone engineer, i could see why the Pi woild be a better choice for this. Though I would love to see how the performance is on a SD 8 Gen 3 that's got some power and thermal headroom.

A smart phone is probably the ideal hardware for this. High res camera, 4G networking and ai accelerator all in 1 device. Just plug it into the solar power board.

hey i just preprocessed the data ,building tokens,vocab,and everything from scratch using pytorch and nltk i and trained a model which predict the next word and i saved the model for later use, then i decide to train the model on a different dataset so i loaded the previous model to train on a new dataset but there is an error for vocab size mismatch because the dataset i used to train the model first time is different from the dataset i used second time to train so i am just wondering what can i do to train the model on different dataset by using my own model trained on some data. i am using the transformer architecture for this i hope i explained it well because my english is not that good i am using word level tokenizer

Moving parts without wear? A Engineers dream come true.

Thanks dude

So glad I found your channel

Hey man did u figure out how to connect the ups to a solar panel? I got the same ups and was js wondering thanks!

the uranium adds flavor.

Nevermind my question was answered lolololol

Are you converting the strings input from chat into g code?

What a cool project! Would love to pick your brain some time about something similar I'm working on.

what do you like better? jetson nano or raspberry based?

LGTM

Finally found the part of KZbin where things happen

Do you mentor?

Yeah that's super cool

Hello, thank you for this tutorial, I managed to find a 1 F 5V capacitor I wanted to install it on my Jetson Nano but from the video I can't see where you connected the positive terminal could you send me a diagram? thanks

0 fps, a lot of hassle with the power supply system and low efficiency.... For what? 🗿

What does it do

This was my Final Year Project, although my angle was to have it detect at night near human settlements since wild animals usually come out around that time. Cool stuff, mine ended at prototyping phase due to limited budget, but its kinda cool to see you take things much further.

How expensive was the training?

Cool stuff! Keep going man 😀👍

Wtf was that dick smith 😂

That's nice if you would leave it somwhere where there are a lot of animals it could create a big dataset xD

wonder how many species it detected total

Texas?

Drone content when 👀👀👀

This is awesome stuff, thank you for sharing

would this be possible to do with a Bambulab A1? I'm trying to figure out ways on how I could control it by code but I got no clue honestly