-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
MASTER THREAD: V 2019.40.1.1 170kW charging, Neural Net for Auto Wipers, Auto Lane Change quicker
- Thread starter Kmartyn
- Start date
it’s just a small bit of data transmitted every few minutes so i’m guessing way less than what it takes to use the gps.
I haven't received it. Checking Teslafi, it looks like only 10 SR+ have got the update so far.
Only 1 per TeslaFi. I usually get it around 600 installs (overall) on TeslaFi so a few hundred more and we shall see. I’m hoping to find the update notification when I wake up and I’ll be running around town tomorrow for stuff and to run the battery down so I can hit the V3 Supercharger here in Las Vegas to try that 170kW goodness.
Attachments
jebinc
Well-Known Member
I was one of he lucky (unlucky?) ones to receive this earlier in the week. No issues to report, but there must be some as the release seems to have stalled.
I received mine on Thanksgiving day. SR+ with FSD.
The recent rain has given a quick test for the new wipers. Much better than going on the fastest setting for too long.
Haven't tried supercharging yet.
Chris350
Member
Eno Deb
Active Member
Did my first drive with this firmware today. Auto wipers worked well in light to medium rain. I wasn't able to test the scenario where it most often failed with the previous versions though (spray from cars in front of you when the road is wet from heavy rain). The auto lane change is really more assertive, very nice. In one case though it was so assertive that I went "whoa, hold your horses". That was when I triggered a lane change into a lane that was wider than usual because of another lane merging, and the car swerved rather abruptly trying to find the center in the new lane. But generally I think this is a big improvement.
That was when I triggered a lane change into a lane that was wider than usual because of another lane merging, and the car swerved rather abruptly trying to find the center in the new lane. But generally I think this is a big improvement.It's not using the NN to operate in the car, the NN was just used to develop a model that the car applies to the video. It was doing this before, but the model is just more refined now. I doubt there is any change at all to power consumption.
This post didn't make any sense.
The NN certainly operates in the car, by running data from the video feed through the NN model. This requires GPU cycles, which increases power consumption. How much depends on the size of their deep rain NN network (larger networks require more calculations). The NN image run is a set of "neurons" with pretrained threshold values.
What is not done in the car is training the NN model. Training requires significantly more processor power and is done on the Tesla servers using training material (apparently 1 million pictures of rain droplets according to Tesla).
Off topic clarification: I assume they trained a new NN model with more robust inputs and labeling, but then deploy that NN to run on the car, to replace what they have now. Andrzej is big on touting how you can get drastically better classifier algorithms on constant-ish compute effort/power (a crappily trained Nn and a well trained NN of similar topologies will use same-ish power). Also, these simple models run on very low power for inference especially with dedicated hw and edge-scenario optimizations. Im not sure if beyond a “what speed should I run at” classifier they also collect user corrections for reinforcement learning style feedback. I wonder what things @verygreen can discover are being sent to the mothership.
MagnusMako
Member
I can't wait to see how good this neural net wiper system works. One thing I absolutely hate is when I'm driving down the highway in the rain while talking on the phone through the car and there's no way to adjust the wiper speed because the phone call display is overlapping it. I've had it where suddenly the rain will pick up speed and I needed to adjust the wipers but couldn't. Had to hang up the call so I could get to the button in the menu to increase it. This is probably one of the worst oversights on Tesla's interface if you ask me.
i was just trying to say that the new model isn't learning on the fly it's just applying the model that was learned from the training, which I doubt is that intensive.
Have a look at this video on how they added person detection to the HW on the $20 wyze cam, they have a model that runs on the wyze cam which is impressive. They used deep learning to create the model and then it runs on the cheap wyze cam hardware.
I'm assuming that the rain detection AI probably uses a similar training and application method.
I'm assuming that the rain detection AI probably uses a similar training and application method.
Tholland
Plaid/Black&ebony/21in wheels/OD- 5/31/21
Just so you know, you can adjust the wiper while on a call. You just swipe the call display to the right I believe. I’ve done it before. The dialer appears next the the wheel psi screen when in use. So you just have to keep swiping until you get to the home page where the wiper are the. Adjust. It is a nuisance but it can be done.
MagnusMako
Member
Tholland
Plaid/Black&ebony/21in wheels/OD- 5/31/21
Unless they removed it. I swear you could do that. I know I have done it a few times. It took a few times. But I’ll try again in the morning and try to video it as well
This makes sense and is correct.
However the computer power required for running a pre-trained NN is directly proportional to the number of neurons in the net. The number of neurons depends on what they needed to achieve a decent accuracy for detecting rain (which I assume is more than their first sucky attempt at this).
I guess @verygreen or someone else will see how large the network actually is once they get their hands on this firmware.
Similar threads
