jimmy_d
Deep Learning Dork
I wonder if either way would affect how long your wipers last. I always cringe on a dry wipe, though in truth I have no idea if it really matters.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
I wonder if either way would affect how long your wipers last. I always cringe on a dry wipe, though in truth I have no idea if it really matters.
How long your wipers last, yeah, dry wipes are definitely no good. Especially if your windshield is hot (e.g. a false positive in the summer).
How long your wiper motors/arm links last, that's the scarier thing. IIRC, "chatter" (e.g. when you have rain-X and wipe prematurely and the blade bounces across your windshield) are absolutely awful to the link that connects your wiper arms together. Fast speeds tend to push metal fatigue to a complete wiper failure too.
lunitiks
Cool James & Black Teacher
I had that happen to my other car not long ago. ice/snow buildup on the wipers. was too lazy to remove. a nut unscrewed and i had to open everything and tighten
jimmy_d
Deep Learning Dork
Quite true. And you can always train to a scalar output, but somehow people don't like that with nonlinear classifiers.
It just seems like using a log probability classifier with softmax has become the default way stuff gets done. I guess I should go back and look to see if there was a nonlinearity applied to that one output neuron. Maybe I missed something there.
Joe F
Disruption is hard.
Now there you go, talking common sense, ruining a perfectly good bashing of Tesla and their inoperative wiper solution. How dare you!
Joe F
Disruption is hard.
Now there you go, talking common sense, ruining a perfectly good bashing of Tesla and their inoperative wiper solution. How dare you!
Note of clarification: Posts after @Daniellane do not fall into the bashing camp and are added value. I should read ahead first.
Joe F
Disruption is hard.
Funniest thing is, the car I drove when I was in high school had a wiper system from his company, and he would always yell at me whenever I put the wiper in the fast setting. Talk about believing in what you worked on
I drove a '59 Rambler Classic. It had a vacuum wiper motor that ran off the intake manifold. Wipers stopped dead on the slightest acceleration.
whitex
Well-Known Member
If I try to sum up, they could choose one of two approaches:
1) start a preset wiping frequency for x seconds. (Needs multiple outputs of rain intensity and a set number of wiper speeds, 1/8, 1/4, 1/2, 1/1, 2/1)
2) single wipe on demand (rain=true/false, fixed sampling frequency, max speed 2 wipes/sec)
Or?
1) start a preset wiping frequency for x seconds. (Needs multiple outputs of rain intensity and a set number of wiper speeds, 1/8, 1/4, 1/2, 1/1, 2/1)
2) single wipe on demand (rain=true/false, fixed sampling frequency, max speed 2 wipes/sec)
Or?
3) If they can show a solid correlation between high score on their single output and rain intensity, they could use multiple thresholds on that value to run the different speeds.
And, I guess, to get really crazy with it: 4) An RNN(likely necessary to make sure it adjusts based on rain over time) outputting a wiper speed.
Now, there is nothing for me to be anxious about or debate when you guys are so reasonable and correct and analytical.
I will go exploring the world of rap with @buttershrimp instead.
Oh good, I found something.
No, we believe from research by @verygreen and @jimmy_d that on the contrary the NN is outputting yes/no binary answers - it used to output levels at some stage, but that was removed...
AP1 on the other hand, using the off-the-shelf chip that it does, outputs levels.
No, we believe from research by @verygreen and @jimmy_d that on the contrary the NN is outputting yes/no binary answers - it used to output levels at some stage, but that was removed...
AP1 on the other hand, using the off-the-shelf chip that it does, outputs levels.
jimmy_d
Deep Learning Dork
If I implied at some point that there were binary answers coming out please allow me to correct that error. I believed that the output configuration matched a probabilistic output (which is standard for a classifier). So in that case the output would then be a scalar floating point value between 0 and 1 indicating the NN's evaluation of the likelihood that what it is seeing is rain. The output also happens to be named "rain_prob_5", which I take to be shorthand for "rain probabilities 5". Why 5? Vestigial variable name from an earlier version that had 5 categories.
In 42 the same NN output five different values and was also named "rain_prob_5". My guess at the time was that 42's output was 5 different probabilities that represented the relative likelihoods that the rain situation fell into one of five categories. In 44 they got rid of 4 categories but didn't change the variable name.
Prompted by @MarcusMaximus earlier comment I went back to look at the final output stages and realized that they are linear (unlike every other convolutional layer in the fisheye network). That means each of the categories in rain_prob_5 (previously 5 but now just one) is very likely being trained as a scalar.
So what this means is that the output is a number of some sort. It could still be a probability that the NN is seeing rain, but it could also be some indication of how heavy the rain is or, potentially, anything else that you can measure with a single number.
lunitiks
Cool James & Black Teacher
This is confusing as hell.
My theory: The final layer has only one possible output if the previous layers determines that rain is likely. So then the output will be «rain».
The «rain» output is relayed to BCM which, depending on your stalk setting, asks your wiper ecu/motor to engage.
So the stalk settings are merely programmed to ask BCM to engage «immediately» (Auto 2) or ... well, «a bit later» (Auto 1). Or not at all (Off).
I may be very, very wrong on this, but this is my understanding right now (still trying to learn).
I think it should be possible to field test somehow. Need water hose and a stop watch
My theory: The final layer has only one possible output if the previous layers determines that rain is likely. So then the output will be «rain».
The «rain» output is relayed to BCM which, depending on your stalk setting, asks your wiper ecu/motor to engage.
So the stalk settings are merely programmed to ask BCM to engage «immediately» (Auto 2) or ... well, «a bit later» (Auto 1). Or not at all (Off).
I may be very, very wrong on this, but this is my understanding right now (still trying to learn).
I think it should be possible to field test somehow. Need water hose and a stop watch
What's surprising is that auto-wipers are working correctly for non-AP2 cars. The primary difference with AP2 should have been the method to detect if it's raining - and how much it's raining - to determine how quickly the wiper motor should be running. The non-AP2 cars don't have the issue of wiping water inside an open car door.
Then again, it's not surprising - sometimes the software coming from Tesla can be a real head scratcher - like when they decided to completely remove the time-of-day clock from the console display or someone thought it was a great idea to remove the first letter scroll bar from the media player - and we're still waiting for them to get Shuffle mode with the USB media player to actually play random songs.
So we probably shouldn't be that surprised that they missed this obvious issue with the auto-wiper software - which should be something they can easily fix, and now that it's been reported publicly, very likely they'll push that update out quickly (next few days???).
Then again, it's not surprising - sometimes the software coming from Tesla can be a real head scratcher - like when they decided to completely remove the time-of-day clock from the console display or someone thought it was a great idea to remove the first letter scroll bar from the media player - and we're still waiting for them to get Shuffle mode with the USB media player to actually play random songs.
So we probably shouldn't be that surprised that they missed this obvious issue with the auto-wiper software - which should be something they can easily fix, and now that it's been reported publicly, very likely they'll push that update out quickly (next few days???).
If they keep up their "every even week of the year" cadence, there should be another update around the end of this week(presumably 2018.2)
Uncle Paul
Well-Known Member
Since this first release of the rain sensing wipers is labeled as a Beta, I would imagine that it will be improved over time, and with over the air updates for all users.
