HW2.5 capabilities

[alcibiades]

Also somewhat unexpectedly the changelog is completely empty for everybody which I don't think I saw before usually there were always conditional changes listed.

What are "conditional changes" in this context? Just code mumbo-jumbo or something a moron like myself can understand?

 

[lunitiks]

@verygreen - You commented that the PID was written in Python which, assuming you're not having fun with us, is a bit shocking

Can you please translate this to English? Why is Python code shocking? I'm not following

 

[MP3Mike]

@verygreen - You commented that the PID was written in Python which, assuming you're not having fun with us, is a bit shocking

@sandpiper I think you have confused your information @chillaban commented that Comma.ai used Python for their steering control. I haven't seen anyone say that Tesla uses Python.

 

[MP3Mike]

What are "conditional changes" in this context?

It means that the changes are displayed conditionally. For example only display this change if the car is in China. Or only display this change if the car has no AP hardware, etc.

 

[chillaban]

Right. I was talking about Comma and not Tesla.

 

[BigD0g]

@verygreen - You commented that the PID was written in Python which, assuming you're not having fun with us, is a bit shocking. Do you know what platform the NN was developed on? Tensorflow, Theano, Torch, or something of Tesla's own creation?

Based on Kapanarthy's twitter, I suspect they are using PyTorch...
Andrej Karpathy on Twitter

 

[verygreen]

@verygreen - You commented that the PID was written in Python which, assuming you're not having fun with us, is a bit shocking. Do you know what platform the NN was developed on? Tensorflow, Theano, Torch, or something of Tesla's own creation?

PID? I am not even sure what are you talking about.
I listed the convolution kernels they use somewhere, but I am not an NN expert to know much in this area, somebody identified some google code in those.

 

[JeffK]

Based on Kapanarthy's twitter, I suspect they are using PyTorch...
Andrej Karpathy on Twitter

They were using Neon for at least some of it for sure as evidenced in one of @verygreen's older posts
Is Tesla using machine learning for AP?
some openai stuff too. openai-gemm, for example, also relies on Neon.

 

[conman]

I have some data on this, with the latest update I received .34 xyz, two times I pulled off an exit ramp at 70 mph with untracked cars at the stop sign at the end of the exit and AP2 started slowing about 150 - 200' away from the car and was very smooth, almost too conservative, but the fact that I was essentially flying and it picked up and started to slow for untracked stopped cars was really good. Also, had the same experience with a few stop lights and untracked cars with the same behavior. I've only had the build for 24hrs or so and more testing is needed, but so far it's looking good.

However, it still struggles with the damn curves, and wants to drift into the outside lane, but I'll take what I can get.

Oh and it's now showing two cars in front.

Thanks. I was on .32.6 and only just received notification of a new update now which is currently installing. I'll see what it does shortly

EDIT: Looks like I got the fairly wide release .34 I assume this is the first one that has started logging? I plan to use AP+AS as much as possible if that's so.

 

[fasteddie7]

However, it still struggles with the damn curves, and wants to drift into the outside lane, but I'll take what I can get.

Oh and it's now showing two cars in front.

I had quite the opposite experience with AP1. My Tesla used to barrel into curves and slam on the breaks halfway into it but due to the speed it would cross the line onto the shoulder. The odd thing was that if auto steer was disabled and only TACC was on, it slowed early and I steered the curve, but it would accelerate way too early while coming out. Today after the update it slowed down right before and took the curve, then slowly accelerate out. I was shocked. (All the same curves I’ve driven for months on my commute)

 

[mrkisskiss]

some openai stuff too. openai-gemm, for example, also relies on Neon.

Well - at the risk of being pedantic... NEON is basically an set of ARM vector accelerators that the compiler uses... it's not a language per se. The equivalent on Intel would be their performance primitives - IPP, or vDSP in Apple land.

I'm sure that a lot of the AP ML stuff is written in Python - it's pretty standard for ML stuff.

 

[JeffK]

Well - at the risk of being pedantic... NEON is basically an set of ARM vector accelerators that the compiler uses... it's not a language per se. The equivalent on Intel would be their performance primitives - IPP, or vDSP in Apple land.

I'm sure that a lot of the AP ML stuff is written in Python - it's pretty standard for ML stuff.

Pedantic, yes, but also slightly inaccurate and out of context.
The poster I was replying to and the poster that user was replying to were naming frameworks and not languages.

Oh, and if you haven't followed the news, Intel bought Nervana Systems last year.
neon™ - Intel Nervana

 

[sandpiper]

@sandpiper I think you have confused your information @chillaban commented that Comma.ai used Python for their steering control. I haven't seen anyone say that Tesla uses Python.

Doh!!! Apparently I was double confuzzled. I mis-remembered both the name of the commenter and the company!

 

[mrkisskiss]

Pedantic, yes, but also slightly inaccurate and out of context.
The poster I was replying to and the poster that user was replying to were naming frameworks and not languages.

Oh, and if you haven't followed the news, Intel bought Nervana Systems last year.
neon™ - Intel Nervana

Hehe, touché - my apologies!

 

[chillaban]

NN in 17.34 is the same as before.
A new binary appeared named "clip_logger", I guess I really need to install it to try it out now.

Also somewhat unexpectedly the changelog is completely empty for everybody which I don't think I saw before usually there were always conditional changes listed.

Curious, did you see if there's anything significant in 2017.34 or what that binary does?

Like a lot of others, I also seem to log more upload traffic from 2017.34 than ever before.

 

[verygreen]

Curious, did you see if there's anything significant in 2017.34 or what that binary does?

Like a lot of others, I also seem to log more upload traffic from 2017.34 than ever before.

the binary seems to be a debugging aid they forgot to remove as I now see, it lacks some other infrastructure, though and it seems to assume a sata drive plugged into usb-c for the logging.

The most significant thing I see now if that vision task is really crashy. In the 2x 2min test drive I just took I had like 8 crashes.

Also saw a new log message in the logs that I have not seen before:

RAW: Retrieving HW ID for back-up camera on this type of vehicle is not supported
RAW: Error initializing sparse logging buffer consumer for camera at position 8: Could not open sync file on time

Hopefully that does not mean I'll no longer have backup camera images in the snapshots.

 

[chillaban]

the binary seems to be a debugging aid they forgot to remove as I now see, it lacks some other infrastructure, though and it seems to assume a sata drive plugged into usb-c for the logging.

The most significant thing I see now if that vision task is really crashy. In the 2x 2min test drive I just took I had like 8 crashes.

Also saw a new log message in the logs that I have not seen before:

RAW: Retrieving HW ID for back-up camera on this type of vehicle is not supported
RAW: Error initializing sparse logging buffer consumer for camera at position 8: Could not open sync file on time

Hopefully that does not mean I'll no longer have backup camera images in the snapshots.

Thanks as always! It's really odd that they slowly trickle out a firmware like this and then push it to everyone….

 

[Bladerskb]

You realize that what you just posted says very clearly that radar can detect an object that could be a deer thus proving @stopcrazypp correct even though you claim it's incorrect. Read the picture you posted.

Radar is used to see different sizes of aircraft all the time...

It is YOU that need to read that picture.

Note it says "Know there's a large object that COULD be a deer" then it says "can't differentiate objects such as a deer from a big rock."

 

[JeffK]

It is YOU that need to read that picture.

Note it says "Know there's a large object that COULD be a deer" then it says "can't differentiate objects such as a deer from a big rock."

It doesn't need to differentiate. You don't want to hit a big rock or a deer.

 

[Bladerskb]

Don't give me that BS. Yes a picture is a collection of numbers, so is lidar data, and radar data. Attach a computer to a camera and you can tell the difference between all sorts of things. To say otherwise is just a lie and you contradict yourself in previous posts.

Incorrect. Lidar returns 3d point coordinates that requires no additional processing. A picture is simply numbers that need cutting-edge sophisticated machine learning of deep neural network to understand.

Lidar and Radar has existed and used for hundreds of years.

Lidar has been used for dozens of years to detect and classify objects. Why? because it returns 3d data points which you can then simply plot on a graph using a simple python library called numpy, using the xyz coordinates or any other library or even manually by hand.

Its would be like plotting a graph back in elementary school. the same can't be said for picture however.

picture data just got started being reliably used to classify objects with the advent of deep learning in 2012.
Before that not even the most powerful super computer in the world could tell you that a picture is a cat or a dog.

Chris Bishop: Even the most sophisticated computers can't tell a dog

Radar doesn't need the same corrections to measure distance through fog, snow, rain, etc. Lidar has to throw away data from laser reflections on nearby objects whereas radar can go through water.

Yes it does, everything needs corrections. That's a vast difference from processed data to remove any known artifacts and noise vs. raw data. All hardware data are processed.

No one said radar is the primary sensor, all current autonomous vehicles use cameras. Waymo's implementation uses eight cameras and 360 degree radar, in addition to it's lidar that it likes to show in presentations.

Then stop arguing radar over lidar when its useless in comparison!

As far as lidar working in heavy rain and snow do you have a source that explicitly says heavy rain or heavy snow?

Ford says they have their lidar working in blizzard situation.