HW2.5 capabilities

[AnxietyRanger]

Your statement implies that they have claimed what you claim on their behalf. AFAIK this is false. Please show me where a Mobileye spokesman claims that they have been working on deep learning neural networks since 2012.

I can not show it at this time. What I can say is that I have come to this understanding personally, to the extent I thought it was common knowledge. I believe it was some or one of their presentations where this was mentioned. As it is not the nature of info that would appear in easily sought press releases, it would take tons of presentation watching to look for it. So I have not looked for the evidence at all and I don't think it is good use of my time to do so, so I don't expect any of you to take my word for it.

You can consider it unproven and non-factual.

Your statement also implies that I am giving Tesla the benefit of the doubt. I believe this is also false. I will be happy to recant any statements where you can show me to be doing this.

I think, overall, there is a bias to look at Tesla optimistically and MobilEye pessimistically on this thread, yes. That has been the sense I've gotten out of it (and @Bladerskb apparently got the sense in spades ). It is not any one particular thing.

I'm not trying to be difficult here. I would sincerely like to know the source of your certainty. Here I am giving you the benefit of the doubt by assuming that such information exists but that I have simply not been able to find it. I don't think it's too much to ask that you share it with me.

Completely fair. It just becomes a bit futile on my end, because I do sense a certain bias facing any response I make, so it doesn't really seem like a good way to spend my time looking for things that in the end probably end up being interpreted differently based on biases anyway.

When it comes to assessing what companies do behind the scenes, it is a fools game when it comes to partisan arguments, that much has three years on TMC taught me. There is always so much ambiguity in the partial information that can be attained, that it is fodder for fights only. So one draws the line somewhere. I am not trying to be difficult either, but explaining what my general stance on this is and that is the extent I'm willing to go.

I understand it means some (many?) will just ignore these points as unproven - and so be it. Fair enough. Maybe at some point in the future it makes sense for some of you why I felt about this conversation this way, maybe it won't. Just wanted to share.

 

[buttershrimp]

Just an update to this thread. Software v 2017.38 f87c64d5 was updated last night to 2017.40.1 e29b97f. I am out of town (driving MB S550e) for the next couple of weeks so I cannot comment on the changes. If anyone else has received this new 2017.40.1 update, perhaps they can comment on any improvements with HW 2.5 capabilities. The only issue we have had with our MX 100D delivered on 9/29 with the .38 version has been Homelink stops working and music resume stops working. Now that we have received our first over the air update, we really feel like Tesla owners. Have fun dissecting v2017.40.1.

Yikes... nothing will make you appreciate your Tesla, like Mercedes Benz drive pilot.

 

[jimmy_d]

What was the low frequency wobble?

Ah - the first annoying thing I noticed when I upgraded from AP1 to AP2 was that AP2 would wander even on fairly straight lanes. AP1 never did this. I'm pretty susceptible to motion sickness so this behavior is toxic to me. After watching it for a couple of hours on a long drive I tried measuring it and found it was a pretty consistent 0.2hz oscillation. Anyway, I played around with it a bit to see how I could induce/suppress it. Varying speeds, following other cars, changing travel direction relative to the sun could cause it to come and go, but other than that it was pretty consistent.

Anyway - with 40 that seems to be less for me than it was on 38.

 

[jimmy_d]

I posted my initial thoughts on 40.1 over in AP 1.0 vs AP 2.0 Functionality Tracking

 

[calisnow]

Ah - the first annoying thing I noticed when I upgraded from AP1 to AP2 was that AP2 would wander even on fairly straight lanes. AP1 never did this. I'm pretty susceptible to motion sickness so this behavior is toxic to me. After watching it for a couple of hours on a long drive I tried measuring it and found it was a pretty consistent 0.2hz oscillation. Anyway, I played around with it a bit to see how I could induce/suppress it. Varying speeds, following other cars, changing travel direction relative to the sun could cause it to come and go, but other than that it was pretty consistent.

Anyway - with 40 that seems to be less for me than it was on 38.

Ah ok. Yeah I never described it as articulately as you have but it did have that. I don't notice it now - I'm on 17.34 still.

 

[verygreen]

A friend finally got 17.40.1

I can see that NN finally changed.

 

[d21mike]

A friend finally got 17.40.1

I can see that NN finally changed.

Could you explain what you mean by "I can see the NN finally changed"? I have seen a number of posts about NN which I think means Neural Net? Thanks....

 

[verygreen]

Could you explain what you mean by "I can see the NN finally changed"? I have seen a number of posts about NN which I think means Neural Net? Thanks....

Well, the firmware update comes with a pre-trained NeuralNet image (the NN is trained externally), and so we can see if it changed or not just by looking at the update.

 

[conman]

A friend finally got 17.40.1

I can see that NN finally changed.

Sweet. That probably explains the very different behaviour I noticed. The winding urban road test at night was the worst failure I've had in ages, however the highway experience is by far the best I've experienced.

 

[BigD0g]

@verygreen how big is it now?

 

[d21mike]

Well, the firmware update comes with a pre-trained NeuralNet image (the NN is trained externally), and so we can see if it changed or not just by looking at the update.

Cool, thanks for that.

 

[verygreen]

@verygreen how big is it now?

~15M

 

[jimmy_d]

A friend finally got 17.40.1

I can see that NN finally changed.

So maybe I'm not imagining things. Always hard to be sure when you're talking about incremental behavior on an NN.

BTW - I measured the size on the download and it looks like it was about 320MB overall (according to my network monitor)

 

[jimmy_d]

~15M

What number is that? File on disk / image in memory / other? NNs compress fairly well so the file stored on a drive might not be an accurate indication of the size of the network itself if it's a compressed parameter set.

If the network is being quantized to 8bit, which is pretty common recently, then 15MB is 15M parameters, which is a substantial network to be sure but it's on the small side right now. If they are using FP16 or FP32 (possible for a work in progress, but less likely if it's being treated as a mature image) then it's even smaller. Normally I'd guess that 15MB would represent a compressed or pruned parameter set on disk, but of course I have no idea what they are actually doing.

 

[J1mbo]

Would this be the first NN update since Karpathy joined?

 

[lunitiks]

Good question @J1mbo!

Allright I'm just going to ask it: What the heck does a NN file even look like?

 

[verygreen]

What number is that? File on disk / image in memory / other? NNs compress fairly well so the file stored on a drive might not be an accurate indication of the size of the network itself if it's a compressed parameter set.

It does not appear to be compressed since it compresses to 11M with gzip (I think it was compressing even better before).
That's just on-disk file.

If the network is being quantized to 8bit, which is pretty common recently, then 15MB is 15M parameters, which is a substantial network to be sure but it's on the small side right now. If they are using FP16 or FP32 (possible for a work in progress, but less likely if it's being treated as a mature image) then it's even smaller. Normally I'd guess that 15MB would represent a compressed or pruned parameter set on disk, but of course I have no idea what they are actually doing.

It takes 2 channel 416x640 image, 8 bit. Inside it starts with fp32 convolution but eventually drops to all integers before branching out another fp32 (confidence?) and a few ints as outputs.

 

[conman]

Would this be the first NN update since Karpathy joined?

I recall @verygreen saying it shrunk at one point though I'm not sure if that was after Karpathy joined or not.

 

[jimmy_d]

Good question @J1mbo!

Allright I'm just going to ask it: What the heck does a NN file even look like?

The file format can vary, of course. There are various frameworks that have their own format. And within a framework there are formats that have a lot of metadata and others that are just a binary file of parameters (sometimes also called weights). To describe a network completely you need some metadata that describes the topology (the shape for each layer and how it connects to adjacent layers), the nonlinearity used between adjacent layers, and special features like memory cells, max/averaging/pooling and so forth. Finally and most importantly you need a file of parameters, which is the multiplicative weight applied to the value that propagates from the output of one neuron to the input of the next. There's a separate value for each pair of interconnected neurons in a network and recent high performance neural networks for vision usually have in the neighborhood of tens of millions to hundreds of millions of parameters.

Generally this set of parameters will be the overwhelming majority of the contents of an NN file (think 99.999%) and it'll just be binary numbers. For development you generally use floating point numbers (because they are easy to work with) but to deploy commercial networks you usually 'quantize' the parameters to allow fixed point representations, which are both smaller and faster to compute. A very common size for quantization these days is 8bits, in which case the size of an uncompressed parameter file is the same as the number of parameters. However, using a compressed parameter set can lead to 10x or greater file size reductions, though you have to uncompress it for use, of course.

From the Tesla NN file, assuming it isn't encrypted, it might be possible to determine the format, and given that, the network size and shape.

 

[lunitiks]