Machine Learning/Object Detection PC Specs
Discussion
Just wondering if anyone had any experience of this?
I am looking to Spec up a PC act as the training PC.
Minimum i7 Probably 10th Gen, maybe an i9 if the bit coin miners haven't bought them all .
32GB RAM (DDR4 probably)
GPU is almost a necessity for quick running (it should be running in real time eventually) so probably at least a 12GB Nvidia.
Obviously, there is the option of looking for a spec equivalent to the Tensor book, but I would rather understand the choice rather than going in partially sighted.
It will be used as a Label reader just to prove the technology in a single location (yes, I agree there are easier ways), the reason we are going down this route is... in the long term it will be linked into the product release standards; packaging, labelling etc.
I will be using Tensor Flow (probably Win10) although speaking to a few people Linux may be more stable?
I understand that we are almost looking at effectively a good spec gaming PC, but as it is Object detection rather than gaming, are any chip sets better than others for this application?
At the moment it will be just in one location, but as we progress we may need to adapt it into other locations so there may be several models to cope with the varying environments and challenges. As I get each model sorted, I will probably moving them to a server; freeing up the training PC for new and exciting projects!
There is no worry that if it doesn't work we have a high spec PC hanging around doing nothing, I will need to lock it away and access remotely as there are so many people salivating at the possibility!
Your thoughts?
I am looking to Spec up a PC act as the training PC.
Minimum i7 Probably 10th Gen, maybe an i9 if the bit coin miners haven't bought them all .
32GB RAM (DDR4 probably)
GPU is almost a necessity for quick running (it should be running in real time eventually) so probably at least a 12GB Nvidia.
Obviously, there is the option of looking for a spec equivalent to the Tensor book, but I would rather understand the choice rather than going in partially sighted.
It will be used as a Label reader just to prove the technology in a single location (yes, I agree there are easier ways), the reason we are going down this route is... in the long term it will be linked into the product release standards; packaging, labelling etc.
I will be using Tensor Flow (probably Win10) although speaking to a few people Linux may be more stable?
I understand that we are almost looking at effectively a good spec gaming PC, but as it is Object detection rather than gaming, are any chip sets better than others for this application?
At the moment it will be just in one location, but as we progress we may need to adapt it into other locations so there may be several models to cope with the varying environments and challenges. As I get each model sorted, I will probably moving them to a server; freeing up the training PC for new and exciting projects!
There is no worry that if it doesn't work we have a high spec PC hanging around doing nothing, I will need to lock it away and access remotely as there are so many people salivating at the possibility!
Your thoughts?
mikef said:
AWS, GC or Azure all offer specialised ML environments We have a couple of hundred data scientists and wouldn’t think of doing ML on a desktop
Agreed, a desktop isn't the place to do it. Cloud environments will save you so much time (especially if you are still not decided on points such as Win10 vs Linux).If you *must* be on-premise, then look at server hardware, not desktop.
biggiles said:
mikef said:
AWS, GC or Azure all offer specialised ML environments We have a couple of hundred data scientists and wouldn’t think of doing ML on a desktop
Agreed, a desktop isn't the place to do it. Cloud environments will save you so much time (especially if you are still not decided on points such as Win10 vs Linux).If you *must* be on-premise, then look at server hardware, not desktop.
Thanks for the advice; it's appreciated
.I read your intro and thought “duh, AWS”, you have my sympathies on the restrictions. Stuff like Rekognition is incredibly clever and easy to use, I had my son using it for his class project ...
Indeed, you want as good a gaming PC as you can get, minus the light show. I would go large on the memory - probably 64 GB or more which will narrow the motherboard choice a bit. You’re going to end up wanting to run VMs on it at some stage, and then all your memory will vanish. And good luck getting a graphics card - my old 1660TI is now apparently worth more than I paid for it.
Indeed, you want as good a gaming PC as you can get, minus the light show. I would go large on the memory - probably 64 GB or more which will narrow the motherboard choice a bit. You’re going to end up wanting to run VMs on it at some stage, and then all your memory will vanish. And good luck getting a graphics card - my old 1660TI is now apparently worth more than I paid for it.
Indeed. I attended some good sessions at the last real life AWS:Reinvent in Vegas:
All sessions linked here: http://aws-reinvent-audio.s3-website.us-east-2.ama...
Maybe time to look for new Overlords of IT if ML is your thing?
- Labelling with SageMaker: https://www.youtube.com/watch?v=6WJxzKsIFKA
- AWS instance types for ML: https://www.youtube.com/watch?v=vRB9Uncsia8
- Scaling Tensorflow on AWS: https://www.youtube.com/watch?v=iW0RASdjnOk
All sessions linked here: http://aws-reinvent-audio.s3-website.us-east-2.ama...
Maybe time to look for new Overlords of IT if ML is your thing?
GrizzlyBear said:
biggiles said:
mikef said:
AWS, GC or Azure all offer specialised ML environments We have a couple of hundred data scientists and wouldn’t think of doing ML on a desktop
Agreed, a desktop isn't the place to do it. Cloud environments will save you so much time (especially if you are still not decided on points such as Win10 vs Linux).If you *must* be on-premise, then look at server hardware, not desktop.
Thanks for the advice; it's appreciated
.Perhaps when they see the capital cost of an AI-spec server they might reconsider their decision.
Show them this Nvidia DGX A100 'PC' and I'm sure they'll suggest AWS is better solution.
Edited by mmm-five on Monday 26th April 14:49
mmm-five said:
Show them this Nvidia DGX A100 'PC' and I'm sure they'll suggest AWS is better solution
Yeah, do it. The cloud environments are not dissimilar, but you only spin up as big a cluster as you need (Databricks etc handles that) and for as long as the training job runs
Are they different parts/ labels every time?
Damaged labels or freshly applied?
I worked in a factory who built their own optical defect detection systems and I don't think the hardware was that flash... but it was programmed to do one job. Towards the end I am sure that cameras were available with what we were doing on PC on the camera itself which would trigger the eject mechanism rather than the pc.
I must be be behind the times, but if the labels are consistent and fresh why is ML required? I can imagine that you teach it defects and after a while it makes it's own decision on what to accept rather than a pre-determined blob size and greyscale value.
We'd typically take a photo of a part, different parts had different strobe intensity, maybe an infra red setup for certain items. Potentially create a binary filter to expose defects. Then locate a feature to align the thing, then look for blobs of pixels outside an agreed darkness/size to detect a defect, or search a certain area with a text recognition tool, number of blobs within certain region etc.
Actually thinking about it... training it rather than programming it would probably be a dream!!
Damaged labels or freshly applied?
I worked in a factory who built their own optical defect detection systems and I don't think the hardware was that flash... but it was programmed to do one job. Towards the end I am sure that cameras were available with what we were doing on PC on the camera itself which would trigger the eject mechanism rather than the pc.
I must be be behind the times, but if the labels are consistent and fresh why is ML required? I can imagine that you teach it defects and after a while it makes it's own decision on what to accept rather than a pre-determined blob size and greyscale value.
We'd typically take a photo of a part, different parts had different strobe intensity, maybe an infra red setup for certain items. Potentially create a binary filter to expose defects. Then locate a feature to align the thing, then look for blobs of pixels outside an agreed darkness/size to detect a defect, or search a certain area with a text recognition tool, number of blobs within certain region etc.
Actually thinking about it... training it rather than programming it would probably be a dream!!
Glade said:
Are they different parts/ labels every time?
Damaged labels or fresh
Ah, I had assumed that The OP was referring to labelling in the Machine Learning sense... Damaged labels or fresh
https://aws.amazon.com/sagemaker/groundtruth/what-...
biggiles said:
"Push back" on them?? It's cheaper and generally more secure than your on-premise IT. Very few firms "nowadays" have a reasonable excuse to say "no Cloud". Even governments and Swiss banks use cloud now. Maybe MI5 doesn't.
Perhaps when they see the capital cost of an AI-spec server they might reconsider their decision.
CIA is on AWSPerhaps when they see the capital cost of an AI-spec server they might reconsider their decision.
mikef said:
Glade said:
Are they different parts/ labels every time?
Damaged labels or fresh
Ah, I had assumed that The OP was referring to labelling in the Machine Learning sense... Damaged labels or fresh
https://aws.amazon.com/sagemaker/groundtruth/what-...
Edited by Glade on Tuesday 27th April 22:31
Sorry, that wasn't thinking, they are printed labels, print quality is consistent (I won't even consider it until easy fixes like that are sorted), label placement is not so consistent, package size varies. I have suggested several cheaper (and potentially easier) solutions), but they have their hearts set on a object detection solution, OK your money.
I suspect their desire for Object detection solution is for it to be expanded into automated quality checks etc in time, which would make sense.
e.
Looks like things are starting to come together, so we will see how it goes. Thanks for all your support, 5 years of stress here we come.
I suspect their desire for Object detection solution is for it to be expanded into automated quality checks etc in time, which would make sense.
maffski said:
Why not suggest cloud for development & evaluation - which should appeal to the accountants.
You would think that wouldn't you, and especially as the accountants know me. I would also prefer the cloud as if the project falls on its face; when I get dragged onto another project, we don't have an asset gathering dust... oh well not my money!biggiles said:
"Push back" on them?? It's cheaper and generally more secure than your on-premise IT. Very few firms "nowadays" have a reasonable excuse to say "no Cloud". Even governments and Swiss banks use cloud now. Maybe MI5 doesn't.
Perhaps when they see the capital cost of an AI-spec server they might reconsider their decision.
I agree, went right to the top and the decision was "no cloud please, but make it happen" so pick your battles Mr Bear, and get all the decisions in writing to cover your aPerhaps when they see the capital cost of an AI-spec server they might reconsider their decision.
e.Looks like things are starting to come together, so we will see how it goes. Thanks for all your support, 5 years of stress here we come.
Would a system like this work?
https://www.keyence.co.uk/products/vision/vision-s...
I have just been browsing their CX-V catalogue and the tech has come on a lot since I was using it... they even have their own auto learning feature.
https://www.keyence.co.uk/products/vision/vision-s...
If it's physical part/label inspection this type of thing is tried and tested technology - not sure what the benefit of developing your own machine learning is... especially if the label print is not ambiguous. (And even if it was a more traditional vision system could manage it).
https://www.keyence.co.uk/products/vision/vision-s...
I have just been browsing their CX-V catalogue and the tech has come on a lot since I was using it... they even have their own auto learning feature.
https://www.keyence.co.uk/products/vision/vision-s...
If it's physical part/label inspection this type of thing is tried and tested technology - not sure what the benefit of developing your own machine learning is... especially if the label print is not ambiguous. (And even if it was a more traditional vision system could manage it).
Edited by Glade on Saturday 8th May 22:34
Gassing Station | Computers, Gadgets & Stuff | Top of Page | What's New | My Stuff