I am a Beginner of Deep Learning. I have had a hands-on experience on Machine Learning. Now I am gonna do some project on Deep Learning based Object Detection using CNN model. For that I need a decent PC with GPU in a Budget Price. What are the minimum requirements for a GPU and CPU?. I can only buy NVIDIA GTX 1050 (4 GB) or 1060 (6 GB).
For decent speed, what CPU specifications do I need ?
What type of data sets we can use with Single GPU?
There’s a good discussion in this medium article.
I doubt you’ll find a parts list but you’ll better understand a lot of things you need to optimize for.
Also, if you are very new to deep learning and on a limited budget I recommend AWS spot instances. I hear paperspace and crestle are good places to start as well.
My first card was a gtx 1050 ti. It worked fine for most of the examples. You can train sample sets on your CPU pretty easily as well. Instructions are in the videos and mentioned on github in the fastai repo README.md
As opposed to building your own machine definitely consider some of the clouds offerings. The cloud will enable you to gain access to better computation power like V100 which have Tensor cores enabling faster computation. Additionally, another reason to go with the cloud is it constantly updates with new hardware(SSDs, CPUs, GPUs).
Thank you replying me telarson.
I got committed a project on DL based Object detection. I need a standalone system. So, the minimum hardware requirements to do small projects with a limited data sets. Because, I’m going to deploy it on Raspberry Pi. For that too much size of data set is no use.
My thinking on hardware requirements is:
- Intel I5-Processor (3.2 Ghz)
- GTX 1060 / 1050ti (4/6 GB) GPU
- RAM 8/16 GB?
- Storage SSD 250GB or HDD 7200 Rpm (1 TB)
These are the requirements I have noted. Can you optimize it?
Can you tell us from where you would be purchasing your system? India? Also are you into DIY or Will go for system integrators like Boxx or Ant PC for the solution? However later is expensive yet reliable while DIY all depends on skillsets.
Now coming to the configuaration:
-> Intel I5-Processor (4.0 Ghz) --> Buy Intel Core i5 8400 but the prices are quite high you can event go for AMD Ryzen 1600.
–> GTX 1060 / 1050ti (4/6 GB) GPU --> Higher is better, Go for 1060 6GB
—> RAM 8/16 GB? —> Again if budget allows go for 16GB with atleast 2400Mhz Frequency else go for 8GB as of now but of 2400Mhz from Brands like G,Skill or Corsair with heatsink and good timings like Ripjaws Series of G.Skill
—> Storage SSD 250GB or HDD 7200 Rpm (1 TB) —> Go with 120GB SSD + 1TB HDD Cpmbo…
–> Cabinet: Buy anything as of now like from iball/Zebronics if you are purchasing from India.
—> PSU: 500W will get the job done provided it has 75% Efficiency and from Brands like Cooler Master or Corsair.
—> OS: Ubuntu which is again free go for 18.04LTS
Rest pheripherals I assume you have. Hope this helps. Do share with us build images once you done with the same.
Thank you for giving me a suggestion. Your post helped me. Now I can build my system based on that as you said benchmarks and budget price. I will send a build images once i have done. I would like to ask another question. I am planning to deploy in Raspberry Pi 3B+. So, which model or architecture would help me to better prediction results on object detection. I have found some models like SSD MobileNET and YOLO V2. But I don’t know the benchmarks of those models with these above hardware as Sahil Mentioned. Could anyone suggest me to explain the benchmarks of those models.
As per my research SSD MobileNet adds additional feature maps from different on Yolo but again it is more memory intensive and but again SSD is better,. I don’t have much information about Raspberry so sorry I can assist you with the same.
Thank you Sahil. Do you have any prototype or tutorial on how to start with SSD MobileNet. I need a some kind of prototype to start working with Deep Learning. If you know any kind of tutorial or some Github , you can mail me email@example.com or post any link that. Please help me to start my deep learning.
Sure. I shall connect with my contacts and will share with you if I find one All The Best.
Once again Thank you Sahil.