search

EdjeElectronics / TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10

How to train a TensorFlow Object Detection Classifier for multiple object detection on Windows
Apache License 2.0
2.92k stars 1.3k forks source link

Graphics Card Recommendation #422

Open SouradipBh opened 4 years ago

SouradipBh commented 4 years ago

I am running video-detection application using FRCNN- inception v2 model( using Tensorflow backend). I have to run around 50 parallel video detection simultaneously. Please advice which Graphics card or Combination of Graphics should be able to handle the load of Graphics.

My options are : - i) 2 x RTX-TITAN ii) 4 x RTX 2080ti iii) 8 x GTX 1080ti

JulianOrteil commented 4 years ago

@SouradipBh That is an insanely heaving load. Are you hoping for real-time object detection? With that much data, I don't think you could come even close to real-time without spending A LOT of money.

As for your choices, I'd actually recommend looking at the Tesla GPUs that NVIDIA offers. They are designed for AI and object detection. I'd recommend getting a couple if you can afford it. Beware: they are extremely expensive.

Also, another thing you should look into something called TensorRT (and/or TensorRT Inference Server if you plan to have execution on a server or server-like computer). It optimizes your networks to bring down their size which increases their throughput and speed. Just be careful when doing the optimization. Not all the ops that are being performed in the model are supported, so you may need to implement your own op conversions to get it to work.

turowicz commented 4 years ago

Look into converting your model to TensorRT and then leverage the RTX card's tensor cores in order to squeeze out the performance. You should get 5-10fps for 50 HD/FHD feeds on a single Titan RTX.