Detectnet v2, tlt-infer and performance

Hello everybody,

Our objective is to know the performance of a detectnet v2 model (to compare it to a ssd or a faster rcnn).

We can’t find any benchmark with detectnet v2.
Can we say it’s a fast model (like ssd) or slower but more accurate (like faster rcnn)?

We have trained a detectectnet v2 faster resnet18 model:

  • Image resolution: 1280x1024
  • Number of classes: 2
  • GPU used: rxt 2080 ti

We got respectable Average precision at the end of the training.
Now we want to see the results on test images with tlt-infer but it’s crashing because of the lack of gpu memory.

Is detectnet v2 a model which consumes a lot of GPU ram, or did we misconfigure something?

thanks

Hi steventel,
TLT detectnet v2 network is faster than TLT ssd and TLT faster-rcnn network.
For KITTI dataset, TLT get the better mAP than TLT ssd and TLT faster-rcnn network.

For OOM, what is the log?

Reference: W tensorflow/core/framework/op_kernel.cc:1401] OP_REQUIRES failed at pack_op.cc:88 : Resource exhausted: OOM when allocating tensor with shape[32,3,2160,4096] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc