Tianxiaomo / pytorch-YOLOv4

PyTorch ,ONNX and TensorRT implementation of YOLOv4
Apache License 2.0
4.48k stars 1.49k forks source link

Why in the config files it is recommended batch size-size=1? #399

Open vilmara opened 3 years ago

vilmara commented 3 years ago

Hi all, I need to deploy the model with dynamic batching on DS-Triton, but the YOLOV4 example in DeepStream says Following properties are always recommended: # batch-size(Default=1)

I run a test with yolov3 comparing BS=8 vs BS=1 and it exposed poor performance 0.24X , running TensorRT engine with DeepStream 5.1:

Throughput FPS (avg) | INT8

BS =1 → PERF: 246.29 (245.98) BS =8 → PERF: 60.31 (60.63)

Is this multi-batch poor performance related to the statement "Note that for multi-batch, overhead is large owing to NMS is not used."?, what does it mean?. What do you recommend to work with BS>1?