YOLOv5 Study: Speed vs Batch-Size #6649
glenn-jocher
started this conversation in
Ideas
Replies: 2 comments 3 replies
-
|
The results above is for Pytorch model format? |
Beta Was this translation helpful? Give feedback.
1 reply
-
|
@glenn-jocher I try the yolov5s with different batch size in T4 GPU, however it seems no gains in inference time on average of each image . I use cmd like
Have you try the batch size inference in T4 GPU? @glenn-jocher |
Beta Was this translation helpful? Give feedback.
2 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
Batch Size study here. Larger batches contribute to improved per-image inference speeds. Study run on Colab Pro+ with A100 40GB GPU. I used
val.py --data coco.yaml --task study, and just updated the code a bit to run the study over batch size instead of image size.Google sheet with complete results.
https://docs.google.com/spreadsheets/d/1Nm3jofjdgKja0AZHV8Jk_m8TgcF7jenCSA06DuEG2C0
Inference Time vs Batch Size
FPS vs Batch Size
Speed gains vs
batch-size 1Interesting takeaway is that smaller models benefit disproportionately from large batch sizes. These are the FPS gains vs batch-size 1 from the study:
Beta Was this translation helpful? Give feedback.
All reactions