ModelOpt (INT8) · TensorRT
- torchvision
- INT8
- L4
- Labels
- Image
- Segmentation
- Build
- Embedl-optimized
- Precision
- INT8
- Runtime
- TensorRT 11.2.1 with Embedl optimizations
deeplabv3_mobilenet_v3_large is a semantic segmentation model from torchvision: it assigns a class to every pixel of an image.
Methodology
- Latency: trtexec on the NVIDIA L4 with TensorRT 11.2.1: batch 1, median and p99 over the timed iterations.
- Accuracy: the COCO segmentation validation set, 2,000 images, on the NVIDIA L4.
- Build: ModelOpt INT8 quantization, calibrated on CocoSegmentationCalibDataset (64 samples), compiled with TensorRT 11.2.1 on the NVIDIA L4.