Skip to content

Latest commit

 

History

History
71 lines (46 loc) · 3.21 KB

README.md

File metadata and controls

71 lines (46 loc) · 3.21 KB
Raw

Engine Creation Instructions

Generating TensorRT Engines

Default

  1. cd models; mkdir -p onnx

  2. Download the component ONNX files

    • Listed here; save them to the onnx directory within this folder.

  3. Run bash create_default_engines.sh

    • Models generated with FP32 precision: image encoder
    • Models generated with FP16 precision: image decoder, depth estimation model, face detection model, gaze estimation model, object detection model

Optimized

  1. cd models; mkdir -p tensorrt/int8/caches

  2. Download INT8 calibration caches

    • Listed here; save them to tensorrt/int8/caches.
    • Depending on your download method, the filenames may contain the gazesam_int8_calib_caches_ prefix. To remove this prefix, run rename 's/^gazesam_int8_calib_caches_//' gazesam_int8_calib_caches_*.cache (while cd'ed into tensorrt/int8/caches).

  3. Run bash create_optimized_engines.sh

    • Models generated with FP32 precision: image encoder
    • Models generated with FP16 precision: image decoder, depth estimation model, face detection model
    • Models generated with INT8 precision: gaze estimation model, object detection model

Generating ONNX models

Note that default ONNX models are available, so this section is likely not going to be relevant to you unless you'd like to generate your own ONNXes. If you plan to generate an ONNX model and later use it to compile an engine, please remember to replace our defaults with your new file!

Instructions below indicate how to recreate our ONNX models.

Face detection model

Gaze estimation model

  • L2CS-Net model, downloaded by running this script and choosing the l2cs_net_1x3x448x448.onnx variation.

Depth estimation model

  • Depth-Anything-M model, downloaded by following these instructions. We use vitb_14 by default.

YOLO object detection model

  • python create_onnx/create_yolo.py --model-size [s | m | l] --runtime [trt | onnx].
  • Set the runtime flag to trt (controls the NMS format) if you plan to compile a TensorRT engine from it. We use the yolo-nas-m model.

EfficientViT encoder model

  • Download efficientvit-sam-l0.pt

  •   python applications/efficientvit_sam/deployment/onnx/export_encoder.py \
  --model efficientvit-sam-l0 \
  --output demo/gazesam/models/onnx/evit_encoder_l0.onnx

EfficientVIT decoder model

  •   python applications/efficientvit_gazesam/models/create_onnx/create_evit_decoder.py \
  --output demo/gazesam/models/onnx/evit_decoder_l0.onnx \
  --model-type efficientvit-sam-l0 \
  --opset 17 \
  --return-single-mask