Introduction

Anomalib inference with TensorRT (python).

Requirements

albumentations==1.3.1

anomalib==1.0.0.dev0

anomalib.egg==info

numpy==1.23.1

omegaconf==2.3.0

opencv_python==4.8.1.78

opencv_python_headless==4.8.1.78

pycuda==2023.1

tensorrt==8.5.1.7

Usage

step 1

Train and export anomalib models to onnx.

step 2

Convert onnx to trt engine.

Example:

trtexec --onnx=efficient_ad.onnx --saveEngine=efficient_ad.engine --minShapes=input:1x3x256x256 --optShapes=input:4x3x256x256 --maxShapes=input:8x3x256x256

step 3

Do inference.

Example:

python inference.py --batchsize 4 --weights weights/efficient_ad.engine --metadata data/metadata_transistor_efficient_ad.json --input D:/surface_defect_datasets/mvtec_anomaly_detection/transistor/test --output result --visualize 1 --task segmentation --visualization_mode full --show 0

python inference.py --batchsize 1 --weights weights/dfkde.engine --metadata data/metadata_transistor_dfkde.json --input D:/surface_defect_datasets/mvtec_anomaly_detection/transistor/test --output result --visualize 1 --task classification --visualization_mode full --show 0

Note: some bugs occur when doing inference with dfkde model (batchsize > 1).

References

https://github.com/openvinotoolkit/anomalib

https://github.com/NagatoYuki0943/anomalib-tensorrt-cpp

Todo

int8 calibration