smocilac / swiftnet_inference

BUILD TENSORRT TEST:

mkdir build
cd build
cmake ..
make

RUN TENSORRT:

./tensorrt_test -h

RUN ACCURACY TEST:

From https://www.cityscapes-dataset.com/downloads/ download:

• leftImg8bit_trainvaltest.zip (11GB)
• gtFine_trainvaltest.zip (241MB)

Download and extract to accuracy-test/datasets/Cityscapes

cd accuracy-test
python trt_inference.py

CONVERT SWIFTNET MODEL TO ONNX:

cd model
nano swiftnet_to_onnx.py 
# change pre-trained model path and input resolution (must be the same as in models/util.py)
python swiftnet_to_onnx.py

RUN TORCH EVALUATION:

cd model
nano torch_timer_eval.py
# change pre-trained model path and input resolution (must be the same as in models/util.py)
python torch_timer_eval.py

DOWNLOAD WEIGHTS:

wget http://elbereth.zemris.fer.hr/swiftnet/swiftnet_ss_cs.pt -P weights/

SWIFTNET MODEL PUBLIC REPOSITORY:

https://github.com/orsic/swiftnet/