prepare_dataset.md

Prepare Dataset

There are 3 steps to get everything installed

• Creating the TensorRT version of HRNet
• Passing the datasets through the Semantic Map Generation Pipeline
• Generate the custom annotation files as in MapTRv2

a. Creating the TensorRT version of HRNet

Download the assets (weights and config) here link, put it in a folder, this will be your ASSETS_PATH.

We use the docker image naruarjun/tensorrt-nvidia-semseg

# Pull docker image
docker pull naruarjun/tensorrt-nvidia-semseg

# Create a docker container
docker run --name hrnet-1 -t -d --gpus all --shm-size=16gb --net=host --volume="/media:/media:rw" --volume="/home:/home:rw" naruarjun/tensorrt-nvidia-semseg

# Enter the container
docker exec -it hrnet-1 /bin/bash

# Clone forked and modified Nvidia Semantic Segmentation Monorepo
git clone https://github.com/naruarjun/TensorRT-tutorial.git
cd TensorRT-tutorial

# Change ASSETS_PATH(Where the segmentation weights and config are stored) in config.py Line 52

# Convert to ONNX
# dataset-name mapillary
# checkpoint name mapillary_ocrnet.HRNet_Mscale_fast-rattlesnake.pth
python convert-onnx.py --dataset {dataset-name} --cv 0 --bs_val 1 --eval folder --eval_folder ./imgs/test_imgs --dump_assets --dump_all_images --n_scales 0.5,1.0 --snapshot ASSETS_PATH/seg_weights/path-to-pth.pth --arch ocrnet.HRNet_Mscale --result_dir logs/onnx-export/log-name

# Create TensorRT engine to be used
trtexec --onnx=hrnet.onnx --fp16 --workspace=64 --buildOnly --saveEngine=hrnet.engine

b. Passing the datasets through the Semantic Map Generation Pipeline

We use the docker image naruarjun/tensorrt-nvidia-semseg

# For AV2
# Change paths in semantic_mapping/src/node_config/
cd SemVecNet
python python tools/av2_semantic_mapping.py

c. Generate the custom annotation files as in MapTRv2

We use the docker image naruarjun/semvecnet:v2

This is exactly as it is in MapTRv2