This is my fork of LSS.

clone:

git clone [email protected]:Eslzzyl/lift-splat-shoot.git -b main

Install:

1. Follow the instructions in https://pytorch.org/ to install newest pytorch and torchvision.

2. Install nuscenes-devkit:

   pip install nuscenes-devkit

   In order to use nuscenes-devkit with Python 3.12, you can use my fork: https://github.com/Eslzzyl/nuscenes-devkit

3. Install other dependencies:

   pip install -r requirements.txt

Prepare Data:

You should download the nuScenes dataset from its official website and unpack it to a place with enough disk space (no less than 500 GB for trainval subset). Theoretically, mini subset and trainval subset are both supported. I only tested this code for trainval subset.

Train:

It is recommended to copy the train.sh to a new my_train.sh script, and modify my_train.sh according to your hardware conditions. Current setting in train.sh should be able to train on a 8 * RTX 2080Ti server equipped with no less than 100GB CPU memory. You should also modify the nuScenes dataset root path and TensorBoard log path according to your need.

Then start the training:

bash my_train.sh

A pretrained EfficientNet checkpoint will be automatically downloaded from pytorch hub. This operation will be performed only once.

Use TensorBoard:

tensorboard --logdir=./runs

Visualize:

bash visual.sh

Evaluate:

bash eval.sh