trailnet-pytorch

This repo is created for re-producing Pytorch TrailNet of deep trail following which was published in ICRA2018.

I. Prerequisite

OS: Ubuntu 16.04/18.04
GPU with Nvidia driver >= 384.xx (supprt cuda9.0 or newer)
Docker
nvidia-docker2 (older version)

II. Test

$ cd trailnet-pytorch
$ source docker_run.sh [cuda9 | cuda10]     # Depend on your nvidia-driver version

# In the docker container
$ source jupyter_run_no_browser.sh

The jupyter token will be shown in the terminal. Please copy the text after the IP info, like '?8889?/token=aabbcc009966...' Open the browser, type your GPU machine's IP and paste the token info. the sample result is shown below.

Choose trailnet-pytorch-evaluation.ipynb and run, the notebook script will download test dataset and pre-trained model converted from caffemodel. Finally, evaluating the model by test dataset.

III. Re-train

Because the pre-trained model of TrailNet was converted from caffemodel, there are several implementation differences between Caffe and Pytorch, e.g. If input size cannot be divided by convolution size (consider kernel size, padding and stride), Implementation by Caffe will do zero-padding automatically but Pytorch do not. The testing accuracy from caffemodel is about 93.89%. Converting caffemodel to Pytorch model caused lower accurcy (86.94%) that tested in Pytorch framework, so we need to re-train pytorch model to improve the accurcy. If you are interested in Caffe implementation of TrailNet, please check out this repo.

Follow above instruction to run jupyter notebook in the Docker container, then choose trailnet-pytorch-train.ipynb and run, the notebook script will download the dataset and create training dataset and validation dataset from two existing txt files. It will also download the model converted from caffemodel for re-training. Finally, output the model (*.pth) to ./models folder.

If you want to create own datalist to split dataset to training and validation dataset, please take a look at create_custom_dataset.ipynb. This notebook will split dataset into two sets with specific size ratio, and export them to train_dataset.txt and validation_dataset.txt.

Far from this (50 epoch re-training), I have trained the model that test accurcy is 95%, I will continune improve this by revising training strategy, such as cross-validation.

IV. Pre-trained model

Trailnet 3 outputs model converted from caffemodel
trailnet_3class_real_vr_mix_color.h5

Trailnet 3 outputs model after one epoch re-training
trailnet_3class_epoch0_loss0.02158.pth

V. Issues

1. High test accuracy but lower validation accuracy after re-training 1 epoch

Because we use two different datasets to train and test model respectively, instead of splitting one dataset to conduct model training/testing. Maybe after one epoch re-training, this model is more suitable for testing dataset but the validation dataset. We also provided Trailnet pre-trained model for Pytorch framework, please check out **Pre-trained model** section.

wkyouncnu / trailnet-pytorch