/manhole-cover-retrainer

Primary LanguagePython

manhole-cover-retrainer

This repository trains the resnet152 model to predict manhole covers in 13 different classes:

  • Rost/Strassenrost
  • Vollguss/Pickelloch belueftet
  • Gussbeton/Pickelloch geschlossen
  • Vollguss/Pickelloch geschlossen
  • Gussbeton/Pickelloch belueftet
  • Vollguss/Handgriff geschlossen
  • Gussbeton/Handgriff seitlich
  • Rost/Einlauf rund
  • Rost/Strassenrost gewoelbt
  • Vollguss/Aufklappbar
  • Gussbeton/Handgriff mitte
  • Vollguss/Handgriff geschlossen, verschraubt
  • Andere/-

Installation

  1. Clone project locally
git clone git@github.com:FiratSaritas/manhole-cover-retrainer.git

  1. Take the images from the repository manhole-cover-labelling from the folder ./data/images_transformed and added it to the folder ./data/images_transformed in this repository.

  2. Take the labels.csv file that was created by repository manhole-cover-labelling in the folder ./data/ and rename it to "new_labels.csv" and add it to the folder ./data/ in that repository.

  3. Download old Images

Download images as Folder (images_transformed) from Google Drive and add it to the folder ./images_transformed here: (https://drive.google.com/drive/folders/1y5T1-WUZB1Vsp87aBiU6hDxagiY2mGgi?usp=sharing)

  1. Install required packages

create a new enviorment with conda.

conda create --name retrainer
activate retrainer

Install Cuda:

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

Usage

  1. run the code:
python retrainer.py

Output:

  • When you have trained the model, you get a .pth file model.pth
  1. You can now replace this model with the old model in the repository manhole-cover-model

Credits

This is a continuation of the repository manhole-cover-classification and the actual work was created on the other repository.