Skip to content

DaltonLens/DaltonLens-Research

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DaltonLens Machine Learning

This repository collects ongoing experiments, and is by no mean production ready.

Charts enhancement by removing anti-aliasing

The main goal is to remove line anti-aliasing so colors become flat and can be easily segmented in DaltonLens.

Setup with conda

conda config --set auto_activate_base false
conda config --set channel_priority strict
conda install mamba

conda create --name dl python=3.7
conda config --add channels pytorch
mamba install pytorch torchvision torchaudio cudatoolkit matplotlib tensorboard jupyterlab ipykernel tqdm

# Register the conda env for jupyter notebooks
python -m ipykernel install --user --name=dl

conda activate dl

Then install the packages in edit mode:

pip install -e charts

It may require DaltonLens-Python to be installed.

Training a model

Edit and run charts/next_train.py.

There were experiments with fast.ai and pytorch-lightning (in the attic), but the latest only uses vanilla pytorch.

Logs can be visualized with tensorboard --logdir=charts/logs.

Evolution of the images across epochs can be visualized with e.g. zv charts/logs/xxx/**/*.png.

Training data

Training data is expected under charts/inputs/train.

  • charts/inputs/generate/: scripts to generate data with OpenCV drawings and matplotlib are under .

  • charts/inputs/arxiv/: scripts to collect data from arxiv figures. See charts/inputs/arxiv/README.md.

A backup of the last generated data is available on google drive.

Testing a model

TODO: these scripts needs to be adapted to the latest Gated/Masked version.

scripts/export_to_torchscript.py generates an onnx file from a checkpoint.

scripts/test_onnx.py loads pretrained models with various backends, like cv2.dnn and onnxruntime .

About

Machine learning research material, early experiments.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published