DECODE FISH is a deep learning based algorithm to localize fluorescent spots in RNA-FISH (smFISH) data. It allows for fast and precise localization even under difficult conditions of low SNR and high density.
The method is an extension of the DECODE algorithm to 3D data (https://www.biorxiv.org/content/10.1101/2020.10.26.355164v1).
We recommend to use a conda virtual environment. DECODE FISH requires a GPU with CUDA capability of 3.7 or higher and at least 8 GB RAM.
git clone https://github.com/TuragaLab/decode_fish.git
cd decode_fish
conda env create -f requirements.yaml
conda activate decode_fish_dev
python setup.py developAnalyzing a data set with DECODE FISH has two steps, training the network and running the prediction. These can be executed these scripts:
python decode_fish/train.py +experiment=your_experiment
python decode_fish/predict.py out_file='results.csv' model_path='your_experiment/model.pkl' image_path='your_recordings/*tif'
Training is parametrized by a .yaml config file. Default parameters are stored in the config/train.yaml. Before starting set the 'base_dir' to the path where you cloned this repository. Also create a directory where you wan't to save your trained models and change the 'out_put.save_dir'
Please refer to the experiment.ipynb notebook to see the complete workflow of setting dataset specific training parameters, running the training, performing predictions and inspecting the results on an example dataset. We recommed copying and editing this notebook for your own experiments.
We use Weights & Biases to track training progress. Many performance metrics and intermediary results are continuously logged and uploaded. However, you'll have to create a (free) account. To disable wandb logging set 'output.wandb_mode' = disabled