./main.py -h
You have to provide to the program : - either the FWHM of the PSF or the wavelength and the numerical aperture - the sensor's pixels size in nanometers
The corresponding options are respectively --FWHM or --NA --wavelength
Try to apply the software on a small crop of your sample and optimize manually the regularization parameters -g_x_l1 and -g_b_grad. There are related to the proportion of signal compared to the backgroud and the smoothness of the background.
To find the best regularization it is better not to use the discriminator, which means -g_D 0 and -n_d 0. First it will be faster and in any case you have to achieve first convergence with L2 distance before using the discriminator distance to improve results.
./main.py images.tif --FWHM 318 -g_x_l1 0.8 -g_D 0 -n_d 0 -n_e 1000 -x 0 0 50 50
Once you found the best parameters -g_x_l1 and g_b_grad for regularization then you can start the program with discriminator on the full sample with more epochs.
./main.py images.tif --FWHM 318 -g_x_l1 0.8 -g_D 1 -n_d 1 -n_e 5000
You can save the results and different curves using the -o option and -l -m -v or -r
Two different sets of parameters are provided in the tests/ folder. Their only difference is the use of discriminator or not. You can run them using the following commands :
./main.py tests/with_discriminator.json
./main.py tests/without_discriminator.json
The results are automatically saved in the corresponding folder.
Requires python 3 and pytorch >= 1.10
Install required python modules (Linux) :
sudo pip3 install torch torchvision matplotlib tifffile read_lif tqdm czifile
See https://pytorch.org to install PyTorch on different platforms
fft_conv from https://github.com/fkodom/fft-conv-pytorch