MIU²Net

Weak Lensing Mass Inversion

MIU²Net stands for Mass Inversion U²Net. It uses deep learning to convert weak lensing shear ($\gamma$) maps to convergence ($\kappa$) maps, which trace the projected dark matter distribution.

We develop MIU²Net as a deep learning framework for weak lensing mass inversion. MIU²Net includes observations effects like shape noise, reduced shear, and data masks in the training.

Installation

The main MIU²Net package depends on the following packages:

• pytorch
• numpy
• scipy
• astropy

Prior to installing MIU²Net, make sure to install PyTorch according to your OS and compute platform. We recommend installing pytorch 1.12.0 and torchvision 0.13.0 to avoid unexpected dependency errors. The main MIU²Net package includes the full training and testing code for deep learning.

To reconstruct convergence maps using traditional (non- deep learning) methods, we have modified the cosmostat package developed at the CosmoStat Lab in CEA Paris-Saclay, so that we can use traditional $\kappa$ map reconstructions alongside deep learning for comparison, or even use traditional methods during network training. Avaiilable reconstruction methods from cosmostat include:

• Kaiser-Squires (KS) deconvolution
• Wiener Filtering (WF)
• sparse reconstruction (Lanusse et al. 2016, Glimpse)
• MCALens (Starck et al.)

To use these methods within MIU²Net, you should install Sparse2D developed by the CosmoStat Lab. This is not required by the deep learning framework.

Training MIU²Net

To train a MIU²Net model with noise corresponding to 20 galaxies per square arcmin, reduced shear, and 0-20% randomized masked pixels:

cd ./miu2net/main
python -u train.py --gpu-ids 6,7 --cpu 32 -b 128 -g 20 -e 2000 --mixed-precision --load pretrain_k2d_e472_c5r2_huber_reduced_m20_g20 --reduced-shear --mask-frac 0.2 --rand-mask-frac --freq-loss freq1d --beta 3.0 | tee k2d_trainlog.txt

Testing MIU²Net

rmg20 (noise galaxy = 20, mask frac = 20%, reduced shear):

python pred.py k2d_e1893_c5r2_huber_freq1d_reduced_mrand20_g20 -g 20 --num 500 --cpu 16 --reduced-shear --mask-frac 0.2
python make_master_cubes_multiproc.py -g 20 --cpu 32

rg20 (noise galaxy = 20, no masks, reduced shear):

python pred.py k2d_e1893_c5r2_huber_freq1d_reduced_mrand20_g20 -g 20 --num 500 --cpu 16 --reduced-shear
python make_master_cubes_multiproc.py -g 20 --cpu 32

cosmology 2, rmg20

python pred.py k2d_e1893_c5r2_huber_freq1d_reduced_mrand20_g20 -g 20 --num 500 --cpu 16 --reduced-shear --mask-frac 0.2 --dir /share/lirui/Wenhan/WL/cosmology2 --cosmo2
python make_master_cubes_multiproc.py -g 20 --cpu 32 --cosmo2

cosmology 2, rg20

python pred.py k2d_e1893_c5r2_huber_freq1d_reduced_mrand20_g20 -g 20 --num 500 --cpu 16 --reduced-shear --dir /share/lirui/Wenhan/WL/cosmology2 --cosmo2
python make_master_cubes_multiproc.py -g 20 --cpu 32 --cosmo2

testing for power spectrum, rg20

python pred_pspec.py k2d_e1893_c5r2_huber_freq1d_reduced_mrand20_g20 -g 20 --num-avg 500 --num-it 30 --cpu 32 --reduced-shear --noise-seed 0

Training DeepMass

python -u train_deepmass.py --gpu-ids 4 --cpu 32 -b 16 -g 20 --reduced-shear --mask-frac 0.2 --rand-mask-frac --wiener only | tee deepmass_trainlog.txt

Testing DeepMass

rmg20

python pred_deepmass.py deepmass_e349_c5r2_reduced_mrand20_g20 -g 20 --num 500 --cpu 16 --mask-frac 0.2 --wiener only

rg20

python pred_deepmass.py deepmass_e349_c5r2_reduced_mrand20_g20 -g 20 --num 500 --cpu 16 --wiener only

cosmology 2, rmg20

python pred_deepmass.py deepmass_e349_c5r2_reduced_mrand20_g20 -g 20 --num 500 --cpu 16 --dir /share/lirui/Wenhan/WL/cosmology2 --mask-frac 0.2 --wiener only --cosmo2

cosmology 2, rg20

python pred_deepmass.py deepmass_e349_c5r2_reduced_mrand20_g20 -g 20 --num 500 --cpu 16 --dir /share/lirui/Wenhan/WL/cosmology2 --wiener only --cosmo2