DeepWiener

Neural network to simulate the Wiener Filter (WF) for Cosmic Microwave Background (CMB) polarization maps with inhomogeneous noise applied. In addition, the power spectrum estimation algorithm is included

Description of the codes:

Brief overview of the codes inclued in source:

• CG_inho.py: Conjugate Gradient algorithm for Wiener Filter calculation.
• wf_noise.py: Module used to generate flat sky maps with Nifty7.
• make_dataset.py: Module that creates the dataset for each iteration.
• losses.py: Module with the loss functions used to train the neural network.
• train_functions.py: DeepWiener neural network and training.
• PowerSpectrum.py: Module to estimate the noise bias term and fisher matrix of the power spectrum.

Requirements:

• Tensorflow 2.X
• CAMB
• Nifty7

Usage:

The implemented codes here perform the WF reconstruction of CMB polarization maps with inhomogeneous noise applied. The inhomogeneous noise is presented by a variance map extracted from Planck. Two masks are presented with different fraction of sky.

The dictionary input include these parameters:

• modelb: put equal to 0 if only the loss function J(Q,U) is used, put equal to 1 if the loss function J(B) is used.
• iter: iteration number. Until 4 for the Mask1 case and until 5 for the Mask2 case. Put equal to 0 if modelb is equal to 1. If the iteration number is higher than 1 be aware of having the results of the previous iterations.
• name_data_folder: path where the dataset is stored.
• name_train, name_valid, name_test: filenames of the training, validation and test set that will be saved in the data folder.
• nsims_train, nsims_valid, nsims_test: number of maps for each set.
• mask: put equal to 0 if the Mask1 is used, put equal to 1 if the Mask2 is used.
• name_result_folder: path where the results are stored.
• factor: normalization factor for the dataset in each iteration.
• name_model: name of the models that will be saved in the results folder in (.h5 format).
• name_path: checkpoints name.
• filters: number of filters used for the convolution in each layer.
• namequ_file_bl, nameb_file_bl: name of the noise bias calculated with PowerSpectrum.py, for models trained with J(Q,U) and J(B), that will be saved in the results folder.
• namequ_fisher_mean, nameb_fisher_mean: name of the fisher matrix calculated with PowerSpectrum.py, for models trained with J(Q,U) and J(B), that will be saved in the results folder.
• nsamples_bias, nsamples_fisher: number of simulations used to estimate the noise bias term and the fisher matrix.
• nbins: number of bins.

To run the software:

1. Edit the input dictionary input_inho_mask.dict with the desired example.
2. Run run_dataset.py /path/input_inho_mask.dict to create the dataset.
3. Run train_functions.py /path/input_inho_mask.dict to train the neural network.
4. Run eval_model.py /path/input_inho_mask.dict to evaluate the trained models and obtain the WF predictions.
5. Run fisher_and_noise_bias.py /path/input_inho_mask.dict to calculate the noise bias and the fisher matrix for the power spectrum estimation.

Contact

Feel free to contact me at belen@fcaglp.unlp.edu.ar for comments, questions and suggestions.