Normalizing Flows for Light Curve Approximation

Collaborators:

• Sehmimul Hoque
• Vishnu Mugundan
• Maimouna Sow
• Siddhardha Penmetsa

Install Dependencies

• numpy
• matplotlib
• torch
• fulu
• pandas
• sklearn
• joblib
• torchvision

Code to install requirements:

pip3 install -r requirements.txt

Running main.py

To run main.py, use

python3 main.py

Get more information about the other arguments by running

python3 main.py --help

An example call of the program is

python main.py --param param/param.json -nf 10 -v -p -o -s

1. -nf num_objects runs Normalizing Flows to approximate num_objects light curves, and runs the CNN on the objects. In the above example, the code will produce 10 objects using Normalizing Flows, and run CNN on the 10 objects. num_objects must be at least 10. Be warned that even with multiprocessing implemented, the code takes a while to run.
2. -v increases the verbosity of the program.
3. -p produces plots for the light curves of the objects generated by Normalizing Flows.
4. -s shuffles the objects generated by Normalizing Flows.
5. -o option to be used when running -nf. Trains the CNN on the data generated from 600 objects and predicts on the objects generated by the normalizing flows run.

Remarks

It is suggested that you train the CNN using the data of 600 objects already obtained by us and then run NF for about 15-50 objects depending on the computing resources. NF takes significant amount of time to compile, and can be tracked in either verbose mode or in report.txt. The program trains and samples from the NF in parallel. Running -nf 24 takes around 20 minutes. It took 15 hours to get the output for 600 objects.

It is recommended to use the option -o while running -nf as the CNN predicts better when there are more objects. Furthermore, for every 10 objects generated by the NF, the CNN trains on 6, validates on 2, and only predicts on 2 objects.

Hyperparameters in param/param.json

NF

• num_samples: number of samples taken from the Multivariate Gaussian prior for inverse transforms
• lr: Learning rate of the Adam Optimizer
• n_epochs: Number of epochs for updating the parameters of the unknown model
• display_epochs: Number of epochs for printing the loss

CNN

• n_epochs: Number of epochs for updating the parameters of the unknown model
• display_epochs: Number of epochs for printing the loss
• weight_decay: regularization technique applied to the weights of CNN
• lr: Learning rate of the gradient descent

Data

Files

• main.py : main running script
• requirements.txt : requirements file
• src/normalizing_flows.py : normalizing flows module
• src/CNN.py : CNN module
• src/plot_utls.py : plot utilities for normalizing flows
• src/cnn_metrics.py : cnn metrics utilities
• src/nf_metrics.py : normalizing flows metrics utilities
• data/ANTARES_NEW.csv : dataset containing relevant data for flux interpolation and flux type classification of 1870 objects
• data/images.json : flux and flux error for 600 objects using the normalzing flows algorithm (CNN can be trained on this data)
• data/labels.json : expected labels of the 600 objects in images.json
• param/param,json : hyperparameters
• plots: contains plots of several flux interpolations using normalizing flows

Outputs

• nfmetrics.csv : contains metrics of flux approximation using normalizing flows
• report.txt : final report containing metrics for normalizing flows, training epochs of the CNN and metrics of CNN
• src/X_matrix.json : output of normalizing flows
• src/y_vector.json : expected labels of objects in src/X_matrix.json
• nf_run_plots: generated plots of several flux interpolations using normalizing flows