PyLightcurve-torch

An exoplanet transit modelling package for deep learning applications in Pytorch.

See this open publication in the Publications of the Astronomical Society of the Pacific for more details and official citation and jupyter notebook tutorials here.

The code for orbit and flux drop computation is adapted from Pylightcurve.

The module pylightcurve_torch.functional.py contains the functions implemented in Pytorch and computing the orbital positions, transit durations and flux drops. (see PyLightcurve repository for more information about the numerical models used).

A TransitModule class is implemented in pylightcurve_torch.nn.py with the following features:

• Computes time series of planetary positions and primary/secondary flux drops
• it inherits torch.nn.Module class to benefit from its parameters optimisation and management capabilities and facilitated combination with neural networks
• native GPU compatibility

Installation

$ pip install pylightcurve-torch

Basic use

from pylightcurve_torch import TransitModule

tm = TransitModule(time, **transit_params)

flux_drop = tm()

If needs be, the returned torch.Tensor can be converted to a numpy.ndarrray using flux_drop.numpy() torch method or flux.cpu().numpy() if the computation took place on a gpu.

Transit parameters

Below is a summary table of the planetary orbital and transit parameters use in PyLightcurve-torch:

A short version of each parameter has been introduced, while maintaining a compatibility with origin PyLightcurve parameter names. All the parameters except method are converted to torch.Parameters when passed to a ``TransitModule```, with double dtype.

Differentiation

One of the main benefits of having a pytorch implementation for modelling transits is offered by its automatic differentiation feature with torch.autograd, stemming from autograd library.

Here is an example of basic usage:

...
tm.fit_param('rp')                  # activates the gradient computation for parameter 'rp'
err = loss(flux, **data)            # loss computation in pytorch 
err.backward()                      # gradients computation 
tm.rp.grad                          # access to computed gradient for parameter 'rp'

More Pytorch support

Several utility methods inherited from PyTorch modules are listed below, simplifying operations on all module's defined tensor parameters.

tm = TransitModule()

# Parameters access (iterators)
tm.parameters()
tm.named_parameters()

# dtype conversions
tm.float()
tm.double()

# Gradient local deactivation
with torch.no_grad():
    flux_no_grad = tm()

# device conversion
tm.cpu()
tm.cuda()

Running performance tests

In addition to traditional unit tests, computation performance tests can be executed this way:

 python tests/performance.py --plot

This will measure the computation time for computing forward transits as a function of transit duration, time vector length or batch size. If data have been savec previously, these will be plotted to with the name of the corresponding tag.