This is a neural network emulator of the code Radex for radiative transfer predictions (https://home.strw.leidenuniv.nl/~moldata/radex.html). The model is analogous to the Radex emulator in (https://github.com/drd13/emulchem), only slightly extended and more flexible to use.
Emuradex predicts flux and optical depth, based on four inputs: temperature [K], density of H$_2$ [cm$^{-3}$], line width [km/s] and molecular column density [cm$^{-2}$].
The model here is produced with tensorflow.keras (or tf.keras) library.
Please download the emuradex repository to use it.
In emuradex/code/ there is calc_lik_space.py for making predictions with Radex. The file requires the inputs:
input.csv
param_input.py
The first can be substituted for input1.csvorinput3.csvthat hold data for 1- and 3-phase observations. Theparams_input.pyhas to be substituted correspondingly forparams_input1.pyorparams_input3.py. All files are in the same emuradex/code` directory. Then, run, specifying the number of samples to make predictions for:
python calc_lik_space.py nsamples
To start making predictions with the model, you need to do:
import emuradex
specie = emuradex.Radex("CS", trans=1)
preds = specie.preduct_flux(features)
where features.shape is (nsamples, 4) and preds.shape is (nsamples, 2)
1. The model returned by tf.keras.models.load_model() is ready to use.
2. The predict() method (for tf.keras.Model object) is optimised for large scale input to work with data batches. Processing small amounts of data (point samples of size that fits into one batch; batch size with which the network was trained can be viewed by calling model...; by default predict() batch_size=32) is slow with predict() and should instead be done with the call() or __call__() methods for faster execution (https://www.tensorflow.org/api_docs/python/tf/keras/Model). For example:
model(X),model.call(tf.convert_to_tensor(X)),
as opposed to:
model.predict(X).
The model was trained with the notebook network_training.ipynb in Google Colab. The files for training are going to be provided...