The workflow is set up to run the fitting on many sources efficiently by splitting the full catalog into a number of smaller files. This allows distributing the fitting across cores. There are manual steps to allow for the refitting, fixing issues, etc without rerunning everything. This workflow has been tested on large (e.g., PHAT) and small (e.g. METAL) datasets.
For visualizations of the code and files, see BEAST Graphical Models.
Setup a working location. For reference, there are examples in
In this location, you will need a
beast_settings.txt parameter file.
These parameters are described in the BEAST setup documentation.
The BEAST also has some tools for converting catalogs between file formats, see other tools.
Source Density or Background Level¶
The data generally need to have source density information added as it is common for the observation model (scatter and bias) to be strongly dependent on source density due to crowding/confusion noise. The background may also be important in some situations, so there is code to calculate it as well.
Adding source density to observations¶
Create a new version of the observations that includes a column with the source density. The user chooses one band to use as the reference, and chooses the magnitude range of sources to use for calculating the source density (generally, this would be the range over which the catalog is complete). The user can also choose a band for which sources that have ‘[band]_FLAG == 99’ are ignored.
Three files are created. The prefix is derived from the name of the input photometry catalog.
[prefix]_source_den_image.fits: an image of the source density map
[prefix]_with_sourceden.fits: photometry catalog with an additional column that has the source density
[prefix]_sourceden_map.hd5: contains information about the map grid
Command to create the observed catalog with source density column with a pixel scale of 5 arcsec using the ‘phot_catalog.fits’ catalog.
$ python -m beast.tools.create_background_density_map sourceden \ -catfile phot_catalog.fits --pixsize 5.
Adding background to observations¶
Create a new version of the observations that includes a column with the background level. This is done by calculating the median background for stars that fall in each spatial bin. The code will output a new catalog, an hdf5 file with the background maps and grid information, and some diagnostic plots.
Command to create the observed catalog with background column with a 15x15 pixel array using the ‘phot_catalog.fits’ catalog and the ‘image.fits’ reference image.
$ python -m beast.tools.create_background_density_map background \ -catfile phot_catalog.fits --npix 15 -reference image.fits
To check if the background (or source density) map makes sense, the ‘tileplot’ subcommand of the same script can be used. If the output of one of the previous commands was ‘map_name.hd5’, then use
$ python -m beast.tools.create_background_density_map tileplot map_name.hd5 \ -image image.fits --colorbar 'background'
Generate the full physics model grid. This is needed for both the fitting and for generating the artificial star test (AST) inputs. Note that you may want to use a coarser model grid for the ASTs.
If you’re creating a model grid that’s so large it may not read into memory, you can use subgrids, which splits the grid into more manageable pieces.
To create a physics model grid with 5 subgrids:
$ python -m beast.tools.run.create_physicsmodel beast_settings.txt --nsubs=5
If you’re running the BEAST on a survey in which different fields have different
filters, you may wish to save time by creating a master grid with all possible
filters and just copying out the subset of filters you need for each field. To
do this, create a
beast_settings.txt file with all relevant filters listed in
basefilters, and run
create_physicsmodel as above. Then use
remove_filters to create each modified grid. The list of filters to remove
will be determined by what’s present in the input catalog file. If you’re using
subgrids, repeat the command for each subgrid.
$ python -m beast.tools.remove_filters.py catfile.fits \ --physgrid master_physgrid.hd5 --physgrid_outfile new_physgrid.hd5
If you would like to examine some or all of the grid values in a physics model,
you can use the
read_sed_data function in
function can also be set to just extract the list of parameter names.
Artificial Star Tests¶
The observation model is based on artificial star tests (ASTs). More details about the BEAST AST code components can be found at Artificial Star Input Lists.
The BEAST selects SEDs from the physics model grid with a technique that minimizes the number of ASTs needed to allow the construction of a good toothpick observation model. For each band, the range of fluxes in the model grid is split into bins (default=40, set by ast_n_flux_bins in beast_settings), and models are randomly selected. The model is retained if there are fewer than the set number of models (default=50, set by ast_n_per_flux_bin in beast_settings) in each of the relevant flux bins.
$ python -m beast.tools.run.make_ast_inputs beast_settings.txt
While not recommended, it is possible to randomly select SEDs from the physics model grid.
$ python -m beast.tools.run.make_ast_inputs beast_settings.txt --random_seds
In case the user needs to supplement the existing input ASTs, it is possible to select additional model SEDs.
$ python -m beast.tools.run.make_ast_inputs beast_settings.txt --suppl_seds
How the sources are placed in the image is determined by the ast_source_density_table
ast_source_density_table is set to
filebase_sourceden_map.hd5: The source density or background image is split into bins, and for each bin, all selected SEDs are randomly replicated within pixels of that bin. These bins are determined by the beast_settings parameters, and can have linear (default) or log spacing, where the user can determine the number or width of the bins (set using sd_binmode, sd_binwidth and sd_Nbins in beast_settings). Alternatively, the user can input a custom list of bin edges, which will override the other binning settings. This same binning scheme is used later to split the catalogs (next step).
ast_source_density_table = None: Randomly choose a star from the photometry catalog, and place the artificial star nearby. Repeat until all SEDs have been placed.
These ASTs should be processed with the same code that was used to extract the source photometry.
You may wish to remove artifacts from the photometry catalog. If you do so, the same criteria must be applied to the AST catalog.
The code to edit catalogs can do three different things:
Remove objects without full imaging coverage. Note that the overlap is determined by eliminating sources with a flux of precisely 0 in any band. However, any sources with a flux of 0 in all bands are not removed, since that would indicate that an artificial star was not recovered (this criterion does not affect standard photometry catalogs, which do not have any sources with flux=0 in all bands).
Remove flagged sources. This eliminates any source with
[filter]_FLAG=99in the specified filter. If that source has flux<0, it is not removed, because those sources are set by
dolphotto have flag=99 regardless of quality.
Create ds9 region files. If set, it will create a ds9 region file where good sources are green and removed sources are magenta.
Command to edit the files, both to remove flagged sources and eliminate sources that don’t have full imaging coverage, and to create ds9 region files:
$ python -m beast.tools.cut_catalogs \ phot_catalog_with_sourceden.fits phot_catalog_cut.fits \ --input_ast_file ast_catalog.fits \ --output_ast_file ast_catalog_cut.fits \ --partial_overlap --region_file --flagged --flag_filter F475W
The observed catalog should be split into separate files for each source density bin. These bins are determined by the beast_settings parameters, and can have linear (default) or log spacing, where the user can determine the number or width of the bins (set using sd_binmode, sd_binwidth and sd_Nbins in beast_settings). Alternatively, the user can input a custom list of bin edges, which will override the other binning settings. In addition, each source density catalog can be split into a set of sub-files to have at most ‘n_per_file’ sources (or, if there are very few stars in a source density bin, at least ‘min_n_subfile’ sub-files). The sources are sorted by the ‘sort_col’ flux before splitting to put sources with similar brightness together. This splitting into sub files sorted by flux allows for trimming the BEAST physics+observation model, removing objects that are too bright or too faint to fit any of the sources in the file; more sub-files mean a narrower range of flux in each one, so more is trimmed and fitting is faster. In addition, this allows for running the BEAST fitting in parallel with each sub-file on a different core.
Command to split both the catalog and AST files by source density:
$ python -m beast.tools.split_catalog_using_map beast_settings.txt \ phot_catalog_cut.fits ast_catalog_cut.fits phot_catalog_sourceden_map.hd5 \ --n_per_file 6250 --min_n_subfile 3 --sort_col F475W_RATE
The observation model is generally based on artificial star tests (ASTs). ASTs are artificial sources inserted into the observations and extracted with the same software that was used for the observed photometry catalog. This ensures that the observation model has the same selection function as the data.
There are 3 different flavors of observation models.
‘Splinter’: A very simple (and likely not very good) model that assumes the noise is a fraction of the model SED flux and there is no bias. No ASTs are used.
‘Toothpick’: The AST results are assumed to be independent between different bands (even if they are not). The AST results are binned in log(flux) bins and the average bias and standard deviation is tabulated and used to compute the bias and noise for each model in the physics grid.
‘Truncheon’: The covariance between bands is measured using the AST results. The input AST SEDs are assumed to have been chosen from the BEAST physics model grid and are expected to sparsely sample the full model grid. The ASTs should be run simultaneously with all bands and it assumed that there are multiple ASTs run for the same model. The covariance between the bands is approximated with a multi-variate Gaussian. The bias and a multi-variate Gaussian is computed for each model in the physics grid by interpolating between the sparse grid computed from the AST results.
The code to compute the observation can be done with or without subgridding, and with or without source density splitting. Here are some examples:
$ # with source density splitting and no subgridding $ python -m beast.tools.run.create_obsmodel beast_settings.txt --use_sd --nsubs 1 $ # with source density splitting and 5 subgrids $ python -m beast.tools.run.create_obsmodel beast_settings.txt --use_sd --nsubs 5 $ # no source density splitting or subgrids $ python -m beast.tools.run.create_obsmodel beast_settings.txt --nsubs 1
If you would like to examine some of all of the values in the observation model,
you can use the
read_noise_data function in
Trimming for speed¶
The physics+observation model can be trimmed of sources that are so bright or so faint (compared to min/max flux in the observation file) that they will by definition produce effectively zero likelihood fits. Such trimming will speed up the fitting.
The source density split sub files are organized such that the range of fluxes is minimized in each sub file. This allows for trimming and faster fitting.
The trimming can take significant time to run. In addition, reading in the full physics+observation model can be slow and such reading can be minimized by producing multiple trimmed models with a single read. A specific tool is provided to setup batch files for this trimming and to do the actual trimming.
This code sets up batch files for submission to the ‘at’ queue on linux or similar systems (such as slurm). The projectname (e.g., ‘PHAT’) provides a portion of the batch file names. The datafile and astfile are the observed photometry file (not sub files) and file with the ASTs in them. The optional input seds_fname can be used to specify the file with the physics model grid, which overrides the default filename when you wish to use one model grid for multiple fields. A subdirectory in the project directory is created with a joblist file for submission to the batch queue and smaller files used by the trimming code.
The joblist file can be split into smaller files if submission to multiple cores is desired. Use the ‘num_subtrim’ commandline tool. The optional ‘nice’ input allows you to prepend a ‘nice’ option, especially useful if you’re utilizing shared computing resources.
$ python -m beast.tools.setup_batch_beast_trim projectname phot_catalog_cut.fits \ ast_catalog_cut.fits --num_subtrim 5 --nice 19
If you’re doing a BEAST run that utilizes both subgrids and background/source
density splitting, a handy wrapper will generate each combination of file names
setup_batch_beast_trim for you:
$ python -m beast.tools.run.make_trim_scripts beast_settings.txt \ --num_subtrim 5 --nice 19
Once the batch files are created, then the joblist can be submitted to the queue. The beast/tools/trim_many_via_obsdata.py code is called and trimmed versions of the physics and observation models are created in the project directory.
$ at -f project/trim_batch_jobs/XX_joblist now
The fitting is done for each sub file separately. Code in the tools directory can be used to create the needed set of batch files for submission to a queue. In addition, this code will check and see if the fitting has already been done or was interrupted for the sub files. Only sub files that have not been fit or where the fitting was interrupted will be added to the batch files. The number of sub files to be run on each core is a command line argument (the runs will are serial on the core).
$ python -m beast.tools.setup_batch_beast_fit.py --num_percore 2 --nice 19 \ --use_sd 1 --nsubs 5 --pdf2d_param_list Av M_ini logT
The jobs can be submitted to the batch queue via:
$ at -f projectname/fit_batch_jobs/beast_batch_fit_X.joblist now
The fitting yields several output files (which are described in detail here):
*_stats.fits: Statistics for each of the fitted and derived parameters, including the 16th/50th/84th percentiles, mean, and expectation value
*_pdf1d.fits: Marginalized 1D PDFs for each of the fitted and derived parameters
*_pdf2d.fits: Marginalized 2D PDFs for pairs of parameters. If
pdf2d_param_listis set to
None, 2D PDFs will not be generated. The default set is the 7 main BEAST parameters, but any parameters in the grid can be chosen.
*_lnp.hd5: Sparsely sampled log likelihoods
The contents of the
lnp file can be easily accessed with the
tools/read_beast_data.py, which converts the hdf5 file structure
into a dictionary. If you need the SED grid values associated with the saved
lnP points, use the
get_lnp_grid_vals function in the same file.
Create the merged stats file¶
The stats files (catalog of fit parameters) can then be merged into a single file for the field. The 1D PDF and lnP files are merged across subgrids, but not yet across source density or background bins. Merging 2D PDFs has not yet been implemented.
$ python -m beast.tools.run.merge_files beast_settings.txt --use_sd 1
If fitting using subgrids, it’s possible to do a partial merge. This is useful if you’d like to look at fitting results before all stars have finished fitting for all subgrids. This find stars that have fits in all subgrids and merges just those. Partial merge is only currently implemented for stats and 1D PDFs.
$ python -m beast.tools.run.merge_files beast_settings.txt --nsubs 5 --partial
Reorganize the results into spatial region files¶
The output files from the BEAST with this workflow are organized by source density and brightness. This is not ideal for finding sources of interest or performing ensemble processing. A more useful organization is by spatial region. The large amount of BEAST output information makes it best to have individual files for each spatial region. Code to do this spatial reordering is provided in two parts. The 1st spatially reorders the results for each source density/brightness BEAST run into files for each spatial region. The 2nd condenses the multiple individual files for each spatial region into the minimal set (stats, pdf1d, and lnp).
Divide each source density/brightness file into files of spatial regions with 10”x10” pixels.
$ python -m beast.tools.reorder_beast_results_spatial --stats_filename filebase_stats.fits --region_filebase filebase_ --output_filebase spatial/filebase --reg_size 10.0
Condense the multiple files for each spatial region into the minimal set. Each spatial region will have files containing the stats, pdf1d, and lnp results for the stars in that region.
$ python -m beast.tools.condense_beast_results_spatial --filedir spatial
You may wish to use these files as inputs for the MegaBEAST.
This is a wrapper for each of the commands described above:
You may choose to run each of the above commands individually, but this conveniently packages them into one file. If you use this wrapper, you should edit several items in the file:
field_names: used to identify photometry files and create BEAST files
gst_filter_names: labels for the filters used in your photometry file (e.g., ‘X_RATE’)
beast_filter_names: the corresponding long names used by the BEAST
settings for the source density map: pixel size, filter, magnitude range
settings for the background map: pixel dimensions, reference image
settings for splitting the catalog by source density: filter, number of sources per file
settings for the trimming/fitting batch scripts: number of files, nice level
You can (and should!) read about the individual functions above before running beast_production_wrapper:
$ python beast_production_wrapper
The first thing it does is use beast_settings_template.txt to create a field-specific beast settings file. You will need to modify the beast_settings_template.py file to specify the required parameters for generating models and fitting data. The settings will be utilized as needed in the functions called by the wrapper. Four of the settings fields (project, obsfile, filters, and basefilters) will be filled in by beast_production_wrapper.py, so ensure that the other fields in beast_settings_template.py have the desired values.
The wrapper will proceed through each of the functions above. At three points, you will need to manually run things independently of the wrapper. It will not continue running subsequent functions until it finds that the necessary steps have been taken.
Creating ASTs (if a fake star catalog doesn’t exist)
running the batch trimming scripts
running the batch fitting scripts
Once you have completed each of these, run the wrapper again. It will skip past the steps that it has already processed, and resume at the point where you left off. In the case of the batch scripts, if you only partially completed them, it will re-generate new scripts for the remaining trimming/fitting (and tell you which ones are new), and pause again.
Many of the steps described above require considerable computational resources,
especially if your grid is large. If you’re running on XSEDE
or another system that uses the slurm queue, you may wish to use
write_sbatch_file.py. This will create a job file that can be submitted with
More information about how this file is constructed can be found in the TACC user guide
Here is an example call to
write_sbatch_file.py that shows some of its
$ # create submission script $ python -m beast.tools.write_sbatch_file \ 'sbatch_file.script' './path/to/job/beast_batch_fit_X.joblist' \ '/path/to/files/projectname/' \ --modules 'module load anaconda3' 'source activate beast_v1.4' \ --queue LM --run_time 2:30:00 --mem 250GB
This creates a file
sbatch_file.script with these contents:
#!/bin/bash #SBATCH -J beast # Job name #SBATCH -p LM # Queue name #SBATCH -t 2:30:00 # Run time (hh:mm:ss) #SBATCH --mem 250GB # Requested memory # move to appropriate directory cd /path/to/files/projectname/ # Load any necessary modules # Loading modules in the script ensures a consistent environment. module load anaconda3 source activate beast_v1.4 # Launch a job ./path/to/job/beast_batch_fit_X.joblist
Then the file can be submitted:
$ sbatch sbatch_file.script