NWChem/doc/user/python.tex

%
% $Id$
%
\label{sec:python}

Python (version 1.5.1) programs may be embedded into the NWChem input
and used to control the execution of NWChem.  Python is a very
powerful and widely used scripting language that provides useful
things such as variables, conditional branches and loops, and is also
readily extended.  Example applications include scanning potential
energy surfaces, computing properties in a variety of basis sets,
optimizing the energy w.r.t. parameters in the basis set, computing
polarizabilities with finite field, and simple molecular dynamics.

Look in the NWChem \verb+contrib+ directory for useful scripts and
examples. Visit the Python web-site 
\htmladdnormallink{http://www.python.org}{http://www.python.org}
for a full manual and lots of useful code and resources.  

\section{How to input and run a Python program inside NWChem}

A Python program is input into NWChem inside a Python compound directive.
\begin{verbatim}
  python [print|noprint]
    ...
  end
\end{verbatim}
The \verb+END+ directive must be flush against the left
margin (see the Troubleshooting section for the reason why).

The program is by default printed to standard output when read, but
this may be disabled with the \verb+noprint+ keyword.  Python uses
indentation to indicate scope (and the initial level of indentation
must be zero), whereas NWChem uses optional indentation only to make
the input more readable.  For example, in Python, the contents of a
loop, or conditionally-executed block of code must be indented further
than the surrounding code.  Also, Python attaches special meaning to
several symbols also used by NWChem.  For these reasons, the input
inside a \verb+PYTHON+ compound directive is read verbatim except that
if the first line of the Python program is indented, the same amount
of indentation is removed from all subsequent lines.  This is so that
a program may be indented inside the \verb+PYTHON+ input block for
improved readability of the NWChem input, while satisfying the
constraint that when given to Python the first line has zero
indentation.

E.g., the following two sets of input specify the same Python program.
\begin{verbatim}
  python
    print 'Hello'
    print 'Goodbye'
  end

  python
  print 'Hello'
  print 'Goodbye'
  end
\end{verbatim}
whereas this program is in error since the indentation of the second
line is less than that of the first.
\begin{verbatim}
  python
    print 'Hello'
  print 'Goodbye'
  end
\end{verbatim}

The Python program is not executed until the following directive
is encountered
\begin{verbatim}
  task python
\end{verbatim}
which is to maintain consistency with the behavior of NWChem in general.
{\em The program is executed by all nodes.}  This enables the full functionality and speed of NWChem to be accessible from Python, but there are some gotchas
\begin{itemize}
\item Print statements and other output will be executed by all nodes
so you will get a lot more output than probably desired unless the
output is restricted to just one node (by convention node zero).
\item The calls to NWChem functions are all collective (i.e., all
nodes must execute them).  If these calls are not made collectively
your program may deadlock (i.e., cease to make progress).

\item When writing to the database (\verb+rtdb_put()+) it is the data
from node zero that is written.

\item NWChem overrides certain default signal handlers so care
must be taken when creating processes (see Section \ref{sec:sigchld}).
\end{itemize}

\section{NWChem extensions}

Since we have little experience using Python, the NWChem-Python
interface might change in a non-backwardly compatible fashion as we
discover better ways of providing useful functionality.  We would
appreciate suggestions about useful things that can be added to the
NWChem-Python interface.  In principle, nearly any Fortran or C
routine within NWChem can be extended to Python, but we are also
interested in ideas that will enable users to build completely new
things.  For instance, how about being able to define your own energy
functions that can be used with the existing optimizers or dynamics
package?

Python has been extended with a module named \verb+"nwchem"+ which is
automatically imported and contains the following NWChem-specific
commands.  They all handle NWChem-related errors by raising the
exception \verb+"NWChemError"+, which may be handled in the standard
Python manner (see Section \ref{sec:pyerr}). 
\begin{itemize}
\item \verb+input_parse(string)+ --- invokes the standard NWChem input
parser with the data in \verb+string+ as input.  Note that the usual
behavior of NWChem will apply --- the parser only reads input up to
either end of input or until a \verb+TASK+ directive is encountered
(the task directive is {\em not} executed by the parser).

\item \verb+task_energy(theory)+ --- returns the energy as if computed
with the NWChem directive \verb+TASK ENERGY <THEORY>+.

\item \verb+task_gradient(theory)+ --- returns a tuple
\verb+(energy,gradient)+ as if computed with the NWChem
directive \verb+TASK GRADIENT <THEORY>+.

\item \verb+task_optimize(theory)+ --- returns a tuple
\verb+(energy,gradient)+ as if computed with the NWChem
directive \verb+TASK OPTIMIZE <THEORY>+.  The energy and gradient
will be those at the last point in the optimization and consistent
with the current geometry in the database.

\item \verb+task_freq(theory)+ --- returns a tuple
directive \verb+TASK FREQ <THEORY>+.  They will be consistent
with the current geometry in the database. 

\item \verb+task_property(theory)+ --- returns the energy
as if computed with the NWChem
directive \verb+TASK PROPERIY <THEORY>+.

\item \verb+task_stress(theory)+ --- returns a tuple
\verb+(energy,gradient,stress)+ as if computed with the NWChem
directive \verb+TASK STRESS <THEORY>+.  The energy and gradient
will be consistent with the current geometry in the database.

\item \verb+task_lstress(theory)+ --- returns a tuple
\verb+(energy,gradient,lstress)+.  The energy and gradient
will be consistent with the current geometry in the database.

\item \verb+task_hessian(theory)+ --- performs a hessian calculation
as if requested with the NWChem
directive \verb+TASK HESSIAN <THEORY>+.  They will be consistent
with the current geometry in the database.

\item \verb+task_saddle(theory)+ --- returns a tuple
\verb+(energy,gradient)+ as if computed with the NWChem
directive \verb+TASK SADDLE <THEORY>+.  The energy and gradient
will be those at the last point in the search and consistent
with the current geometry in the database.

\item \verb+task_coulomb(theory)+ --- returns the
couloumb energy.
It will be consistent with the current geometry in the database.

\item \verb+task_coulomb_ref(theory)+ --- returns the
couloumb reference energy.
It will be consistent with the current geometry in the database.

\item \verb+ga_nodeid()+ --- returns the GA number of the parallel
process.

\item \verb+ga_groupid()+ --- returns the GA group ID.



\item \verb+pgroup_create(num)+ --- create num process groups.  Must be called by all members of current group.

\item \verb+pgroup_destroy()+ --- destroy current group.  Must be called by all members of current group.


\item \verb+pgroup_sync()+ --- a barrier that all processes in a goup must get to before they are allowed to continue.

\item \verb+pgroup_sync_all()+ --- sync all processes regardless of process groups.


\item \verb+pgroup_global_op()+ --- a global operation within a group.
 This is like the MPI DGOP/IGOP commands.
 This input list or numbers or single number must be the same on all nodes.
 If no operation is included then ``+''  is assumed.

\item \verb+pgroup_global_op_all()+ --- a global operation regardless of process groups.

\item \verb+pgroup_global_op_zero()+ --- a global operation among the zero nodes of the current set of groups.  ie. one level up from the lowest level.


\item \verb+pgroup_broadcast()+ --- a broadcast within a group.  This is like the MPI broadcst command.
  Node zero of the group always does the talking.
  All nodes must pass in the same sized object (list of numbers or a single number)
  and will recieve the same sized object.

\item \verb+pgroup_broadcast_all()+ ---  a broadcast regardless of process groups.

\item \verb+pgroup_broadcast_zero()+ ---  a broadcast among the zero nodes of the current set of groups.  ie. one level up from the lowest level.


\item \verb+pgroup_nnodes()+ --- the number of nodes within the current group.

\item \verb+pgroup_nodeid()+ --- the node ID within the current group.

\item \verb+pgroup_groupid()+ --- the group ID of the current group.

\item \verb+pgroup_ngroups()+ --- the number of groups at the current group level.


\item \verb+rtdb_print(print_values)+ --- prints the contents of the
RTDB.  If \verb+print_values+ is 0, only the keys are printed, if it
is 1 then the values are also printed.

\item \verb+rtdb_put(name, values)+ or
\verb+rtdb_put(name, values, type)+ --- puts the values into the
database with the given name.  In the first form, the type is inferred
from the first value, and in the second form the type is specified
using the last argument as one of \verb+INT+, \verb+DBL+,
\verb+LOGICAL+, or \verb+CHAR+.

\item \verb+rtdb_get(name) + --- returns the data from the database
associated with the given name.
\end{itemize}

An example below (Section \ref{sec:pygeom}) explains, in lieu of a
Python wrapper for the geometry object, how to obtain the Cartesian
molecular coordinates directly from the database.

\section{Examples}

Several examples will provide the best explanation of how the extensions
are used, and how Python might prove useful.

\subsection{Hello world}

\begin{verbatim}
  python
    print 'Hello world from process ', ga_nodeid()
  end

  task python
\end{verbatim}

This input prints the traditional greeting from each parallel process.

\subsection{Scanning a basis exponent}
\begin{verbatim}
  geometry units au
    O 0 0 0; H 0 1.430 -1.107; H 0 -1.430 -1.107
  end

  python
    exponent = 0.1
    while (exponent <= 2.01):
       input_parse('''
          basis noprint
             H library 3-21g; O library 3-21g; O d; %f 1.0
          end
       ''' % (exponent))
       print ' exponent = ', exponent, ' energy = ', task_energy('scf')
       exponent = exponent + 0.1
  end

  print none

  task python
\end{verbatim}

This program augments a 3-21g basis for water with a d-function on
oxygen and varies the exponent from 0.1 to 2.0 in steps of 0.1,
printing the exponent and energy at each step.  

The geometry is input as usual, but the basis set input is embedded
inside a call to \verb+input_parse()+ in the Python program.  The
standard Python string substitution is used to put the current value of
the exponent into the basis set (replacing the \verb+%f+) before being
parsed by NWChem.  The energy is returned by \verb+task_energy('scf')+
and printed out.  The \verb+print none+ in the NWChem input switches
off all NWChem output so all you will see is the output from your
Python program.

Note that execution in parallel may produce unwanted output since
all process execute the print statement inside the Python program.

Look in the NWChem \verb+contrib+ directory for a routine that makes
the above task easier.

\subsection{Scanning a basis exponent revisited.}
\label{sec:scan2}

\begin{verbatim}
  geometry units au
    O 0 0 0; H 0 1.430 -1.107; H 0 -1.430 -1.107
  end

  print none

  python
    if (ga_nodeid() == 0): plotdata = open("plotdata",'w')

    def energy_at_exponent(exponent):
       input_parse('''
          basis noprint
             H library 3-21g; O library 3-21g; O d; %f 1.0
          end
       ''' % (exponent))
       return task_energy('scf')

    exponent = 0.1
    while exponent <= 2.01:
       energy = energy_at_exponent(exponent)
       if (ga_nodeid() == 0):
          print ' exponent = ', exponent, ' energy = ', energy
          plotdata.write('%f %f\n' % (exponent , energy))
       exponent = exponent + 0.1

    if (ga_nodeid() == 0): plotdata.close()
  end

  task python
\end{verbatim}

This input performs exactly the same calculation as the previous one,
but uses a slightly more sophisticated Python program, also writes
the data out to a file for easy visualization with a package such as
\verb+gnuplot+, and protects write statements to prevent 
duplicate output in a parallel job.  The only significant differences
are in the Python program.  A file called \verb+"plotdata"+ is opened,
and then a procedure is defined which given an exponent returns the
energy.  Next comes the main loop that scans the exponent through the
desired range and prints the results to standard output and to the
file.  When the loop is finished the additional output file is closed.

\subsection{Scanning a geometric variable}

\begin{verbatim}
  python
    geometry = '''
      geometry noprint; symmetry d2h
         C 0 0 %f; H 0  0.916 1.224
      end
    '''
    x = 0.6
    while (x < 0.721):
      input_parse(geometry % x)
      energy = task_energy('scf')
      print ' x = %5.2f   energy = %10.6f' % (x, energy)
      x = x + 0.01
  end

  basis; C library 6-31g*; H library 6-31g*; end

  print none

  task python
\end{verbatim}

This scans the bond length in ethene from 1.2 to 1.44 in steps 
of 0.2 computing the energy at each geometry.  Since it is using 
$D_{2h}$ symmetry the program actually uses a variable (\verb+x+) that is
half the bond length.

Look in the NWChem \verb+contrib+ directory for a routine that makes
the above task easier.

\subsection{Scan using the BSSE counterpoise corrected energy}

\begin{verbatim}
  basis spherical
    Ne library cc-pvdz; BqNe library Ne cc-pvdz
    He library cc-pvdz; BqHe library He cc-pvdz
  end

  mp2; tight; freeze core atomic; end

  print none

  python noprint
    supermolecule = 'geometry noprint;   Ne 0 0 0;   He 0 0 %f; end\n'
    fragment1     = 'geometry noprint;   Ne 0 0 0; BqHe 0 0 %f; end\n'
    fragment2     = 'geometry noprint; BqNe 0 0 0;   He 0 0 %f; end\n'

    def energy(geometry):
      input_parse(geometry + 'scf; vectors atomic; end\n')
      return task_energy('mp2')

    def bsse_energy(z):
      return energy(supermolecule % z) - \
             energy(fragment1 % z) - \
             energy(fragment2 % z)
    z = 3.3
    while (z < 4.301):
      e = bsse_energy(z)
      if (ga_nodeid() == 0):
        print ' z = %5.2f   energy = %10.7f ' % (z, e)
      z = z + 0.1
  end

  task python
\end{verbatim}

This example scans the He---Ne bond-length from 3.3 to 4.3 and prints out
the BSSE counterpoise corrected MP2 energy.

The basis set is specified as usual, noting that we will need
functions on ghost centers to do the counterpoise correction.  The
Python program commences by defining strings containing the geometry
of the super-molecule and two fragments, each having one variable to be
substituted.  Next, a function is defined to compute the energy given
a geometry, and then a function is defined to compute the counterpoise
corrected energy at a given bond length.  Finally, the bond length is
scanned and the energy printed.  When computing the energy, the atomic
guess has to be forced in the SCF since by default it will attempt to
use orbitals from the previous calculation which is not appropriate
here.

Since the counterpoise corrected energy is a linear combination of
other standard energies, it is possible to compute the analytic
derivatives term by term.  Thus, combining this example and the next
could yield the foundation of a BSSE corrected geometry optimization
package.

\subsection{Scan the geometry and compute the energy and gradient}

\begin{verbatim}
  basis noprint; H library sto-3g; O library sto-3g; end

  python noprint
    print '   y     z     energy                gradient'
    print ' ----- ----- ---------- ------------------------------------'
    y = 1.2
    while y <= 1.61:
       z = 1.0
       while z <= 1.21:
          input_parse('''
             geometry noprint units atomic
                O 0   0   0
                H 0  %f -%f
                H 0 -%f -%f
             end
          ''' % (y, z, y, z))

          (energy,gradient) = task_gradient('scf')

          print ' %5.2f %5.2f %9.6f' % (y, z, energy),
          i = 0
          while (i < len(gradient)):
             print '%5.2f' % gradient[i],
             i = i + 1
          print ''
          z = z + 0.1
       y = y + 0.1
  end

  print none

  task python
\end{verbatim}

This program illustrates evaluating the energy and gradient
by calling \verb+task_gradient()+.  A water molecule is scanned
through several $C_{2v}$ geometries by varying the y and z coordinates
of the two hydrogen atoms.  At each geometry the coordinates, energy 
and gradient are printed.

The basis set (sto-3g) is input as usual.  The two while loops vary
the y and z coordinates.  These are then substituted into a geometry
which is parsed by NWChem using \verb+input_parse()+.  The energy and
gradient are then evaluated by calling \verb+task_gradient()+ which
returns a tuple containing the energy (a scalar) and the gradient (a
vector or list).  These are printed out exploiting the Python
convention that a print statement ending in a comma does not print
end-of-line.

\subsection{Reaction energies varying the basis set}

\begin{verbatim}
  mp2; freeze atomic; end

  print none

  python
    energies = {}
    c2h4 = 'geometry noprint; symmetry d2h; \
            C 0 0 0.672; H 0 0.935 1.238; end\n'
    ch4  = 'geometry noprint; symmetry td; \
            C 0 0 0; H 0.634 0.634 0.634; end\n'
    h2   = 'geometry noprint; H 0 0 0.378; H 0 0 -0.378; end\n'

    def energy(basis, geometry):
      input_parse('''
        basis spherical noprint
          c library %s ; h library %s 
        end
      ''' % (basis, basis))
      input_parse(geometry)
      return task_energy('mp2')

    for basis in ('sto-3g', '6-31g', '6-31g*', 'cc-pvdz', 'cc-pvtz'):
       energies[basis] =   2*energy(basis, ch4) \
                         - 2*energy(basis, h2) - energy(basis, c2h4)
       if (ga_nodeid() == 0): print basis, ' %8.6f' % energies[basis]
  end 

  task python
\end{verbatim}

In this example the reaction energy for 
$2H_2 + C_2H_4 \rightarrow 2CH_4$ is evaluated using MP2 in several
basis sets.  The geometries are fixed, but could be re-optimized in
each basis.  To illustrate the useful associative arrays in Python,
the reaction energies are put into the associative array
\verb+energies+ --- note its declaration at the top of the program.

\subsection{Using the database}

\begin{verbatim}
  python
    rtdb_put("test_int2", 22)
    rtdb_put("test_int", [22, 10, 3],    INT)
    rtdb_put("test_dbl", [22.9, 12.4, 23.908],  DBL)
    rtdb_put("test_str", "hello", CHAR)
    rtdb_put("test_logic", [0,1,0,1,0,1], LOGICAL)
    rtdb_put("test_logic2", 0, LOGICAL)

    rtdb_print(1)

    print "test_str    = ", rtdb_get("test_str")
    print "test_int    = ", rtdb_get("test_int")
    print "test_in2    = ", rtdb_get("test_int2")
    print "test_dbl    = ", rtdb_get("test_dbl")
    print "test_logic  = ", rtdb_get("test_logic")
    print "test_logic2 = ", rtdb_get("test_logic2")
  end

  task python
\end{verbatim}

This example illustrates how to access the database from Python.

\subsection{Handling exceptions from NWChem}
\label{sec:pyerr}

\begin{verbatim}
  geometry; he 0 0 0; he 0 0 2; end
  basis; he library 3-21g; end
  scf; maxiter 1; end

  python
    try:
      task_energy('scf')
    except NWChemError, message:
      print 'Error from NWChem ... ', message
  end

  task python
\end{verbatim}

The above test program shows how to handle exceptions generated by
NWChem by forcing an SCF calculation on $He_2$ to fail due to
insufficient iterations.

If an NWChem command fails it will raise the exception
\verb+"NWChemError"+ (case sensitive) unless the error was fatal.
If the exception is not caught, then it will cause the entire Python
program to terminate with an error.  This Python program catches the
exception, prints out the message, and then continues as if all was
well since the exception has been handled.  

If your Python program detects an error, raise an unhandled
exception.  Do not call \verb+exit(1)+ since this may circumvent
necessary clean-up of the NWChem execution environment.

\subsection{Accessing geometry information --- a temporary hack}
\label{sec:pygeom}

In an ideal world the geometry and basis set objects would have full
Python wrappers, but until then a back-door solution will have to
suffice.  We've already seen how to use \verb+input_parse()+ to put
geometry (and basis) data into NWChem, so it only remains to get the
geometry data back after it has been updated by a geometry optimzation
or some other operation.  

The following Python procedure retrieves the coordinates in the
same units as initially input for a geometry of a given name.
Its full source is included in the NWChem \verb+contrib+ directory.
\begin{verbatim}
  def geom_get_coords(name):
    try:
      actualname = rtdb_get(name)
    except NWChemError:
      actualname = name
    coords = rtdb_get('geometry:' + actualname + ':coords')
    units  = rtdb_get('geometry:' + actualname + ':user units')
    if (units == 'a.u.'):
      factor = 1.0
    elif (units == 'angstroms'):
      factor = rtdb_get('geometry:'+actualname+':angstrom_to_au')
    else:
      raise NWChemError,'unknown units'
    i = 0
    while (i < len(coords)):
      coords[i] = coords[i] / factor
      i = i + 1
    return coords
\end{verbatim}

A geometry (see Section \ref{sec:geom}) with name \verb+NAME+ has its
coordinates (in atomic units) stored in the database entry
\verb+geometry:NAME:coords+.  A minor wrinkle here is that 
indirection is possible (and used by the optimizers) so that we must
first check if \verb+NAME+ actually points to another name.  In the
program this is done in the first \verb+try...except+ sequence.  With
the actual name of the geometry, we can get the coordinates.  Any
exceptions are passed up to the caller.  The rest of the code is just
to convert back into the initial input units --- only atomic units 
or \angstroms\ are handled in this simple example.  Returned 
is a list of the atomic coordinates in the same units as your
initial input.

The routine is used as follows
\begin{verbatim}
    coords = geom_get_coords('geometry')
\end{verbatim}
or, if you want better error handling 
\begin{verbatim}
    try:
      coords = geom_get_coords('geometry')
    except NWChemError,message:
      print 'Coordinates for geometry not found ', message
    else:
      print coords    
\end{verbatim}

This is very dirty and definitely not supported from one release to
another, but, browsing the output of \verb+rtdb_print()+ at the end of
a calculation is a good way to find stuff.  To be on safer ground,
look in the programmers manual since some of the high-level routines
do pass data via the database in a well-defined and supported manner.
{\em Be warned} --- you must be very careful if you try to modify data
in the database. The input parser does many important things that are
not immediately apparent (e.g., ensure the geometry is consistent with
the point group, mark the SCF as not converged if the SCF options are
changed, \ldots).  Where at all possible your Python program should
generate standard NWChem input and pass it to \verb+input_parse()+
rather than setting parameters directly in the database.

\subsection{Scaning a basis exponent yet again --- plotting and 
handling child processes}
\label{sec:sigchld}

\begin{verbatim}
  geometry units au
    O 0 0 0; H 0 1.430 -1.107; H 0 -1.430 -1.107
  end

  print none

  python
    import Gnuplot, time, signal

    def energy_at_exponent(exponent):
       input_parse('''
          basis noprint
             H library 3-21g; O library 3-21g; O d; %f 1.0
          end
       ''' % (exponent))
       return task_energy('scf')

    data = []
    exponent = 0.5
    while exponent <= 0.6:
       energy = energy_at_exponent(exponent)
       print ' exponent = ', exponent, ' energy = ', energy
       data = data + [[exponent,energy]]
       exponent = exponent + 0.02

    if (ga_nodeid() == 0):
       signal.signal(signal.SIGCHLD, signal.SIG_DFL)
       g = Gnuplot.Gnuplot()
       g('set data style linespoints')
       g.plot(data)
       time.sleep(30)  # 30s to look at the plot

  end

  task python
\end{verbatim}

This illustrates how to handle signals from terminating child
processes and how to generate simple plots on UNIX systems.  The
example from Section \ref{sec:scan2} is modified so that instead of
writing the data to a file for subsequent visualization, it is saved
for subsequent visualization with Gnuplot (you'll need both Gnuplot
and the corresponding package for Python in your \verb+PYTHONPATH+.
Look at \htmladdnormallink{http://monsoon.harvard.edu/~mhagger/download}{http://monsoon.harvard.edu/~mhagger/download}).

The issue is that NWChem traps various signals from the O/S that
usually indicate bad news in order to provide better error handling
and reliable clean-up of shared, parallel resources.  One of these
signals is \verb+SIGCHLD+ which is generated whenever a child process
terminates.  If you want to create child processes within Python, then
the NWChem handler for \verb+SIGCHLD+ must be replaced with the
default handler.  There seems to be no easy way to restore the
NWChem handler after the child has completed, but this should have
no serious side effect.

\section{Troubleshooting}

Common problems with Python programs inside NWChem.

\begin{enumerate}
\item You get the message
\begin{verbatim}
      0:python_input: indentation must be >= that of first line: 4
\end{verbatim}
This indicates that NWChem thinks that a line is less indented than
the first line.  If this is not the case then perhaps there is a tab
in your input which NWChem treats as a single space character but
appears to you as more spaces. Try running \verb+untabify+ in Emacs.
The problem could also be the \verb+END+ directive that terminates the
\verb+PYTHON+ compound directive --- since Python also has an
\verb+end+ statement.  To avoid confusion the \verb+END+ directive
for NWChem {\em must} be at the start of the line.

\item Your program hangs or deadlocks --- most likely you have a piece
of code that is restricted to executing on a subset of the processors
(perhaps just node 0) but is calling (perhaps indirectly) a function
that must execute on all nodes.  

\end{enumerate}