'-p': topology_file
},
output={'-o': gmx.File(suffix='.tpr')})
# Note: Before gmx.File, users still have to manage filenames
# The above would have `output={'-o': [initial_tpr for _ in range(N)]}`
# Note: initial_tpr has a single output that can be automatically broadcast now or later.
# Broadcast to the read_tpr operation:
#simulation_input = gmx.read_tpr([initial_tpr for _ in range(N)])
# Wait to broadcast until the next operation:
simulation_input = gmx.read_tpr(initial_tpr.output.file['-o'])
# Array inputs imply array outputs.
input_array = gmx.modify_input(simulation_input,
params={'tau-t': [t / 10.0 for t in range(N)]})
md = gmx.mdrun(input_array) # An array of simulations
rmsf = gmx.commandline_operation('gmx',
'rmsf',
input={
'-f': md.output.trajectory,
'-s': initial_tpr
},
output={'-o': gmx.File(suffix='.xvg')})
output_files = gmx.gather(rmsf.output.file['-o'])
gmx.run()
print('Output file list:')
print(', '.join(output_files.result()))
# For subgraphs, inputs can be accessed as variables and are copied to the next
# iteration (not typical for gmxapi operation input/output).
train = gmx.subgraph(variables={'conformation': initial_input})
# References to the results of operations know which (sub)graph they live in.
# The `with` block activates and deactivates the scope of the subgraph in order
# to constrain the section of this script in which references are valid.
# Otherwise, a user could mistakenly use a reference that only points to the
# result of the first iteration of a "while" loop. If the `with` block succeeds,
# then the outputs of `train` are afterwards fully specified.
with train:
myplugin.training_restraint(label='training_potential',
params=my_dict_params)
modified_input = gmx.modify_input(input=initial_input,
structure=train.conformation)
md = gmx.mdrun(input=modified_input, potential=train.training_potential)
# Alternate syntax to facilitate adding multiple potentials:
# md.interface.potential.add(train.training_potential)
brer_tools.training_analyzer(label='is_converged',
params=train.training_potential.output.alpha)
train.conformation = md.output.conformation
# At the end of the `with` block, `train` is no longer the active graph, and
# gmx.exceptions.ScopeError will be raised if `modified_input`, or `md` are used
# in other graph scopes (without first reassigning, of course)
# More discussion at https://github.com/kassonlab/gmxapi/issues/205
# In the default work graph, add a node that depends on `condition` and
# wraps subgraph.
train_loop = gmx.while_loop(operation=train,
condition=gmx.logical_not(train.is_converged))
'pair_distance2': gmx.NDArray(0., shape=restraint2_params['nbins']),
})
with converge:
# ensemble_restraint is implemented using gmxapi ensemble allReduce operations
# that do not need to be expressed in this procedural interface.
potential1 = myplugin.ensemble_restraint(
label='ensemble_restraint_1',
params=restraint1_params,
input={'pair_distance': converge.pair_distance1})
potential2 = myplugin.ensemble_restraint(
label='ensemble_restraint_2',
params=restraint2_params,
input={'pair_distance': converge.pair_distance2})
md = gmx.mdrun(initial_input)
md.interface.potential.add(potential1)
md.interface.potential.add(potential2)
# Compare the distribution from the current iteration to the experimental
# data and look for a threshold of low J-S divergence
# We perform the calculation using all of the ensemble data.
js_1 = calculate_js(
input={
'params': restraint1_params,
'simulation_distances': gmx.gather(potential1.output.pair_distance)
})
js_2 = calculate_js(
input={
'params': restraint2_params,
'simulation_distances': gmx.gather(potential2.output.pair_distance)