More about the Environment¶
Creating an Environment¶
In most use cases you will interact with the Environment to do your numerical simulations. The environment is your handyman for your numerical experiments, it sets up new trajectories, keeps log files and can be used to distribute your simulations onto several CPUs.
You start your simulations by creating an environment object:
>>> env = Environment(trajectory='trajectory', comment='A useful comment')
You can pass the following arguments. Note usually you only have to change very few of these because most of the time the default settings are sufficient.
trajectory
The first argument trajectory can either be a string or a given trajectory object. In case of a string, a new trajectory with that name is created. You can access the new trajectory via v_trajectory property. If a new trajectory is created, the comment and dynamically imported classes are added to the trajectory.
add_time
Whether the current time in format XXXX_XX_XX_XXhXXmXXs is added to the trajectory name if the trajectory is newly created.
comment
The comment that will be added to a newly created trajectory.
dynamic_imports
Only considered if a new trajectory is created.
The argument dynamic_imports is important if you have written your own parameter or result classes, you can pass these either as class variables MyCustomParameterClass or as strings leading to the classes in your package: 'mysim.myparameters.MyCustomParameterClass'. If you have several classes, just put them in a list dynamic_imports=[MyCustomParameterClass, MyCustomResultClass]. In case you want to load a custom class from disk and the trajectory needs to know how they are built.
It is VERY important, that every class name is UNIQUE. So you should not have two classes named 'MyCustomParameterClass' in two different python modules! The identification of the class is based only on its name and not its path in your packages.
automatic_storing
If True the trajectory will be stored at the end of the simulation and single runs will be stored after their completion. Be aware of data loss if you set this to False and not manually store everything.
log_folder
The log_folder specifies where all log files will be stored. The environment will create a sub-folder with the name of the trajectory and the name of the environment where all txt files will be put. The environment will create a major logfile (main.txt) incorporating all messages of the current log level and beyond and a log file that only contains warnings and errors errors_and_warnings.txt.
Moreover, if you use multiprocessing, there will be a log file for every single run and process named run_XXXXXXXX_process_YYYY.txt with XXXXXXXX the run id and YYYY the process id. It contains all log messages produced by the corresponding process within the single run.
If you don’t want the logging message stored to the file system set to None.
If you don’t set a log level elsewhere before, the standard level will be INFO (if you have no clue what I am talking about, take a look at the logging module).
logger_names
List or tuple of logger names to which the logging settings apply, i.e which log messages are stored to disk by pypet. Default is root ('',), i.e. all logging messages are logged to the folder specified above. For instance, if you only want pypet to save messages created by itself and not by your own loggers use logger_names='(pypet,)'. If you only want to store message from your custom loggers, you could pass the names of your loggers, like logger_names=('MyCustomLogger1', 'MyCustomLogger2', ...). Or, for example, if you only want to store messages from stdout and stderr set logger_names to ('STDOUT','STDERR').
log_levels
List or tuple of log levels with the same length as logger_names. If the length is 1 and loger_names has more than 1 entry, the log level is used for all loggers.
Default is level (logging.INFO,). If you choose (logging.DEBUG,) more verbose statements will be displayed. Set to None if you don’t want to set log-levels or if you already specified log-levels somewhere else.
log_stdout
Whether the output of stdout and stderr should be recorded into the log files. Disable if only logging statement should be recorded. Note if you work with an interactive console like IPython, it is a good idea to set log_stdout=False to avoid messing up the console output.
multiproc
multiproc specifies whether or not to use multiprocessing (take a look at Multiprocessing). Default is False.
ncores
If multiproc is True, this specifies the number of processes that will be spawned to run your experiment. Note if you use 'QUEUE' mode (see below) the queue process is not included in this number and will add another extra process for storing. If you have psutil installed, you can set ncores=0 to let psutil determine the number of CPUs available.
use_pool
If you choose multiprocessing you can specify whether you want to spawn a new process for every run or if you want a fixed pool of processes to carry out your computation.
If you use a pool, all your data and the tasks you compute must be picklable! If you never heard about pickling or object serialization, you might want to take a loot at the pickle module.
Thus, if your simulation data cannot be pickled (which is the case for some BRIAN networks, for instance), choose use_pool=False and continuable=False (see below).
cpu_cap
If multiproc=True and use_pool=False you can specify a maximum CPU utilization between 0.0 (excluded) and 1.0 (included) as fraction of maximum capacity. If the current CPU usage is above the specified level (averaged across all cores), pypet will not spawn a new process and wait until activity falls below the threshold again. Note that in order to avoid dead-lock at least one process will always be running regardless of the current utilization. If the threshold is crossed a warning will be issued. The warning won’t be repeated as long as the threshold remains crossed.
For example let us assume you chose cpu_cap=0.7, ncores=3, and currently on average 80 percent of your CPU are used. Moreover, at the moment only 2 processes are computing single runs simultaneously. Due to the usage of 80 percent of your CPU, pypet will wait until CPU usage drops below (or equal to) 70 percent again until it starts a third process to carry out another single run.
The parameters memory_cap and swap_cap are analogous. These three thresholds are combined to determine whether a new process can be spawned. Accordingly, if only one of these thresholds is crossed, no new processes will be spawned.
To disable the cap limits simply set all three values to 1.0.
You need the psutil package to use this cap feature. If not installed and you choose cap values different from 1.0 a ValueError is thrown.
memory_cap
Cap value of RAM usage. If more RAM than the threshold is currently in use, no new processes are spawned.
swap_cap
Analogous to memory_cap but the swap memory is considered.
wrap_mode
If multiproc is True, specifies how storage to disk is handled via the storage service. Since PyTables HDF5 is not thread safe, the HDF5 storage service needs to be wrapped with a helper class to allow the interaction with multiple processes.
There are two options:
pypet.pypetconstants.MULTIPROC_MODE_QUEUE: (‘QUEUE’)
Another process for storing the trajectory is spawned. The sub processes running the individual single runs will add their results to a multiprocessing queue that is handled by an additional process.
pypet.pypetconstants.MULTIPROC_MODE_LOCK: (‘LOCK’)
Each individual process takes care about storage by itself. Before carrying out the storage, a lock is placed to prevent the other processes to store data.
If you don’t want wrapping at all use pypet.pypetconstants.MULTIPROC_MODE_NONE (‘NONE’).
If you have no clue what I am talking about, you might want to take a look at multiprocessing in python to learn more about locks, queues and thread safety and so forth.
clean_up_runs
In case of single core processing, whether all results under results.runs.run_XXXXXXXX and derived_parameters.runs.run_XXXXXXXX should be removed after the completion of the run. Note in case of multiprocessing this happens anyway since the trajectory container will be destroyed after finishing of the process.
Moreover, if set to True after post-processing run data is also cleaned up.
immediate_postproc
If you use post- and multiprocessing, you can immediately start analysing the data as soon as the trajectory runs out of tasks, i.e. is fully explored but the final runs are not completed. Thus, while executing the last batch of parameter space points, you can already analyse the finished runs. This is especially helpful if you perform some sort of adaptive search within the parameter space.
The difference to normal post-processing is that you do not have to wait until all single runs are finished, but your analysis already starts while there are still runs being executed. This can be a huge time saver especially if your simulation time differs a lot between individual runs. Accordingly, you don’t have to wait for a very long run to finish to start post-processing.
Note that after the execution of the final run, your post-processing routine will be called again as usual.
continuable
Whether the environment should take special care to allow to resume or continue crashed trajectories. Default is False.
You need to install dill to use this feature. dill will make snapshots of your simulation function as well as the passed arguments. Be aware that dill is still rather experimental!
Assume you run experiments that take a lot of time. If during your experiments there is a power failure, you can resume your trajectory after the last single run that was still successfully stored via your storage service.
The environment will create several .ecnt and .rcnt files in a folder that you specify (see below). Using this data you can continue crashed trajectories.
In order to resume trajectories use f_continue().
Your individual single runs must be completely independent of one another to allow continuing to work. Thus, they should not be based on shared data that is manipulated during runtime (like a multiprocessing manager list) in the positional and keyword arguments passed to the run function.
If you use postprocessing, the expansion of trajectories and continuing of trajectories is not supported properly. There is no guarantee that both work together.
continue_folder
The folder where the continue files will be placed. Note that pypet will create a sub-folder with the name of the environment.
delete_continue
If true, pypet will delete the continue files after a successful simulation.
storage_service
Pass a given storage service or a class constructor (default is HDF5StorageService) if you want the environment to create the service for you. The environment will pass additional keyword arguments you provide directly to the constructor. If the trajectory already has a service attached, the one from the trajectory will be used. For the additional keyword arguments, see below.
git_repository
If your code base is under git version control you can specify the path (relative or absolute) to the folder containing the .git directory. See also Git Integration.
git_message
Message passed onto git command.
do_single_runs
Whether you intend to actually to compute single runs with the trajectory. If you do not intend to carry out single runs (probably because you loaded an old trajectory for data analysis), than set to False and the environment won’t add config information like number of processors to the trajectory.
lazy_debug
If lazy_debug=True and in case you debug your code (aka you use pydevd and the expression 'pydevd' in sys.modules is True), the environment will use the LazyStorageService instead of the HDF5 one. Accordingly, no files are created and your trajectory and results are not saved. This allows faster debugging and prevents pypet from blowing up your hard drive with trajectories that you probably not want to use anyway since you just debug your code.
If you use the standard HDF5StorageService you can pass the following additional keyword arguments to the environment. These are handed over to the service:
filename
The name of the hdf5 file. If none is specified, the default ./hdf5/the_name_of_your_trajectory.hdf5 is chosen. If filename contains only a path like filename='./myfolder/', it is changed to filename='./myfolder/the_name_of_your_trajectory.hdf5'.
file_title
Title of the hdf5 file (only important if file is created new)
overwrite_file
If the file already exists it will be overwritten. Otherwise the trajectory will simply be added to the file and already existing trajectories are not deleted.
encoding
Encoding for unicode characters. The default 'utf8' is highly recommended.
complevel
You can specify your compression level. 0 means no compression and 9 is the highest compression level. By default the level is set to 9 to reduce the size of the resulting HDF5 file. See PyTables Compression for a detailed explanation.
complib
The library used for compression. Choose between zlib, blosc, and lzo. Note that ‘blosc’ and ‘lzo’ are usually faster than ‘zlib’ but it may be the case that you can no longer open your hdf5 files with third-party applications that do not rely on PyTables.
shuffle
Whether or not to use the shuffle filters in the HDF5 library. This normally improves the compression ratio.
fletcher32
Whether or not to use the Fletcher32 filter in the HDF5 library. This is used to add a checksum on hdf5 data.
pandas_format
How to store pandas data frames. Either in ‘fixed’ (‘f’) or ‘table’ (‘t’) format. Fixed format allows fast reading and writing but disables querying the hdf5 data and appending to the store (with other 3rd party software other than pypet).
purge_duplicate_comments
If you add a result via f_add_result() or a derived parameter f_add_derived_parameter() and you set a comment, normally that comment would be attached to each and every instance. This can produce a lot of unnecessary overhead if the comment is the same for every result over all runs. If hdf5.purge_duplicate_comments=True than only the comment of the first result or derived parameter instance created is stored, or comments that differ from this first comment. You might want to take a look at HDF5 Purging of duplicate Comments.
summary_tables
Whether summary tables should be created. These give overview about ‘derived_parameters_runs_summary’, and ‘results_runs_summary’. They give an example about your results by listing the very first computed result. If you want to purge_duplicate_comments you will need the summary_tables. You might want to check out HDF5 Overview Tables.
small_overview_tables
Whether the small overview tables should be created. Small tables are giving overview about ‘config’, ‘parameters’, ‘derived_parameters_trajectory’, ‘results_trajectory’.
large_overview_tables
Whether to add large overview tables. These encompass information about every derived parameter and result and the explored parameters in every single run. If you want small HDF5 files set to False (default).
results_per_run
Expected results you store per run. If you give a good/correct estimate, storage to HDF5 file is much faster in case you want large_overview_tables.
Default is 0, i.e. the number of results is not estimated!
derived_parameters_per_run
Analogous to the above.
Config Data added by the Environment¶
The Environment will automatically add some config settings to your trajectory. Thus, you can always look up how your trajectory was run. This encompasses many of the above named parameters as well as some information about the environment. This additional information includes a timestamp and a SHA-1 hash code that uniquely identifies your environment. If you use git integration (Git Integration), the SHA-1 hash code will be the one from your git commit. Otherwise the code will be calculated from the trajectory name, the current time, and your current pypet version.
The environment will be named environment_XXXXXXX_XXXX_XX_XX_XXhXXmXXs. The first seven X are the first seven characters of the SHA-1 hash code followed by a human readable timestamp.
All information about the environment can be found in your trajectory under config.environment.environment_XXXXXXX_XXXX_XX_XX_XXhXXmXXs. Your trajectory could potentially be run by several environments due to merging or extending an existing trajectory. Thus, you will be able to track how your trajectory was built over time.
Logging¶
Per default the environment will created loggers and stores all logged messages to log files. This includes also everything written to the standard streams stdout and stderr, like print statements, for instance. To disable logging of the standard streams set log_stdout=False. Note that you should always do this in case you use an interactive console like IPython. Otherwise your console output will be garbled due to the redirection of standard streams.
You can find the log files in the log_folder you specified. They are placed in a sub-folder with the name of the current trajectory and the current environment. To disable logging to files simply set log_folder=None.
Moreover, you can fine tune what is supposed to be logged to files. By passing a list of logger_names you tell pypet which loggers should be recorded to the files (default is the root logger, i.e. all loggers, logger_names=('',)). Accordingly, you can set individual log levels for all these loggers via a log_levels list of same length (default is log_levels=(logging.INFO,)).
After your experiments are finished you can disable logging to files via f_disable_logging(). This also restores the standard streams.
Furthermore, an environment can also be used as a context manager such that logging is automatically disabled in the end:
import logging
from pypet import Environment
with Environment(trajectory='mytraj',
log_folder='logs',
logger_names=('',),
log_levels=(logging.DEBUG),
log_stdout=True) as env:
traj = env.v_trajectory
# do your complex experiment...
This is equivalent to:
import logging
from pypet import Environment
env = Environment(trajectory='mytraj',
log_folder='logs',
logger_names=('',),
log_levels=(logging.DEBUG,),
log_stdout=True)
traj = env.v_trajectory
# do your complex experiment...
env.f_disable_logging()
Multiprocessing¶
For an example on multiprocessing see Multiprocessing.
The following code snippet shows how to enable multiprocessing with 4 CPUs, a pool, and a queue.
env = Environment(self, trajectory='trajectory',
comment='',
dynamic_imports=None,
log_folder='../log/',
use_hdf5=True,
filename='../experiments.h5',
file_title='experiment',
multiproc=True,
ncores=4,
use_pool=True,
wrap_mode='QUEUE')
Setting use_pool=True will create a pool of ncores worker processes which perform your simulation runs.
IMPORTANT: Python multiprocessing does not work well with multi-threading of openBLAS. If your simulation relies on openBLAS, you need to make sure that multi-threading is disabled. For disabling set the environment variables OPENBLAS_NUM_THREADS=1 and OMP_NUM_THREADS=1 before starting python and using pypet. For instance, numpy and matplotlib (!) use openBLAS to solve linear algebra operations. If your simulation relies on these packages, make sure the environment variables are changed appropriately. Otherwise your program might crash or get stuck in an infinite loop.
IMPORTANT: In order to allow multiprocessing with a pool (or in general under Windows), all your data and objects of your simulation need to be serialized with pickle. But don’t worry, most of the python stuff you use is automatically picklable.
If you come across the situation that your data cannot be pickled (which is the case for some BRIAN networks, for example), don’t worry either. Set use_pool=False (and also continuable=False) and for every simulation run pypet will spawn an entirely new subprocess. The data is than passed to the subprocess by forking on OS level and not by pickling. However, this only works under Linux. If you use Windows and choose use_pool=False you still need to rely on pickle because Windows does not support forking of python processes.
Moreover, if you enable multiprocessing and disable pool usage, besides the maximum number of utilized processors ncores, you can specify usage cap levels with cpu_cap, memory_cap, and swap_cap as fractions of the maximum capacity. Values must be chosen larger than 0.0 and smaller or equal to 1.0. If any of these thresholds is crossed no new processes will be started by pypet. For instance, if you want to use 3 cores aka ncores=3 and set a memory cap of memory_cap=0.9 and let’s assume that currently only 2 processes are started with currently 95 percent of you RAM are occupied. Accordingly, pypet will not start the third process until RAM usage drops again below (or equal to) 90 percent.
Moreover, to prevent dead-lock pypet will regardless of the cap values always start at least one process. To disable the cap levels, simply set all three to 1.0 (which is default, anyway). pypet does not check if the processes themselves obey the cap limit. Thus, if one of the process that computes your single runs needs more RAM/Swap or CPU power than the cap value, this is its very own problem. The process will not be terminated by pypet. The process will only cause pypet to not start new processes until the utilization falls below the threshold again. In order to use this cap feature you need the psutil package.
Note that HDF5 is not thread safe, so you cannot use the standard HDF5 storage service out of the box. However, if you want multiprocessing, the environment will automatically provide wrapper classes for the HDF5 storage service to allow safe data storage. There are two different modes that are supported. You can choose between them via setting wrap_mode. You can select between 'QUEUE' and 'LOCK' wrapping. If you have your own service that is already thread safe you can also choose 'NONE' to skip wrapping.
If you chose the 'QUEUE' mode, there will be an additional process spawned that is the only one writing to the HDF5 file. Everything that is supposed to be stored is send over a queue to the process. This has the advantage that your worker processes are only busy with your simulation and are not bothered with writing data to a file. More important, they don’t spend time waiting for other processes to release a thread lock to allow file writing. The disadvantage are that you can only store but not load data and storage relies a lot on pickling of data, so often your entire trajectory is send over the queue.
If you chose the 'LOCK' mode, every process will place a lock before it opens the HDF5 file for writing data. Thus, only one process at a time stores data. The advantages are the possibility to load data and that your data does not need to be send over a queue over and over again. Yet, your simulations might take longer since processes have to wait often for each other to release locks.
Finally, there also exist a lightweight multiprocessing environment MultiprocContext. It allows to use trajectories in a multiprocess safe setting without the need of a full Environment. For instance, you might use this if you also want to analyse the trajectory with multiprocessing. You can find an example here: Lightweight Multiprocessing.
Git Integration¶
The environment can make use of version control. If you manage your code with git, you can trigger automatic commits with the environment to get a proper snapshot of the code you actually use. This ensures that your experiments are repeatable. In order to use the feature of git integration, you additionally need GitPython.
To trigger an automatic commit simply pass the arguments git_repository and git_message to the Environment constructor. git_repository specifies the path to the folder containing the .git directory. git_message is optional and adds the corresponding message to the commit. Note that the message will always be augmented with some short information about the trajectory you are running. The commit SHA-1 hash and some other information about the commit will be added to the config subtree of your trajectory, so you can easily recall that commit from git later on.
The automatic commit functionality will only commit changes in files that are currently tracked by your git repository, it will not add new files. So make sure to put new files into your repository before running an experiment. Moreover, a commit will only be triggered if your working copy contains changes. If there are no changes detected, information about the previous commit will be added to the trajectory. By the way, the autocommit function is similar to calling $ git add -u and $ git commit -m 'Some Message' in your console.
Sumatra Integration¶
The environment can make use of a Sumatra experimental lab-book.
Just pass the argument sumatra_project - which should specify the path to your root sumatra folder - to the Environment constructor. You can additionally pass a sumatra_reason, a string describing the reason for you sumatra simulation. pypet will automatically add the name, comment, and the names of all explored parameters to the reason. You can also pick a sumatra_label, set this to None if you want Sumatra to pick a label for you. Moreover, pypet automatically adds all parameters to the sumatra record. The explored parameters are added with their full range instead of the default values.
In contrast to the automatic git commits (see above), which are done as soon as the environment is created, a sumatra record is only created and stored if you actually perform single runs. Hence, records are stored if you use one of following three functions: f_run(), or f_pipeline(), or f_continue() and your simulation succeeds and does not crash.
HDF5 Overview Tables¶
The HDF5StorageService creates summarizing information about your trajectory that can be found in the overview group within your HDF5 file. These overview tables give you a nice summary about all parameters and results you needed and computed during your simulations.
The following tables are created depending of your choice of large_overview_tables and small_overview_tables:
An info table listing general information about your trajectory
A runs table summarizing the single runs
The branch tables:
parameters_overview
Containing all parameters, and some information about comments, length etc.
config_overview,
As above, but config parameters
results_runs
All results of all individual runs, to reduce memory size only a short value summary and the name is given. Per default this table is switched off, to enable it pass large_overview_tables=True to your environment.
results_runs_summary
Only the very first result with a particular name is listed. For instance if you create the result ‘my_result’ in all runs only the result of run_00000000 is listed with detailed information.
If you use this table, you can purge duplicate comments, see HDF5 Purging of duplicate Comments.
results_trajectory
All results created not within single runs
derived_parameters_trajectory
derived_parameters_runs
derived_parameters_runs_summary
All three are analogous to the result overviews above
The explored_parameters overview table showing the explored parameter ranges
In each subtree results.run_XXXXXXXX there will be another explored parameter table summarizing the values in each run. Per default these tables are switched off, to enable it pass large_overview_tables=True to your environment.
HDF5 Purging of duplicate Comments¶
Adding a result with the same name and same comment in every single run, may create a lot of overhead. Since the very same comment would be stored in every node in the HDF5 file. To get rid of this overhead use the option purge_duplicate_comments=True and summary_tables=True.
For instance, during a single run you call traj.f_add_result('my_result', 42, comment='Mostly harmless!') and the result will be renamed to results.runs.run_00000000.my_result. After storage of the result into your HDF5 file, you will find the comment 'Mostly harmless!' in the corresponding HDF5 group node. If you call traj.f_add_result('my_result',-55, comment='Mostly harmless!') in another run again, let’s say run_00000001, the name will be mapped to results.runs.run_00000001.my_result. But this time the comment will not be saved to disk, since 'Mostly harmless!' is already part of the very first result with the name ‘my_result’.
Furthermore, if you reload your data from the example above, the result instance results.runs.run_00000001.my_result won’t have a comment only the instance results.runs.run_00000000.my_result.
IMPORTANT: If you use multiprocessing, the storage service will take care that the comment for the result or derived parameter with the lowest run index will be considered, regardless of the order of the finishing of your runs. Note that this only works properly if all comments are the same. Otherwise the comment in the overview table might not be the one with the lowest run index. Moreover, if you merge trajectories (see ref:more-on-merging) there is no support for purging comments in the other trajectory. All comments of the other trajectory’s results and derived parameters will be kept and merged into your current one.
IMPORTANT Purging of duplicate comments requires overview tables. Since there are no overview tables for group nodes, this feature does not work for comments in group nodes. So try to avoid to adding the same comments over and over again in group nodes within single runs.
Running an Experiment¶
In order to run an experiment, you need to define a job or a top level function that specifies your simulation. This function gets as first positional argument the: Trajectory container (see More on Trajectories), and optionally other positional and keyword arguments of your choice.
def myjobfunc(traj, *args, **kwargs)
#Do some sophisticated simulations with your trajectory
...
return 'fortytwo'
In order to run this simulation, you need to hand over the function to the environment. You can also specify the additional arguments and keyword arguments using f_run():
env.f_run(myjobfunc, *args, **kwargs)
The argument list args and keyword dictionary kwargs are directly handed over to the myjobfunc during runtime.
The f_run() will return a list of tuples. Whereas the first tuple entry is the index of the corresponding run and the second entry of the tuple is the result returned by your run function (for the example above this would simply always be the string 'fortytwo', i.e. ((0, 'fortytwo'), (1, 'fortytwo'),...)). In case you use multiprocessing these tuples are not in the order of the run indices but in the order of their finishing time!
Adding Post-Processing¶
You can add a post-processing function that is called after the execution of all the single runs via f_add_postprocessing().
Your post processing function must accept the trajectory container as the first argument, a list of tuples (containing the run indices and results), and arbitrary positional and keyword arguments. In order to pass arbitrary arguments to your post-processing function, simply pass these first to f_add_postprocessing().
For example:
def mypostprocfunc(traj, result_list, extra_arg1, extra_arg2):
# do some postprocessing here
...
Whereas in your main script you can call
env.f_add_postproc(mypostprocfunc, 42, extra_arg2=42.5)
which will later on pass 42 as extra_arg1 and 42.4 as extra_arg2. It is the very same principle as before for your run function. The post-processing function will be called after the completion of all single runs.
Moreover, please note that your trajectory usually does not contain the data computed during the single runs, since this has been removed after the single runs to save RAM. If your post-processing needs access to this data, you can simply load it via one of the many loading functions (f_load_child(), f_load_item(), f_load()) or even turn on Automatic Loading.
Note that your post-processing function should not return any results, since these will simply be lost. However, there is one particular result that can be returned, see below.
Expanding your Trajectory via Post-Processing¶
If your post-processing function expands the trajectory via f_expand() or if your post-processing function returns a dictionary of lists that can be interpreted to expand the trajectory, pypet will start the single runs again and explore the expanded trajectory. Of course, after this expanded exploration, your post-processing function will be called again. Likewise, you could potentially expand again, and after the next expansion post-processing will be executed again (and again, and again, and again, I guess you get it). Thus, you can use post-processing for an adaptive search within your parameter space.
IMPORTANT: All changes you apply to your trajectory, like setting auto-loading or changing fast access, are propagated to the new single runs. So try to undo all changes before finishing the post-processing if you plan to trigger new single runs.
Expanding your Trajectory and using Multiprocessing¶
If you use multiprocessing and you want to adaptively expand your trajectory, it can be a waste of precious time to wait until all runs have finished. Accordingly, you can set the argument immediate_postproc to True when you create your environment. Then your post-processing function is called as soon as pypet runs out of jobs for single runs. Thus, you can expand your trajectory while the last batch of single runs is still being executed.
To emphasize this a bit more and to not be misunderstood: Your post-processing function is not called as soon as a single run finishes and the first result is available but as soon as there are no more single runs available to start new processes. Still, that does not mean you have to wait until all single runs are finished (as for normal post-processing), but you can already add new single runs to the trajectory while the final n runs are still being executed. Where n is determined by the number of cores (ncores) and probably the cap values you have chosen (see Multiprocessing).
pypet will not start a new process for your post-processing. Your post-processing function is executed in the main process (this makes writing actual post-processing functions much easier because you don’t have to wrap your head around dead-locks). Accordingly, post-processing should be rather quick in comparison to your single runs, otherwise post-processing will become the bottleneck in your parallel simulations.
Using a Experiment Pipeline¶
Your numerical experiments usually work like the following: You add some parameters to your trajectory, you mark a few of these for exploration, and you pass your main function to the environment via f_run(). Accordingly, this function will be executed with all parameter combinations. Maybe you want some post-processing in the end and that’s about it. However, sometimes even the addition of parameters can be fairly complex. Thus, you want this part under the supervision of an environment, too. For instance, because you have a Sumatra lab-book and adding of parameters should also account as runtime. Thus, to have your entire experiment and not only the exploration of the parameter space managed by pypet you can use the f_pipeline() function, see also Post-Processing and Pipelining (from the Tutorial).
You have to pass a so called pipeline function to f_pipeline() that defines your entire experiment. Accordingly, your pipeline function is only allowed to take a single parameter, that is the trajectory container. Next, your pipeline function can fill in some parameters and do some pre-processing. Afterwards your pipeline function needs to return the run function, the corresponding arguments and potentially a post-processing function with arguments. To be more precise your pipeline function needs to return two tuples with at most 3 entries each, for example:
def myjobfunc(traj, extra_arg1, extra_arg2, extra_arg3)
# do some sophisticated simulation stuff
solve_p_equals_np(traj, extra_arg1)
disproof_spock(traj, extra_arg2, extra_arg3)
...
def mypostproc(traj, postproc_arg1, postproc_arg2, postproc_arg3)
# do some analysis here
...
exploration_dict={'ncards' : [100, 200]}
if maybe_i_should_explore_more_cards:
return exploration_dict
else
return None
def mypipeline(traj):
# add some parameters
traj.f_add_parameter('poker.ncards', 7, comment='Usually we play 7-card-stud')
...
# Explore the trajectory
traj.f_explore({'ncards': range(42)})
# Finally return the tuples
args = (myarg1, myarg2) # myargX can be anything form ints to strings to complex objects
kwargs = {'extra_arg3': myarg3}
postproc_args = (some_other_arg1,) # Check out the comma here! Important to make it a tuple
postproc_kwargs = {'postproc_arg2' : some_other_arg2,
'postproc_arg3' : some_other_arg3}
return (myjobfunc, args, kwargs), (mypostproc, postproc_args, postproc_kwargs)
The first entry of the first tuple is you run or top-level execution function, followed by a list or tuple defining the positional arguments and, thirdly, a dictionary defining the keyword arguments. The second tuple has to contain the post-processing function and positional arguments and keyword arguments. If you do not have any positional arguments pass an empty tuple (), if you do not have any keyword arguments pass an empty dictionary {}.
If you do not need postprocessing at all, your pipeline function can simply return the run function followed by the positional and keyword arguments:
def mypipeline(traj):
#...
return myjobfunc, args, kwargs
Continuing or Resuming a Crashed Experiment¶
In order to use this feature you need dill. Careful, dill is rather experimental and still in alpha status!
If all of your data can be handled by dill, you can use the config parameter continuable=True passed to the Environment constructor. This will create a continue directory (name specified by you via continue_folder) and a sub-folder with the name of the trajectory. This folder is your safety net for data loss due to a computer crash. If for whatever reason your day or week-long lasting simulation was interrupted, you can resume it without recomputing already obtained results. Note that this works only if the HDF5 file is not corrupted and for interruptions due to computer crashes, like power failure etc. If your simulations crashed due to errors in your code, there is no way to restore that!
You can resume a crashed trajectory via f_continue() with the name of the continue folder (not the subfolder) and the name of the trajectory:
env = Environment(continuable=True)
env.f_continue(trajectory_name='my_traj_2015_10_21_04h29m00s',
continue_folder='./experiments/continue/')
The neat thing here is, that you create a novel environment for the continuation. Accordingly, you can set different environmental settings, like changing the number of cores, etc. You cannot change any HDF5 settings or even change the whole storage service.
When does continuing not work?
Continuing will not work if your top-level simulation function or the arguments passed to your simulation function are altered between individual runs. For instance, if you use multiprocessing and you want to write computed data into a shared data list (like multiprocessing.Manager().list(), see Sharing Data during Multiprocessing), these changes will be lost and cannot be captured by the continue snapshots.
A work around here would be to not manipulate the arguments but pass these values as results of your top-level simulation function. Everything that is returned by your top-level function will be part of the snapshots and can be reconstructed after a crash.
Continuing might not work if you use post-processing that expands the trajectory. Since you are not limited in how you manipulate the trajectory within your post-processing, there are potentially many side effects that remain undetected by the continue snapshots. You can try to use both together, but there is no guarantee whatsoever that continuing a crashed trajectory and post-processing with expanding will work together.