paraproc.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-

# Copyright (c) 2018 herrlich10@gmail.com
#
# Permission is hereby granted, free of charge, to any person obtaining a copy 
# of this software and associated documentation files (the "Software"), to deal 
# in the Software without restriction, including without limitation the rights 
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 
# copies of the Software, and to permit persons to whom the Software is 
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included 
# in all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 
# SOFTWARE.

from __future__ import print_function, division, absolute_import, unicode_literals
import sys, os, shlex, time, textwrap, re, warnings
import subprocess, multiprocessing, queue, threading, ctypes, uuid
import numpy as np

__author__ = 'herrlich10 <herrlich10@gmail.com>'
__version__ = '0.1.7'

# The following are copied from six
# =================================
if sys.version_info[0] == 3:
    string_types = (str,)
    from io import StringIO
else:
    string_types = (basestring,)
    from StringIO import StringIO

def add_metaclass(metaclass):
    """Class decorator for creating a class with a metaclass."""
    def wrapper(cls):
        orig_vars = cls.__dict__.copy()
        slots = orig_vars.get('__slots__')
        if slots is not None:
            if isinstance(slots, str):
                slots = [slots]
            for slots_var in slots:
                orig_vars.pop(slots_var)
        orig_vars.pop('__dict__', None)
        orig_vars.pop('__weakref__', None)
        return metaclass(cls.__name__, cls.__bases__, orig_vars)
    return wrapper
# =================================


def format_duration(duration, format='standard'):
    '''Format duration (in seconds) in a more human friendly way.
    '''
    if format == 'short':
        units = ['d', 'h', 'm', 's']
    elif format == 'long':
        units = [' days', ' hours', ' minutes', ' seconds']
    else: # Assume 'standard'
        units = [' day', ' hr', ' min', ' sec']
    values = [int(duration//86400), int(duration%86400//3600), int(duration%3600//60), duration%60]
    for K in range(len(values)): # values[K] would be the first non-zero value
        if values[K] > 0:
            break
    formatted = ((('%d' if k<len(values)-1 else '%.3f') % values[k]) + units[k] for k in range(len(values)) if k >= K)
    return ' '.join(formatted)


def cmd_for_exec(cmd, shell=False):
    ''' Format cmd appropriately for execution according to whether shell=True.

    Split a cmd string into a list, if shell=False.
    Join a cmd list into a string, if shell=True.
    Do nothing to callable.

    Parameters
    ----------
    cmd : str, list, or callable
    shell : bool
    '''
    # If shell=kwargs, its true value is inferred.
    if isinstance(shell, dict):
        shell = ('shell' in shell and shell['shell'])
    if not callable(cmd):
        if shell: # cmd string is required
            if not isinstance(cmd, string_types):
                cmd = ' '.join(cmd)
        else: # cmd list is required
            if isinstance(cmd, string_types):
                cmd = shlex.split(cmd) # Split by space, preserving quoted substrings
    return cmd


def cmd_for_disp(cmd):
    '''Format cmd for printing.
    
    Parameters
    ----------
    cmd : str, list, or callable
    '''
    if not callable(cmd):
        if isinstance(cmd, string_types):
            cmd = shlex.split(cmd) # Remove insignificant whitespaces
        cmd = ' '.join(shlex.quote(s) for s in cmd)
    return cmd


ERROR_PATTERN = r'error|^\*{2}\s'


def check_output_for_errors(output, error_pattern=None, error_whitelist=None, verbose=1, label=''):
    '''
    User can skip error checking by setting error_pattern=''
    '''
    if error_pattern is None:
        error_pattern = ERROR_PATTERN
    n_errors = 0
    if error_pattern != '': # User can skip error checking by setting error_pattern=''
        if isinstance(error_pattern, string_types): # User can provide compiled regex if case sensitivity is desired
            error_pattern = re.compile(error_pattern, re.IGNORECASE)
        if isinstance(error_whitelist, string_types):
            error_whitelist = re.compile(error_whitelist, re.IGNORECASE)
        for line in output:
            if error_pattern.search(line) and (error_whitelist is None or not error_whitelist.search(line)):
                if verbose > 0:
                    print(label, line, end='')
                n_errors += 1
    return n_errors


def check_output_for_goal(output, goal_pattern=None):
    if goal_pattern is None:
        return True
    if isinstance(goal_pattern, string_types): # User can provide compiled regex if case sensitivity is desired
        goal_pattern = re.compile(goal_pattern, re.IGNORECASE)
    for line in output:
        if goal_pattern.search(line):
            return True
    return False


def run(cmd, check=True, error_pattern=None, error_whitelist=None, goal_pattern=None, shell=False, verbose=2):
    '''Run an external command line.
    
    This function is similar to subprocess.run introduced in Python 3.5, but
    provides a slightly simpler and perhaps more convenient API.

    Parameters
    ----------
    cmd : str or list
    '''
    cmd = cmd_for_exec(cmd, shell=shell)
    cmd_str = cmd_for_disp(cmd)
    if verbose > 0:
        print('>>', cmd_str)
    p = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, shell=shell)
    res = {'cmd': cmd_str, 'pid': p.pid, 'output': [], 'start_time': time.time()}
    for line in iter(p.stdout.readline, b''): # The 2nd argument is sentinel character (there will be no ending empty line)
        res['output'].append(line.decode('utf-8'))
        if verbose > 1:
            print(res['output'][-1], end='')
    p.stdout.close() # Notify the child process that the PIPE has been broken
    res['returncode'] = p.wait()
    res['stop_time'] = time.time()
    if verbose > 0:
        print('>> Command finished in {0}.'.format(format_duration(res['stop_time'] - res['start_time'])))
    if check and (res['returncode'] or check_output_for_errors(res['output'], error_pattern=error_pattern, 
                                           error_whitelist=error_whitelist, verbose=verbose)):
        print('>> Please pay attention to the above errors.')
        raise RuntimeError(f'Error occurs when executing the following command (returncode={p.returncode}):\n{cmd_str}')
    if check and not check_output_for_goal(res['output'], goal_pattern=goal_pattern):
        raise RuntimeError(f'Expected goal pattern "{goal_pattern}" does not found! Something must be wrong!')
    return res


STDOUT = sys.stdout
STDERR = sys.stderr

class TeeOut(StringIO):
    def __init__(self, err=False, tee=True):
        super().__init__()
        self.err = err
        self.tee = tee

    def write(self, s):
        super().write(s)
        if self.err: # Always output error message
            STDERR.write(s)
        elif self.tee:
            STDOUT.write(s)


class PooledCaller(object):
    '''
    Execute multiple command line programs, as well as python callables, 
    asynchronously and parallelly across a pool of processes.
    '''
    def __init__(self, pool_size=None, verbose=1):
        self.ctx = multiprocessing.get_context('fork')
        if pool_size is None:
            # self.pool_size = multiprocessing.cpu_count() * 3 // 4
            self.pool_size = self.ctx.cpu_count() * 3 // 4
        else:
            self.pool_size = pool_size
        self.verbose = verbose
        self.ps = []
        self.cmd_queue = [] # Queue for commands and callables, as well as any additional args
        self._n_cmds = 0 # Auto increased counter for generating cmd idx
        self._idx2pid = {}
        self._pid2job = {} # Hold all jobs for each wait()
        self._log = [] # Hold all jobs across waits (entire execution history for this PooledCaller instance)
        self._fulfilled = {} # Fulfilled dependencies across waits (a faster API compared with self._log)
        # self.res_queue = multiprocessing.Queue() # Queue for return values of executed python callables
        self.res_queue = self.ctx.Queue() # Queue for return values of executed python callables
 
    def run(self, cmd, *args, _depends=None, _retry=None, _dispatch=False, _error_pattern=None, _error_whitelist=None, _suppress_warning=False, _block=False, **kwargs):
        '''Asynchronously run command or callable (queued execution, return immediately).
        
        See subprocess.Popen() for more information about the arguments.

        Multiple commands can be separated with ";" and executed sequentially 
        within a single subprocess in linux/mac, only if shell=True.
        
        Python callable can also be executed in parallel via multiprocessing.
        Note that although return values of the callable are retrieved via PIPE,
        sometimes it could be advantageous to directly save the computation 
        results into a shared file (e.g., an HDF5 file), esp. when they're large.
        In the later case, a proper lock mechanism via multiprocessing.Lock() 
        is required.

        Parameters
        ----------
        cmd : list, str, or callable
            Computation in command line programs is handled with subprocess.
            Computation in python callable is handled with multiprocessing.
        shell : bool
            If provided, must be a keyword argument.
            If shell is True, the command will be executed through the shell.
        *args : 
            If cmd is a callable, `*args` are passed to the callable as its arguments.
        **kwargs : 
            If cmd is a callable, `**kwargs` are passed to the callable as its keyword arguments.
            If cmd is a list or str, `**kwargs` are passed to subprocess.Popen().
        _depends : list
            A list of jobs (identified by their uuid) that have to be done 
            before this job can be scheduled.
        _retry: int
            Number of retry before accepting failure (if detecting non-zero return code).
        _dispatch : bool
            Dispatch the job immediately, which will run in the background without blocking.
        _error_pattern : str
        _suppress_warning : bool
        _block : bool
            if True, call wait() internally and block.

        Returns
        -------
        _uuid : str
            The uuid of current job (which can be used as future jobs' dependency)
        '''
        cmd = cmd_for_exec(cmd, shell=kwargs)
        _uuid = uuid.uuid4().hex[:8]
        if _retry is None:
            _retry = 0
        self.cmd_queue.append((self._n_cmds, cmd, args, kwargs, _uuid, _depends, _retry, 
            _error_pattern, _error_whitelist, _suppress_warning))
        self._n_cmds += 1 # Accumulate by each call to run(), and reset after wait()
        if _dispatch:
            self.dispatch()
        if _block:
            self.wait()
        return _uuid

    def run1(self, cmd, *args, _error_pattern=None, _error_whitelist=None, _suppress_warning=False, **kwargs):
        self.run(cmd, *args, _error_pattern=_error_pattern, _error_whitelist=_error_whitelist, 
            _suppress_warning=_suppress_warning, **kwargs)
        return self.wait()

    def _callable_wrapper(self, idx, cmd, *args, **kwargs):
        out = TeeOut(tee=(self.verbose > 1))
        err = TeeOut(err=True)
        sys.stdout = out # This substitution only affect spawned process
        sys.stderr = err
        res = None # Initialized in case of exception
        try:
            res = cmd(*args, **kwargs)
        except Exception as e:
            print('>> Error occurs in job#{0}'.format(idx), file=err)
            print('** ERROR:', e, file=err) # AFNI style error message
            raise e # Re-raise and let parent process to handle it
        finally:
            # Grab all output at the very end of the process (assume that there aren't too much of them)
            # TODO: This could be a potential bug...
            # https://ryanjoneil.github.io/posts/2014-02-14-capturing-stdout-in-a-python-child-process.html
            output = out.getvalue().splitlines(True) + err.getvalue().splitlines(True)
            self.res_queue.put([idx, res, output]) # Communicate return value and output (Caution: The underlying pipe has limited size. Have to get() soon in wait().)

    def _async_reader(self, idx, f, output_list, speed_up, suppress_warning=False):
        while True: # We can use event to tell the thread to stop prematurely, as demonstrated in https://stackoverflow.com/questions/323972/is-there-any-way-to-kill-a-thread
            line = f.readline()
            line = line.decode('utf-8')
            if line: # This is not lock protected, because only one thread (i.e., this thread) is going to write
                output_list.append(line)
                if (line.startswith('*') or line.startswith('\x1b[7m')) and not suppress_warning: # Always print AFNI style WARNING and ERROR through stderr unless explicitly suppressed
                    # '\x1b[7m' and '\x1b[0m' are 'reverse' and 'reset' respectively (https://gist.github.com/abritinthebay/d80eb99b2726c83feb0d97eab95206c4)
                    print('>> Something happens in job#{0}'.format(idx), file=sys.stderr)
                    print(line, end='', file=sys.stderr)
                elif self.verbose > 1:
                    print(line, end='')
            else: # Empty line signifies the end of the spawned process
                break
            if not speed_up.is_set():
                time.sleep(0.1) # Don't need to poll for output too aggressively during run time

    def dispatch(self):
        # If there are free slot and more jobs
        # while len(self.ps) < self.pool_size and len(self.cmd_queue) > 0:
        if len(self.ps) < self.pool_size and len(self.cmd_queue) > 0:
            idx, cmd, args, kwargs, _uuid, _depends, _retry, _error_pattern, _error_whitelist, _suppress_warning = self.cmd_queue.pop(0)
            if _depends is None or all([dep in self._fulfilled for dep in _depends]): # No dependency or all fulfilled
                # Create a job process only after it is popped from the queue
                job = {'idx': idx, 'cmd': cmd, 'args': args, 'kwargs': kwargs, 'uuid':  _uuid, 
                    'depends': _depends, 'retry': _retry, 'error_pattern': _error_pattern, 'error_whitelist': _error_whitelist, 
                    'suppress_warning': _suppress_warning, 'output': []} 
                if self.verbose > 0:
                    print('>> job#{0}: {1}'.format(idx, cmd_for_disp(job['cmd'])))
                if callable(cmd):
                    # TODO: Add an if-else branch here if shared memory doesn't work for wrapper 
                    # p = multiprocessing.Process(target=self._callable_wrapper, args=(idx, cmd) + args, kwargs=kwargs)
                    p = self.ctx.Process(target=self._callable_wrapper, args=(idx, cmd) + args, kwargs=kwargs)
                    p.start()
                else:
                    # Use PIPE to capture output and error message
                    p = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, **kwargs)
                    # Capture output without blocking (the main thread) by using a separate thread to do the blocking readline()
                    job['speed_up'] = threading.Event()
                    job['watcher'] = threading.Thread(target=self._async_reader, args=(idx, p.stdout, 
                        job['output'], job['speed_up'], job['suppress_warning']), daemon=True)
                    job['watcher'].start()
                self.ps.append(p)
                job['start_time'] = time.time()
                job['pid'] = p.pid
                job['successor'] = None
                job['log_idx'] = len(self._log)
                self._idx2pid[idx] = p.pid
                self._pid2job[p.pid] = job
                self._log.append(job)
            else: # Re-queue the job whose dependencies are not fully fulfilled to the END of the queue
                self.cmd_queue.append((idx, cmd, args, kwargs, _uuid, _depends, _retry, _error_pattern, _error_whitelist, _suppress_warning))

    def _async_get_res(self, res_list):
        try:
            res = self.res_queue.get(block=False) # idx, return_value, output
        except queue.Empty:
            pass
        else:
            res_list.append(res[:2])
            if len(res) > 2: # For callable only
                job = self._pid2job[self._idx2pid[res[0]]]
                job['output'] = res[2]

    def wait(self, pool_size=None, return_codes=False, return_jobs=False, raise_when_failed=True):
        '''
        Wait for all jobs in the queue to finish.
        
        Returns
        -------
        return_values : list
            Return values of executed python callable. Always `None` for command.
        codes : list (only when return_codes=True)
            The return code of the child process for each job.
        jobs : list (only when return_jobs=True)
            Detailed information about each child process, including captured stdout and stderr.
        '''
        if isinstance(pool_size, string_types) and pool_size == 'balanced':
            # Make sure each volley has roughly equal number of jobs
            n = len(self.cmd_queue)
            pool_size = int(np.ceil(n/np.ceil(n/self.pool_size)))
        if pool_size is not None:
            # Allow temporally adjust pool_size for current batch of jobs
            old_size = self.pool_size
            self.pool_size = pool_size
        start_time = time.time()
        ress = []
        while len(self.ps) > 0 or len(self.cmd_queue) > 0:
            # Dispatch jobs if possible
            self.dispatch()
            # Poll workers' state
            for p in self.ps:
                job = self._pid2job[p.pid]
                if isinstance(p, subprocess.Popen):
                    if p.poll() is not None: # If the process is terminated
                        job['stop_time'] = time.time()
                        job['returncode'] = p.returncode
                        job['speed_up'].set()
                        job['watcher'].join() # Retrieve all remaining output before closing PIPE
                        p.stdout.close() # Notify the child process that the PIPE has been broken
                        self.ps.remove(p)
                        if self.verbose > 0:
                            print('>> job#{0} finished (return {1}) in {2}.'.format(job['idx'], job['returncode'], format_duration(job['stop_time']-job['start_time'])))
                        if job['returncode'] != 0: # Failed
                            if job['retry'] > 0: # Need retry
                                # Insert a new cmd (as if we automatically run it again)
                                self.cmd_queue.append((self._n_cmds, job['cmd'], job['args'], job['kwargs'], job['uuid'], 
                                    job['depends'], job['retry']-1, job['error_pattern'], job['suppress_warning']))
                                job['successor'] = self._n_cmds
                                self._n_cmds += 1
                            else: # No more retry, accept failure...
                                msg = f">> job#{job['idx']} failed!\n Full output:\n {''.join(job['output'])}"
                                if raise_when_failed:
                                    raise RuntimeError(msg)
                                else:
                                    warnings.warn(msg, RuntimeWarning)
                        else: # Successful
                            self.res_queue.put([job['idx'], None]) # Return None to mimic callable behavior
                            self._fulfilled[job['uuid']] = job['log_idx'] # Marked as fulfilled, even with error (TODO: or shall I break all??)
                        # These helper objects may not be useful for the end users
                        for key in ['watcher', 'speed_up', 'args', 'kwargs']:
                            job.pop(key) 
                    else:
                        pass
                # elif isinstance(p, multiprocessing.Process):
                elif isinstance(p, self.ctx.Process):
                    if not p.is_alive(): # If the process is terminated
                        job['stop_time'] = time.time()
                        job['returncode'] = p.exitcode # subprocess.Popen and multiprocessing.Process use different names for this
                        self.ps.remove(p)
                        if self.verbose > 0:
                            print('>> job#{0} finished (return {1}) in {2}.'.format(job['idx'], job['returncode'], format_duration(job['stop_time']-job['start_time'])))
                        # TODO: retry mechanism for callable
                        self._fulfilled[job['uuid']] = job['log_idx'] # Marked as fulfilled
                        # Remove potentially very large data
                        for key in ['args', 'kwargs']:
                            job.pop(key) 
                    else:
                        pass
            time.sleep(0.1)
            # Dequeuing, see https://stackoverflow.com/questions/10028809/maximum-size-for-multiprocessing-queue-item
            self._async_get_res(ress)
        # Handle return values by callable cmd
        while not self.res_queue.empty():
            self._async_get_res(ress)
        ress = [res[1] for res in sorted(ress, key=lambda res: res[0])]
        # Handle return codes by children processes
        jobs = sorted([job for job in self._pid2job.values() if job['successor'] is None], key=lambda job: job['idx'])
        codes = [job['returncode'] for job in jobs]
        if self.verbose > 0:
            duration = time.time() - start_time
            print('>> All {0} jobs done in {1}.'.format(self._n_cmds, format_duration(duration)))
            if np.any(codes):
                print('returncodes: {0}'.format(codes))
                first_error = np.nonzero(codes)[0][0]
                print(f">> Output for job#{first_error} was as follows:\n------------------------------")
                print(jobs[first_error]['output'])
            else:
                print('all returncodes are 0.')
            if self.all_successful(jobs=jobs):
                print('>> All {0} jobs finished successfully.'.format(len(jobs)))
            else:
                print('>> Please pay attention to the above errors.')
        # Reset object states
        self._n_cmds = 0
        self._idx2pid = {}
        self._pid2job = {}
        if pool_size is not None:
            self.pool_size = old_size
        res = (ress,) + ((codes,) if return_codes else ()) + ((jobs,) if return_jobs else ())
        if len(res) == 1:
            return res[0]
        else:
            return res

    def all_successful(self, jobs=None, verbose=None):
        if jobs is None:
            jobs = self._log
        if verbose is None:
            verbose = self.verbose
        # Check return codes
        all_zero = not np.any([job['returncode'] for job in jobs])
        # Check output
        n_errors = sum([check_output_for_errors(job['output'], error_pattern=job['error_pattern'], 
            error_whitelist=job['error_whitelist'], verbose=verbose, label='[job#{0}]'.format(job['idx'])) 
            for job in jobs])
        return all_zero and n_errors == 0

    def idss(self, total, batch_size=None):
        if batch_size is None:
            batch_size = int(np.ceil(total / self.pool_size / 10))
        return (range(k, min(k+batch_size, total)) for k in range(0, total, batch_size))

    def __call__(self, job_generator, **kwargs):
        # This is similar to the joblib.Parallel signature, which is the only way to
        # pass both args and kwargs for inner execution.
        # >>> pc(pc.run(f"3dvolreg -prefix ... {func}{run}.nii") for run in runs)
        # 
        # It also allows each call to deal with a batch of jobs for better performance, 
        # if the callable is purposely designed to do so, which is especially useful 
        # when there are a huge amount of small jobs.
        # >>> pc(pc.run(compute_depth, ids, *args) for ids in pc.idss(len(depths)))
        n_jobs = 0
        for _ in job_generator: # Queue all jobs from the generator
            n_jobs += 1
        if self.verbose > 0:
            print('>> Start with a total of {0} jobs...'.format(n_jobs))
        return self.wait(**kwargs) # Wait all jobs to finish


class ArrayWrapper(type):
    '''
    This is the metaclass for classes that wrap an np.ndarray and delegate 
    non-reimplemented operators (among other magic functions) to the wrapped array.
    '''
    def __init__(cls, name, bases, dct):
        def make_descriptor(name):
            '''
            Implementation notes
            --------------------
            1. Method (or non-data) descriptors are objects that define __get__() method
               but not __set__() method. Refer to [here](https://docs.python.org/3.6/howto/descriptor.html).
            2. The magic methods of an object (e.g., arr.__add__) are descriptors, not callable.
               So here we must return a property (with getter only), not a lambda.
            3. Strangely, the whole thing must be wrapped in a nested function. See [here](
               https://stackoverflow.com/questions/9057669/how-can-i-intercept-calls-to-pythons-magic-methods-in-new-style-classes).
            4. The wrapped array must be named self.arr
            '''
            return property(lambda self: getattr(self.arr, name))

        type.__init__(cls, name, bases, dct)
        ignore = 'class mro new init setattr getattr getattribute'
        ignore = set('__{0}__'.format(name) for name in ignore.split())
        for name in dir(np.ndarray):
            if name.startswith('__'):
                if name not in ignore and name not in dct:
                    setattr(cls, name, make_descriptor(name))


# TODO: 1. Use ctx instead of multiprocessing. 2. Use multiprocessing.shared_memory
@add_metaclass(ArrayWrapper) # Compatibility code from six
class SharedMemoryArray(object):
    '''
    This class can be used as a usual np.ndarray, but its data buffer
    is allocated in shared memory (under Cached Files in memory monitor), 
    and can be passed across processes without any data copy/duplication, 
    even when write access happens (which is lock-synchronized).

    The idea is to allocate memory using multiprocessing.Array, and  
    access it from current or another process via a numpy.ndarray view, 
    without actually copying the data.
    So it is both convenient and efficient when used with multiprocessing.

    This implementation also demonstrates the power of composition + metaclass,
    as opposed to the canonical multiple inheritance.
    '''
    def __init__(self, dtype, shape, initializer=None, lock=True):
        self.dtype = np.dtype(dtype)
        self.shape = shape
        if initializer is None:
            # Preallocate memory using multiprocessing is the preferred usage
            self.shared_arr = multiprocessing.Array(self.dtype2ctypes[self.dtype], int(np.prod(self.shape)), lock=lock)
        else:
            self.shared_arr = multiprocessing.Array(self.dtype2ctypes[self.dtype], initializer, lock=lock)
        if not lock:
            self.arr = np.frombuffer(self.shared_arr, dtype=self.dtype).reshape(self.shape)
        else:
            self.arr = np.frombuffer(self.shared_arr.get_obj(), dtype=self.dtype).reshape(self.shape)
 
    @classmethod
    def zeros(cls, shape, dtype=float, lock=True):
        '''
        Return a new array of given shape and dtype, filled with zeros.

        This is the preferred usage, which avoids holding two copies of the
        potentially very large data simultaneously in the memory.
        '''
        return cls(dtype, shape, lock=lock)

    @classmethod
    def from_array(cls, arr, lock=True):
        '''
        Initialize a new shared-memory array with an existing array.
        '''
        # return cls(arr.dtype, arr.shape, arr.ravel(), lock=lock) # Slow and memory inefficient, why?
        a = cls.zeros(arr.shape, dtype=arr.dtype, lock=lock)
        a[:] = arr # This is a more efficient way of initialization
        return a

    def __getattr__(self, attr):
        if attr in self._SHARED_ARR_ATTRIBUTES:
            return getattr(self.shared_arr, attr)
        else:
            return getattr(self.arr, attr)

    def __dir__(self):
        return list(self.__dict__.keys()) + self._SHARED_ARR_ATTRIBUTES + dir(self.arr)

    _SHARED_ARR_ATTRIBUTES = ['acquire', 'release', 'get_lock']

    # At present, only numerical dtypes are supported.
    dtype2ctypes = {
        bool: ctypes.c_bool,
        int: ctypes.c_long,
        float: ctypes.c_double,
        np.dtype('bool'): ctypes.c_bool,
        np.dtype('int64'): ctypes.c_long,
        np.dtype('int32'): ctypes.c_int,
        np.dtype('int16'): ctypes.c_short,
        np.dtype('int8'): ctypes.c_byte,
        np.dtype('uint64'): ctypes.c_ulong,
        np.dtype('uint32'): ctypes.c_uint,
        np.dtype('uint16'): ctypes.c_ushort,
        np.dtype('uint8'): ctypes.c_ubyte,
        np.dtype('float64'): ctypes.c_double,
        np.dtype('float32'): ctypes.c_float,
        }