def test_pool_with_memmap_array_view(tmpdir): """Check that subprocess can access and update shared memory array""" assert_array_equal = np.testing.assert_array_equal # Fork the subprocess before allocating the objects to be passed pool_temp_folder = tmpdir.mkdir('pool').strpath p = MemmapingPool(10, max_nbytes=2, temp_folder=pool_temp_folder) try: filename = tmpdir.join('test.mmap').strpath a = np.memmap(filename, dtype=np.float32, shape=(3, 5), mode='w+') a.fill(1.0) # Create an ndarray view on the memmap instance a_view = np.asarray(a) assert not isinstance(a_view, np.memmap) assert has_shareable_memory(a_view) p.map(inplace_double, [(a_view, (i, j), 1.0) for i in range(a.shape[0]) for j in range(a.shape[1])]) # Both a and the a_view have been updated assert_array_equal(a, 2 * np.ones(a.shape)) assert_array_equal(a_view, 2 * np.ones(a.shape)) # Passing memmap array view to the pool should not trigger the # creation of new files on the FS assert os.listdir(pool_temp_folder) == [] finally: p.terminate() del p
def test_memmaping_on_dev_shm(): """Check that MemmapingPool uses /dev/shm when possible""" p = MemmapingPool(3, max_nbytes=10) try: # Check that the pool has correctly detected the presence of the # shared memory filesystem. pool_temp_folder = p._temp_folder folder_prefix = '/dev/shm/joblib_memmaping_pool_' assert_true(pool_temp_folder.startswith(folder_prefix)) assert_true(os.path.exists(pool_temp_folder)) # Try with a file larger than the memmap threshold of 10 bytes a = np.ones(100, dtype=np.float64) assert_equal(a.nbytes, 800) p.map(id, [a] * 10) # a should have been memmaped to the pool temp folder: the joblib # pickling procedure generate a .pkl and a .npy file: assert_equal(len(os.listdir(pool_temp_folder)), 2) b = np.ones(100, dtype=np.float64) assert_equal(b.nbytes, 800) p.map(id, [b] * 10) # A copy of both a and b are not stored in the shared memory folder assert_equal(len(os.listdir(pool_temp_folder)), 4) finally: # Cleanup open file descriptors p.terminate() del p # The temp folder is cleaned up upon pool termination assert_false(os.path.exists(pool_temp_folder))
def test_memmaping_on_dev_shm(): """Check that MemmapingPool uses /dev/shm when possible""" p = MemmapingPool(3, max_nbytes=10) try: # Check that the pool has correctly detected the presence of the # shared memory filesystem. pool_temp_folder = p._temp_folder folder_prefix = '/dev/shm/joblib_memmaping_pool_' assert pool_temp_folder.startswith(folder_prefix) assert os.path.exists(pool_temp_folder) # Try with a file larger than the memmap threshold of 10 bytes a = np.ones(100, dtype=np.float64) assert a.nbytes == 800 p.map(id, [a] * 10) # a should have been memmaped to the pool temp folder: the joblib # pickling procedure generate one .pkl file: assert len(os.listdir(pool_temp_folder)) == 1 # create a new array with content that is different from 'a' so that # it is mapped to a different file in the temporary folder of the # pool. b = np.ones(100, dtype=np.float64) * 2 assert b.nbytes == 800 p.map(id, [b] * 10) # A copy of both a and b are now stored in the shared memory folder assert len(os.listdir(pool_temp_folder)) == 2 finally: # Cleanup open file descriptors p.terminate() del p # The temp folder is cleaned up upon pool termination assert not os.path.exists(pool_temp_folder)
def test_pool_memmap_with_big_offset(tmpdir): # Test that numpy memmap offset is set correctly if greater than # mmap.ALLOCATIONGRANULARITY, see # https://github.com/joblib/joblib/issues/451 and # https://github.com/numpy/numpy/pull/8443 for more details. fname = tmpdir.join('test.mmap').strpath size = 5 * mmap.ALLOCATIONGRANULARITY offset = mmap.ALLOCATIONGRANULARITY + 1 obj = make_memmap(fname, mode='w+', shape=size, dtype='uint8', offset=offset) p = MemmapingPool(2, temp_folder=tmpdir.strpath) result = p.apply_async(identity, args=(obj, )).get() assert isinstance(result, np.memmap) assert result.offset == offset np.testing.assert_array_equal(obj, result)
def test_memmaping_pool_for_large_arrays_in_return(): """Check that large arrays are not copied in memory in return""" assert_array_equal = np.testing.assert_array_equal # Build an array reducers that automaticaly dump large array content # but check that the returned datastructure are regular arrays to avoid # passing a memmap array pointing to a pool controlled temp folder that # might be confusing to the user # The MemmapingPool user can always return numpy.memmap object explicitly # to avoid memory copy p = MemmapingPool(3, max_nbytes=10, temp_folder=TEMP_FOLDER) try: res = p.apply_async(np.ones, args=(1000, )) large = res.get() assert_false(has_shareable_memory(large)) assert_array_equal(large, np.ones(1000)) finally: p.terminate() del p
def test_memmaping_pool_for_large_arrays_in_return(tmpdir): """Check that large arrays are not copied in memory in return""" assert_array_equal = np.testing.assert_array_equal # Build an array reducers that automaticaly dump large array content # but check that the returned datastructure are regular arrays to avoid # passing a memmap array pointing to a pool controlled temp folder that # might be confusing to the user # The MemmapingPool user can always return numpy.memmap object explicitly # to avoid memory copy p = MemmapingPool(3, max_nbytes=10, temp_folder=tmpdir.strpath) try: res = p.apply_async(np.ones, args=(1000,)) large = res.get() assert not has_shareable_memory(large) assert_array_equal(large, np.ones(1000)) finally: p.terminate() del p
def test_memmaping_pool_for_large_arrays_disabled(tmpdir): """Check that large arrays memmaping can be disabled""" # Set max_nbytes to None to disable the auto memmaping feature p = MemmapingPool(3, max_nbytes=None, temp_folder=tmpdir.strpath) try: # Check that the tempfolder is empty assert os.listdir(tmpdir.strpath) == [] # Try with a file largish than the memmap threshold of 40 bytes large = np.ones(100, dtype=np.float64) assert large.nbytes == 800 p.map(check_array, [(large, i, 1.0) for i in range(large.shape[0])]) # Check that the tempfolder is still empty assert os.listdir(tmpdir.strpath) == [] finally: # Cleanup open file descriptors p.terminate() del p
def test_pool_with_memmap_array_view(): """Check that subprocess can access and update shared memory array""" assert_array_equal = np.testing.assert_array_equal # Fork the subprocess before allocating the objects to be passed pool_temp_folder = os.path.join(TEMP_FOLDER, 'pool') os.makedirs(pool_temp_folder) p = MemmapingPool(10, max_nbytes=2, temp_folder=pool_temp_folder) try: filename = os.path.join(TEMP_FOLDER, 'test.mmap') a = np.memmap(filename, dtype=np.float32, shape=(3, 5), mode='w+') a.fill(1.0) # Create an ndarray view on the memmap instance a_view = np.asarray(a) assert_false(isinstance(a_view, np.memmap)) assert_true(has_shareable_memory(a_view)) p.map(inplace_double, [(a_view, (i, j), 1.0) for i in range(a.shape[0]) for j in range(a.shape[1])]) # Both a and the a_view have been updated assert_array_equal(a, 2 * np.ones(a.shape)) assert_array_equal(a_view, 2 * np.ones(a.shape)) # Passing memmap array view to the pool should not trigger the # creation of new files on the FS assert_equal(os.listdir(pool_temp_folder), []) finally: p.terminate() del p
def test_pool_with_memmap(): """Check that subprocess can access and update shared memory memmap""" assert_array_equal = np.testing.assert_array_equal # Fork the subprocess before allocating the objects to be passed pool_temp_folder = os.path.join(TEMP_FOLDER, 'pool') os.makedirs(pool_temp_folder) p = MemmapingPool(10, max_nbytes=2, temp_folder=pool_temp_folder) try: filename = os.path.join(TEMP_FOLDER, 'test.mmap') a = np.memmap(filename, dtype=np.float32, shape=(3, 5), mode='w+') a.fill(1.0) p.map(inplace_double, [(a, (i, j), 1.0) for i in range(a.shape[0]) for j in range(a.shape[1])]) assert_array_equal(a, 2 * np.ones(a.shape)) # Open a copy-on-write view on the previous data b = np.memmap(filename, dtype=np.float32, shape=(5, 3), mode='c') p.map(inplace_double, [(b, (i, j), 2.0) for i in range(b.shape[0]) for j in range(b.shape[1])]) # Passing memmap instances to the pool should not trigger the creation # of new files on the FS assert_equal(os.listdir(pool_temp_folder), []) # the original data is untouched assert_array_equal(a, 2 * np.ones(a.shape)) assert_array_equal(b, 2 * np.ones(b.shape)) # readonly maps can be read but not updated c = np.memmap(filename, dtype=np.float32, shape=(10, ), mode='r', offset=5 * 4) assert_raises(AssertionError, p.map, check_array, [(c, i, 3.0) for i in range(c.shape[0])]) # depending on the version of numpy one can either get a RuntimeError # or a ValueError assert_raises((RuntimeError, ValueError), p.map, inplace_double, [(c, i, 2.0) for i in range(c.shape[0])]) finally: # Clean all filehandlers held by the pool p.terminate() del p
def get_score(data, labels, fold_pairs, name, model, param): """ Function to get score for a classifier. Parameters ---------- data: array_like Data from which to derive score. labels: array_like or list Corresponding labels for each sample. fold_pairs: list of pairs of array_like A list of train/test indicies for each fold dhjelm(Why can't we just use the KFold object?) name: str Name of classifier. model: WRITEME param: WRITEME Parameters for the classifier. Returns ------- classifier: WRITEME fScore: WRITEME """ assert isinstance(name, str) logger.info("Classifying %s" % name) ksplit = len(fold_pairs) if name not in NAMES: raise ValueError("Classifier %s not supported. " "Did you enter it properly?" % name) # Redefine the parameters to be used for RBF SVM (dependent on # training data) if True: #better identifier here logger.info("Attempting to use grid search...") fScore = [] for i, fold_pair in enumerate(fold_pairs): print("Classifying a %s the %d-th out of %d folds..." % (name, i + 1, len(fold_pairs))) classifier = get_classifier(name, model, param, data[fold_pair[0], :]) area = classify(data, labels, fold_pair, classifier) fScore.append(area) else: logger.warn("Multiprocessing splits not tested yet.") pool = Pool(processes=min(ksplit, PROCESSORS)) classify_func = lambda f: classify( data, labels, fold_pairs[f], classifier=get_classifier( name, model, param, data=data[fold_pairs[f][0], :])) fScore = pool.map(functools.partial(classify_func, xrange(ksplit))) pool.close() pool.join() return classifier, fScore
def test_memmaping_pool_for_large_arrays(): """Check that large arrays are not copied in memory""" assert_array_equal = np.testing.assert_array_equal # Check that the tempfolder is empty assert_equal(os.listdir(TEMP_FOLDER), []) # Build an array reducers that automaticaly dump large array content # to filesystem backed memmap instances to avoid memory explosion p = MemmapingPool(3, max_nbytes=40, temp_folder=TEMP_FOLDER) try: # The tempory folder for the pool is not provisioned in advance assert_equal(os.listdir(TEMP_FOLDER), []) assert_false(os.path.exists(p._temp_folder)) small = np.ones(5, dtype=np.float32) assert_equal(small.nbytes, 20) p.map(check_array, [(small, i, 1.0) for i in range(small.shape[0])]) # Memory has been copied, the pool filesystem folder is unused assert_equal(os.listdir(TEMP_FOLDER), []) # Try with a file larger than the memmap threshold of 40 bytes large = np.ones(100, dtype=np.float64) assert_equal(large.nbytes, 800) p.map(check_array, [(large, i, 1.0) for i in range(large.shape[0])]) # The data has been dumped in a temp folder for subprocess to share it # without per-child memory copies assert_true(os.path.isdir(p._temp_folder)) dumped_filenames = os.listdir(p._temp_folder) assert_equal(len(dumped_filenames), 2) # Check that memmory mapping is not triggered for arrays with # dtype='object' objects = np.array(['abc'] * 100, dtype='object') results = p.map(has_shareable_memory, [objects]) assert_false(results[0]) finally: # check FS garbage upon pool termination p.terminate() assert_false(os.path.exists(p._temp_folder)) del p
def test_workaround_against_bad_memmap_with_copied_buffers(): """Check that memmaps with a bad buffer are returned as regular arrays Unary operations and ufuncs on memmap instances return a new memmap instance with an in-memory buffer (probably a numpy bug). """ assert_array_equal = np.testing.assert_array_equal p = MemmapingPool(3, max_nbytes=10, temp_folder=TEMP_FOLDER) try: # Send a complex, large-ish view on a array that will be converted to # a memmap in the worker process a = np.asarray(np.arange(6000).reshape((1000, 2, 3)), order='F')[:, :1, :] # Call a non-inplace multiply operation on the worker and memmap and # send it back to the parent. b = p.apply_async(_worker_multiply, args=(a, 3)).get() assert_false(has_shareable_memory(b)) assert_array_equal(b, 3 * a) finally: p.terminate() del p
def test_workaround_against_bad_memmap_with_copied_buffers(tmpdir): """Check that memmaps with a bad buffer are returned as regular arrays Unary operations and ufuncs on memmap instances return a new memmap instance with an in-memory buffer (probably a numpy bug). """ assert_array_equal = np.testing.assert_array_equal p = MemmapingPool(3, max_nbytes=10, temp_folder=tmpdir.strpath) try: # Send a complex, large-ish view on a array that will be converted to # a memmap in the worker process a = np.asarray(np.arange(6000).reshape((1000, 2, 3)), order='F')[:, :1, :] # Call a non-inplace multiply operation on the worker and memmap and # send it back to the parent. b = p.apply_async(_worker_multiply, args=(a, 3)).get() assert not has_shareable_memory(b) assert_array_equal(b, 3 * a) finally: p.terminate() del p
def get_score(data, labels, fold_pairs, name, model, param): """ Function to get score for a classifier. Parameters ---------- data: array-like Data from which to derive score. labels: array-like or list. Corresponding labels for each sample. fold_pairs: list of pairs of array-like A list of train/test indicies for each fold (Why can't we just use the KFold object?) name: string Name of classifier. model: WRITEME param: WRITEME Parameters for the classifier. """ assert isinstance(name, str) logger.info("Classifying %s" % name) ksplit = len(fold_pairs) if name not in NAMES: raise ValueError("Classifier %s not supported. " "Did you enter it properly?" % name) # Redefine the parameters to be used for RBF SVM (dependent on # training data) if True: #better identifier here logger.info("Attempting to use grid search...") fScore = [] for i, fold_pair in enumerate(fold_pairs): print ("Classifying a %s the %d-th out of %d folds..." % (name, i+1, len(fold_pairs))) classifier = get_classifier(name, model, param, data[fold_pair[0], :]) area = classify(data, labels, fold_pair, classifier) fScore.append(area) else: warnings.warn("Multiprocessing splits not tested yet.") pool = Pool(processes=min(ksplit, PROCESSORS)) classify_func = lambda f : classify( data, labels, fold_pairs[f], classifier=get_classifier( name, model, param, data=data[fold_pairs[f][0], :])) fScore = pool.map(functools.partial(classify_func, xrange(ksplit))) pool.close() pool.join() return classifier, fScore
def initialize(self, n_parallel): self.n_parallel = n_parallel if self.pool is not None: print("Warning: terminating existing pool") self.pool.terminate() self.queue.close() self.worker_queue.close() self.G = SharedGlobal() if n_parallel > 1: self.queue = mp.Queue() self.worker_queue = mp.Queue() self.pool = MemmappingPool( self.n_parallel, temp_folder="/tmp", )
def initialize(self, n_parallel): self.n_parallel = n_parallel if self.pool is not None: print("Warning: terminating existing pool") self.pool.terminate() self.queue.close() self.worker_queue.close() self.G = SharedGlobal() if n_parallel > 1: self.queue = mp.Queue() self.worker_queue = mp.Queue() self.pool = MemmapingPool( self.n_parallel, temp_folder="/tmp", )
def test_pool_with_memmap(tmpdir_path): """Check that subprocess can access and update shared memory memmap""" assert_array_equal = np.testing.assert_array_equal # Fork the subprocess before allocating the objects to be passed pool_temp_folder = os.path.join(tmpdir_path, 'pool') os.makedirs(pool_temp_folder) p = MemmapingPool(10, max_nbytes=2, temp_folder=pool_temp_folder) try: filename = os.path.join(tmpdir_path, 'test.mmap') a = np.memmap(filename, dtype=np.float32, shape=(3, 5), mode='w+') a.fill(1.0) p.map(inplace_double, [(a, (i, j), 1.0) for i in range(a.shape[0]) for j in range(a.shape[1])]) assert_array_equal(a, 2 * np.ones(a.shape)) # Open a copy-on-write view on the previous data b = np.memmap(filename, dtype=np.float32, shape=(5, 3), mode='c') p.map(inplace_double, [(b, (i, j), 2.0) for i in range(b.shape[0]) for j in range(b.shape[1])]) # Passing memmap instances to the pool should not trigger the creation # of new files on the FS assert os.listdir(pool_temp_folder) == [] # the original data is untouched assert_array_equal(a, 2 * np.ones(a.shape)) assert_array_equal(b, 2 * np.ones(b.shape)) # readonly maps can be read but not updated c = np.memmap(filename, dtype=np.float32, shape=(10,), mode='r', offset=5 * 4) assert_raises(AssertionError, p.map, check_array, [(c, i, 3.0) for i in range(c.shape[0])]) # depending on the version of numpy one can either get a RuntimeError # or a ValueError assert_raises((RuntimeError, ValueError), p.map, inplace_double, [(c, i, 2.0) for i in range(c.shape[0])]) finally: # Clean all filehandlers held by the pool p.terminate() del p
class StatefulPool(object): def __init__(self): self.n_parallel = 1 self.pool = None self.queue = None self.worker_queue = None self.G = SharedGlobal() def initialize(self, n_parallel): self.n_parallel = n_parallel if self.pool is not None: print("Warning: terminating existing pool") self.pool.terminate() self.queue.close() self.worker_queue.close() self.G = SharedGlobal() if n_parallel > 1: self.queue = mp.Queue() self.worker_queue = mp.Queue() self.pool = MemmapingPool(self.n_parallel, temp_folder="/tmp") def run_each(self, runner, args_list=None): if args_list is None: args_list = [tuple()] * self.n_parallel assert len(args_list) == self.n_parallel if self.n_parallel > 1: results = self.pool.map_async(worker_run_each, [(runner, args) for args in args_list]) for _ in range(self.n_parallel): self.worker_queue.get() for _ in range(self.n_parallel): self.queue.put(None) return results.get() else: return [runner(self.G, *args_list[0])]
def test_memmaping_pool_for_large_arrays_disabled(): """Check that large arrays memmaping can be disabled""" # Set max_nbytes to None to disable the auto memmaping feature p = MemmapingPool(3, max_nbytes=None, temp_folder=TEMP_FOLDER) try: # Check that the tempfolder is empty assert_equal(os.listdir(TEMP_FOLDER), []) # Try with a file largish than the memmap threshold of 40 bytes large = np.ones(100, dtype=np.float64) assert_equal(large.nbytes, 800) p.map(check_array, [(large, i, 1.0) for i in range(large.shape[0])]) # Check that the tempfolder is still empty assert_equal(os.listdir(TEMP_FOLDER), []) finally: # Cleanup open file descriptors p.terminate() del p
class StatefulPool: def __init__(self): self.n_parallel = 1 self.pool = None self.queue = None self.worker_queue = None self.G = SharedGlobal() def initialize(self, n_parallel): self.n_parallel = n_parallel if self.pool is not None: print("Warning: terminating existing pool") self.pool.terminate() self.queue.close() self.worker_queue.close() self.G = SharedGlobal() if n_parallel > 1: self.queue = mp.Queue() self.worker_queue = mp.Queue() self.pool = MemmapingPool( self.n_parallel, temp_folder="/tmp", ) def terminate(self): if self.pool: self.pool.terminate() def run_each(self, runner, args_list=None): """ Run the method on each worker process, and collect the result of execution. The runner method will receive 'g' as its first argument, followed by the arguments in the args_list, if any :return: """ assert not inspect.ismethod(runner), ( "run_each() cannot run a class method. Please ensure that runner " "is a function with the prototype def foo(g, ...), where g is an " "object of type garage.sampler.stateful_pool.SharedGlobal") if args_list is None: args_list = [tuple()] * self.n_parallel assert len(args_list) == self.n_parallel if self.n_parallel > 1: results = self.pool.map_async(_worker_run_each, [(runner, args) for args in args_list]) for i in range(self.n_parallel): self.worker_queue.get() for i in range(self.n_parallel): self.queue.put(None) return results.get() return [runner(self.G, *args_list[0])] def run_map(self, runner, args_list): assert not inspect.ismethod(runner), ( "run_map() cannot run a class method. Please ensure that runner " "is a function with the prototype 'def foo(g, ...)', where g is " "an object of type garage.sampler.stateful_pool.SharedGlobal") if self.n_parallel > 1: return self.pool.map(_worker_run_map, [(runner, args) for args in args_list]) else: ret = [] for args in args_list: ret.append(runner(self.G, *args)) return ret def run_imap_unordered(self, runner, args_list): assert not inspect.ismethod(runner), ( "run_imap_unordered() cannot run a class method. Please ensure " "that runner is a function with the prototype 'def foo(g, ...)', " "where g is an object of type " "garage.sampler.stateful_pool.SharedGlobal") if self.n_parallel > 1: for x in self.pool.imap_unordered(_worker_run_map, [(runner, args) for args in args_list]): yield x else: for args in args_list: yield runner(self.G, *args) def run_collect(self, collect_once, threshold, args=None, show_prog_bar=True): """ Run the collector method using the worker pool. The collect_once method will receive 'g' as its first argument, followed by the provided args, if any. The method should return a pair of values. The first should be the object to be collected, and the second is the increment to be added. This will continue until the total increment reaches or exceeds the given threshold. Sample script: def collect_once(g): return 'a', 1 stateful_pool.run_collect(collect_once, threshold=3) # should return ['a', 'a', 'a'] :param collector: :param threshold: :return: """ assert not inspect.ismethod(collect_once), ( "run_collect() cannot run a class method. Please ensure that " "collect_once is a function with the prototype 'def foo(g, ...)', " "where g is an object of type " "garage.sampler.stateful_pool.SharedGlobal") if args is None: args = tuple() if self.pool: manager = mp.Manager() counter = manager.Value('i', 0) lock = manager.RLock() results = self.pool.map_async( _worker_run_collect, [(collect_once, counter, lock, threshold, args)] * self.n_parallel) if show_prog_bar: pbar = ProgBarCounter(threshold) last_value = 0 while True: time.sleep(0.1) with lock: if counter.value >= threshold: if show_prog_bar: pbar.stop() break if show_prog_bar: pbar.inc(counter.value - last_value) last_value = counter.value return sum(results.get(), []) else: count = 0 results = [] if show_prog_bar: pbar = ProgBarCounter(threshold) while count < threshold: result, inc = collect_once(self.G, *args) results.append(result) count += inc if show_prog_bar: pbar.inc(inc) if show_prog_bar: pbar.stop() return results return []
class StatefulPool(object): def __init__(self): self.n_parallel = 1 self.pool = None self.queue = None self.worker_queue = None self.G = SharedGlobal() def initialize(self, n_parallel): self.n_parallel = n_parallel if self.pool is not None: print("Warning: terminating existing pool") self.pool.terminate() self.queue.close() self.worker_queue.close() self.G = SharedGlobal() if n_parallel > 1: self.queue = mp.Queue() self.worker_queue = mp.Queue() self.pool = MemmapingPool( self.n_parallel, temp_folder="/tmp", ) def run_each(self, runner, args_list=None): """ Run the method on each worker process, and collect the result of execution. The runner method will receive 'G' as its first argument, followed by the arguments in the args_list, if any :return: """ if args_list is None: args_list = [tuple()] * self.n_parallel assert len(args_list) == self.n_parallel if self.n_parallel > 1: #return [runner(self.G, *args_list[i]) for i in range(self.n_parallel)] results = self.pool.map_async( _worker_run_each, [(runner, args) for args in args_list] ) for i in range(self.n_parallel): self.worker_queue.get() for i in range(self.n_parallel): self.queue.put(None) return results.get() return [runner(self.G, *args_list[0])] def run_map(self, runner, args_list): if self.n_parallel > 1: return self.pool.map(_worker_run_map, [(runner, args) for args in args_list]) else: ret = [] for args in args_list: ret.append(runner(self.G, *args)) return ret def run_imap_unordered(self, runner, args_list): if self.n_parallel > 1: for x in self.pool.imap_unordered(_worker_run_map, [(runner, args) for args in args_list]): yield x else: for args in args_list: yield runner(self.G, *args) def run_collect(self, collect_once, threshold, args=None, show_prog_bar=True, multi_task=False): """ Run the collector method using the worker pool. The collect_once method will receive 'G' as its first argument, followed by the provided args, if any. The method should return a pair of values. The first should be the object to be collected, and the second is the increment to be added. This will continue until the total increment reaches or exceeds the given threshold. Sample script: def collect_once(G): return 'a', 1 stateful_pool.run_collect(collect_once, threshold=3) # => ['a', 'a', 'a'] :param collector: :param threshold: :return: """ if args is None: args = tuple() if self.pool and multi_task: manager = mp.Manager() counter = manager.Value('i', 0) lock = manager.RLock() inputs = [(collect_once, counter, lock, threshold, arg) for arg in args] results = self.pool.map_async( _worker_run_collect, inputs, ) if show_prog_bar: pbar = ProgBarCounter(threshold) last_value = 0 while True: time.sleep(0.1) with lock: if counter.value >= threshold: if show_prog_bar: pbar.stop() break if show_prog_bar: pbar.inc(counter.value - last_value) last_value = counter.value finished_results = results.get() # TODO - for some reason this is buggy. return {i:finished_results[i] for i in range(len(finished_results))} elif multi_task: assert False # not supported elif self.pool: manager = mp.Manager() counter = manager.Value('i', 0) lock = manager.RLock() results = self.pool.map_async( _worker_run_collect, [(collect_once, counter, lock, threshold, args)] * self.n_parallel ) if show_prog_bar: pbar = ProgBarCounter(threshold) last_value = 0 while True: time.sleep(0.1) with lock: if counter.value >= threshold: if show_prog_bar: pbar.stop() break if show_prog_bar: pbar.inc(counter.value - last_value) last_value = counter.value return sum(results.get(), []) else: count = 0 results = [] if show_prog_bar: pbar = ProgBarCounter(threshold) while count < threshold: result, inc = collect_once(self.G, *args) results.append(result) count += inc if show_prog_bar: pbar.inc(inc) if show_prog_bar: pbar.stop() return results
def get_score(data, labels, fold_pairs, name, model, param, numTopVars, rank_per_fold=None, parallel=True, rand_iter=-1): """ Function to get score for a classifier. Parameters ---------- data: array_like Data from which to derive score. labels: array_like or list Corresponding labels for each sample. fold_pairs: list of pairs of array_like A list of train/test indicies for each fold dhjelm(Why can't we just use the KFold object?) name: str Name of classifier. model: WRITEME param: WRITEME Parameters for the classifier. parallel: bool Whether to run folds in parallel. Default: True Returns ------- classifier: WRITEME allConfMats: Confusion matrix for all folds and all variables sets and best performing parameter set ([numFolds, numVarSets]) """ assert isinstance(name, str) logging.info("Classifying %s" % name) ksplit = len(fold_pairs) # if name not in NAMES: # raise ValueError("Classifier %s not supported. " # "Did you enter it properly?" % name) # Redefine the parameters to be used for RBF SVM (dependent on # training data) if "SGD" in name: param["n_iter"] = [25] # [np.ceil(10**3 / len(fold_pairs[0][0]))] classifier = get_classifier(name, model, param, rand_iter=rand_iter) if name == "RBF SVM": #This doesn't use labels, but looks as ALL data logging.info("RBF SVM requires some preprocessing." "This may take a while") # is_data_computed_gamma = True # if not is_data_computed_gamma: # Sahil commented the code below that computes the gamma choices from data. # The computed gamma choices seem too low thereby making SVM very slow. Instead, trying out fixed values. print param gamma = param['gamma'] gamma = np.array(gamma) print 'gamma', gamma else: #Euclidean distances between samples # sahil switched from the first call to second one for computing the dist as the first one is giving error. # dist = pdist(StandardScaler().fit(data), "euclidean").ravel() dist = pdist(RobustScaler().fit_transform(data), "euclidean").ravel() print 'dist', dist #Estimates for sigma (10th, 50th and 90th percentile) sigest = np.asarray(np.percentile(dist, [10, 50, 90])) print 'sigest', sigest #Estimates for gamma (= -1/(2*sigma^2)) gamma = 1. / (2 * sigest**2) print 'gamma', gamma # # #Set SVM parameters with these values # sahil changed the code a bit to remove a bug # param = [{"kernel": ["rbf"], # "gamma": gamma.tolist(), # "C": np.logspace(-2,2,5).tolist()}] param = { "kernel": ["rbf"], "gamma": gamma.tolist(), "C": np.logspace(-2, 2, 5).tolist() } # if name not in ["Decision Tree", "Naive Bayes"]: if param: if hasattr(classifier, 'param_grid'): # isinstance(classifier, GridSearchCV): print 'param', param N_p = np.prod([len(l) for l in param.values()]) elif isinstance(classifier, RandomizedSearchCV): N_p = classifier.n_iter else: N_p = 1 # is_cv = isinstance(classifier, GridSearchCV) or \ # isinstance(classifier, RandomizedSearchCV) # print('Name: {}, ksplit: {}, N_p: {}'.format(name, ksplit, N_p)) if (not parallel) or ksplit <= N_p or \ (name == "Random Forest") or ("SGD" in name): logging.info("Attempting to use grid search...") classifier.n_jobs = PROCESSORS classifier.pre_dispatch = 1 # np.floor(PROCESSORS/24) allConfMats = [] allTotalErrs = [] allFittedClassifiers = [] for i, fold_pair in enumerate(fold_pairs): confMats = [] totalErrs = [] fitted_classifiers = [] logging.info("Classifying a %s the %d-th out of %d folds..." % (name, i + 1, len(fold_pairs))) if rank_per_fold is not None: rankedVars = rank_per_fold[i] else: rankedVars = np.arange(data.shape[1]) # for numVars in numTopVars: logging.info('Classifying for top %i variables' % numVars) # # print 'rankedVars', rankedVars # confMat, totalErr, fitted_classifier = classify( data[:, rankedVars[:numVars]], labels, fold_pair, classifier) confMats.append(confMat) totalErrs.append(totalErr) fitted_classifiers.append(fitted_classifier) # recheck the structure of area and fScore variables allConfMats.append(confMats) allTotalErrs.append(totalErrs) allFittedClassifiers.append(fitted_classifiers) else: print 'parallel computing going on (debug Sahil ...) ..........................' # classifier.n_jobs = PROCESSORS logging.info("Multiprocessing folds for classifier {}.".format(name)) pool = Pool(processes=min(ksplit, PROCESSORS)) out_list = pool.map( per_split_classifier(data, labels, classifier, numTopVars), zip(rank_per_fold, fold_pairs)) pool.close() pool.join() #allConfMats = [el[0] for el in out_list] #allTotalErrs = [el[1] for el in out_list] #allFittedClassifiers = [el[2] for el in out_list] allConfMats, allTotalErrs, allFittedClassifiers = tuple(zip(*out_list)) return classifier, allConfMats, allTotalErrs, allFittedClassifiers
def get_rank_per_fold(data, labels, fold_pairs, ranking_function=ttest_ind, save_path=None, load_file=True, parallel=True): ''' Applies rank_vars to each test set in list of fold pairs Inputs: data: array features for all samples labels: array label vector of each sample fold_pair: list list pairs of index arrays containing train and test sets ranking_function: function object, default: ttest_ind function to apply for ranking features ranking_function: function ranking function to use, default: ttest_ind save_path: dir to load and save ranking files load_file: bool Whether to try to load an existing file, default: True parallel: bool True if multicore processing is desired, default: True Outputs: rank_per_fod: list List of ranked feature indexes for each fold pair ''' file_loaded = False if load_file: if isinstance(save_path, str): fname = path.join( save_path, "{}_{}_folds.mat".format(ranking_function.__name__, len(fold_pairs))) try: rd = scipy.io.loadmat(fname, mat_dtype=True) rank_per_fold = rd['rank_per_fold'] file_loaded = True except: pass else: print('No rank file path: Computing from scratch without saving') if not file_loaded: if not parallel: rank_per_fold = [] for fold_pair in fold_pairs: rankedVars = rank_vars(data[fold_pair[0], :], labels[fold_pair[0]], ranking_function) rank_per_fold.append(rankedVars) else: pool = Pool(processes=min(len(fold_pairs), PROCESSORS)) rank_per_fold = pool.map( Ranker(data, labels, ranking_function, rank_vars), fold_pairs) pool.close() pool.join() if isinstance(save_path, str): fname = path.join( save_path, "{}_{}_folds.mat".format(ranking_function.__name__, len(fold_pairs))) with open(fname, 'wb') as f: scipy.io.savemat(f, {'rank_per_fold': rank_per_fold}) return rank_per_fold