sentence1
stringlengths
52
3.87M
sentence2
stringlengths
1
47.2k
label
stringclasses
1 value
def run_evaluation(self, stream_name: str) -> None: """ Run the main loop with the given stream in the prediction mode. :param stream_name: name of the stream to be evaluated """ def prediction(): logging.info('Running prediction') self._run_zeroth_epoch([stream_name]) logging.info('Prediction done\n\n') self._try_run(prediction)
Run the main loop with the given stream in the prediction mode. :param stream_name: name of the stream to be evaluated
entailment
def major_vote(all_votes: Iterable[Iterable[Hashable]]) -> Iterable[Hashable]: """ For the given iterable of object iterations, return an iterable of the most common object at each position of the inner iterations. E.g.: for [[1, 2], [1, 3], [2, 3]] the return value would be [1, 3] as 1 and 3 are the most common objects at the first and second positions respectively. :param all_votes: an iterable of object iterations :return: the most common objects in the iterations (the major vote) """ return [Counter(votes).most_common()[0][0] for votes in zip(*all_votes)]
For the given iterable of object iterations, return an iterable of the most common object at each position of the inner iterations. E.g.: for [[1, 2], [1, 3], [2, 3]] the return value would be [1, 3] as 1 and 3 are the most common objects at the first and second positions respectively. :param all_votes: an iterable of object iterations :return: the most common objects in the iterations (the major vote)
entailment
def _load_models(self) -> None: """Maybe load all the models to be assembled together and save them to the ``self._models`` attribute.""" if self._models is None: logging.info('Loading %d models', len(self._model_paths)) def load_model(model_path: str): logging.debug('\tloading %s', model_path) if path.isdir(model_path): model_path = path.join(model_path, CXF_CONFIG_FILE) config = load_config(model_path) config['model']['inputs'] = self._inputs config['model']['outputs'] = self._outputs return create_model(config, output_dir=None, dataset=self._dataset, restore_from=path.dirname(model_path)) self._models = list(map(load_model, self._model_paths))
Maybe load all the models to be assembled together and save them to the ``self._models`` attribute.
entailment
def run(self, batch: Batch, train: bool=False, stream: StreamWrapper=None) -> Batch: """ Run feed-forward pass with the given batch using all the models, aggregate and return the results. .. warning:: :py:class:`Ensemble` can not be trained. :param batch: batch to be processed :param train: ``True`` if this batch should be used for model update, ``False`` otherwise :param stream: stream wrapper (useful for precise buffer management) :return: aggregated results dict :raise ValueError: if the ``train`` flag is set to ``True`` """ if train: raise ValueError('Ensemble model cannot be trained.') self._load_models() # run all the models batch_outputs = [model.run(batch, False, stream) for model in self._models] # aggregate the outputs aggregated = {} for output_name in self._outputs: output_values = [batch_output[output_name] for batch_output in batch_outputs] if self._aggregation == 'mean': aggregated[output_name] = np.mean(output_values, axis=0) elif self._aggregation == 'major_vote': output_values_arr = np.array(output_values) output = major_vote(output_values_arr.reshape((output_values_arr.shape[0], -1))) aggregated[output_name] = np.array(output).reshape(output_values_arr[0].shape) return aggregated
Run feed-forward pass with the given batch using all the models, aggregate and return the results. .. warning:: :py:class:`Ensemble` can not be trained. :param batch: batch to be processed :param train: ``True`` if this batch should be used for model update, ``False`` otherwise :param stream: stream wrapper (useful for precise buffer management) :return: aggregated results dict :raise ValueError: if the ``train`` flag is set to ``True``
entailment
def parse_fully_qualified_name(fq_name: str) -> Tuple[Optional[str], str]: """ Parse the given fully-quallified name (separated with dots) to a tuple of module and class names. :param fq_name: fully qualified name separated with dots :return: ``None`` instead of module if the given name contains no separators (dots). """ last_dot = fq_name.rfind('.') if last_dot != -1: return fq_name[:last_dot], fq_name[last_dot + 1:] else: return None, fq_name
Parse the given fully-quallified name (separated with dots) to a tuple of module and class names. :param fq_name: fully qualified name separated with dots :return: ``None`` instead of module if the given name contains no separators (dots).
entailment
def get_attribute(module_name: str, attribute_name: str): """ Get the specified module attribute. It most cases, it will be a class or function. :param module_name: module name :param attribute_name: attribute name :return: module attribute """ assert isinstance(module_name, str) assert isinstance(attribute_name, str) _module = importlib.import_module(module_name) return getattr(_module, attribute_name)
Get the specified module attribute. It most cases, it will be a class or function. :param module_name: module name :param attribute_name: attribute name :return: module attribute
entailment
def create_object(module_name: str, class_name: str, args: Iterable=(), kwargs: Dict[str, Any]=_EMPTY_DICT): """ Create an object instance of the given class from the given module. Args and kwargs are passed to the constructor. This mimics the following code: .. code-block:: python from module import class return class(*args, **kwargs) :param module_name: module name :param class_name: class name :param args: args to be passed to the object constructor :param kwargs: kwargs to be passed to the object constructor :return: created object instance """ return get_attribute(module_name, class_name)(*args, **kwargs)
Create an object instance of the given class from the given module. Args and kwargs are passed to the constructor. This mimics the following code: .. code-block:: python from module import class return class(*args, **kwargs) :param module_name: module name :param class_name: class name :param args: args to be passed to the object constructor :param kwargs: kwargs to be passed to the object constructor :return: created object instance
entailment
def list_submodules(module_name: str) -> List[str]: # pylint: disable=invalid-sequence-index """ List full names of all the submodules in the given module. :param module_name: name of the module of which the submodules will be listed """ _module = importlib.import_module(module_name) return [module_name+'.'+submodule_name for _, submodule_name, _ in pkgutil.iter_modules(_module.__path__)]
List full names of all the submodules in the given module. :param module_name: name of the module of which the submodules will be listed
entailment
def find_class_module(module_name: str, class_name: str) \ -> Tuple[List[str], List[Tuple[str, Exception]]]: # pylint: disable=invalid-sequence-index """ Find sub-modules of the given module that contain the given class. Moreover, return a list of sub-modules that could not be imported as a list of (sub-module name, Exception) tuples. :param module_name: name of the module to be searched :param class_name: searched class name :return: a tuple of sub-modules having the searched class and sub-modules that could not be searched """ matched_submodules = [] erroneous_submodules = [] for submodule_name in list_submodules(module_name): try: # the sub-module to be included may be erroneous and we need to continue submodule = importlib.import_module(submodule_name) if hasattr(submodule, class_name): matched_submodules.append(submodule_name) except Exception as ex: # pylint: disable=broad-except erroneous_submodules.append((submodule_name, ex)) return matched_submodules, erroneous_submodules
Find sub-modules of the given module that contain the given class. Moreover, return a list of sub-modules that could not be imported as a list of (sub-module name, Exception) tuples. :param module_name: name of the module to be searched :param class_name: searched class name :return: a tuple of sub-modules having the searched class and sub-modules that could not be searched
entailment
def get_class_module(module_name: str, class_name: str) -> Optional[str]: """ Get a sub-module of the given module which has the given class. This method wraps `utils.reflection.find_class_module method` with the following behavior: - raise error when multiple sub-modules with different classes with the same name are found - return None when no sub-module is found - warn about non-searchable sub-modules .. note:: This function logs! :param module_name: module to be searched :param class_name: searched class name :return: sub-module with the searched class or None """ matched_modules, erroneous_modules = find_class_module(module_name, class_name) for submodule, error in erroneous_modules: logging.warning('Could not inspect sub-module `%s` due to `%s` ' 'when searching for `%s` in sub-modules of `%s`.', submodule, type(error).__name__, class_name, module_name) if len(matched_modules) == 1: return matched_modules[0] if len(matched_modules) > 1: # check if all the module attributes point to the same class first_class = getattr(importlib.import_module(matched_modules[0]), class_name) for matched_module in matched_modules: another_class = getattr(importlib.import_module(matched_module), class_name) if another_class is not first_class: raise ValueError('Found more than one sub-module when searching for `{}` in sub-modules of `{}`. ' 'Please specify the module explicitly. Found sub-modules: `{}`' .format(class_name, module_name, matched_modules)) return matched_modules[0] return None
Get a sub-module of the given module which has the given class. This method wraps `utils.reflection.find_class_module method` with the following behavior: - raise error when multiple sub-modules with different classes with the same name are found - return None when no sub-module is found - warn about non-searchable sub-modules .. note:: This function logs! :param module_name: module to be searched :param class_name: searched class name :return: sub-module with the searched class or None
entailment
def entry_point() -> None: """**cxflow** entry point.""" # make sure the path contains the current working directory sys.path.insert(0, os.getcwd()) parser = get_cxflow_arg_parser(True) # parse CLI arguments known_args, unknown_args = parser.parse_known_args() # show help if no subcommand was specified. if not hasattr(known_args, 'subcommand'): parser.print_help() quit(1) # set up global logger logger = logging.getLogger('') logger.setLevel(logging.DEBUG if known_args.verbose else logging.INFO) logger.handlers = [] # remove default handlers # set up STDERR handler stderr_handler = logging.StreamHandler(sys.stderr) stderr_handler.setFormatter(logging.Formatter(CXF_LOG_FORMAT, datefmt=CXF_LOG_DATE_FORMAT)) logger.addHandler(stderr_handler) if known_args.subcommand == 'train': train(config_path=known_args.config_file, cl_arguments=unknown_args, output_root=known_args.output_root) elif known_args.subcommand == 'resume': resume(config_path=known_args.config_path, restore_from=known_args.restore_from, cl_arguments=unknown_args, output_root=known_args.output_root) elif known_args.subcommand == 'predict': logging.warning('Predict command is deprecated and will be removed, use ``cxflow eval predict ...`` instead') predict(config_path=known_args.config_path, restore_from=known_args.restore_from, cl_arguments=unknown_args, output_root=known_args.output_root) elif known_args.subcommand == 'eval': evaluate(model_path=known_args.model_path, stream_name=known_args.stream_name, config_path=known_args.config, cl_arguments=unknown_args, output_root=known_args.output_root) elif known_args.subcommand == 'dataset': invoke_dataset_method(config_path=known_args.config_file, method_name=known_args.method, cl_arguments=unknown_args, output_root=known_args.output_root) elif known_args.subcommand == 'gridsearch': grid_search(script=known_args.script, params=known_args.params, dry_run=known_args.dry_run) elif known_args.subcommand == 'ls': list_train_dirs(known_args.dir, known_args.recursive, known_args.all, known_args.long, known_args.verbose) elif known_args.subcommand == 'prune': prune_train_dirs(known_args.dir, known_args.epochs, known_args.subdirs)
**cxflow** entry point.
entailment
def get_cxflow_arg_parser(add_common_arguments: bool=False) -> ArgumentParser: """ Create the **cxflow** argument parser. :return: an instance of the parser """ # create parser main_parser = ArgumentParser('cxflow', description='cxflow: lightweight framework for machine learning with ' 'focus on modularization, re-usability and rapid experimenting.', epilog='For more info see <https://cognexa.github.io/cxflow>') main_parser.add_argument('--version', action='version', help='Print cxflow version and quit.', version='cxflow {}'.format(pkg_resources.get_distribution('cxflow').version)) subparsers = main_parser.add_subparsers(help='cxflow commands') # create train sub-parser train_parser = subparsers.add_parser('train', description='Start cxflow training from the ``config_file``.') train_parser.set_defaults(subcommand='train') train_parser.add_argument('config_file', help='path to the config file') # create resume sub-parser resume_parser = subparsers.add_parser('resume', description='Resume cxflow training from the ``config_path``.') resume_parser.set_defaults(subcommand='resume') resume_parser.add_argument('config_path', help='path to the config file or the directory in which it is stored') resume_parser.add_argument('restore_from', nargs='?', default=None, help='information passed to the model constructor (backend-specific); ' 'usually a directory in which the trained model is stored') # create predict sub-parser (deprecated) predict_parser = subparsers.add_parser('predict', description='Run prediction with the given ``config_path``.') predict_parser.set_defaults(subcommand='predict') predict_parser.add_argument('config_path', help='path to the config file or the directory in which it is stored') predict_parser.add_argument('restore_from', nargs='?', default=None, help='information passed to the model constructor (backend-specific); usually a ' 'directory in which the trained model is stored') # create eval sub-parser eval_parser = subparsers.add_parser('eval', description='Evaluate the given model on the specified data stream.') eval_parser.set_defaults(subcommand='eval') eval_parser.add_argument('stream_name', help='stream name to be evaluated') eval_parser.add_argument('model_path', help='model path to be evaluated') eval_parser.add_argument('--config', '-c', nargs='?', default=None, help='optional config path to be used') # create dataset sub-parser dataset_parser = subparsers.add_parser('dataset', description='Invoke arbitrary dataset method.') dataset_parser.set_defaults(subcommand='dataset') dataset_parser.add_argument('method', help='name of the method to be invoked') dataset_parser.add_argument('config_file', help='path to the config file') # create grid-search sub-parser gridsearch_parser = subparsers.add_parser('gridsearch', description='Do parameter grid search (experimental).') gridsearch_parser.set_defaults(subcommand='gridsearch') gridsearch_parser.add_argument('script', help='Script to be grid-searched') gridsearch_parser.add_argument('params', nargs='*', help='Params to be tested. Format: name:type=[value1,value2]. ' 'Type is optional') gridsearch_parser.add_argument('--dry-run', action='store_true', help='Only print command output instead ' 'of executing it right away') # create ls sub-parser ls_parser = subparsers.add_parser('ls', description='List training log dirs in the given path.') ls_parser.set_defaults(subcommand='ls') ls_parser.add_argument('dir', nargs='?', default=CXF_DEFAULT_LOG_DIR, help='path to the log directory to be listed') ls_parser.add_argument('-l', '--long', action='store_true', help='use long listing format') ls_parser.add_argument('-a', '--all', action='store_true', help='include trainings with no epochs done') ls_parser.add_argument('-r', '--recursive', action='store_true', help='list all the dirs recursively, stop at training dirs') ls_parser.add_argument('-v', '--verbose', action='store_true', help='print more verbose output, applicable only when a single train dir is listed') # create prune sub-parser prune_parser = subparsers.add_parser('prune', description='Prune training log dirs in the given path without finished epochs.') prune_parser.set_defaults(subcommand='prune') prune_parser.add_argument('dir', nargs='?', default=CXF_DEFAULT_LOG_DIR, help='path to the log directory to be pruned') prune_parser.add_argument('-e', '--epochs', default=1, type=int, help='keep only training log dirs having at least this many completed epochs, default 1') prune_parser.add_argument('-s', '--subdirs', action='store_true', help='delete all subdirectories in training directories') # add common arguments if add_common_arguments: for parser in [main_parser, train_parser, resume_parser, predict_parser, dataset_parser, eval_parser]: parser.add_argument('--output_root', '-o', default='./log', help='output directory') parser.add_argument('--verbose', '-v', action='store_true', help='increase verbosity to level DEBUG') return main_parser
Create the **cxflow** argument parser. :return: an instance of the parser
entailment
def _get_stream(self) -> Iterator: """Possibly create and return raw dataset stream iterator.""" if self._stream is None: self._stream = iter(self._get_stream_fn()) return self._stream
Possibly create and return raw dataset stream iterator.
entailment
def _enqueue_batches(self, stop_event: Event) -> None: """ Enqueue all the stream batches. If specified, stop after ``epoch_size`` batches. .. note:: Signal the epoch end with ``None``. Stop when: - ``stop_event`` is risen - stream ends and epoch size is not set - specified number of batches is enqueued .. note:: This is used only with ``buffer`` > 0. :param stop_event: event signaling stop instruction """ while True: self._stream = self._get_stream() while True: # Acquire the semaphore before processing the next batch # but immediately release it so that other threads # are not blocked when they decide to acquire it again. with self._semaphore: pass # It always takes a short moment before the native call actually # releases the GIL and we are free to compute. The following sleep # is here to compensate for this short moment - we don't want to # slow down the native call before the GIL is released. time.sleep(CXF_BUFFER_SLEEP) try: batch = next(self._stream) except StopIteration: break self._queue.put(batch) self._batch_count += 1 if stop_event.is_set(): return if self._epoch_limit_reached(): self._queue.put(None) self._batch_count = 0 return self._stream = None # yield a new iterator next time if self._epoch_size <= 0: # for non-fixed size epochs self._queue.put(None) self._batch_count = 0 return
Enqueue all the stream batches. If specified, stop after ``epoch_size`` batches. .. note:: Signal the epoch end with ``None``. Stop when: - ``stop_event`` is risen - stream ends and epoch size is not set - specified number of batches is enqueued .. note:: This is used only with ``buffer`` > 0. :param stop_event: event signaling stop instruction
entailment
def _dequeue_batch(self) -> Optional[Batch]: """ Return a single batch from queue or ``None`` signaling epoch end. :raise ChildProcessError: if the enqueueing thread ended unexpectedly """ if self._enqueueing_thread is None: raise ValueError('StreamWrapper `{}` with buffer of size `{}` was used outside with-resource environment.' .format(self._name, self._buffer_size)) if not self._enqueueing_thread.is_alive() and self._queue.empty(): self._start_thread() while True: try: batch = self._queue.get(timeout=2) self._queue.task_done() break except Empty: if not self._enqueueing_thread.is_alive(): try: # the enqueueing thread may just finished properly so lets check the queue eagerly batch = self._queue.get_nowait() self._queue.task_done() break except Empty: # so we failed to retrieve a batch and the enqueueing thread is dead # there is no hope, something must went wrong raise ChildProcessError('Enqueueing thread ended unexpectedly.') return batch
Return a single batch from queue or ``None`` signaling epoch end. :raise ChildProcessError: if the enqueueing thread ended unexpectedly
entailment
def _next_batch(self) -> Optional[Batch]: """ Return a single batch or ``None`` signaling epoch end. .. note:: Signal the epoch end with ``None``. Stop when: - stream ends and epoch size is not set - specified number of batches is returned :return: a single batch or ``None`` signaling epoch end """ if self._epoch_limit_reached(): self._batch_count = 0 return None try: batch = next(self._get_stream()) self._batch_count += 1 return batch except StopIteration: self._stream = None # yield a new iterator next time if self._epoch_size > 0: # underlying stream ended but our fixed size epoch did not batch = next(self._get_stream()) # get another stream and return its 1st batch self._batch_count += 1 return batch else: self._batch_count = 0 return None
Return a single batch or ``None`` signaling epoch end. .. note:: Signal the epoch end with ``None``. Stop when: - stream ends and epoch size is not set - specified number of batches is returned :return: a single batch or ``None`` signaling epoch end
entailment
def _start_thread(self): """Start an enqueueing thread.""" self._stopping_event = Event() self._enqueueing_thread = Thread(target=self._enqueue_batches, args=(self._stopping_event,)) self._enqueueing_thread.start()
Start an enqueueing thread.
entailment
def _stop_thread(self): """Stop the enqueueing thread. Keep the queue content and stream state.""" self._stopping_event.set() queue_content = [] try: # give the enqueueing thread chance to put a batch to the queue and check the stopping event while True: queue_content.append(self._queue.get_nowait()) except Empty: pass self._enqueueing_thread.join() try: queue_content.append(self._queue.get_nowait()) # collect the very last item except Empty: pass self._queue = Queue(max(len(queue_content), self._buffer_size)) # queue content may be bigger than queue size for batch in queue_content: self._queue.put(batch)
Stop the enqueueing thread. Keep the queue content and stream state.
entailment
def _after_n_epoch(self, epoch_id: int, **_) -> None: """ Save the model every ``n_epochs`` epoch. :param epoch_id: number of the processed epoch """ SaveEvery.save_model(model=self._model, name_suffix=str(epoch_id), on_failure=self._on_save_failure)
Save the model every ``n_epochs`` epoch. :param epoch_id: number of the processed epoch
entailment
def save_model(model: AbstractModel, name_suffix: str, on_failure: str) -> None: """ Save the given model with the given name_suffix. On failure, take the specified action. :param model: the model to be saved :param name_suffix: name to be used for saving :param on_failure: action to be taken on failure; one of :py:attr:`SAVE_FAILURE_ACTIONS` :raise IOError: on save failure with ``on_failure`` set to ``error`` """ try: logging.debug('Saving the model') save_path = model.save(name_suffix) logging.info('Model saved to: %s', save_path) except Exception as ex: # pylint: disable=broad-except if on_failure == 'error': raise IOError('Failed to save the model.') from ex elif on_failure == 'warn': logging.warning('Failed to save the model.')
Save the given model with the given name_suffix. On failure, take the specified action. :param model: the model to be saved :param name_suffix: name to be used for saving :param on_failure: action to be taken on failure; one of :py:attr:`SAVE_FAILURE_ACTIONS` :raise IOError: on save failure with ``on_failure`` set to ``error``
entailment
def _get_value(self, epoch_data: EpochData) -> float: """ Retrieve the value of the monitored variable from the given epoch data. :param epoch_data: epoch data which determine whether the model will be saved or not :raise KeyError: if any of the specified stream, variable or aggregation is not present in the ``epoch_data`` :raise TypeError: if the variable value is not a dict when aggregation is specified :raise ValueError: if the variable value is not a scalar """ if self._stream_name not in epoch_data: raise KeyError('Stream `{}` was not found in the epoch data.\nAvailable streams are `{}`.' .format(self._stream_name, epoch_data.keys())) stream_data = epoch_data[self._stream_name] if self._variable not in stream_data: raise KeyError('Variable `{}` for stream `{}` was not found in the epoch data. ' 'Available variables for stream `{}` are `{}`.' .format(self._variable, self._stream_name, self._stream_name, stream_data.keys())) value = stream_data[self._variable] if self._aggregation: if not isinstance(value, dict): raise TypeError('Variable `{}` is expected to be a dict when aggregation is specified. ' 'Got `{}` instead.'.format(self._variable, type(value).__name__)) if self._aggregation not in value: raise KeyError('Specified aggregation `{}` was not found in the variable `{}`. ' 'Available aggregations: `{}`.'.format(self._aggregation, self._variable, value.keys())) value = value[self._aggregation] if not np.isscalar(value): raise ValueError('Variable `{}` value is not a scalar.'.format(value)) return value
Retrieve the value of the monitored variable from the given epoch data. :param epoch_data: epoch data which determine whether the model will be saved or not :raise KeyError: if any of the specified stream, variable or aggregation is not present in the ``epoch_data`` :raise TypeError: if the variable value is not a dict when aggregation is specified :raise ValueError: if the variable value is not a scalar
entailment
def _is_value_better(self, new_value: float) -> bool: """ Test if the new value is better than the best so far. :param new_value: current value of the objective function """ if self._best_value is None: return True if self._condition == 'min': return new_value < self._best_value if self._condition == 'max': return new_value > self._best_value
Test if the new value is better than the best so far. :param new_value: current value of the objective function
entailment
def after_epoch(self, epoch_data: EpochData, **_) -> None: """ Save the model if the new value of the monitored variable is better than the best value so far. :param epoch_data: epoch data to be processed """ new_value = self._get_value(epoch_data) if self._is_value_better(new_value): self._best_value = new_value SaveEvery.save_model(model=self._model, name_suffix=self._OUTPUT_NAME, on_failure=self._on_save_failure)
Save the model if the new value of the monitored variable is better than the best value so far. :param epoch_data: epoch data to be processed
entailment
def print_progress_bar(done: int, total: int, prefix: str = '', suffix: str = '') -> None: """ Print a progressbar with the given prefix and suffix, without newline at the end. param done: current step in computation param total: total count of steps in computation param prefix: info text displayed before the progress bar param suffix: info text displayed after the progress bar """ percent = '{0:.1f}'.format(100 * (done / float(total))) base_len = shutil.get_terminal_size().columns - 7 - len(prefix) - len(suffix) base_len = min([base_len, 50]) min_length = base_len - 1 - len('{}/{}={}'.format(total, total, '100.0')) length = base_len - len('{}/{}={}'.format(done, total, percent)) if min_length > 0: filled_len = int(min_length * done // total) bar = '='*filled_len + '-'*(min_length - filled_len) spacing = ' '*(length - min_length) print('\r{}: |{}|{}{}/{}={}% {}'.format(prefix, bar, spacing, done, total, percent, suffix), end='\r') else: short_progress = '\r{}: {}/{}'.format(prefix, done, total) if len(short_progress) <= shutil.get_terminal_size().columns: print(short_progress, end='\r') else: print(['-', '\\', '|', '/'][done % 4], end='\r')
Print a progressbar with the given prefix and suffix, without newline at the end. param done: current step in computation param total: total count of steps in computation param prefix: info text displayed before the progress bar param suffix: info text displayed after the progress bar
entailment
def get_formatted_time(seconds: float) -> str: """ Convert seconds to the time format ``H:M:S.UU``. :param seconds: time in seconds :return: formatted human-readable time """ seconds = round(seconds) m, s = divmod(seconds, 60) h, m = divmod(m, 60) return '{:d}:{:02d}:{:02d}'.format(h, m, s)
Convert seconds to the time format ``H:M:S.UU``. :param seconds: time in seconds :return: formatted human-readable time
entailment
def after_batch(self, stream_name: str, batch_data: Batch) -> None: """ Display the progress and ETA for the current stream in the epoch. If the stream size (total batch count) is unknown (1st epoch), print only the number of processed batches. """ if self._current_stream_name is None or self._current_stream_name != stream_name: self._current_stream_name = stream_name self._current_stream_start = None erase_line() self._current_batch_count[stream_name] += 1 current_batch = self._current_batch_count[stream_name] # total batch count is available if stream_name in self._total_batch_count: # compute ETA total_batches = self._total_batch_count[stream_name] if self._current_stream_start: measured_batches = current_batch - 1 avg_batch_time = (time.time() - self._current_stream_start) / measured_batches eta_sec = avg_batch_time * (total_batches - current_batch) eta = get_formatted_time(eta_sec) else: self._current_stream_start = time.time() eta = '' print_progress_bar(current_batch, total_batches, prefix=stream_name, suffix=eta) # total batch count is not available (1st epoch) else: short_progress = '{}: {}'.format(stream_name, current_batch) if len(short_progress) <= shutil.get_terminal_size().columns: print(short_progress, end='\r') else: print(['-', '\\', '|', '/'][current_batch % 4], end='\r')
Display the progress and ETA for the current stream in the epoch. If the stream size (total batch count) is unknown (1st epoch), print only the number of processed batches.
entailment
def after_epoch(self, **_) -> None: """ Reset progress counters. Save ``total_batch_count`` after the 1st epoch. """ if not self._total_batch_count_saved: self._total_batch_count = self._current_batch_count.copy() self._total_batch_count_saved = True self._current_batch_count.clear() self._current_stream_start = None self._current_stream_name = None erase_line()
Reset progress counters. Save ``total_batch_count`` after the 1st epoch.
entailment
def after_epoch_profile(self, epoch_id, profile: TimeProfile, train_stream_name: str, extra_streams: Iterable[str]) -> None: """ Summarize and log the given epoch profile. The profile is expected to contain at least: - ``read_data_train``, ``eval_batch_train`` and ``after_batch_hooks_train`` entries produced by the train stream (if train stream name is `train`) - ``after_epoch_hooks`` entry :param profile: epoch timings profile :param extra_streams: enumeration of additional stream names """ read_data_total = 0 eval_total = 0 train_total = sum(profile.get('eval_batch_{}'.format(train_stream_name), [])) hooks_total = sum(profile.get('after_epoch_hooks', [])) for stream_name in chain(extra_streams, [train_stream_name]): read_data_total += sum(profile.get('read_batch_' + stream_name, [])) hooks_total += sum(profile.get('after_batch_hooks_' + stream_name, [])) if stream_name != train_stream_name: eval_total += sum(profile.get('eval_batch_' + stream_name, [])) logging.info('\tT read data:\t%f', read_data_total) logging.info('\tT train:\t%f', train_total) logging.info('\tT eval:\t%f', eval_total) logging.info('\tT hooks:\t%f', hooks_total)
Summarize and log the given epoch profile. The profile is expected to contain at least: - ``read_data_train``, ``eval_batch_train`` and ``after_batch_hooks_train`` entries produced by the train stream (if train stream name is `train`) - ``after_epoch_hooks`` entry :param profile: epoch timings profile :param extra_streams: enumeration of additional stream names
entailment
def confusion_matrix(expected: np.ndarray, predicted: np.ndarray, num_classes: int) -> np.ndarray: """ Calculate and return confusion matrix for the predicted and expected labels :param expected: array of expected classes (integers) with shape `[num_of_data]` :param predicted: array of predicted classes (integers) with shape `[num_of_data]` :param num_classes: number of classification classes :return: confusion matrix (cm) with absolute values """ assert np.issubclass_(expected.dtype.type, np.integer), " Classes' indices must be integers" assert np.issubclass_(predicted.dtype.type, np.integer), " Classes' indices must be integers" assert expected.shape == predicted.shape, "Predicted and expected data must be the same length" assert num_classes > np.max([predicted, expected]), \ "Number of classes must be at least the number of indices in predicted/expected data" assert np.min([predicted, expected]) >= 0, " Classes' indices must be positive integers" cm_abs = np.zeros((num_classes, num_classes), dtype=np.int32) for pred, exp in zip(predicted, expected): cm_abs[exp, pred] += 1 return cm_abs
Calculate and return confusion matrix for the predicted and expected labels :param expected: array of expected classes (integers) with shape `[num_of_data]` :param predicted: array of predicted classes (integers) with shape `[num_of_data]` :param num_classes: number of classification classes :return: confusion matrix (cm) with absolute values
entailment
def _build_grid_search_commands(script: str, params: typing.Iterable[str]) -> typing.Iterable[typing.List[str]]: """ Build all grid search parameter configurations. :param script: String of command prefix, e.g. ``cxflow train -v -o log``. :param params: Iterable collection of strings in standard **cxflow** param form, e.g. ``'numerical_param=[1, 2]'`` or ``'text_param=["hello", "cio"]'``. """ param_space = OrderedDict() for arg in params: assert '=' in arg name = arg[:arg.index('=')] options = arg[arg.index('=') + 1:] options = ast.literal_eval(options) assert isinstance(options, list), options param_space[name] = options param_names = param_space.keys() commands = [] for values in itertools.product(*[param_space[name] for name in param_names]): command = str(script).split() for name, value in zip(param_names, values): command.append(str(name) + '="' + str(value) + '"') commands.append(command) return commands
Build all grid search parameter configurations. :param script: String of command prefix, e.g. ``cxflow train -v -o log``. :param params: Iterable collection of strings in standard **cxflow** param form, e.g. ``'numerical_param=[1, 2]'`` or ``'text_param=["hello", "cio"]'``.
entailment
def grid_search(script: str, params: typing.Iterable[str], dry_run: bool=False) -> None: """ Build all grid search parameter configurations and optionally run them. :param script: String of command prefix, e.g. ``cxflow train -v -o log``. :param params: Iterable collection of strings in standard **cxflow** param form, e.g. ``'numerical_param=[1, 2]'`` or ``'text_param=["hello", "cio"]'``. :param dry_run: If set to ``True``, the built commands will only be printed instead of executed. """ commands = _build_grid_search_commands(script=script, params=params) if dry_run: logging.warning('Dry run') for command in commands: logging.info(command) else: for command in commands: try: completed_process = subprocess.run(command) logging.info('Command `%s` completed with exit code %d', command, completed_process.returncode) except Exception as _: # pylint: disable=broad-except logging.error('Command `%s` failed.', command)
Build all grid search parameter configurations and optionally run them. :param script: String of command prefix, e.g. ``cxflow train -v -o log``. :param params: Iterable collection of strings in standard **cxflow** param form, e.g. ``'numerical_param=[1, 2]'`` or ``'text_param=["hello", "cio"]'``. :param dry_run: If set to ``True``, the built commands will only be printed instead of executed.
entailment
def stream_info(self) -> None: """Check and report source names, dtypes and shapes of all the streams available.""" stream_names = [stream_name for stream_name in dir(self) if 'stream' in stream_name and stream_name != 'stream_info'] logging.info('Found %s stream candidates: %s', len(stream_names), stream_names) for stream_name in stream_names: try: stream_fn = getattr(self, stream_name) logging.info(stream_name) batch = next(iter(stream_fn())) rows = [] for key, value in batch.items(): try: value_arr = np.array(value) row = [key, value_arr.dtype, value_arr.shape] if value_arr.dtype.kind in 'bui': # boolean, unsigned, integer row.append('{} - {}'.format(value_arr.min(), value_arr.max())) elif value_arr.dtype.kind is 'f': row.append('{0:.2f} - {1:.2f}'.format(value_arr.min(), value_arr.max())) except ValueError: # np broadcasting failed (ragged array) value_arr = None row = [key, '{}'.format(type(value[0]).__name__), '({},)'.format(len(list(value)))] if value_arr is None or \ (value_arr.ndim > 0 and value_arr.shape[1:] != np.array(value_arr[0]).shape): logging.warning('*** stream source `%s` appears to be ragged (non-rectangular) ***', key) rows.append(row) for line in tabulate.tabulate(rows, headers=['name', 'dtype', 'shape', 'range'], tablefmt='grid').split('\n'): logging.info(line) except Exception: logging.warning('Exception was raised during checking stream `%s`, ' '(stack trace is displayed only with --verbose flag)', stream_name) logging.debug(traceback.format_exc())
Check and report source names, dtypes and shapes of all the streams available.
entailment
def parse_arg(arg: str) -> typing.Tuple[str, typing.Any]: """ Parse CLI argument in format ``key=value`` to ``(key, value)`` :param arg: CLI argument string :return: tuple (key, value) :raise: yaml.ParserError: on yaml parse error """ assert '=' in arg, 'Unrecognized argument `{}`. [name]=[value] expected.'.format(arg) key = arg[:arg.index('=')] value = yaml.load(arg[arg.index('=') + 1:]) return key, value
Parse CLI argument in format ``key=value`` to ``(key, value)`` :param arg: CLI argument string :return: tuple (key, value) :raise: yaml.ParserError: on yaml parse error
entailment
def load_config(config_file: str, additional_args: typing.Iterable[str]=()) -> dict: """ Load config from YAML ``config_file`` and extend/override it with the given ``additional_args``. :param config_file: path the YAML config file to be loaded :param additional_args: additional args which may extend or override the config loaded from the file. :return: configuration as dict """ config = load_yaml(config_file) for key_full, value in [parse_arg(arg) for arg in additional_args]: key_split = key_full.split('.') key_prefix = key_split[:-1] key = key_split[-1] conf = config for key_part in key_prefix: conf = conf[key_part] conf[key] = value return reload(config)
Load config from YAML ``config_file`` and extend/override it with the given ``additional_args``. :param config_file: path the YAML config file to be loaded :param additional_args: additional args which may extend or override the config loaded from the file. :return: configuration as dict
entailment
def find_config(config_path: str) -> str: """ Derive configuration file path from the given path and check its existence. The given path is expected to be either 1. path to the file 2. path to a dir, in such case the path is joined with ``CXF_CONFIG_FILE`` :param config_path: path to the configuration file or its parent directory :return: validated configuration file path """ if path.isdir(config_path): # dir specified instead of config file config_path = path.join(config_path, CXF_CONFIG_FILE) assert path.exists(config_path), '`{}` does not exist'.format(config_path) return config_path
Derive configuration file path from the given path and check its existence. The given path is expected to be either 1. path to the file 2. path to a dir, in such case the path is joined with ``CXF_CONFIG_FILE`` :param config_path: path to the configuration file or its parent directory :return: validated configuration file path
entailment
def fallback(message: str, ex: Exception) -> None: """ Fallback procedure when a cli command fails. :param message: message to be logged :param ex: Exception which caused the failure """ logging.error('%s', message) logging.exception('%s', ex) sys.exit(1)
Fallback procedure when a cli command fails. :param message: message to be logged :param ex: Exception which caused the failure
entailment
def _configure_dataset(self, data_root: str=None, download_urls: Iterable[str]=None, **kwargs) -> None: """ Save the passed values and use them as a default property implementation. :param data_root: directory to which the files will be downloaded :param download_urls: list of URLs to be downloaded """ self._data_root = data_root self._download_urls = download_urls
Save the passed values and use them as a default property implementation. :param data_root: directory to which the files will be downloaded :param download_urls: list of URLs to be downloaded
entailment
def resume(config_path: str, restore_from: Optional[str], cl_arguments: Iterable[str], output_root: str) -> None: """ Load config from the directory specified and start the training. :param config_path: path to the config file or the directory in which it is stored :param restore_from: backend-specific path to the already trained model to be restored from. If ``None`` is passed, it is inferred from the configuration file location as the directory it is located in. :param cl_arguments: additional command line arguments which will update the configuration :param output_root: output root in which the training directory will be created """ config = None try: config_path = find_config(config_path) restore_from = restore_from or path.dirname(config_path) config = load_config(config_file=config_path, additional_args=cl_arguments) validate_config(config) logging.debug('\tLoaded config: %s', config) except Exception as ex: # pylint: disable=broad-except fallback('Loading config failed', ex) run(config=config, output_root=output_root, restore_from=restore_from)
Load config from the directory specified and start the training. :param config_path: path to the config file or the directory in which it is stored :param restore_from: backend-specific path to the already trained model to be restored from. If ``None`` is passed, it is inferred from the configuration file location as the directory it is located in. :param cl_arguments: additional command line arguments which will update the configuration :param output_root: output root in which the training directory will be created
entailment
def after_epoch_profile(self, epoch_id: int, profile: TimeProfile, train_stream_name: str, extra_streams: Iterable[str]) -> None: """ After epoch profile event. This event provides opportunity to process time profile of the finished epoch. :param epoch_id: finished epoch id :param profile: dictionary of lists of event timings that were measured during the epoch :param extra_streams: enumeration of additional stream names """ pass
After epoch profile event. This event provides opportunity to process time profile of the finished epoch. :param epoch_id: finished epoch id :param profile: dictionary of lists of event timings that were measured during the epoch :param extra_streams: enumeration of additional stream names
entailment
def load_yaml(yaml_file: str) -> Any: """ Load YAML from file. :param yaml_file: path to YAML file :return: content of the YAML as dict/list """ with open(yaml_file, 'r') as file: return ruamel.yaml.load(file, ruamel.yaml.RoundTripLoader)
Load YAML from file. :param yaml_file: path to YAML file :return: content of the YAML as dict/list
entailment
def yaml_to_file(data: Mapping, output_dir: str, name: str) -> str: """ Save the given object to the given path in YAML. :param data: dict/list to be dumped :param output_dir: target output directory :param name: target filename :return: target path """ dumped_config_f = path.join(output_dir, name) with open(dumped_config_f, 'w') as file: yaml.dump(data, file, Dumper=ruamel.yaml.RoundTripDumper) return dumped_config_f
Save the given object to the given path in YAML. :param data: dict/list to be dumped :param output_dir: target output directory :param name: target filename :return: target path
entailment
def yaml_to_str(data: Mapping) -> str: """ Return the given given config as YAML str. :param data: configuration dict :return: given configuration as yaml str """ return yaml.dump(data, Dumper=ruamel.yaml.RoundTripDumper)
Return the given given config as YAML str. :param data: configuration dict :return: given configuration as yaml str
entailment
def make_simple(data: Any) -> Any: """ Substitute all the references in the given data (typically a mapping or sequence) with the actual values. This is useful, if you loaded a yaml with RoundTripLoader and you need to dump part of it safely. :param data: data to be made simple (dict instead of CommentedMap etc.) :return: simplified data """ return yaml.load(yaml.dump(data, Dumper=ruamel.yaml.RoundTripDumper), Loader=ruamel.yaml.Loader)
Substitute all the references in the given data (typically a mapping or sequence) with the actual values. This is useful, if you loaded a yaml with RoundTripLoader and you need to dump part of it safely. :param data: data to be made simple (dict instead of CommentedMap etc.) :return: simplified data
entailment
def reload(data: Any) -> Any: """ Dump and load yaml data. This is useful to avoid many anchor parsing bugs. When you edit a yaml config, reload it to make sure the changes are propagated to anchor expansions. :param data: data to be reloaded :return: reloaded data """ return yaml.load(yaml.dump(data, Dumper=ruamel.yaml.RoundTripDumper), Loader=ruamel.yaml.RoundTripLoader)
Dump and load yaml data. This is useful to avoid many anchor parsing bugs. When you edit a yaml config, reload it to make sure the changes are propagated to anchor expansions. :param data: data to be reloaded :return: reloaded data
entailment
def after_epoch(self, epoch_id: int, epoch_data: EpochData) -> None: """ Call :py:meth:`_on_plateau_action` if the ``long_term`` variable mean is lower/greater than the ``short_term`` mean. """ super().after_epoch(epoch_id=epoch_id, epoch_data=epoch_data) self._saved_loss.append(epoch_data[self._stream][self._variable][OnPlateau._AGGREGATION]) long_mean = np.mean(self._saved_loss[-self._long_term:]) short_mean = np.mean(self._saved_loss[-self._short_term:]) if self._objective == 'min' and long_mean < short_mean: self._on_plateau_action(epoch_id=epoch_id, epoch_data=epoch_data) elif self._objective == 'max' and long_mean > short_mean: self._on_plateau_action(epoch_id=epoch_id, epoch_data=epoch_data)
Call :py:meth:`_on_plateau_action` if the ``long_term`` variable mean is lower/greater than the ``short_term`` mean.
entailment
def _is_nan(self, variable: str, data) -> bool: """ Recursively search passed data and find NaNs. :param variable: name of variable to be checked :param data: data object (dict, list, scalar) :return: `True` if there is a NaN value in the data; `False` otherwise. :raise ValueError: if the variable value is of unsupported type and ``on_unknown_type`` is set to ``error`` """ if isinstance(data, np.ndarray) or isinstance(data, list): return any(np.isnan(data)) or (self._stop_on_inf and any(np.isinf(data))) elif np.isscalar(data): return np.isnan(data) or (self._stop_on_inf and np.isinf(data)) elif isinstance(data, dict): return any([self._is_nan(key, value) for key, value in data.items()]) else: message = 'Variable `{}` of type `{}` can not be checked for NaNs.'.format(variable, type(data)) if self._on_unkown_type == 'warn': logging.warning(message) elif self._on_unkown_type == 'error': raise ValueError(message) return False
Recursively search passed data and find NaNs. :param variable: name of variable to be checked :param data: data object (dict, list, scalar) :return: `True` if there is a NaN value in the data; `False` otherwise. :raise ValueError: if the variable value is of unsupported type and ``on_unknown_type`` is set to ``error``
entailment
def _check_nan(self, epoch_data: EpochData) -> None: """ Raise an exception when some of the monitored data is NaN. :param epoch_data: epoch data checked :raise KeyError: if the specified variable is not found in the stream :raise ValueError: if the variable value is of unsupported type and ``self._on_unknown_type`` is set to ``error`` """ for stream_name in epoch_data.keys(): stream_data = epoch_data[stream_name] variables = self._variables if self._variables is not None else stream_data.keys() for variable in variables: if variable not in stream_data: raise KeyError('Variable `{}` to be nan-checked was not found in the batch data for stream `{}`. ' 'Available variables are `{}`.'.format(variable, stream_name, stream_data.keys())) value = stream_data[variable] if self._is_nan(variable, value): raise TrainingTerminated('Variable `{}` is NaN.'.format(variable))
Raise an exception when some of the monitored data is NaN. :param epoch_data: epoch data checked :raise KeyError: if the specified variable is not found in the stream :raise ValueError: if the variable value is of unsupported type and ``self._on_unknown_type`` is set to ``error``
entailment
def after_epoch(self, epoch_data: EpochData, **kwargs) -> None: """ If initialized to check after each epoch, stop the training once the epoch data contains a monitored variable equal to NaN. :param epoch_data: epoch data to be checked """ if self._after_epoch: self._check_nan(epoch_data)
If initialized to check after each epoch, stop the training once the epoch data contains a monitored variable equal to NaN. :param epoch_data: epoch data to be checked
entailment
def after_batch(self, stream_name: str, batch_data) -> None: """ If initialized to check after each batch, stop the training once the batch data contains a monitored variable equal to NaN. :param stream_name: name of the stream to be checked :param batch_data: batch data to be checked """ if self._after_batch: self._check_nan({stream_name: batch_data})
If initialized to check after each batch, stop the training once the batch data contains a monitored variable equal to NaN. :param stream_name: name of the stream to be checked :param batch_data: batch data to be checked
entailment
def after_epoch(self, epoch_id: int, **kwargs) -> None: """ Call ``_after_n_epoch`` method every ``n_epochs`` epoch. :param epoch_id: number of the processed epoch """ if epoch_id % self._n_epochs == 0: self._after_n_epoch(epoch_id=epoch_id, **kwargs)
Call ``_after_n_epoch`` method every ``n_epochs`` epoch. :param epoch_id: number of the processed epoch
entailment
def path_total_size(path_: str) -> int: """Compute total size of the given file/dir.""" if path.isfile(path_): return path.getsize(path_) total_size = 0 for root_dir, _, files in os.walk(path_): for file_ in files: total_size += path.getsize(path.join(root_dir, file_)) return total_size
Compute total size of the given file/dir.
entailment
def humanize_filesize(filesize: int) -> Tuple[str, str]: """Return human readable pair of size and unit from the given filesize in bytes.""" for unit in ['', 'K', 'M', 'G', 'T', 'P', 'E', 'Z']: if filesize < 1024.0: return '{:3.1f}'.format(filesize), unit+'B' filesize /= 1024.0
Return human readable pair of size and unit from the given filesize in bytes.
entailment
def is_train_dir(dir_: str) -> bool: """Test if the given dir contains training artifacts.""" return path.exists(path.join(dir_, CXF_CONFIG_FILE)) and \ path.exists(path.join(dir_, CXF_TRACE_FILE)) and \ path.exists(path.join(dir_, CXF_LOG_FILE))
Test if the given dir contains training artifacts.
entailment
def walk_train_dirs(root_dir: str) -> Iterable[Tuple[str, Iterable[str]]]: """ Modify os.walk with the following: - return only root_dir and sub-dirs - return only training sub-dirs - stop recursion at training dirs :param root_dir: root dir to be walked :return: generator of (root_dir, training sub-dirs) pairs """ if is_train_dir(root_dir): yield '', [root_dir] return for dir_, subdirs, _ in os.walk(root_dir, topdown=True): # filter train sub-dirs train_subdirs = [subdir for subdir in subdirs if is_train_dir(path.join(dir_, subdir))] # stop the recursion at the train sub-dirs for subdir in train_subdirs: subdirs.remove(subdir) yield dir_, train_subdirs
Modify os.walk with the following: - return only root_dir and sub-dirs - return only training sub-dirs - stop recursion at training dirs :param root_dir: root dir to be walked :return: generator of (root_dir, training sub-dirs) pairs
entailment
def _print_trainings_long(trainings: Iterable[Tuple[str, dict, TrainingTrace]]) -> None: """ Print a plain table with the details of the given trainings. :param trainings: iterable of tuples (train_dir, configuration dict, trace) """ long_table = [] for train_dir, config, trace in trainings: start_datetime, end_datetime = trace[TrainingTraceKeys.TRAIN_BEGIN], trace[TrainingTraceKeys.TRAIN_END] if start_datetime: age = format_timedelta(datetime.now() - start_datetime) + ' ago' if end_datetime: duration = format_timedelta(end_datetime - start_datetime) else: duration = CXF_NA_STR else: age = CXF_NA_STR duration = CXF_NA_STR epochs_done = trace[TrainingTraceKeys.EPOCHS_DONE] if trace[TrainingTraceKeys.EPOCHS_DONE] else 0 long_table.append([path.basename(train_dir)] + list(map(lambda fq_name: fq_name.split('.')[-1], get_classes(config))) + [age, duration, epochs_done]) print(tabulate(long_table, tablefmt='plain'))
Print a plain table with the details of the given trainings. :param trainings: iterable of tuples (train_dir, configuration dict, trace)
entailment
def _ls_print_listing(dir_: str, recursive: bool, all_: bool, long: bool) -> List[Tuple[str, dict, TrainingTrace]]: """ Print names of the train dirs contained in the given dir. :param dir_: dir to be listed :param recursive: walk recursively in sub-directories, stop at train dirs (--recursive option) :param all_: include train dirs with no epochs done (--all option) :param long: list more details including model name, model and dataset classes, age, duration and epochs done (--long option) :return: list of found training tuples (train_dir, configuration dict, trace) """ all_trainings = [] for root_dir, train_dirs in walk_train_dirs(dir_): if train_dirs: if recursive: print(root_dir + ':') trainings = [(train_dir, load_config(path.join(train_dir, CXF_CONFIG_FILE), []), TrainingTrace.from_file(path.join(train_dir, CXF_TRACE_FILE))) for train_dir in [os.path.join(root_dir, train_dir) for train_dir in train_dirs]] if not all_: trainings = [train_dir for train_dir in trainings if train_dir[2][TrainingTraceKeys.EPOCHS_DONE]] if long: print('total {}'.format(len(trainings))) _print_trainings_long(trainings) else: for train_dir, _, _ in trainings: print(path.basename(train_dir)) all_trainings.extend(trainings) if recursive: print() if not recursive: break return all_trainings
Print names of the train dirs contained in the given dir. :param dir_: dir to be listed :param recursive: walk recursively in sub-directories, stop at train dirs (--recursive option) :param all_: include train dirs with no epochs done (--all option) :param long: list more details including model name, model and dataset classes, age, duration and epochs done (--long option) :return: list of found training tuples (train_dir, configuration dict, trace)
entailment
def _ls_print_summary(all_trainings: List[Tuple[str, dict, TrainingTrace]]) -> None: """ Print trainings summary. In particular print tables summarizing the number of trainings with - particular model names - particular combinations of models and datasets :param all_trainings: a list of training tuples (train_dir, configuration dict, trace) """ counts_by_name = defaultdict(int) counts_by_classes = defaultdict(int) for _, config, _ in all_trainings: counts_by_name[get_model_name(config)] += 1 counts_by_classes[get_classes(config)] += 1 print_boxed('summary') print() counts_table = [[name, count] for name, count in counts_by_name.items()] print(tabulate(counts_table, headers=['model.name', 'count'], tablefmt='grid')) print() counts_table = [[classes[0], classes[1], count] for classes, count in counts_by_classes.items()] print(tabulate(counts_table, headers=['model.class', 'dataset.class', 'count'], tablefmt='grid')) print()
Print trainings summary. In particular print tables summarizing the number of trainings with - particular model names - particular combinations of models and datasets :param all_trainings: a list of training tuples (train_dir, configuration dict, trace)
entailment
def _ls_print_verbose(training: Tuple[str, dict, str]) -> None: """ Print config and artifacts info from the given training tuple (train_dir, configuration dict, trace). :param training: training tuple (train_dir, configuration dict, trace) """ train_dir, config, _ = training print_boxed('config') print(yaml_to_str(config)) print() print_boxed('artifacts') _, dirs, files = next(os.walk(train_dir)) artifacts = [('d', dir) for dir in dirs] + \ [('-', file_) for file_ in files if file_ not in [CXF_CONFIG_FILE, CXF_LOG_FILE, CXF_TRACE_FILE]] artifacts = [(type_, name) + humanize_filesize(path_total_size(path.join(train_dir, name))) for type_, name in artifacts] print(tabulate(artifacts, tablefmt='plain', floatfmt='3.1f')) print()
Print config and artifacts info from the given training tuple (train_dir, configuration dict, trace). :param training: training tuple (train_dir, configuration dict, trace)
entailment
def list_train_dirs(dir_: str, recursive: bool, all_: bool, long: bool, verbose: bool) -> None: """ List training dirs contained in the given dir with options and outputs similar to the regular `ls` command. The function is accessible through cxflow CLI `cxflow ls`. :param dir_: dir to be listed :param recursive: walk recursively in sub-directories, stop at train dirs (--recursive option) :param all_: include train dirs with no epochs done (--all option) :param long: list more details including model name, odel and dataset class, age, duration and epochs done (--long option) :param verbose: print more verbose output with list of additional artifacts and training config, applicable only when a single train dir is listed (--verbose option) """ if verbose: long = True if dir_ == CXF_DEFAULT_LOG_DIR and not path.exists(CXF_DEFAULT_LOG_DIR): print('The default log directory `{}` does not exist.\n' 'Consider specifying the directory to be listed as an argument.'.format(CXF_DEFAULT_LOG_DIR)) quit(1) if not path.exists(dir_): print('Specified dir `{}` does not exist'.format(dir_)) quit(1) all_trainings = _ls_print_listing(dir_, recursive, all_, long) if long and len(all_trainings) > 1: if not recursive: print() _ls_print_summary(all_trainings) if verbose and len(all_trainings) == 1: if not recursive: print() _ls_print_verbose(all_trainings[0])
List training dirs contained in the given dir with options and outputs similar to the regular `ls` command. The function is accessible through cxflow CLI `cxflow ls`. :param dir_: dir to be listed :param recursive: walk recursively in sub-directories, stop at train dirs (--recursive option) :param all_: include train dirs with no epochs done (--all option) :param long: list more details including model name, odel and dataset class, age, duration and epochs done (--long option) :param verbose: print more verbose output with list of additional artifacts and training config, applicable only when a single train dir is listed (--verbose option)
entailment
def after_batch(self, stream_name: str, batch_data: Batch): """ Extend the accumulated variables with the given batch data. :param stream_name: stream name; e.g. ``train`` or any other... :param batch_data: batch data = stream sources + model outputs :raise KeyError: if the variables to be aggregated are missing :raise TypeError: if the variable value is not iterable (e.g. it is only a scalar) """ for variable in self._variables: if variable in batch_data: value = batch_data[variable] if not hasattr(value, '__iter__'): raise TypeError('Variable `{}` to be accumulated is not iterable.'.format(variable)) self._accumulator[stream_name][variable] += list(value) else: raise KeyError('Variable `{}` to be accumulated was not found in the batch data. ' 'Available variables are `{}`.'.format(variable, batch_data.keys()))
Extend the accumulated variables with the given batch data. :param stream_name: stream name; e.g. ``train`` or any other... :param batch_data: batch data = stream sources + model outputs :raise KeyError: if the variables to be aggregated are missing :raise TypeError: if the variable value is not iterable (e.g. it is only a scalar)
entailment
def invoke_dataset_method(config_path: str, method_name: str, output_root: str, cl_arguments: Iterable[str]) -> None: """ Create the specified dataset and invoke its specified method. :param config_path: path to the config file or the directory in which it is stored :param method_name: name of the method to be invoked on the specified dataset :param cl_arguments: additional command line arguments which will update the configuration :param output_root: output root in which the training directory will be created """ config = dataset = method = output_dir = None try: config_path = find_config(config_path) config = load_config(config_file=config_path, additional_args=cl_arguments) assert 'dataset' in config, '`dataset` section not present in the config' logging.debug('\tLoaded config: %s', config) except Exception as ex: # pylint: disable=broad-except fallback('Loading config failed', ex) try: dataset = create_dataset(config) except Exception as ex: # pylint: disable=broad-except fallback('Creating dataset failed', ex) try: method = getattr(dataset, method_name) except AttributeError as ex: fallback('Method `{}` not found in the dataset'.format(method_name), ex) try: method() except Exception as ex: # pylint: disable=broad-except fallback('Exception occurred during method `{}` invocation'.format(method_name), ex)
Create the specified dataset and invoke its specified method. :param config_path: path to the config file or the directory in which it is stored :param method_name: name of the method to be invoked on the specified dataset :param cl_arguments: additional command line arguments which will update the configuration :param output_root: output root in which the training directory will be created
entailment
def _signal_handler(self, *_) -> None: """ On the first signal, increase the ``self._num_signals`` counter. Call ``sys.exit`` on any subsequent signal. """ if self._num_signals == 0: logging.warning('Interrupt signal caught - training will be terminated') logging.warning('Another interrupt signal will terminate the program immediately') self._num_signals += 1 else: logging.error('Another interrupt signal caught - terminating program immediately') sys.exit(2)
On the first signal, increase the ``self._num_signals`` counter. Call ``sys.exit`` on any subsequent signal.
entailment
def create_output_dir(config: dict, output_root: str, default_model_name: str='Unnamed') -> str: """ Create output_dir under the given ``output_root`` and - dump the given config to YAML file under this dir - register a file logger logging to a file under this dir :param config: config to be dumped :param output_root: dir wherein output_dir shall be created :param default_model_name: name to be used when `model.name` is not found in the config :return: path to the created output_dir """ logging.info('Creating output dir') # create output dir model_name = default_model_name if 'name' not in config['model']: logging.warning('\tmodel.name not found in config, defaulting to: %s', model_name) else: model_name = config['model']['name'] if not os.path.exists(output_root): logging.info('\tOutput root folder "%s" does not exist and will be created', output_root) os.makedirs(output_root) # keep trying to create new output dir until it succeeds # this is neccessary due to improbable yet possible output dir name conflicts while True: try: output_dir = path.join(output_root, '{}_{}_{}'.format(datetime.now().strftime('%Y-%m-%d-%H-%M-%S'), model_name, get_random_name())) os.mkdir(output_dir) break except OSError as ex: if ex.errno != errno.EEXIST: raise ex logging.info('\tOutput dir: %s', output_dir) # create file logger file_handler = logging.FileHandler(path.join(output_dir, CXF_LOG_FILE)) file_handler.setFormatter(logging.Formatter(CXF_LOG_FORMAT, datefmt=CXF_LOG_DATE_FORMAT)) logging.getLogger().addHandler(file_handler) return output_dir
Create output_dir under the given ``output_root`` and - dump the given config to YAML file under this dir - register a file logger logging to a file under this dir :param config: config to be dumped :param output_root: dir wherein output_dir shall be created :param default_model_name: name to be used when `model.name` is not found in the config :return: path to the created output_dir
entailment
def create_dataset(config: dict, output_dir: Optional[str]=None) -> AbstractDataset: """ Create a dataset object according to the given config. Dataset config section and the `output_dir` are passed to the constructor in a single YAML-encoded string. :param config: config dict with dataset config :param output_dir: path to the training output dir or None :return: dataset object """ logging.info('Creating dataset') dataset_config = make_simple(config)['dataset'] assert 'class' in dataset_config, '`dataset.class` not present in the config' dataset_module, dataset_class = parse_fully_qualified_name(dataset_config['class']) if 'output_dir' in dataset_config: raise ValueError('The `output_dir` key is reserved and can not be used in dataset configuration.') dataset_config = {'output_dir': output_dir, **config['dataset']} del dataset_config['class'] dataset = create_object(dataset_module, dataset_class, args=(yaml_to_str(dataset_config),)) logging.info('\t%s created', type(dataset).__name__) return dataset
Create a dataset object according to the given config. Dataset config section and the `output_dir` are passed to the constructor in a single YAML-encoded string. :param config: config dict with dataset config :param output_dir: path to the training output dir or None :return: dataset object
entailment
def create_model(config: dict, output_dir: Optional[str], dataset: AbstractDataset, restore_from: Optional[str]=None) -> AbstractModel: """ Create a model object either from scratch of from the checkpoint in ``resume_dir``. Cxflow allows the following scenarios 1. Create model: leave ``restore_from=None`` and specify ``class``; 2. Restore model: specify ``restore_from`` which is a backend-specific path to (a directory with) the saved model. :param config: config dict with model config :param output_dir: path to the training output dir :param dataset: dataset object implementing the :py:class:`cxflow.datasets.AbstractDataset` concept :param restore_from: from whence the model should be restored (backend-specific information) :return: model object """ logging.info('Creating a model') model_config = config['model'] # workaround for ruamel.yaml expansion bug; see #222 model_config = dict(model_config.items()) assert 'class' in model_config, '`model.class` not present in the config' model_module, model_class = parse_fully_qualified_name(model_config['class']) # create model kwargs (without `class` and `name`) model_kwargs = {'dataset': dataset, 'log_dir': output_dir, 'restore_from': restore_from, **model_config} del model_kwargs['class'] if 'name' in model_kwargs: del model_kwargs['name'] try: model = create_object(model_module, model_class, kwargs=model_kwargs) except (ImportError, AttributeError) as ex: if restore_from is None: # training case raise ImportError('Cannot create model from the specified model module `{}` and class `{}`.'.format( model_module, model_class)) from ex else: # restore cases (resume, predict) logging.warning('Cannot create model from the specified model class `%s`.', model_config['class']) assert 'restore_fallback' in model_config, '`model.restore_fallback` not present in the config' logging.info('Trying to restore with fallback `%s` instead.', model_config['restore_fallback']) try: # try fallback class fallback_module, fallback_class = parse_fully_qualified_name(model_config['restore_fallback']) model = create_object(fallback_module, fallback_class, kwargs=model_kwargs) except (ImportError, AttributeError) as ex: # if fallback module/class specified but it fails raise ImportError('Cannot create model from the specified restore_fallback `{}`.'.format( model_config['restore_fallback'],)) from ex logging.info('\t%s created', type(model).__name__) return model
Create a model object either from scratch of from the checkpoint in ``resume_dir``. Cxflow allows the following scenarios 1. Create model: leave ``restore_from=None`` and specify ``class``; 2. Restore model: specify ``restore_from`` which is a backend-specific path to (a directory with) the saved model. :param config: config dict with model config :param output_dir: path to the training output dir :param dataset: dataset object implementing the :py:class:`cxflow.datasets.AbstractDataset` concept :param restore_from: from whence the model should be restored (backend-specific information) :return: model object
entailment
def create_hooks(config: dict, model: AbstractModel, dataset: AbstractDataset, output_dir: str) -> Iterable[AbstractHook]: """ Create hooks specified in ``config['hooks']`` list. Hook config entries may be one of the following types: .. code-block:: yaml :caption: A hook with default args specified only by its name as a string; e.g. hooks: - LogVariables - cxflow_tensorflow.WriteTensorBoard .. code-block:: yaml :caption: A hook with custom args as a dict name -> args; e.g. hooks: - StopAfter: n_epochs: 10 :param config: config dict :param model: model object to be passed to the hooks :param dataset: dataset object to be passed to hooks :param output_dir: training output dir available to the hooks :return: list of hook objects """ logging.info('Creating hooks') hooks = [] if 'hooks' in config: for hook_config in config['hooks']: if isinstance(hook_config, str): hook_config = {hook_config: {}} assert len(hook_config) == 1, 'Hook configuration must have exactly one key (fully qualified name).' hook_path, hook_params = next(iter(hook_config.items())) if hook_params is None: logging.warning('\t\t Empty config of `%s` hook', hook_path) hook_params = {} # workaround for ruamel.yaml expansion bug; see #222 hook_params = dict(hook_params.items()) hook_module, hook_class = parse_fully_qualified_name(hook_path) # find the hook module if not specified if hook_module is None: hook_module = get_class_module(CXF_HOOKS_MODULE, hook_class) logging.debug('\tFound hook module `%s` for class `%s`', hook_module, hook_class) if hook_module is None: raise ValueError('Can`t find hook module for hook class `{}`. ' 'Make sure it is defined under `{}` sub-modules.' .format(hook_class, CXF_HOOKS_MODULE)) # create hook kwargs hook_kwargs = {'dataset': dataset, 'model': model, 'output_dir': output_dir, **hook_params} # create new hook try: hook = create_object(hook_module, hook_class, kwargs=hook_kwargs) hooks.append(hook) logging.info('\t%s created', type(hooks[-1]).__name__) except (ValueError, KeyError, TypeError, NameError, AttributeError, AssertionError, ImportError) as ex: logging.error('\tFailed to create a hook from config `%s`', hook_config) raise ex return hooks
Create hooks specified in ``config['hooks']`` list. Hook config entries may be one of the following types: .. code-block:: yaml :caption: A hook with default args specified only by its name as a string; e.g. hooks: - LogVariables - cxflow_tensorflow.WriteTensorBoard .. code-block:: yaml :caption: A hook with custom args as a dict name -> args; e.g. hooks: - StopAfter: n_epochs: 10 :param config: config dict :param model: model object to be passed to the hooks :param dataset: dataset object to be passed to hooks :param output_dir: training output dir available to the hooks :return: list of hook objects
entailment
def run(config: dict, output_root: str, restore_from: str=None, eval: Optional[str]=None) -> None: """ Run **cxflow** training configured by the passed `config`. Unique ``output_dir`` for this training is created under the given ``output_root`` dir wherein all the training outputs are saved. The output dir name will be roughly ``[model.name]_[time]``. The training procedure consists of the following steps: 1. Set up (create output dir and file logger, dump the loaded config into the output dir) 2. Create dataset (YAML string with ``dataset`` and ``log_dir`` configs are passed to the dataset constructor) 3. Create (or restore) model (dataset, ``log_dir`` and model config is passed to the constructor) 4. Create all the training hooks 5. Create the ``MainLoop`` object 6. Run the main loop If any of the steps fails, the training is terminated. After the training procedure finishes, the output dir will contain the following: - ``train_log.txt`` with entry point and main loop logs (same as the stderr) - dumped YAML config Additional outputs created by hooks, dataset or tensorflow may include: - ``dataset_log.txt`` with info about dataset/stream creation - model checkpoint(s) - TensorBoard log file - TensorFlow event log :param config: configuration :param output_root: dir under which output_dir shall be created :param restore_from: from whence the model should be restored (backend-specific information) :param eval: optional name of the stream to be evaluated """ output_dir = dataset = model = hooks = main_loop = None try: output_dir = create_output_dir(config=config, output_root=output_root) except Exception as ex: # pylint: disable=broad-except fallback('Failed to create output dir', ex) try: dataset = create_dataset(config=config, output_dir=output_dir) except Exception as ex: # pylint: disable=broad-except fallback('Creating dataset failed', ex) try: model = create_model(config=config, output_dir=output_dir, dataset=dataset, restore_from=restore_from) except Exception as ex: # pylint: disable=broad-except fallback('Creating model failed', ex) try: # save the config to file # modify the config so that it contains fallback information config['model']['restore_fallback'] = model.restore_fallback yaml_to_file(data=config, output_dir=output_dir, name=CXF_CONFIG_FILE) except Exception as ex: # pylint: disable=broad-except fallback('Saving config failed', ex) try: hooks = create_hooks(config=config, model=model, dataset=dataset, output_dir=output_dir) except Exception as ex: # pylint: disable=broad-except fallback('Creating hooks failed', ex) try: logging.info('Creating main loop') kwargs = config['main_loop'] if 'main_loop' in config else {} if eval is not None: kwargs['extra_streams'] = [] main_loop = MainLoop(model=model, dataset=dataset, hooks=hooks, **kwargs) except Exception as ex: # pylint: disable=broad-except fallback('Creating main loop failed', ex) if eval is not None: try: with main_loop: logging.info('Running the evaluation of stream `%s`', eval) main_loop.run_evaluation(eval) except Exception as ex: # pylint: disable=broad-except fallback('Running the evaluation failed', ex) else: trace = TrainingTrace(output_dir) try: with main_loop: logging.info('Running the training') trace[TrainingTraceKeys.TRAIN_BEGIN] = datetime.now() main_loop.run_training(trace) trace[TrainingTraceKeys.EXIT_STATUS] = 0 except Exception as ex: # pylint: disable=broad-except trace[TrainingTraceKeys.EXIT_STATUS] = 1 fallback('Running the training failed', ex) except SystemExit as ex: trace[TrainingTraceKeys.EXIT_STATUS] = ex.code finally: trace[TrainingTraceKeys.EPOCHS_DONE] = main_loop.epochs_done trace[TrainingTraceKeys.TRAIN_END] = datetime.now()
Run **cxflow** training configured by the passed `config`. Unique ``output_dir`` for this training is created under the given ``output_root`` dir wherein all the training outputs are saved. The output dir name will be roughly ``[model.name]_[time]``. The training procedure consists of the following steps: 1. Set up (create output dir and file logger, dump the loaded config into the output dir) 2. Create dataset (YAML string with ``dataset`` and ``log_dir`` configs are passed to the dataset constructor) 3. Create (or restore) model (dataset, ``log_dir`` and model config is passed to the constructor) 4. Create all the training hooks 5. Create the ``MainLoop`` object 6. Run the main loop If any of the steps fails, the training is terminated. After the training procedure finishes, the output dir will contain the following: - ``train_log.txt`` with entry point and main loop logs (same as the stderr) - dumped YAML config Additional outputs created by hooks, dataset or tensorflow may include: - ``dataset_log.txt`` with info about dataset/stream creation - model checkpoint(s) - TensorBoard log file - TensorFlow event log :param config: configuration :param output_root: dir under which output_dir shall be created :param restore_from: from whence the model should be restored (backend-specific information) :param eval: optional name of the stream to be evaluated
entailment
def _safe_rmtree(dir_: str): """Wrap ``shutil.rmtree`` to inform user about (un)success.""" try: rmtree(dir_) except OSError: logging.warning('\t\t Skipping %s due to OSError', dir_) else: logging.debug('\t\t Deleted %s', dir_)
Wrap ``shutil.rmtree`` to inform user about (un)success.
entailment
def _prune_subdirs(dir_: str) -> None: """ Delete all subdirs in training log dirs. :param dir_: dir with training log dirs """ for logdir in [path.join(dir_, f) for f in listdir(dir_) if is_train_dir(path.join(dir_, f))]: for subdir in [path.join(logdir, f) for f in listdir(logdir) if path.isdir(path.join(logdir, f))]: _safe_rmtree(subdir)
Delete all subdirs in training log dirs. :param dir_: dir with training log dirs
entailment
def _prune(dir_: str, epochs: int) -> None: """ Delete all training dirs with incomplete training artifacts or with less than specified epochs done. :param dir_: dir with training log dirs :param epochs: minimum number of finished epochs to keep the training logs :return: number of log dirs pruned """ for logdir in [path.join(dir_, f) for f in listdir(dir_) if path.isdir(path.join(dir_, f))]: if not is_train_dir(logdir): _safe_rmtree(logdir) else: trace_path = path.join(logdir, CXF_TRACE_FILE) try: epochs_done = TrainingTrace.from_file(trace_path)[TrainingTraceKeys.EPOCHS_DONE] except (KeyError, TypeError): epochs_done = 0 if not epochs_done or epochs_done < epochs: _safe_rmtree(logdir)
Delete all training dirs with incomplete training artifacts or with less than specified epochs done. :param dir_: dir with training log dirs :param epochs: minimum number of finished epochs to keep the training logs :return: number of log dirs pruned
entailment
def prune_train_dirs(dir_: str, epochs: int, subdirs: bool) -> None: """ Prune training log dirs contained in the given dir. The function is accessible through cxflow CLI `cxflow prune`. :param dir_: dir to be pruned :param epochs: minimum number of finished epochs to keep the training logs :param subdirs: delete subdirs in training log dirs """ if dir_ == CXF_DEFAULT_LOG_DIR and not path.exists(CXF_DEFAULT_LOG_DIR): print('The default log directory `{}` does not exist.\n' 'Consider specifying the directory to be listed as an argument.'.format(CXF_DEFAULT_LOG_DIR)) quit(1) if not path.exists(dir_): print('Specified dir `{}` does not exist'.format(dir_)) quit(1) _prune(dir_, epochs) if subdirs: _prune_subdirs(dir_)
Prune training log dirs contained in the given dir. The function is accessible through cxflow CLI `cxflow prune`. :param dir_: dir to be pruned :param epochs: minimum number of finished epochs to keep the training logs :param subdirs: delete subdirs in training log dirs
entailment
def output_names(self) -> Iterable[str]: """List of model output names.""" self._load_models() return chain.from_iterable(map(lambda m: m.output_names, self._models))
List of model output names.
entailment
def run(self, batch: Batch, train: bool=False, stream: StreamWrapper=None) -> Batch: """ Run all the models in-order and return accumulated outputs. N-th model is fed with the original inputs and outputs of all the models that were run before it. .. warning:: :py:class:`Sequence` model can not be trained. :param batch: batch to be processed :param train: ``True`` if this batch should be used for model update, ``False`` otherwise :param stream: stream wrapper (useful for precise buffer management) :return: accumulated model outputs :raise ValueError: if the ``train`` flag is set to ``True`` """ if train: raise ValueError('Ensemble model cannot be trained.') self._load_models() # run all the models in-order current_batch = dict(copy.deepcopy(batch)) for model in self._models: current_batch.update(model.run(current_batch, False, None)) return {key: current_batch[key] for key in self.output_names}
Run all the models in-order and return accumulated outputs. N-th model is fed with the original inputs and outputs of all the models that were run before it. .. warning:: :py:class:`Sequence` model can not be trained. :param batch: batch to be processed :param train: ``True`` if this batch should be used for model update, ``False`` otherwise :param stream: stream wrapper (useful for precise buffer management) :return: accumulated model outputs :raise ValueError: if the ``train`` flag is set to ``True``
entailment
def _log_variables(self, epoch_data: EpochData): """ Log variables from the epoch data. .. warning:: At the moment, only scalars and dicts of scalars are properly formatted and logged. Other value types are ignored by default. One may set ``on_unknown_type`` to ``str`` in order to log all the variables anyways. :param epoch_data: epoch data to be logged :raise KeyError: if the specified variable is not found in the stream :raise TypeError: if the variable value is of unsupported type and ``self._on_unknown_type`` is set to ``error`` """ for stream_name in epoch_data.keys(): stream_data = epoch_data[stream_name] variables = self._variables if self._variables is not None else stream_data.keys() for variable in variables: if variable not in stream_data: raise KeyError('Variable `{}` to be logged was not found in the batch data for stream `{}`. ' 'Available variables are `{}`.'.format(variable, stream_name, stream_data.keys())) value = stream_data[variable] if np.isscalar(value): logging.info('\t%s %s: %f', stream_name, variable, value) elif isinstance(value, dict): keys = list(value.keys()) if len(keys) == 1: logging.info('\t%s %s %s: %f', stream_name, variable, keys[0], value[keys[0]]) else: logging.info('\t%s %s:', stream_name, variable) for key, val in value.items(): logging.info('\t\t%s: %f', key, val) else: if self._on_unknown_type == 'error': raise TypeError('Variable type `{}` can not be logged. Variable name: `{}`.' .format(type(value).__name__, variable)) elif self._on_unknown_type == 'warn': logging.warning('Variable type `%s` can not be logged. Variable name: `%s`.', type(value).__name__, variable) elif self._on_unknown_type == 'str': logging.info('\t%s %s: %s', stream_name, variable, value)
Log variables from the epoch data. .. warning:: At the moment, only scalars and dicts of scalars are properly formatted and logged. Other value types are ignored by default. One may set ``on_unknown_type`` to ``str`` in order to log all the variables anyways. :param epoch_data: epoch data to be logged :raise KeyError: if the specified variable is not found in the stream :raise TypeError: if the variable value is of unsupported type and ``self._on_unknown_type`` is set to ``error``
entailment
def get_reference_end_from_cigar(reference_start, cigar): ''' This returns the coordinate just past the last aligned base. This matches the behavior of pysam's reference_end method ''' reference_end = reference_start # iterate through cigartuple for i in xrange(len(cigar)): k, n = cigar[i] if k in (0,2,3,7,8): # M, D, N, =, X reference_end += n return reference_end
This returns the coordinate just past the last aligned base. This matches the behavior of pysam's reference_end method
entailment
def set_order_by_clip(self, a, b): ''' Determine which SplitPiece is the leftmost based on the side of the longest clipping operation ''' if self.is_left_clip(a.cigar): self.query_left = b self.query_right = a else: self.query_left = a self.query_right = b
Determine which SplitPiece is the leftmost based on the side of the longest clipping operation
entailment
def is_left_clip(self, cigar): ''' whether the left side of the read (w/ respect to reference) is clipped. Clipping side is determined as the side with the longest clip. Adjacent clipping operations are not considered ''' left_tuple = cigar[0] right_tuple = cigar[-1] left_clipped = self.is_clip_op(left_tuple[0]) right_clipped = self.is_clip_op(right_tuple[0]) return (left_clipped and not right_clipped) or (left_clipped and right_clipped and left_tuple[1] > right_tuple[1])
whether the left side of the read (w/ respect to reference) is clipped. Clipping side is determined as the side with the longest clip. Adjacent clipping operations are not considered
entailment
def mad(arr): """ Median Absolute Deviation: a "Robust" version of standard deviation. Indices variabililty of the sample. https://en.wikipedia.org/wiki/Median_absolute_deviation """ arr = np.ma.array(arr).compressed() # should be faster to not use masked arrays. med = np.median(arr) return np.median(np.abs(arr - med))
Median Absolute Deviation: a "Robust" version of standard deviation. Indices variabililty of the sample. https://en.wikipedia.org/wiki/Median_absolute_deviation
entailment
def _is_streaming_request(self): """check request is stream request or not""" arg2 = self.argstreams[1] arg3 = self.argstreams[2] return not (isinstance(arg2, InMemStream) and isinstance(arg3, InMemStream) and ((arg2.auto_close and arg3.auto_close) or ( arg2.state == StreamState.completed and arg3.state == StreamState.completed)))
check request is stream request or not
entailment
def should_retry_on_error(self, error): """rules for retry :param error: ProtocolException that returns from Server """ if self.is_streaming_request: # not retry for streaming request return False retry_flag = self.headers.get('re', retry.DEFAULT) if retry_flag == retry.NEVER: return False if isinstance(error, StreamClosedError): return True if error.code in [ErrorCode.bad_request, ErrorCode.cancelled, ErrorCode.unhealthy]: return False elif error.code in [ErrorCode.busy, ErrorCode.declined]: return True elif error.code is ErrorCode.timeout: return retry_flag is not retry.CONNECTION_ERROR elif error.code in [ErrorCode.network_error, ErrorCode.fatal, ErrorCode.unexpected]: return retry_flag is not retry.TIMEOUT else: return False
rules for retry :param error: ProtocolException that returns from Server
entailment
def client_for(service, service_module, thrift_service_name=None): """Build a synchronous client class for the given Thrift service. The generated class accepts a TChannelSyncClient and an optional hostport as initialization arguments. Given ``CommentService`` defined in ``comment.thrift`` and registered with Hyperbahn under the name "comment", here's how this might be used: .. code-block:: python from tchannel.sync import TChannelSyncClient from tchannel.sync.thrift import client_for from comment import CommentService CommentServiceClient = client_for('comment', CommentService) tchannel_sync = TChannelSyncClient('my-service') comment_client = CommentServiceClient(tchannel_sync) future = comment_client.postComment( articleId, CommentService.Comment("hi") ) result = future.result() :param service: Name of the Hyperbahn service being called. :param service_module: The Thrift-generated module for that service. This usually has the same name as definied for the service in the IDL. :param thrift_service_name: If the Thrift service has a different name than its module, use this parameter to specify it. :returns: An Thrift-like class, ready to be instantiated and used with TChannelSyncClient. """ assert service_module, 'service_module is required' service = service or '' # may be blank for non-hyperbahn use cases if not thrift_service_name: thrift_service_name = service_module.__name__.rsplit('.', 1)[-1] method_names = get_service_methods(service_module.Iface) def init( self, tchannel, hostport=None, trace=False, protocol_headers=None, ): self.async_thrift = self.__async_client_class__( tchannel=tchannel, hostport=hostport, trace=trace, protocol_headers=protocol_headers, ) self.threadloop = tchannel._threadloop init.__name__ = '__init__' methods = { '__init__': init, '__async_client_class__': async_client_for( service=service, service_module=service_module, thrift_service_name=thrift_service_name, ) } methods.update({ method_name: generate_method(method_name) for method_name in method_names }) return type(thrift_service_name + 'Client', (object,), methods)
Build a synchronous client class for the given Thrift service. The generated class accepts a TChannelSyncClient and an optional hostport as initialization arguments. Given ``CommentService`` defined in ``comment.thrift`` and registered with Hyperbahn under the name "comment", here's how this might be used: .. code-block:: python from tchannel.sync import TChannelSyncClient from tchannel.sync.thrift import client_for from comment import CommentService CommentServiceClient = client_for('comment', CommentService) tchannel_sync = TChannelSyncClient('my-service') comment_client = CommentServiceClient(tchannel_sync) future = comment_client.postComment( articleId, CommentService.Comment("hi") ) result = future.result() :param service: Name of the Hyperbahn service being called. :param service_module: The Thrift-generated module for that service. This usually has the same name as definied for the service in the IDL. :param thrift_service_name: If the Thrift service has a different name than its module, use this parameter to specify it. :returns: An Thrift-like class, ready to be instantiated and used with TChannelSyncClient.
entailment
def generate_method(method_name): """Generate a method for a given Thrift service. Uses the provided TChannelSyncClient's threadloop in order to convert RPC calls to concurrent.futures :param method_name: Method being called. :return: A method that invokes the RPC using TChannelSyncClient """ def call(self, *args, **kwargs): """Forward RPC call to TChannelSyncClient :return concurrent.futures.Future: """ if not self.threadloop.is_ready(): self.threadloop.start() return self.threadloop.submit( getattr(self.async_thrift, method_name), *args, **kwargs ) return call
Generate a method for a given Thrift service. Uses the provided TChannelSyncClient's threadloop in order to convert RPC calls to concurrent.futures :param method_name: Method being called. :return: A method that invokes the RPC using TChannelSyncClient
entailment
def read_full(stream): """Read the full contents of the given stream into memory. :return: A future containing the complete stream contents. """ assert stream, "stream is required" chunks = [] chunk = yield stream.read() while chunk: chunks.append(chunk) chunk = yield stream.read() raise tornado.gen.Return(b''.join(chunks))
Read the full contents of the given stream into memory. :return: A future containing the complete stream contents.
entailment
def maybe_stream(s): """Ensure that the given argument is a stream.""" if isinstance(s, Stream): return s if s is None: stream = InMemStream() stream.close() # we don't intend to write anything return stream if isinstance(s, unicode): s = s.encode('utf-8') if isinstance(s, bytearray): s = bytes(s) if isinstance(s, bytes): stream = InMemStream(s) stream.close() # we don't intend to write anything return stream # s may still conform to the Stream interface. Yay duck typing. return s
Ensure that the given argument is a stream.
entailment
def build_raw_error_message(protocol_exception): """build protocol level error message based on Error object""" message = ErrorMessage( id=protocol_exception.id, code=protocol_exception.code, tracing=protocol_exception.tracing, description=protocol_exception.description, ) return message
build protocol level error message based on Error object
entailment
def build_raw_request_message(self, request, args, is_completed=False): """build protocol level message based on request and args. request object contains meta information about outgoing request. args are the currently chunk data from argstreams is_completed tells the flags of the message :param request: Request :param args: array of arg streams :param is_completed: message flags :return: CallRequestMessage/CallRequestContinueMessage """ request.flags = FlagsType.none if is_completed else FlagsType.fragment # TODO decide what need to pass from request if request.state == StreamState.init: message = CallRequestMessage( flags=request.flags, ttl=request.ttl * 1000, tracing=request.tracing, service=request.service, headers=request.headers, checksum=request.checksum, args=args ) request.state = (StreamState.completed if is_completed else StreamState.streaming) elif request.state == StreamState.streaming: message = CallRequestContinueMessage( flags=request.flags, checksum=request.checksum, args=args ) request.state = (StreamState.completed if is_completed else StreamState.streaming) message.id = request.id return message
build protocol level message based on request and args. request object contains meta information about outgoing request. args are the currently chunk data from argstreams is_completed tells the flags of the message :param request: Request :param args: array of arg streams :param is_completed: message flags :return: CallRequestMessage/CallRequestContinueMessage
entailment
def build_raw_response_message(self, response, args, is_completed=False): """build protocol level message based on response and args. response object contains meta information about outgoing response. args are the currently chunk data from argstreams is_completed tells the flags of the message :param response: Response :param args: array of arg streams :param is_completed: message flags :return: CallResponseMessage/CallResponseContinueMessage """ response.flags = FlagsType.none if is_completed else FlagsType.fragment # TODO decide what need to pass from request if response.state == StreamState.init: message = CallResponseMessage( flags=response.flags, code=response.code, tracing=response.tracing, headers=response.headers, checksum=response.checksum, args=args ) response.state = (StreamState.completed if is_completed else StreamState.streaming) elif response.state == StreamState.streaming: message = CallResponseContinueMessage( flags=response.flags, checksum=response.checksum, args=args ) response.state = (StreamState.completed if is_completed else StreamState.streaming) message.id = response.id return message
build protocol level message based on response and args. response object contains meta information about outgoing response. args are the currently chunk data from argstreams is_completed tells the flags of the message :param response: Response :param args: array of arg streams :param is_completed: message flags :return: CallResponseMessage/CallResponseContinueMessage
entailment
def build_request(self, message): """Build inbound request object from protocol level message info. It is allowed to take incompleted CallRequestMessage. Therefore the created request may not contain whole three arguments. :param message: CallRequestMessage :return: request object """ args = self.prepare_args(message) # TODO decide what to pass to Request from message req = Request( flags=message.flags, ttl=message.ttl / 1000.0, tracing=message.tracing, service=message.service, headers=message.headers, checksum=message.checksum, argstreams=args, id=message.id, ) return req
Build inbound request object from protocol level message info. It is allowed to take incompleted CallRequestMessage. Therefore the created request may not contain whole three arguments. :param message: CallRequestMessage :return: request object
entailment
def build_response(self, message): """Build response object from protocol level message info It is allowed to take incompleted CallResponseMessage. Therefore the created request may not contain whole three arguments. :param message: CallResponseMessage :return: response object """ args = self.prepare_args(message) # TODO decide what to pass to Response from message res = Response( flags=message.flags, code=message.code, headers=message.headers, checksum=message.checksum, argstreams=args, id=message.id, ) return res
Build response object from protocol level message info It is allowed to take incompleted CallResponseMessage. Therefore the created request may not contain whole three arguments. :param message: CallResponseMessage :return: response object
entailment
def build(self, message): """buffer all the streaming messages based on the message id. Reconstruct all fragments together. :param message: incoming message :return: next complete message or None if streaming is not done """ context = None if message.message_type in [Types.CALL_REQ, Types.CALL_RES]: self.verify_message(message) context = self.build_context(message) # streaming message if message.flags == common.FlagsType.fragment: self.message_buffer[message.id] = context # find the incompleted stream num = 0 for i, arg in enumerate(context.argstreams): if arg.state != StreamState.completed: num = i break self.close_argstream(context, num) return context elif message.message_type in [Types.CALL_REQ_CONTINUE, Types.CALL_RES_CONTINUE]: context = self.message_buffer.get(message.id) if context is None: # missing call msg before continue msg raise FatalProtocolError( "missing call message after receiving continue message", message.id, ) # find the incompleted stream dst = 0 for i, arg in enumerate(context.argstreams): if arg.state != StreamState.completed: dst = i break try: self.verify_message(message) except InvalidChecksumError as e: context.argstreams[dst].set_exception(e) raise src = 0 while src < len(message.args): context.argstreams[dst].write(message.args[src]) dst += 1 src += 1 if message.flags != FlagsType.fragment: # get last fragment. mark it as completed assert (len(context.argstreams) == CallContinueMessage.max_args_num) self.message_buffer.pop(message.id, None) context.flags = FlagsType.none self.close_argstream(context, dst - 1) return None elif message.message_type == Types.ERROR: context = self.message_buffer.pop(message.id, None) if context is None: log.info('Unconsumed error %s', message) return None else: error = TChannelError.from_code( message.code, description=message.description, tracing=context.tracing, ) context.set_exception(error) return error else: return message
buffer all the streaming messages based on the message id. Reconstruct all fragments together. :param message: incoming message :return: next complete message or None if streaming is not done
entailment
def fragment(self, message): """Fragment message based on max payload size note: if the message doesn't need to fragment, it will return a list which only contains original message itself. :param message: raw message :return: list of messages whose sizes <= max payload size """ if message.message_type in [Types.CALL_RES, Types.CALL_REQ, Types.CALL_REQ_CONTINUE, Types.CALL_RES_CONTINUE]: rw = RW[message.message_type] payload_space = (common.MAX_PAYLOAD_SIZE - rw.length_no_args(message)) # split a call/request message into an array # with a call/request message and {0~n} continue # message fragment_msg = message.fragment(payload_space) self.generate_checksum(message) yield message while fragment_msg is not None: message = fragment_msg rw = RW[message.message_type] payload_space = (common.MAX_PAYLOAD_SIZE - rw.length_no_args(message)) fragment_msg = message.fragment(payload_space) self.generate_checksum(message) yield message else: yield message
Fragment message based on max payload size note: if the message doesn't need to fragment, it will return a list which only contains original message itself. :param message: raw message :return: list of messages whose sizes <= max payload size
entailment
def verify_message(self, message): """Verify the checksum of the message.""" if verify_checksum( message, self.in_checksum.get(message.id, 0), ): self.in_checksum[message.id] = message.checksum[1] if message.flags == FlagsType.none: self.in_checksum.pop(message.id) else: self.in_checksum.pop(message.id, None) raise InvalidChecksumError( description="Checksum does not match!", id=message.id, )
Verify the checksum of the message.
entailment
def chain(*rws): """Build a ReadWriter from the given list of ReadWriters. .. code-block:: python chain( number(1), number(8), len_prefixed_string(number(2)), ) # == n1:1 n2:8 s~2 Reads/writes from the given ReadWriters in-order. Returns lists of values in the same order as the ReadWriters. :param rws: One or more ReadWriters """ assert rws is not None if len(rws) == 1 and isinstance(rws[0], list): # In case someone does chain([l0, l1, ...]) rws = rws[0] return ChainReadWriter(rws)
Build a ReadWriter from the given list of ReadWriters. .. code-block:: python chain( number(1), number(8), len_prefixed_string(number(2)), ) # == n1:1 n2:8 s~2 Reads/writes from the given ReadWriters in-order. Returns lists of values in the same order as the ReadWriters. :param rws: One or more ReadWriters
entailment
def take(self, stream, num): """Read the given number of bytes from the stream. :param stream: stream to read from :param num: number of bytes to read :raises ReadError: if the stream did not yield the exact number of bytes expected """ s = stream.read(num) slen = len(s) if slen != num: raise ReadError( "Expected %d bytes but got %d bytes." % (num, slen) ) return s
Read the given number of bytes from the stream. :param stream: stream to read from :param num: number of bytes to read :raises ReadError: if the stream did not yield the exact number of bytes expected
entailment
def get_service_methods(iface): """Get a list of methods defined in the interface for a Thrift service. :param iface: The Thrift-generated Iface class defining the interface for the service. :returns: A set containing names of the methods defined for the service. """ methods = inspect.getmembers(iface, predicate=inspect.ismethod) return set( name for (name, method) in methods if not name.startswith('__') )
Get a list of methods defined in the interface for a Thrift service. :param iface: The Thrift-generated Iface class defining the interface for the service. :returns: A set containing names of the methods defined for the service.
entailment
def deprecate(message): """Loudly prints warning.""" warnings.simplefilter('default') warnings.warn(message, category=DeprecationWarning) warnings.resetwarnings()
Loudly prints warning.
entailment
def deprecated(message): """Warn every time a fn is called.""" def decorator(fn): @functools.wraps(fn) def new_fn(*args, **kwargs): deprecate(message) return fn(*args, **kwargs) return new_fn return decorator
Warn every time a fn is called.
entailment
def load(path, service=None, hostport=None, module_name=None): """Loads the Thrift file at the specified path. The file is compiled in-memory and a Python module containing the result is returned. It may be used with ``TChannel.thrift``. For example, .. code-block:: python from tchannel import TChannel, thrift # Load our server's interface definition. donuts = thrift.load(path='donuts.thrift') # We need to specify a service name or hostport because this is a # downstream service we'll be calling. coffee = thrift.load(path='coffee.thrift', service='coffee') tchannel = TChannel('donuts') @tchannel.thrift.register(donuts.DonutsService) @tornado.gen.coroutine def submitOrder(request): args = request.body if args.coffee: yield tchannel.thrift( coffee.CoffeeService.order(args.coffee) ) # ... The returned module contains, one top-level type for each struct, enum, union, exeption, and service defined in the Thrift file. For each service, the corresponding class contains a classmethod for each function defined in that service that accepts the arguments for that function and returns a ``ThriftRequest`` capable of being sent via ``TChannel.thrift``. For more information on what gets generated by ``load``, see `thriftrw <http://thriftrw.readthedocs.org/en/latest/>`_. Note that the ``path`` accepted by ``load`` must be either an absolute path or a path relative to the *the current directory*. If you need to refer to Thrift files relative to the Python module in which ``load`` was called, use the ``__file__`` magic variable. .. code-block:: python # Given, # # foo/ # myservice.thrift # bar/ # x.py # # Inside foo/bar/x.py, path = os.path.join( os.path.dirname(__file__), '../myservice.thrift' ) The returned value is a valid Python module. You can install the module by adding it to the ``sys.modules`` dictionary. This will allow importing items from this module directly. You can use the ``__name__`` magic variable to make the generated module a submodule of the current module. For example, .. code-block:: python # foo/bar.py import sys from tchannel import thrift donuts = = thrift.load('donuts.thrift') sys.modules[__name__ + '.donuts'] = donuts This installs the module generated for ``donuts.thrift`` as the module ``foo.bar.donuts``. Callers can then import items from that module directly. For example, .. code-block:: python # foo/baz.py from foo.bar.donuts import DonutsService, Order def baz(tchannel): return tchannel.thrift( DonutsService.submitOrder(Order(..)) ) :param str service: Name of the service that the Thrift file represents. This name will be used to route requests through Hyperbahn. :param str path: Path to the Thrift file. If this is a relative path, it must be relative to the current directory. :param str hostport: Clients can use this to specify the hostport at which the service can be found. If omitted, TChannel will route the requests through known peers. This value is ignored by servers. :param str module_name: Name used for the generated Python module. Defaults to the name of the Thrift file. """ # TODO replace with more specific exceptions # assert service, 'service is required' # assert path, 'path is required' # Backwards compatibility for callers passing in service name as first arg. if not path.endswith('.thrift'): service, path = path, service module = thriftrw.load(path=path, name=module_name) return TChannelThriftModule(service, module, hostport)
Loads the Thrift file at the specified path. The file is compiled in-memory and a Python module containing the result is returned. It may be used with ``TChannel.thrift``. For example, .. code-block:: python from tchannel import TChannel, thrift # Load our server's interface definition. donuts = thrift.load(path='donuts.thrift') # We need to specify a service name or hostport because this is a # downstream service we'll be calling. coffee = thrift.load(path='coffee.thrift', service='coffee') tchannel = TChannel('donuts') @tchannel.thrift.register(donuts.DonutsService) @tornado.gen.coroutine def submitOrder(request): args = request.body if args.coffee: yield tchannel.thrift( coffee.CoffeeService.order(args.coffee) ) # ... The returned module contains, one top-level type for each struct, enum, union, exeption, and service defined in the Thrift file. For each service, the corresponding class contains a classmethod for each function defined in that service that accepts the arguments for that function and returns a ``ThriftRequest`` capable of being sent via ``TChannel.thrift``. For more information on what gets generated by ``load``, see `thriftrw <http://thriftrw.readthedocs.org/en/latest/>`_. Note that the ``path`` accepted by ``load`` must be either an absolute path or a path relative to the *the current directory*. If you need to refer to Thrift files relative to the Python module in which ``load`` was called, use the ``__file__`` magic variable. .. code-block:: python # Given, # # foo/ # myservice.thrift # bar/ # x.py # # Inside foo/bar/x.py, path = os.path.join( os.path.dirname(__file__), '../myservice.thrift' ) The returned value is a valid Python module. You can install the module by adding it to the ``sys.modules`` dictionary. This will allow importing items from this module directly. You can use the ``__name__`` magic variable to make the generated module a submodule of the current module. For example, .. code-block:: python # foo/bar.py import sys from tchannel import thrift donuts = = thrift.load('donuts.thrift') sys.modules[__name__ + '.donuts'] = donuts This installs the module generated for ``donuts.thrift`` as the module ``foo.bar.donuts``. Callers can then import items from that module directly. For example, .. code-block:: python # foo/baz.py from foo.bar.donuts import DonutsService, Order def baz(tchannel): return tchannel.thrift( DonutsService.submitOrder(Order(..)) ) :param str service: Name of the service that the Thrift file represents. This name will be used to route requests through Hyperbahn. :param str path: Path to the Thrift file. If this is a relative path, it must be relative to the current directory. :param str hostport: Clients can use this to specify the hostport at which the service can be found. If omitted, TChannel will route the requests through known peers. This value is ignored by servers. :param str module_name: Name used for the generated Python module. Defaults to the name of the Thrift file.
entailment
def register(dispatcher, service, handler=None, method=None): """ :param dispatcher: RequestDispatcher against which the new endpoint will be registered. :param Service service: Service object representing the service whose endpoint is being registered. :param handler: A function implementing the given Thrift function. :param method: If specified, name of the method being registered. Defaults to the name of the ``handler`` function. """ def decorator(method, handler): if not method: method = handler.__name__ function = getattr(service, method, None) assert function, ( 'Service "%s" does not define method "%s"' % (service.name, method) ) assert not function.oneway dispatcher.register( function.endpoint, build_handler(function, handler), ThriftRWSerializer(service._module, function._request_cls), ThriftRWSerializer(service._module, function._response_cls), ) return handler if handler is None: return partial(decorator, method) else: return decorator(method, handler)
:param dispatcher: RequestDispatcher against which the new endpoint will be registered. :param Service service: Service object representing the service whose endpoint is being registered. :param handler: A function implementing the given Thrift function. :param method: If specified, name of the method being registered. Defaults to the name of the ``handler`` function.
entailment
def interface_ip(interface): """Determine the IP assigned to us by the given network interface.""" sock = socket.socket(socket.AF_INET, socket.SOCK_DGRAM) return socket.inet_ntoa( fcntl.ioctl( sock.fileno(), 0x8915, struct.pack('256s', interface[:15]) )[20:24] )
Determine the IP assigned to us by the given network interface.
entailment