Skip to content
Fetching contributors…
Cannot retrieve contributors at this time
644 lines (524 sloc) 21.7 KB
Various utilities that don't fit anywhere else.
import importlib
import json
import logging
import os
import pkgutil
import random
import subprocess
import sys
from contextlib import contextmanager
from itertools import islice, zip_longest
from logging import Filter
from pathlib import Path
from typing import (
import numpy
import spacy
import torch
import torch.distributed as dist
from import download as spacy_download
from spacy.language import Language as SpacyModelType
from allennlp.common.checks import log_pytorch_version_info
from allennlp.common.params import Params
from allennlp.common.tqdm import Tqdm
import resource
except ImportError:
# resource doesn't exist on Windows systems
resource = None
logger = logging.getLogger(__name__)
JsonDict = Dict[str, Any]
# If you want to have start and/or end symbols for any reason in your code, we recommend you use
# these, to have a common place to import from. Also, it's important for some edge cases in how
# data is processed for these symbols to be lowercase, not uppercase (because we have code that
# will lowercase tokens for you in some circumstances, and we need this symbol to not change in
# those cases).
START_SYMBOL = "@start@"
END_SYMBOL = "@end@"
PathType = Union[os.PathLike, str]
T = TypeVar("T")
ContextManagerFunctionReturnType = Generator[T, None, None]
def sanitize(x: Any) -> Any:
Sanitize turns PyTorch and Numpy types into basic Python types so they
can be serialized into JSON.
# Import here to avoid circular references
from import Token
if isinstance(x, (str, float, int, bool)):
# x is already serializable
return x
elif isinstance(x, torch.Tensor):
# tensor needs to be converted to a list (and moved to cpu if necessary)
return x.cpu().tolist()
elif isinstance(x, numpy.ndarray):
# array needs to be converted to a list
return x.tolist()
elif isinstance(x, numpy.number):
# NumPy numbers need to be converted to Python numbers
return x.item()
elif isinstance(x, dict):
# Dicts need their values sanitized
return {key: sanitize(value) for key, value in x.items()}
elif isinstance(x, numpy.bool_):
# Numpy bool_ need to be converted to python bool.
return bool(x)
elif isinstance(x, (spacy.tokens.Token, Token)):
# Tokens get sanitized to just their text.
return x.text
elif isinstance(x, (list, tuple)):
# Lists and Tuples need their values sanitized
return [sanitize(x_i) for x_i in x]
elif x is None:
return "None"
elif hasattr(x, "to_json"):
return x.to_json()
raise ValueError(
f"Cannot sanitize {x} of type {type(x)}. "
"If this is your own custom class, add a `to_json(self)` method "
"that returns a JSON-like object."
def group_by_count(iterable: List[Any], count: int, default_value: Any) -> List[List[Any]]:
Takes a list and groups it into sublists of size `count`, using `default_value` to pad the
list at the end if the list is not divisable by `count`.
For example:
>>> group_by_count([1, 2, 3, 4, 5, 6, 7], 3, 0)
[[1, 2, 3], [4, 5, 6], [7, 0, 0]]
This is a short method, but it's complicated and hard to remember as a one-liner, so we just
make a function out of it.
return [list(l) for l in zip_longest(*[iter(iterable)] * count, fillvalue=default_value)]
A = TypeVar("A")
def lazy_groups_of(iterable: Iterable[A], group_size: int) -> Iterator[List[A]]:
Takes an iterable and batches the individual instances into lists of the
specified size. The last list may be smaller if there are instances left over.
iterator = iter(iterable)
while True:
s = list(islice(iterator, group_size))
if len(s) > 0:
yield s
def pad_sequence_to_length(
sequence: List,
desired_length: int,
default_value: Callable[[], Any] = lambda: 0,
padding_on_right: bool = True,
) -> List:
Take a list of objects and pads it to the desired length, returning the padded list. The
original list is not modified.
# Parameters
sequence : List
A list of objects to be padded.
desired_length : int
Maximum length of each sequence. Longer sequences are truncated to this length, and
shorter ones are padded to it.
default_value: Callable, default=lambda: 0
Callable that outputs a default value (of any type) to use as padding values. This is
a lambda to avoid using the same object when the default value is more complex, like a
padding_on_right : bool, default=True
When we add padding tokens (or truncate the sequence), should we do it on the right or
the left?
# Returns
padded_sequence : List
# Truncates the sequence to the desired length.
if padding_on_right:
padded_sequence = sequence[:desired_length]
padded_sequence = sequence[-desired_length:]
# Continues to pad with default_value() until we reach the desired length.
pad_length = desired_length - len(padded_sequence)
# This just creates the default value once, so if it's a list, and if it gets mutated
# later, it could cause subtle bugs. But the risk there is low, and this is much faster.
values_to_pad = [default_value()] * pad_length
if padding_on_right:
padded_sequence = padded_sequence + values_to_pad
padded_sequence = values_to_pad + padded_sequence
return padded_sequence
def add_noise_to_dict_values(dictionary: Dict[A, float], noise_param: float) -> Dict[A, float]:
Returns a new dictionary with noise added to every key in `dictionary`. The noise is
uniformly distributed within `noise_param` percent of the value for every value in the
new_dict = {}
for key, value in dictionary.items():
noise_value = value * noise_param
noise = random.uniform(-noise_value, noise_value)
new_dict[key] = value + noise
return new_dict
def namespace_match(pattern: str, namespace: str):
Matches a namespace pattern against a namespace string. For example, `*tags` matches
`passage_tags` and `question_tags` and `tokens` matches `tokens` but not
if pattern[0] == "*" and namespace.endswith(pattern[1:]):
return True
elif pattern == namespace:
return True
return False
def prepare_environment(params: Params):
Sets random seeds for reproducible experiments. This may not work as expected
if you use this from within a python project in which you have already imported Pytorch.
If you use the scripts/ entry point to training models with this library,
your experiments should be reasonably reproducible. If you are using this from your own
project, you will want to call this function before importing Pytorch. Complete determinism
is very difficult to achieve with libraries doing optimized linear algebra due to massively
parallel execution, which is exacerbated by using GPUs.
# Parameters
params: Params object or dict, required.
A `Params` object or dict holding the json parameters.
seed = params.pop_int("random_seed", 13370)
numpy_seed = params.pop_int("numpy_seed", 1337)
torch_seed = params.pop_int("pytorch_seed", 133)
if seed is not None:
if numpy_seed is not None:
if torch_seed is not None:
# Seed all GPUs with the same seed if available.
if torch.cuda.is_available():
class FileFriendlyLogFilter(Filter):
TQDM and requests use carriage returns to get the training line to update for each batch
without adding more lines to the terminal output. Displaying those in a file won't work
correctly, so we'll just make sure that each batch shows up on its one line.
def filter(self, record):
if "\r" in record.msg:
record.msg = record.msg.replace("\r", "")
if not record.msg or record.msg[-1] != "\n":
record.msg += "\n"
return True
class WorkerLogFilter(Filter):
def __init__(self, rank=-1):
self._rank = rank
def filter(self, record):
if self._rank != -1:
record.msg = f"Rank {self._rank} | {record.msg}"
return True
def prepare_global_logging(
serialization_dir: str, file_friendly_logging: bool, rank: int = 0, world_size: int = 1
) -> None:
# If we don't have a terminal as stdout,
# force tqdm to be nicer.
if not sys.stdout.isatty():
file_friendly_logging = True
# Handlers for stdout/err logging
output_stream_log_handler = logging.StreamHandler(sys.stdout)
error_stream_log_handler = logging.StreamHandler(sys.stderr)
if world_size == 1:
# This case is not distributed training and hence will stick to the older
# log file names
output_file_log_handler = logging.FileHandler(
filename=os.path.join(serialization_dir, "stdout.log")
error_file_log_handler = logging.FileHandler(
filename=os.path.join(serialization_dir, "stderr.log")
# Create log files with worker ids
output_file_log_handler = logging.FileHandler(
filename=os.path.join(serialization_dir, f"stdout_worker{rank}.log")
error_file_log_handler = logging.FileHandler(
filename=os.path.join(serialization_dir, f"stderr_worker{rank}.log")
# This adds the worker's rank to messages being logged to files.
# This will help when combining multiple worker log files using `less` command.
worker_filter = WorkerLogFilter(rank)
formatter = logging.Formatter("%(asctime)s - %(levelname)s - %(name)s - %(message)s")
root_logger = logging.getLogger()
# Remove the already set stream handler in root logger.
# Not doing this will result in duplicate log messages
# printed in the console
if len(root_logger.handlers) > 0:
for handler in root_logger.handlers:
# file handlers need to be handled for tqdm's \r char
file_friendly_log_filter = FileFriendlyLogFilter()
if os.environ.get("ALLENNLP_DEBUG"):
LEVEL = logging.DEBUG
level_name = os.environ.get("ALLENNLP_LOG_LEVEL")
LEVEL = logging._nameToLevel.get(level_name, logging.INFO)
if rank == 0:
# stdout/stderr handlers are added only for the
# master worker. This is to avoid cluttering the console
# screen with too many log messages from all workers.
if file_friendly_logging:
LOADED_SPACY_MODELS: Dict[Tuple[str, bool, bool, bool], SpacyModelType] = {}
def get_spacy_model(
spacy_model_name: str, pos_tags: bool, parse: bool, ner: bool
) -> SpacyModelType:
In order to avoid loading spacy models a whole bunch of times, we'll save references to them,
keyed by the options we used to create the spacy model, so any particular configuration only
gets loaded once.
options = (spacy_model_name, pos_tags, parse, ner)
if options not in LOADED_SPACY_MODELS:
disable = ["vectors", "textcat"]
if not pos_tags:
if not parse:
if not ner:
spacy_model = spacy.load(spacy_model_name, disable=disable)
except OSError:
f"Spacy models '{spacy_model_name}' not found. Downloading and installing."
# Import the downloaded model module directly and load from there
spacy_model_module = __import__(spacy_model_name)
spacy_model = spacy_model_module.load(disable=disable) # type: ignore
LOADED_SPACY_MODELS[options] = spacy_model
return LOADED_SPACY_MODELS[options]
def pushd(new_dir: PathType, verbose: bool = False) -> ContextManagerFunctionReturnType[None]:
Changes the current directory to the given path and prepends it to `sys.path`.
This method is intended to use with `with`, so after its usage, the current directory will be
set to the previous value.
previous_dir = os.getcwd()
if verbose:"Changing directory to {new_dir}") # type: ignore
if verbose:"Changing directory back to {previous_dir}")
def push_python_path(path: PathType) -> ContextManagerFunctionReturnType[None]:
Prepends the given path to `sys.path`.
This method is intended to use with `with`, so after its usage, its value willbe removed from
# In some environments, such as TC, it fails when sys.path contains a relative path, such as ".".
path = Path(path).resolve()
path = str(path)
sys.path.insert(0, path)
# Better to remove by value, in case `sys.path` was manipulated in between.
def import_module_and_submodules(package_name: str) -> None:
Import all submodules under the given package.
Primarily useful so that people using AllenNLP as a library
can specify their own custom packages and have their custom
classes get loaded and registered.
# For some reason, python doesn't always add this by default to your path, but you pretty much
# always want it when using `--include-package`. And if it's already there, adding it again at
# the end won't hurt anything.
with push_python_path("."):
# Import at top level
module = importlib.import_module(package_name)
path = getattr(module, "__path__", [])
path_string = "" if not path else path[0]
# walk_packages only finds immediate children, so need to recurse.
for module_finder, name, _ in pkgutil.walk_packages(path):
# Sometimes when you import third-party libraries that are on your path,
# `pkgutil.walk_packages` returns those too, so we need to skip them.
if path_string and module_finder.path != path_string:
subpackage = f"{package_name}.{name}"
def peak_memory_mb() -> float:
Get peak memory usage for this process, as measured by
max-resident-set size:
Only works on OSX and Linux, returns 0.0 otherwise.
if resource is None or sys.platform not in ("linux", "darwin"):
return 0.0
# TODO(joelgrus): For whatever, our pinned version 0.521 of mypy does not like
# next line, but later versions (e.g. 0.530) are fine with it. Once we get that
# figured out, remove the type: ignore.
peak = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss # type: ignore
if sys.platform == "darwin":
# On OSX the result is in bytes.
return peak / 1_000_000
# On Linux the result is in kilobytes.
return peak / 1_000
def gpu_memory_mb() -> Dict[int, int]:
Get the current GPU memory usage.
Based on
# Returns
`Dict[int, int]`
Keys are device ids as integers.
Values are memory usage as integers in MB.
Returns an empty `dict` if GPUs are not available.
result = subprocess.check_output(
["nvidia-smi", "--query-gpu=memory.used", "--format=csv,nounits,noheader"],
gpu_memory = [int(x) for x in result.strip().split("\n")]
return {gpu: memory for gpu, memory in enumerate(gpu_memory)}
except FileNotFoundError:
# `nvidia-smi` doesn't exist, assume that means no GPU.
return {}
except: # noqa
# Catch *all* exceptions, because this memory check is a nice-to-have
# and we'd never want a training run to fail because of it.
"unable to check gpu_memory_mb() due to occasional failure, continuing", exc_info=True
return {}
def ensure_list(iterable: Iterable[A]) -> List[A]:
An Iterable may be a list or a generator.
This ensures we get a list without making an unnecessary copy.
if isinstance(iterable, list):
return iterable
return list(iterable)
def is_lazy(iterable: Iterable[A]) -> bool:
Checks if the given iterable is lazy,
which here just means it's not a list.
return not isinstance(iterable, list)
def log_frozen_and_tunable_parameter_names(model: torch.nn.Module) -> None:
frozen_parameter_names, tunable_parameter_names = get_frozen_and_tunable_parameter_names(model)"The following parameters are Frozen (without gradient):")
for name in frozen_parameter_names:"The following parameters are Tunable (with gradient):")
for name in tunable_parameter_names:
def get_frozen_and_tunable_parameter_names(
model: torch.nn.Module,
) -> Tuple[Iterable[str], Iterable[str]]:
frozen_parameter_names = (
name for name, parameter in model.named_parameters() if not parameter.requires_grad
tunable_parameter_names = (
name for name, parameter in model.named_parameters() if parameter.requires_grad
return frozen_parameter_names, tunable_parameter_names
def dump_metrics(file_path: Optional[str], metrics: Dict[str, Any], log: bool = False) -> None:
metrics_json = json.dumps(metrics, indent=2)
if file_path:
with open(file_path, "w") as metrics_file:
if log:"Metrics: %s", metrics_json)
def flatten_filename(file_path: str) -> str:
return file_path.replace("/", "_SLASH_")
def is_master(
global_rank: int = None, world_size: int = None, num_procs_per_node: int = None
) -> bool:
Checks if the process is a "master" of its node in a distributed process group. If a
process group is not initialized, this returns `True`.
# Parameters
global_rank : int ( default = None )
Global rank of the process if in a distributed process group. If not
given, rank is obtained using `torch.distributed.get_rank()`
world_size : int ( default = None )
Number of processes in the distributed group. If not
given, this is obtained using `torch.distributed.get_world_size()`
num_procs_per_node: int ( default = None ),
Number of GPU processes running per node
distributed = dist.is_available() and dist.is_initialized()
# In non-distributed case, a "master" process doesn't make any
# sense. So instead of raising an error, returning True would
# make things less painful
if not distributed:
return True
if global_rank is None:
global_rank = dist.get_rank()
if world_size is None:
world_size = dist.get_world_size()
if num_procs_per_node is None and os.environ:
num_procs_per_node = int(os.environ.get("ALLENNLP_PROCS_PER_NODE", world_size))
# rank == 0 would do in a single-node multi-GPU setup. However,
# in a multi-node case, every node has a logical master and hence
# the mod(%) op.
return global_rank % (world_size / num_procs_per_node) == 0
def is_distributed() -> bool:
Checks if the distributed process group is available and has been initialized
return dist.is_available() and dist.is_initialized()
def sanitize_wordpiece(wordpiece: str) -> str:
Sanitizes wordpieces from BERT, RoBERTa or ALBERT tokenizers.
if wordpiece.startswith("##"):
return wordpiece[2:]
elif wordpiece.startswith("Ġ"):
return wordpiece[1:]
elif wordpiece.startswith("▁"):
return wordpiece[1:]
return wordpiece
You can’t perform that action at this time.