311 lines
11 KiB
Python
311 lines
11 KiB
Python
from typing import Dict, Any, Tuple, Callable, List, Optional, IO
|
|
from types import ModuleType
|
|
import os
|
|
import sys
|
|
|
|
from spacy import Language
|
|
from spacy.util import SimpleFrozenList
|
|
from .util import dict_to_dot, dot_to_dict, matcher_for_regex_patterns
|
|
from .util import setup_default_console_logger, LoggerT
|
|
|
|
|
|
# entry point: spacy.ClearMLLogger.v2
|
|
def clearml_logger_v2(
|
|
project_name: str,
|
|
task_name: str,
|
|
remove_config_values: List[str] = SimpleFrozenList(),
|
|
model_log_interval: Optional[int] = None,
|
|
log_dataset_dir: Optional[str] = None,
|
|
log_best_dir: Optional[str] = None,
|
|
log_latest_dir: Optional[str] = None,
|
|
log_custom_stats: Optional[List[str]] = None,
|
|
) -> LoggerT:
|
|
"""Creates a logger that interoperates with the ClearML framework.
|
|
|
|
Args:
|
|
project_name (str):
|
|
The name of the project in the ClearML interface. The project will be created automatically if it doesn't exist yet.
|
|
task_name (str):
|
|
The name of the ClearML task. A task is an experiment that lives inside a project. Can be non-unique.
|
|
remove_config_values (List[str]):
|
|
A list of values to exclude from the config before it is uploaded to ClearML. Defaults to [].
|
|
model_log_interval (Optional[int]):
|
|
Steps to wait between logging model checkpoints to the ClearML dasboard (default: `None`). Will have no effect without also setting `log_best_dir` or `log_latest_dir`. Defaults to None.
|
|
log_dataset_dir (Optional[str]):
|
|
Directory containing the dataset to be logged and versioned as a ClearML Dataset. Defaults to None.
|
|
log_best_dir (Optional[str]):
|
|
Directory containing the best trained model as saved by spaCy, to be logged and versioned as a ClearML artifact. Defaults to None.
|
|
log_latest_dir (Optional[str]):
|
|
Directory containing the latest trained model as saved by spaCy, to be logged and versioned as a ClearML artifact. Defaults to None.
|
|
log_custom_stats (Optional[List[str]]):
|
|
A list of regular expressions that will be applied to the info dictionary passed to the logger. Statistics and metrics that match these regexps will be automatically logged. Defaults to None.
|
|
|
|
Returns:
|
|
LoggerT: Logger instance.
|
|
"""
|
|
clearml = _import_clearml()
|
|
|
|
def setup_logger(
|
|
nlp: Language, stdout: IO = sys.stdout, stderr: IO = sys.stderr
|
|
) -> Tuple[Callable[[Dict[str, Any]], None], Callable[[], None]]:
|
|
match_stat = matcher_for_regex_patterns(log_custom_stats)
|
|
task, best_model, last_model = _setup_clearml(
|
|
clearml,
|
|
nlp,
|
|
project_name,
|
|
task_name,
|
|
log_dataset_dir,
|
|
log_best_dir,
|
|
log_latest_dir,
|
|
remove_config_values,
|
|
)
|
|
|
|
def log_step(info: Optional[Dict[str, Any]]):
|
|
_log_step_clearml(
|
|
info,
|
|
task,
|
|
best_model,
|
|
last_model,
|
|
model_log_interval,
|
|
log_best_dir,
|
|
log_latest_dir,
|
|
)
|
|
_log_custom_stats(clearml, info, match_stat)
|
|
|
|
def finalize():
|
|
_finalize_clearml(task)
|
|
|
|
return log_step, finalize
|
|
|
|
return setup_logger
|
|
|
|
|
|
# entry point: spacy.ClearMLLogger.v1
|
|
def clearml_logger_v1(
|
|
project_name: str,
|
|
task_name: str,
|
|
remove_config_values: List[str] = SimpleFrozenList(),
|
|
model_log_interval: Optional[int] = None,
|
|
log_dataset_dir: Optional[str] = None,
|
|
log_best_dir: Optional[str] = None,
|
|
log_latest_dir: Optional[str] = None,
|
|
) -> LoggerT:
|
|
"""Creates a logger that interoperates with the ClearML framework.
|
|
|
|
Args:
|
|
project_name (str):
|
|
The name of the project in the ClearML interface. The project will be created automatically if it doesn't exist yet.
|
|
task_name (str):
|
|
The name of the ClearML task. A task is an experiment that lives inside a project. Can be non-unique.
|
|
remove_config_values (List[str]):
|
|
A list of values to exclude from the config before it is uploaded to ClearML. Defaults to [].
|
|
model_log_interval (Optional[int]):
|
|
Steps to wait between logging model checkpoints to the ClearML dasboard (default: `None`). Will have no effect without also setting `log_best_dir` or `log_latest_dir`. Defaults to None.
|
|
log_dataset_dir (Optional[str]):
|
|
Directory containing the dataset to be logged and versioned as a ClearML Dataset. Defaults to None.
|
|
log_best_dir (Optional[str]):
|
|
Directory containing the best trained model as saved by spaCy, to be logged and versioned as a ClearML artifact. Defaults to None.
|
|
log_latest_dir (Optional[str]):
|
|
Directory containing the latest trained model as saved by spaCy, to be logged and versioned as a ClearML artifact. Defaults to None.
|
|
|
|
Returns:
|
|
LoggerT: Logger instance.
|
|
"""
|
|
clearml = _import_clearml()
|
|
|
|
def setup_logger(
|
|
nlp: Language, stdout: IO = sys.stdout, stderr: IO = sys.stderr
|
|
) -> Tuple[Callable[[Dict[str, Any]], None], Callable[[], None]]:
|
|
console_log_step, console_finalize = setup_default_console_logger(
|
|
nlp, stdout, stderr
|
|
)
|
|
task, best_model, last_model = _setup_clearml(
|
|
clearml,
|
|
nlp,
|
|
project_name,
|
|
task_name,
|
|
log_dataset_dir,
|
|
log_best_dir,
|
|
log_latest_dir,
|
|
remove_config_values,
|
|
)
|
|
|
|
def log_step(info: Optional[Dict[str, Any]]):
|
|
console_log_step(info)
|
|
_log_step_clearml(
|
|
info,
|
|
task,
|
|
best_model,
|
|
last_model,
|
|
model_log_interval,
|
|
log_best_dir,
|
|
log_latest_dir,
|
|
)
|
|
|
|
def finalize():
|
|
console_finalize()
|
|
_finalize_clearml(task)
|
|
|
|
return log_step, finalize
|
|
|
|
return setup_logger
|
|
|
|
|
|
def _import_clearml() -> ModuleType:
|
|
try:
|
|
import clearml
|
|
|
|
# test that these are available
|
|
from clearml import Task, Dataset, OutputModel # noqa: F401
|
|
except ImportError as exc:
|
|
raise ImportError(
|
|
"The 'clearml' library could not be found - did you install it? "
|
|
"Alternatively, specify the 'ConsoleLogger' in the "
|
|
"'training.logger' config section, instead of the 'ClearMLLogger'."
|
|
) from exc
|
|
return clearml
|
|
|
|
|
|
def _setup_clearml(
|
|
clearml: ModuleType,
|
|
nlp: "Language",
|
|
project_name: str,
|
|
task_name: str,
|
|
log_dataset_dir: Optional[str] = None,
|
|
log_best_dir: Optional[str] = None,
|
|
log_latest_dir: Optional[str] = None,
|
|
remove_config_values: List[str] = SimpleFrozenList(),
|
|
) -> Tuple[Any, Any, Any]:
|
|
config = nlp.config.interpolate()
|
|
config_dot = dict_to_dot(config)
|
|
for field in remove_config_values:
|
|
del config_dot[field]
|
|
config = dot_to_dict(config_dot)
|
|
task = clearml.Task.init(
|
|
project_name=project_name,
|
|
task_name=task_name,
|
|
output_uri=True,
|
|
)
|
|
for config_section, subconfig_or_value in config.items():
|
|
task.connect(subconfig_or_value, name=config_section)
|
|
|
|
if log_dataset_dir:
|
|
dataset = clearml.Dataset.create(
|
|
dataset_project=project_name,
|
|
dataset_name=os.path.basename(log_dataset_dir),
|
|
)
|
|
dataset.add_files(log_dataset_dir)
|
|
dataset.finalize(auto_upload=True)
|
|
task.set_user_properties(
|
|
{
|
|
"name": "Created Dataset ID",
|
|
"value": dataset.id,
|
|
}
|
|
)
|
|
|
|
# Connect 2 models to the task, we will periodically update their weights later on
|
|
if log_best_dir:
|
|
best_model = clearml.OutputModel(
|
|
task=task, framework="spaCy", name="Best Model"
|
|
)
|
|
else:
|
|
best_model = None
|
|
if log_latest_dir:
|
|
last_model = clearml.OutputModel(
|
|
task=task, framework="spaCy", name="Last Model"
|
|
)
|
|
else:
|
|
last_model = None
|
|
|
|
return task, best_model, last_model
|
|
|
|
|
|
def _log_step_clearml(
|
|
info: Optional[Dict[str, Any]],
|
|
task: Any,
|
|
best_model: Optional[Any] = None,
|
|
last_model: Optional[Any] = None,
|
|
model_log_interval: Optional[int] = None,
|
|
log_best_dir: Optional[str] = None,
|
|
log_latest_dir: Optional[str] = None,
|
|
):
|
|
if info is None:
|
|
return
|
|
|
|
score = info.get("score")
|
|
other_scores = info.get("other_scores")
|
|
losses = info.get("losses")
|
|
if score:
|
|
task.get_logger().report_scalar(
|
|
"Score", "Score", iteration=info["step"], value=score
|
|
)
|
|
if losses:
|
|
for metric, metric_value in losses.items():
|
|
task.get_logger().report_scalar(
|
|
title=f"loss_{metric}",
|
|
series=f"loss_{metric}",
|
|
iteration=info["step"],
|
|
value=metric_value,
|
|
)
|
|
if isinstance(other_scores, dict):
|
|
# other_scores is usually a nested dict, so group they by the first key and flatten the rest
|
|
# combine flattened submetrics on the same ClearML graph when they have the same first key
|
|
for metric, metric_value in other_scores.items():
|
|
if isinstance(metric_value, dict):
|
|
sub_metrics_dict = dict_to_dot(metric_value)
|
|
for (
|
|
sub_metric,
|
|
sub_metric_value,
|
|
) in sub_metrics_dict.items():
|
|
# Scalars with the same title get plotted on the same graph as multiple traces
|
|
# This saves a lot of space in the UI
|
|
task.get_logger().report_scalar(
|
|
title=metric,
|
|
series=sub_metric,
|
|
iteration=info["step"],
|
|
value=sub_metric_value,
|
|
)
|
|
elif isinstance(metric_value, (float, int)):
|
|
task.get_logger().report_scalar(
|
|
metric,
|
|
metric,
|
|
iteration=info["step"],
|
|
value=metric_value,
|
|
)
|
|
|
|
if model_log_interval and info.get("output_path"):
|
|
if info["step"] % model_log_interval == 0 and info["step"] != 0:
|
|
if log_latest_dir:
|
|
assert last_model is not None
|
|
last_model.update_weights_package(
|
|
weights_path=log_latest_dir,
|
|
auto_delete_file=False,
|
|
target_filename="last_model",
|
|
)
|
|
if log_best_dir and info["score"] == max(info["checkpoints"])[0]:
|
|
assert best_model is not None
|
|
best_model.update_weights_package(
|
|
weights_path=log_best_dir,
|
|
auto_delete_file=False,
|
|
target_filename="best_model",
|
|
)
|
|
|
|
|
|
def _finalize_clearml(task: Any):
|
|
task.flush(wait_for_uploads=True)
|
|
task.close()
|
|
|
|
|
|
def _log_custom_stats(
|
|
task: Any, info: Optional[Dict[str, Any]], matcher: Callable[[str], bool]
|
|
):
|
|
if info is not None:
|
|
for k, v in info.items():
|
|
if matcher(k):
|
|
task.get_logger().report_scalar(
|
|
title=f"loss_{k}",
|
|
series=f"loss_{k}",
|
|
iteration=info["step"],
|
|
value=v,
|
|
)
|