spaCy/spacy/ml/callbacks.py
Daniël de Kok 5e297aa20e
Add TrainablePipe.{distill,get_teacher_student_loss} (#12016)
* Add `TrainablePipe.{distill,get_teacher_student_loss}`

This change adds two methods:

- `TrainablePipe::distill` which performs a training step of a
   student pipe on a teacher pipe, giving a batch of `Doc`s.
- `TrainablePipe::get_teacher_student_loss` computes the loss
  of a student relative to the teacher.

The `distill` or `get_teacher_student_loss` methods are also implemented
in the tagger, edit tree lemmatizer, and parser pipes, to enable
distillation in those pipes and as an example for other pipes.

* Fix stray `Beam` import

* Fix incorrect import

* Apply suggestions from code review

Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>

* Apply suggestions from code review

Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>

* TrainablePipe.distill: use `Iterable[Example]`

* Add Pipe.is_distillable method

* Add `validate_distillation_examples`

This first calls `validate_examples` and then checks that the
student/teacher tokens are the same.

* Update distill documentation

* Add distill documentation for all pipes that support distillation

* Fix incorrect identifier

* Apply suggestions from code review

Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>

* Add comment to explain `is_distillable`

Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
2023-01-16 10:25:53 +01:00

126 lines
3.7 KiB
Python

from typing import Type, Callable, Dict, TYPE_CHECKING, List, Optional, Set
import functools
import inspect
import types
import warnings
from thinc.layers import with_nvtx_range
from thinc.model import Model, wrap_model_recursive
from thinc.util import use_nvtx_range
from ..errors import Warnings
from ..util import registry
if TYPE_CHECKING:
# This lets us add type hints for mypy etc. without causing circular imports
from ..language import Language # noqa: F401
DEFAULT_NVTX_ANNOTATABLE_PIPE_METHODS = [
"pipe",
"predict",
"set_annotations",
"update",
"rehearse",
"get_loss",
"get_teacher_student_loss",
"initialize",
"begin_update",
"finish_update",
"update",
]
def models_with_nvtx_range(nlp, forward_color: int, backprop_color: int):
pipes = [
pipe
for _, pipe in nlp.components
if hasattr(pipe, "is_trainable") and pipe.is_trainable
]
seen_models: Set[int] = set()
for pipe in pipes:
for node in pipe.model.walk():
if id(node) in seen_models:
continue
seen_models.add(id(node))
with_nvtx_range(
node, forward_color=forward_color, backprop_color=backprop_color
)
return nlp
@registry.callbacks("spacy.models_with_nvtx_range.v1")
def create_models_with_nvtx_range(
forward_color: int = -1, backprop_color: int = -1
) -> Callable[["Language"], "Language"]:
return functools.partial(
models_with_nvtx_range,
forward_color=forward_color,
backprop_color=backprop_color,
)
def nvtx_range_wrapper_for_pipe_method(self, func, *args, **kwargs):
if isinstance(func, functools.partial):
return func(*args, **kwargs)
else:
with use_nvtx_range(f"{self.name} {func.__name__}"):
return func(*args, **kwargs)
def pipes_with_nvtx_range(
nlp, additional_pipe_functions: Optional[Dict[str, List[str]]]
):
for _, pipe in nlp.components:
if additional_pipe_functions:
extra_funcs = additional_pipe_functions.get(pipe.name, [])
else:
extra_funcs = []
for name in DEFAULT_NVTX_ANNOTATABLE_PIPE_METHODS + extra_funcs:
func = getattr(pipe, name, None)
if func is None:
if name in extra_funcs:
warnings.warn(Warnings.W121.format(method=name, pipe=pipe.name))
continue
wrapped_func = functools.partial(
types.MethodType(nvtx_range_wrapper_for_pipe_method, pipe), func
)
# We need to preserve the original function signature so that
# the original parameters are passed to pydantic for validation downstream.
try:
wrapped_func.__signature__ = inspect.signature(func) # type: ignore
except:
# Can fail for Cython methods that do not have bindings.
warnings.warn(Warnings.W122.format(method=name, pipe=pipe.name))
continue
try:
setattr(
pipe,
name,
wrapped_func,
)
except AttributeError:
warnings.warn(Warnings.W122.format(method=name, pipe=pipe.name))
return nlp
@registry.callbacks("spacy.models_and_pipes_with_nvtx_range.v1")
def create_models_and_pipes_with_nvtx_range(
forward_color: int = -1,
backprop_color: int = -1,
additional_pipe_functions: Optional[Dict[str, List[str]]] = None,
) -> Callable[["Language"], "Language"]:
def inner(nlp):
nlp = models_with_nvtx_range(nlp, forward_color, backprop_color)
nlp = pipes_with_nvtx_range(nlp, additional_pipe_functions)
return nlp
return inner