lightning/pytorch_lightning/callbacks/gradient_accumulation_sched...

58 lines
1.9 KiB
Python
Raw Normal View History

r"""
Gradient Accumulator
====================
Change gradient accumulation factor according to scheduling.
"""
from pytorch_lightning.callbacks.base import Callback
from pytorch_lightning.utilities import rank_zero_warn
class GradientAccumulationScheduler(Callback):
r"""
Change gradient accumulation factor according to scheduling.
Args:
scheduling: scheduling in format {epoch: accumulation_factor}
CI: Force docs warnings to be raised as errors (+ fix all) (#1191) * add argument to force warn * fix automodule error * fix permalink error * fix indentation warning * fix warning * fix import warnings * fix duplicate label warning * fix bullet point indentation warning * fix duplicate label warning * fix "import not top level" warning * line too long * fix indentation * fix bullet points indentation warning * fix hooks warnings * fix reference problem with excluded test_tube * fix indentation in print * change imports for trains logger * remove pandas type annotation * Update pytorch_lightning/core/lightning.py * include bullet points inside note * remove old quick start guide (unused) * fix unused warning * fix formatting * fix duplicate label issue * fix duplicate label warning (replaced by class ref) * fix tick * fix indentation warnings * docstring ticks * remove obsolete docstring typing * Revert "remove old quick start guide (unused)" This reverts commit d51bb40695442c8fa11bc9df74f6db56264f7509. * added old quick start guide to navigation * remove unused tutorials file * ignore some modules that got deprecated and are not used anymore * fix duplicate label warning * move examples doc and exclude pl_examples from autodoc * fix formatting for configure_optimizer * fix no blank line warnings * fix "see also" labels and add paramref extension * fix more reference problems * fix multi-gpu reference * fix weird warning * fix indentation and unrecognized characters in code block * fix warning "... not included in toctree" * fix PIL import error * fix duplicate target "here" warning * fix broken link * revert accidentally moved pl_examples * changelog * stdout * note some things to know Co-Authored-By: Jirka Borovec <Borda@users.noreply.github.com> Co-authored-by: J. Borovec <jirka.borovec@seznam.cz> Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
2020-03-20 19:49:01 +00:00
Example::
>>> from pytorch_lightning import Trainer
>>> from pytorch_lightning.callbacks import GradientAccumulationScheduler
# at epoch 5 start accumulating every 2 batches
>>> accumulator = GradientAccumulationScheduler(scheduling={5: 2})
>>> trainer = Trainer(callbacks=[accumulator])
# alternatively, pass the scheduling dict directly to the Trainer
>>> trainer = Trainer(accumulate_grad_batches={5: 2})
"""
def __init__(self, scheduling: dict):
super().__init__()
if not scheduling: # empty dict error
raise TypeError("Empty dict cannot be interpreted correct")
for key in scheduling:
if not isinstance(key, int) or not isinstance(scheduling[key], int):
raise TypeError("All epoches and accumulation factor must be integers")
minimal_epoch = min(scheduling.keys())
if minimal_epoch < 0:
raise IndexError(f"Epochs indexing from 1, epoch {minimal_epoch} cannot be interpreted correct")
if minimal_epoch != 0: # if user didnt define first epoch accumulation factor
scheduling.update({0: 1})
self.scheduling = scheduling
self.epochs = sorted(scheduling.keys())
def on_epoch_start(self, trainer, pl_module):
epoch = trainer.current_epoch
for i in reversed(range(len(self.epochs))):
if epoch >= self.epochs[i]:
trainer.accumulate_grad_batches = self.scheduling.get(self.epochs[i])
break