nnunet/training/network_training/nnUNet_variants/optimizer_and_lr/nnUNetTrainerV2_SGD_ReduceOnPlateau.py

#    Copyright 2020 Division of Medical Image Computing, German Cancer Research Center (DKFZ), Heidelberg, Germany
#
#    Licensed under the Apache License, Version 2.0 (the "License");
#    you may not use this file except in compliance with the License.
#    You may obtain a copy of the License at
#
#        http://www.apache.org/licenses/LICENSE-2.0
#
#    Unless required by applicable law or agreed to in writing, software
#    distributed under the License is distributed on an "AS IS" BASIS,
#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#    See the License for the specific language governing permissions and
#    limitations under the License.


import torch
from nnunet.training.network_training.nnUNetTrainer import nnUNetTrainer
from nnunet.training.network_training.nnUNetTrainerV2 import nnUNetTrainerV2
from torch.optim import lr_scheduler


class nnUNetTrainerV2_SGD_ReduceOnPlateau(nnUNetTrainerV2):
    def __init__(self, plans_file, fold, output_folder=None, dataset_directory=None, batch_dice=True, stage=None,
                 unpack_data=True, deterministic=True, fp16=False):
        super().__init__(plans_file, fold, output_folder, dataset_directory, batch_dice, stage, unpack_data,
                         deterministic, fp16)

    def initialize_optimizer_and_scheduler(self):
        self.optimizer = torch.optim.SGD(self.network.parameters(), self.initial_lr, weight_decay=self.weight_decay,
                                         momentum=0.99, nesterov=True)
        self.lr_scheduler = lr_scheduler.ReduceLROnPlateau(self.optimizer, mode='min', factor=0.2,
                                                           patience=self.lr_scheduler_patience,
                                                           verbose=True, threshold=self.lr_scheduler_eps,
                                                           threshold_mode="abs")

    def maybe_update_lr(self, epoch=None):
        # maybe update learning rate
        if self.lr_scheduler is not None:
            assert isinstance(self.lr_scheduler, (lr_scheduler.ReduceLROnPlateau, lr_scheduler._LRScheduler))

            if isinstance(self.lr_scheduler, lr_scheduler.ReduceLROnPlateau):
                # lr scheduler is updated with moving average val loss. should be more robust
                if self.epoch > 0:  # otherwise self.train_loss_MA is None
                    self.lr_scheduler.step(self.train_loss_MA)
            else:
                self.lr_scheduler.step(self.epoch + 1)
        self.print_to_log_file("lr is now (scheduler) %s" % str(self.optimizer.param_groups[0]['lr']))

    def on_epoch_end(self):
        return nnUNetTrainer.on_epoch_end(self)