# Copyright 2020-present, Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, Simone Calderara.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
from datasets.utils.continual_dataset import ContinualDataset
from torch.nn import functional as F
from models.utils.continual_model import ContinualModel
from utils.args import add_rehearsal_args, ArgumentParser
from utils.buffer import Buffer
from models.star_utils.star_perturber import Perturber, add_perturb_args
[docs]
class DerppSTAR(ContinualModel):
NAME = 'derpp_star'
COMPATIBILITY = ['class-il', 'domain-il', 'task-il', 'general-continual']
def __init__(self, backbone, loss, args, transform, dataset=None):
super(DerppSTAR, self).__init__(backbone, loss, args, transform, dataset)
self.pert = Perturber(self)
self.buffer = Buffer(self.args.buffer_size, self.device)
[docs]
@staticmethod
def get_parser(parser) -> ArgumentParser:
add_rehearsal_args(parser)
# add arguments for STAR
add_perturb_args(parser)
parser.add_argument('--alpha', type=float, required=True,
help='Penalty weight.')
parser.add_argument('--beta', type=float, required=True,
help='Penalty weight.')
return parser
[docs]
def observe(self, inputs, labels, not_aug_inputs, epoch):
self.opt.zero_grad()
loss = 0
if not self.buffer.is_empty():
# STAR here
buf_inputs, buf_labels, _ = self.buffer.get_data(
self.args.minibatch_size, transform=self.transform)
self.pert(buf_inputs, buf_labels)
# normal DER++
buf_inputs, _, buf_logits = self.buffer.get_data(
self.args.minibatch_size, transform=self.transform)
buf_outputs = self.net(buf_inputs)
loss += self.args.alpha * F.mse_loss(buf_outputs, buf_logits)
buf_inputs, buf_labels, _ = self.buffer.get_data(
self.args.minibatch_size, transform=self.transform)
buf_outputs = self.net(buf_inputs)
loss += self.args.beta * self.loss(buf_outputs, buf_labels)
outputs = self.net(inputs)
loss += self.loss(outputs, labels)
loss.backward()
self.opt.step()
self.buffer.add_data(examples=not_aug_inputs,
labels=labels,
logits=outputs.data)
return loss.item()