# Copyright 2022-present, Lorenzo Bonicelli, Pietro Buzzega, Matteo Boschini, Angelo Porrello, 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.
import torch
import logging
from models.utils.continual_model import ContinualModel
from utils.args import add_rehearsal_args, ArgumentParser
from utils.buffer import Buffer
[docs]
class Mer(ContinualModel):
"""Continual Learning via Meta-Experience Replay (Alg 6)."""
NAME = 'mer'
COMPATIBILITY = ['class-il', 'domain-il', 'task-il', 'general-continual']
[docs]
@staticmethod
def get_parser(parser) -> ArgumentParser:
add_rehearsal_args(parser)
parser.set_defaults(batch_size=1)
parser.add_argument('--beta', type=float, required=True,
help='Within-batch update beta parameter.')
parser.add_argument('--gamma', type=float, required=True,
help='Across-batch update gamma parameter.')
parser.add_argument('--batch_num', type=int, default=1,
help='Number of batches extracted from the buffer.')
return parser
def __init__(self, backbone, loss, args, transform, dataset=None):
if args.batch_size != 1:
logging.warning('MER is designed to work with batch_size=1. We will use batch_size=1.')
args.batch_size = 1
super(Mer, self).__init__(backbone, loss, args, transform, dataset=dataset)
self.buffer = Buffer(self.args.buffer_size)
[docs]
def observe(self, inputs, labels, not_aug_inputs, epoch=None):
theta_A0 = self.net.get_params().data.clone()
for i in range(self.args.batch_num):
theta_Wi0 = self.net.get_params().data.clone()
if not self.buffer.is_empty():
buf_inputs, buf_labels = self.buffer.get_data(self.args.minibatch_size,
transform=self.transform, device=self.device)
batch_inputs = torch.cat((buf_inputs, inputs))
batch_labels = torch.cat((buf_labels, torch.tensor([labels]).to(self.device)))
else:
batch_inputs, batch_labels = inputs, torch.tensor([labels]).to(self.device)
# within-batch step
self.opt.zero_grad()
outputs = self.net(batch_inputs)
loss = self.loss(outputs, batch_labels)
loss.backward()
self.opt.step()
# within batch reptile meta-update
new_params = theta_Wi0 + self.args.beta * (self.net.get_params() - theta_Wi0)
self.net.set_params(new_params)
self.buffer.add_data(examples=not_aug_inputs, labels=labels)
# across batch reptile meta-update
new_new_params = theta_A0 + self.args.gamma * (self.net.get_params() - theta_A0)
self.net.set_params(new_new_params)
return loss.item()