"""
A version of A-GEM, leveraging a memory buffer with reservoir sampling.
"""
# 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.
import numpy as np
import torch
from models.agem import project
from models.gem import overwrite_grad, store_grad
from models.utils.continual_model import ContinualModel
from utils.args import add_rehearsal_args, ArgumentParser
from utils.buffer import Buffer
[docs]
class AGemr(ContinualModel):
"""Continual learning via A-GEM, leveraging a reservoir buffer."""
NAME = 'agem_r'
COMPATIBILITY = ['class-il', 'domain-il', 'task-il', 'general-continual']
[docs]
@staticmethod
def get_parser(parser) -> ArgumentParser:
add_rehearsal_args(parser)
return parser
def __init__(self, backbone, loss, args, transform, dataset=None):
super(AGemr, self).__init__(backbone, loss, args, transform, dataset=dataset)
self.buffer = Buffer(self.args.buffer_size)
self.grad_dims = []
for param in self.parameters():
self.grad_dims.append(param.data.numel())
self.grad_xy = torch.Tensor(np.sum(self.grad_dims)).to(self.device)
self.grad_er = torch.Tensor(np.sum(self.grad_dims)).to(self.device)
[docs]
def observe(self, inputs, labels, not_aug_inputs, epoch=None):
self.zero_grad()
p = self.net.forward(inputs)
loss = self.loss(p, labels)
loss.backward()
if not self.buffer.is_empty():
store_grad(self.parameters, self.grad_xy, self.grad_dims)
buf_inputs, buf_labels = self.buffer.get_data(self.args.minibatch_size, device=self.device)
self.net.zero_grad()
buf_outputs = self.net.forward(buf_inputs)
penalty = self.loss(buf_outputs, buf_labels)
penalty.backward()
store_grad(self.parameters, self.grad_er, self.grad_dims)
dot_prod = torch.dot(self.grad_xy, self.grad_er)
if dot_prod.item() < 0:
g_tilde = project(gxy=self.grad_xy, ger=self.grad_er)
overwrite_grad(self.parameters, g_tilde, self.grad_dims)
else:
overwrite_grad(self.parameters, self.grad_xy, self.grad_dims)
self.opt.step()
self.buffer.add_data(examples=not_aug_inputs, labels=labels)
return loss.item()