"""
This module implements the simplest form of rehearsal training: Experience Replay. It maintains a buffer
of previously seen examples and uses them to augment the current batch during training.
Example usage:
model = Er(backbone, loss, args, transform, dataset)
loss = model.observe(inputs, labels, not_aug_inputs, epoch)
"""
# 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 torch
from models.utils.continual_model import ContinualModel
from utils.args import add_rehearsal_args, ArgumentParser
from utils.buffer import Buffer
[docs]
def apply_decay(decay, lr, optimizer, num_iter):
if decay != 1:
learn_rate = lr * (decay ** num_iter)
for param_group in optimizer.param_groups:
param_group['lr'] = learn_rate
[docs]
class ErTricks(ContinualModel):
"""Experience Replay with tricks from `Rethinking Experience Replay: a Bag of Tricks for Continual Learning`."""
NAME = 'er_tricks'
COMPATIBILITY = ['class-il', 'domain-il', 'task-il', 'general-continual']
[docs]
@staticmethod
def get_parser(parser) -> ArgumentParser:
"""
Returns an ArgumentParser object with predefined arguments for the Er model.
"""
add_rehearsal_args(parser)
parser.add_argument('--bic_epochs', type=int, default=50, help='bias injector.')
parser.add_argument('--elrd', type=float, default=1)
parser.add_argument('--sample_selection_strategy', default='labrs', type=str, choices=['reservoir', 'lars', 'labrs'],
help='Sample selection strategy to use: `reservoir`, `lars` (Loss-Aware Reservoir Sampling), `labrs` (Loss-Aware Balanced Reservoir Sampling)')
return parser
def __init__(self, backbone, loss, args, transform, dataset=None):
"""
The ER model maintains a buffer of previously seen examples and uses them to augment the current batch during training.
"""
super(ErTricks, self).__init__(backbone, loss, args, transform, dataset=dataset)
self.buffer = Buffer(self.args.buffer_size, device=self.device, sample_selection_strategy=self.args.sample_selection_strategy)
# BIC
self.bic_params = torch.zeros(2, device=self.device, requires_grad=True)
self.bic_opt = torch.optim.SGD([self.bic_params], lr=0.5)
[docs]
def end_task(self, dataset):
self.net.eval()
for l in range(self.args.bic_epochs):
data = self.buffer.get_data(self.args.buffer_size, transform=dataset.get_normalization_transform())
while data[0].shape[0] > 0:
inputs, labels = data[0][:self.args.batch_size], data[1][:self.args.batch_size]
data = (data[0][self.args.batch_size:], data[1][self.args.batch_size:])
self.bic_opt.zero_grad()
with torch.no_grad():
out = self.net(inputs)
out[:, self.n_past_classes:self.n_seen_classes] *= self.bic_params[1].repeat_interleave(self.n_classes_current_task)
out[:, self.n_past_classes:self.n_seen_classes] += self.bic_params[0].repeat_interleave(self.n_classes_current_task)
loss_bic = self.loss(out, labels)
loss_bic.backward()
self.bic_opt.step()
self.net.train()
[docs]
def observe(self, inputs, labels, not_aug_inputs, epoch=None):
"""
ER trains on the current task using the data provided, but also augments the batch with data from the buffer.
"""
apply_decay(self.args.elrd, self.args.lr, self.opt, self.buffer.num_seen_examples)
real_batch_size = inputs.shape[0]
self.opt.zero_grad()
if not self.buffer.is_empty():
buf_indexes, buf_inputs, buf_labels = self.buffer.get_data(
self.args.minibatch_size, transform=self.transform, device=self.device, return_index=True)
inputs = torch.cat((inputs, buf_inputs))
labels = torch.cat((labels, buf_labels))
outputs = self.net(inputs)
loss_scores = self.loss(outputs, labels, reduction='none')
loss = loss_scores.mean()
loss.backward()
self.opt.step()
if not self.buffer.is_empty():
self.buffer.sample_selection_fn.update(buf_indexes, -loss_scores.detach()[real_batch_size:])
self.buffer.add_data(examples=not_aug_inputs,
labels=labels[:real_batch_size],
sample_selection_scores=-loss_scores.detach()[:real_batch_size])
return loss.item()
[docs]
def forward(self, x: torch.Tensor) -> torch.Tensor:
ret = super(ErTricks, self).forward(x)
if ret.shape[0] > 0:
ret[:, self.n_past_classes:self.n_seen_classes] *= self.bic_params[1].repeat_interleave(self.n_classes_current_task)
ret[:, self.n_past_classes:self.n_seen_classes] += self.bic_params[0].repeat_interleave(self.n_classes_current_task)
return ret