avalanche_extension/avalanche/examples/gem_agem_mnist.py

import torch
import argparse
from avalanche.benchmarks import PermutedMNIST, SplitMNIST
from avalanche.training.strategies import GEM, AGEM
from avalanche.models import SimpleMLP
from avalanche.evaluation.metrics import forgetting_metrics, accuracy_metrics, \
    loss_metrics
from avalanche.logging import InteractiveLogger
from avalanche.training.plugins import EvaluationPlugin

"""
This example tests both GEM and A-GEM on Split MNIST and Permuted MNIST.
GEM is a streaming strategy, that is it uses only 1 training epochs.
A-GEM may use a larger number of epochs.
Both GEM and A-GEM work with small mini batches (usually with 10 patterns).

Warning1: This implementation of GEM and A-GEM does not use task vectors.
Warning2: GEM is much slower than A-GEM.

Results (learning rate is always 0.1):

GEM-PMNIST (5 experiences):
Hidden size 512. 1 training epoch. 512 patterns per experience, 0.5 memory 
strength. Average Accuracy over all experiences at the end of training on the 
last experience: 92.6%

GEM-SMNIST:
Patterns per experience: 256, Memory strength: 0.5, hidden size: 256
Average Accuracy over all experiences at the end of training on the last 
experience: 93.3%

AGEM-PMNIST (5 experiences):
Patterns per experience = sample size: 256. 256 hidden size, 1 training epoch.
Average Accuracy over all experiences at the end of training on the last 
experience: 83.5%

AGEM-SMNIST:
Patterns per experience = sample size: 256, 512, 1024. Performance on previous 
tasks remains very bad in terms of forgetting. Training epochs do not change 
result.
Hidden size 256.
Results for 1024 patterns per experience and sample size, 1 training epoch.
Average Accuracy over all experiences at the end of training on the last 
experience: 67.0%

"""


def main(args):
    model = SimpleMLP(hidden_size=args.hs)
    optimizer = torch.optim.SGD(model.parameters(), lr=args.lr)
    criterion = torch.nn.CrossEntropyLoss()

    # check if selected GPU is available or use CPU
    assert args.cuda == -1 or args.cuda >= 0, "cuda must be -1 or >= 0."
    device = torch.device(f"cuda:{args.cuda}"
                          if torch.cuda.is_available() and
                          args.cuda >= 0 else "cpu")
    print(f'Using device: {device}')

    # create scenario
    if args.scenario == 'pmnist':
        scenario = PermutedMNIST(n_experiences=args.permutations)
    elif args.scenario == 'smnist':
        scenario = SplitMNIST(n_experiences=5, return_task_id=False)
    else:
        raise ValueError("Wrong scenario name. Allowed pmnist, smnist.")

    # choose some metrics and evaluation method
    interactive_logger = InteractiveLogger()

    eval_plugin = EvaluationPlugin(
        accuracy_metrics(minibatch=True, epoch=True,
                         experience=True, stream=True),
        loss_metrics(minibatch=True, epoch=True,
                     experience=True, stream=True),
        forgetting_metrics(experience=True),
        loggers=[interactive_logger])

    # create strategy
    if args.strategy == 'gem':
        strategy = GEM(model, optimizer, criterion, args.patterns_per_exp,
                       args.memory_strength, train_epochs=args.epochs,
                       device=device, train_mb_size=10, evaluator=eval_plugin)
    elif args.strategy == 'agem':
        strategy = AGEM(model, optimizer, criterion, args.patterns_per_exp,
                        args.sample_size, train_epochs=args.epochs,
                        device=device, train_mb_size=10, evaluator=eval_plugin)
    else:
        raise ValueError("Wrong strategy name. Allowed gem, agem.")
    # train on the selected scenario with the chosen strategy
    print('Starting experiment...')
    results = []
    for experience in scenario.train_stream:
        print("Start training on experience ", experience.current_experience)

        strategy.train(experience)
        print("End training on experience ", experience.current_experience)
        print('Computing accuracy on the test set')
        results.append(strategy.eval(scenario.test_stream[:]))


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--strategy', type=str, choices=['gem', 'agem'],
                        default='gem', help='Choose between GEM and A-GEM')
    parser.add_argument('--scenario', type=str,
                        choices=['pmnist', 'smnist'], default='smnist',
                        help='Choose between Permuted MNIST, Split MNIST.')
    parser.add_argument('--patterns_per_exp', type=int, default=256,
                        help='Patterns to store in the memory for each'
                             ' experience')
    parser.add_argument('--sample_size', type=int, default=256,
                        help='Number of patterns to sample from memory when \
                        projecting gradient. A-GEM only.')
    parser.add_argument('--memory_strength', type=float, default=0.5,
                        help='Offset to add to the projection direction. '
                             'GEM only.')
    parser.add_argument('--lr', type=float, default=1e-1, help='Learning rate.')
    parser.add_argument('--hs', type=int, default=256, help='MLP hidden size.')
    parser.add_argument('--epochs', type=int, default=1,
                        help='Number of training epochs.')
    parser.add_argument('--permutations', type=int, default=5,
                        help='Number of experiences in Permuted MNIST.')
    parser.add_argument('--cuda', type=int, default=0,
                        help='Specify GPU id to use. Use CPU if -1.')
    args = parser.parse_args()

    main(args)