avalanche_extension/avalanche/examples/deep_slda.py

################################################################################
# Copyright (c) 2021 ContinualAI.                                              #
# Copyrights licensed under the MIT License.                                   #
# See the accompanying LICENSE file for terms.                                 #
#                                                                              #
# Date: 06-04-2021                                                             #
# Author(s): Tyler Hayes                                                       #
# E-mail: contact@continualai.org                                              #
# Website: avalanche.continualai.org                                           #
################################################################################

"""
This is a simple example on how to use the Deep SLDA strategy.
"""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import argparse
import torch
import warnings
from torchvision import transforms

from avalanche.training.plugins import EvaluationPlugin
from avalanche.evaluation.metrics import accuracy_metrics, loss_metrics, \
    forgetting_metrics
from avalanche.logging import InteractiveLogger
from avalanche.benchmarks.classic import CORe50
from avalanche.training.strategies.deep_slda import StreamingLDA
from avalanche.models import SLDAResNetModel


def main(args):
    # Device config
    device = torch.device(f"cuda:{args.cuda}"
                          if torch.cuda.is_available() and
                          args.cuda >= 0 else "cpu")
    print('device ', device)
    # ---------

    # --- TRANSFORMATIONS
    _mu = [0.485, 0.456, 0.406]  # imagenet normalization
    _std = [0.229, 0.224, 0.225]
    transform = transforms.Compose([
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
        transforms.Normalize(mean=_mu,
                             std=_std)
    ])
    # ---------

    # --- BENCHMARK CREATION
    benchmark = CORe50(scenario=args.scenario, train_transform=transform,
                       eval_transform=transform)

    # ---------

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

    criterion = torch.nn.CrossEntropyLoss()
    model = SLDAResNetModel(device=device, arch='resnet18',
                            imagenet_pretrained=args.imagenet_pretrained)

    # CREATE THE STRATEGY INSTANCE
    cl_strategy = StreamingLDA(model, criterion,
                               args.feature_size, args.n_classes,
                               eval_mb_size=args.batch_size,
                               train_mb_size=args.batch_size,
                               train_epochs=1,
                               shrinkage_param=args.shrinkage,
                               streaming_update_sigma=args.plastic_cov,
                               device=device, evaluator=eval_plugin)

    warnings.warn(
        "The Deep SLDA example is not perfectly aligned with "
        "the paper implementation since it does not use a base "
        "initialization phase and instead starts streming from "
        "pre-trained weights.")

    # TRAINING LOOP
    print('Starting experiment...')
    for i, exp in enumerate(benchmark.train_stream):

        # fit SLDA model to batch (one sample at a time)
        cl_strategy.train(exp)

        # evaluate model on test data
        cl_strategy.eval(benchmark.test_stream)


if __name__ == '__main__':
    parser = argparse.ArgumentParser('SLDA Example with ResNet-18 on CORe50')
    parser.add_argument('--cuda', type=int, default=0,
                        help='Select zero-indexed cuda device. -1 to use CPU.')

    parser.add_argument('--n_classes', type=int, default=50)
    parser.add_argument('--scenario', type=str, default="nc",
                        choices=['ni', 'nc', 'nic', 'nicv2_79', 'nicv2_196',
                                 'nicv2_391'])

    # deep slda model parameters
    parser.add_argument('--imagenet_pretrained', type=bool,
                        default=True)  # initialize backbone with
    # imagenet pre-trained weights
    parser.add_argument('--feature_size', type=int,
                        default=512)  # feature size before output layer
    # (512 for resnet-18)
    parser.add_argument('--shrinkage', type=float,
                        default=1e-4)  # shrinkage value
    parser.add_argument('--plastic_cov', type=bool,
                        default=True)  # plastic covariance matrix
    parser.add_argument('--batch_size', type=int, default=512)

    args = parser.parse_args()
    main(args)