r_walk.py

import torch
import itertools
from argparse import ArgumentParser
from torch.utils.data import DataLoader

from .incremental_learning import Inc_Learning_Appr
from datasets.exemplars_dataset import ExemplarsDataset


class Appr(Inc_Learning_Appr):
    """Class implementing the Riemannian Walk (RWalk) approach described in
    http://openaccess.thecvf.com/content_ECCV_2018/papers/Arslan_Chaudhry__Riemannian_Walk_ECCV_2018_paper.pdf
    """

    def __init__(self, model, device, nepochs=100, lr=0.05, lr_min=1e-4, lr_factor=3, lr_patience=5, clipgrad=10000,
                 momentum=0, wd=0, multi_softmax=False, wu_nepochs=0, wu_lr_factor=1, fix_bn=False, eval_on_train=False,
                 logger=None, exemplars_dataset=None, lamb=1, alpha=0.5, damping=0.1, fim_sampling_type='max_pred',
                 fim_num_samples=-1):
        super(Appr, self).__init__(model, device, nepochs, lr, lr_min, lr_factor, lr_patience, clipgrad, momentum, wd,
                                   multi_softmax, wu_nepochs, wu_lr_factor, fix_bn, eval_on_train, logger,
                                   exemplars_dataset)
        self.lamb = lamb
        self.alpha = alpha
        self.damping = damping
        self.sampling_type = fim_sampling_type
        self.num_samples = fim_num_samples

        # In all cases, we only keep importance weights for the model, but not for the heads.
        feat_ext = self.model.model
        # Page 7: "task-specific parameter importance over the entire training trajectory."
        self.w = {n: torch.zeros(p.shape).to(self.device) for n, p in feat_ext.named_parameters() if p.requires_grad}
        # Store current parameters as the initial parameters before first task starts
        self.older_params = {n: p.clone().detach().to(self.device) for n, p in feat_ext.named_parameters()
                             if p.requires_grad}
        # Store scores and fisher information
        self.scores = {n: torch.zeros(p.shape).to(self.device) for n, p in feat_ext.named_parameters()
                       if p.requires_grad}
        self.fisher = {n: torch.zeros(p.shape).to(self.device) for n, p in feat_ext.named_parameters()
                       if p.requires_grad}

    @staticmethod
    def exemplars_dataset_class():
        return ExemplarsDataset

    @staticmethod
    def extra_parser(args):
        """Returns a parser containing the approach specific parameters"""
        parser = ArgumentParser()
        # Eq. 5 and 8: "regularization hyperparameter lambda being less sensitive to the number of tasks. Whereas,
        # EWC and Path Integral are highly sensitive to lambda, making them relatively less reliable for IL"
        parser.add_argument('--lamb', default=1, type=float, required=False,
                            help='Forgetting-intransigence trade-off (default=%(default)s)')
        # Define how old and new fisher is fused, by default it is a 50-50 fusion
        parser.add_argument('--alpha', default=0.5, type=float, required=False,
                            help='RWalk alpha (default=%(default)s)')
        # Damping parameter as in Path Integral
        parser.add_argument('--damping', default=0.1, type=float, required=False,
                            help='(default=%(default)s)')
        parser.add_argument('--fim_sampling_type', default='max_pred', type=str, required=False,
                            choices=['true', 'max_pred', 'multinomial'],
                            help='Sampling type for Fisher information (default=%(default)s)')
        parser.add_argument('--fim_num_samples', default=-1, type=int, required=False,
                            help='Number of samples for Fisher information (-1: all available) (default=%(default)s)')
        return parser.parse_known_args(args)

    def _get_optimizer(self):
        """Returns the optimizer"""
        if len(self.exemplars_dataset) == 0 and len(self.model.heads) > 1:
            # if there are no exemplars, previous heads are not modified
            params = list(self.model.model.parameters()) + list(self.model.heads[-1].parameters())
        else:
            params = self.model.parameters()
        return torch.optim.SGD(params, lr=self.lr, weight_decay=self.wd, momentum=self.momentum)

    def compute_fisher_matrix_diag(self, trn_loader):
        # Store Fisher Information
        fisher = {n: torch.zeros(p.shape).to(self.device) for n, p in self.model.model.named_parameters()
                  if p.requires_grad}
        # Compute fisher information for specified number of samples -- rounded to the batch size
        n_samples_batches = (self.num_samples // trn_loader.batch_size + 1) if self.num_samples > 0 \
            else (len(trn_loader.dataset) // trn_loader.batch_size)
        # Do forward and backward pass to compute the fisher information
        self.model.train()
        for images, targets in itertools.islice(trn_loader, n_samples_batches):
            outputs = self.model.forward(images.to(self.device))

            if self.sampling_type == 'true':
                # Use the labels to compute the gradients based on the CE-loss with the ground truth
                preds = targets.to(self.device)
            elif self.sampling_type == 'max_pred':
                # Not use labels and compute the gradients related to the prediction the model has learned
                preds = torch.cat(outputs, dim=1).argmax(1).flatten()
            elif self.sampling_type == 'multinomial':
                # Use a multinomial sampling to compute the gradients
                probs = torch.nn.functional.softmax(torch.cat(outputs, dim=1), dim=1)
                preds = torch.multinomial(probs, len(targets)).flatten()

            loss = torch.nn.functional.cross_entropy(torch.cat(outputs, dim=1), preds)
            self.optimizer.zero_grad()
            loss.backward()
            # Page 6: "the Fisher component [...] is the expected square of the loss gradient w.r.t the i-th parameter."
            for n, p in self.model.model.named_parameters():
                if p.grad is not None:
                    fisher[n] += p.grad.pow(2) * len(targets)
        # Apply mean across all samples
        n_samples = n_samples_batches * trn_loader.batch_size
        fisher = {n: (p / n_samples) for n, p in fisher.items()}
        return fisher

    def train_loop(self, t, trn_loader, val_loader):
        """Contains the epochs loop"""

        # add exemplars to train_loader
        if len(self.exemplars_dataset) > 0 and t > 0:
            trn_loader = torch.utils.data.DataLoader(trn_loader.dataset + self.exemplars_dataset,
                                                     batch_size=trn_loader.batch_size,
                                                     shuffle=True,
                                                     num_workers=trn_loader.num_workers,
                                                     pin_memory=trn_loader.pin_memory)

        # FINETUNING TRAINING -- contains the epochs loop
        super().train_loop(t, trn_loader, val_loader)

        # EXEMPLAR MANAGEMENT -- select training subset
        self.exemplars_dataset.collect_exemplars(self.model, trn_loader, val_loader.dataset.transform)

    def post_train_process(self, t, trn_loader):
        """Runs after training all the epochs of the task (after the train session)"""

        # calculate Fisher Information Matrix
        curr_fisher = self.compute_fisher_matrix_diag(trn_loader)

        # Eq. 10: efficiently update Fisher Information Matrix
        for n in self.fisher.keys():
            # Added option to accumulate fisher over time with a pre-fixed growing alpha
            if self.alpha == -1:
                alpha = (sum(self.model.task_cls[:t]) / sum(self.model.task_cls)).to(self.device)
                self.fisher[n] = alpha * self.fisher[n] + (1 - alpha) * curr_fisher[n]
            else:
                self.fisher[n] = self.alpha * self.fisher[n] + (1 - self.alpha) * curr_fisher[n]
        # Page 7: Optimization Path-based Parameter Importance: importance scores computation
        curr_score = {n: torch.zeros(p.shape).to(self.device) for n, p in self.model.model.named_parameters()
                      if p.requires_grad}
        with torch.no_grad():
            curr_params = {n: p for n, p in self.model.model.named_parameters() if p.requires_grad}
            for n, p in self.scores.items():
                curr_score[n] = self.w[n] / (
                        self.fisher[n] * ((curr_params[n] - self.older_params[n]) ** 2) + self.damping)
                self.w[n].zero_()
                # Page 7: "Since we care about positive influence of the parameters, negative scores are set to zero."
                curr_score[n] = torch.nn.functional.relu(curr_score[n])
        # Page 8: alleviating regularization getting increasingly rigid by averaging scores
        for n, p in self.scores.items():
            self.scores[n] = (self.scores[n] + curr_score[n]) / 2

        # Store current parameters for the next task
        self.older_params = {n: p.clone().detach() for n, p in self.model.model.named_parameters() if p.requires_grad}

    def train_epoch(self, t, trn_loader):
        """Runs a single epoch"""
        self.model.train()
        if self.fix_bn and t > 0:
            self.model.freeze_bn()
        for images, targets in trn_loader:
            # store current model
            curr_feat_ext = {n: p.clone().detach() for n, p in self.model.model.named_parameters() if p.requires_grad}

            # Forward current model
            outputs = self.model(images.to(self.device))
            # cross-entropy loss on current task
            loss = torch.nn.functional.cross_entropy(torch.cat(outputs, dim=1), targets.to(self.device))
            self.optimizer.zero_grad()
            loss.backward(retain_graph=True)
            # store gradients without regularization term
            unreg_grads = {n: p.grad.clone().detach() for n, p in self.model.model.named_parameters()
                           if p.grad is not None}
            # apply loss with path integral regularization
            loss = self.criterion(t, outputs, targets.to(self.device))

            # Backward
            self.optimizer.zero_grad()
            loss.backward()
            torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.clipgrad)
            self.optimizer.step()

            # Page 7: "accumulate task-specific parameter importance over the entire training trajectory"
            #  "the parameter importance is defined as the ratio of the change in the loss function to the distance
            #  between the conditional likelihod distributions per step in the parameter space."
            with torch.no_grad():
                for n, p in self.model.model.named_parameters():
                    if n in unreg_grads.keys():
                        self.w[n] -= unreg_grads[n] * (p.detach() - curr_feat_ext[n])

    def criterion(self, t, outputs, targets):
        """Returns the loss value"""
        loss = 0
        if t > 0:
            loss_reg = 0
            # Eq. 9: final objective function
            for n, p in self.model.model.named_parameters():
                if n in self.fisher.keys():
                    loss_reg += torch.sum((self.fisher[n] + self.scores[n]) * (p - self.older_params[n]).pow(2))
            loss += self.lamb * loss_reg
        # Current cross-entropy loss -- with exemplars use all heads
        if len(self.exemplars_dataset) > 0:
            return loss + torch.nn.functional.cross_entropy(torch.cat(outputs, dim=1), targets)
        return loss + torch.nn.functional.cross_entropy(outputs[t], targets - self.model.task_offset[t])