Update: GDumb + minor fixes

README.md

@@ -42,11 +42,17 @@ experiment runs with one task on one dataset, results would be equivalent to 'co
 | non-incremental supervised learning | yes | yes | 1 |
 
 Current available approaches include:
-<div align="center">
-<p align="center"><b>
-  Finetuning • Freezing • Joint
 
-  LwF • iCaRL • EWC • PathInt • MAS • RWalk • EEIL • LwM • DMC • BiC • LUCIR • IL2M
+<div align="center"><p align="center"><b>
+
+  [Finetuning](./src/approach/finetuning.py) • [Freezing](./src/approach/freezing.py) • [Joint](./src/approach/joint.py)
+
+  [LwF](./src/approach/lwf.py) • [iCaRL](./src/approach/icarl.py) • [EWC](./src/approach/ewc.py) • 
+  [PathInt](./src/approach/path_integral.py) • [MAS](./src/approach/mas.py) • [RWalk](./src/approach/r_walk.py) • 
+  [EEIL](./src/approach/eeil.py) • [LwM](./src/approach/lwm.py) • [DMC](./src/approach/dmc.py) • 
+  [BiC](./src/approach/bic.py) • [LUCIR](./src/approach/lucir.py) • [IL2M](./src/approach/il2m.py) •
+  [GDumb](./src/approach/gdumb.py)
+
 </b></p>
 </div>
 

src/approach/README.md

@@ -182,3 +182,10 @@ can be combined with Freezing by using `--freeze-after num_task (int)`. However,
 `--approach il2m`
 [ICCV 2019](https://openaccess.thecvf.com/content_ICCV_2019/papers/Belouadah_IL2M_Class_Incremental_Learning_With_Dual_Memory_ICCV_2019_paper.pdf)
 | [code](https://github.com/EdenBelouadah/class-incremental-learning/tree/master/il2m)
+
+### GDumb
+`--approach gdumb`
+[ECCV 2020](https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123470511.pdf)
+| [code](https://github.com/drimpossible/GDumb)
+
+* `--regularization`: Use regularization (default='cutmix')

src/approach/gdumb.py

@@ -0,0 +1,107 @@
+import torch
+import numpy as np
+from copy import deepcopy
+from argparse import ArgumentParser
+from torch.utils.data.dataloader import default_collate
+
+from utils import cutmix_data
+from .incremental_learning import Inc_Learning_Appr
+from datasets.exemplars_dataset import ExemplarsDataset
+
+
+class Appr(Inc_Learning_Appr):
+    """Class implementing the GDumb approach
+    described in https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123470511.pdf
+    Original code available at https://github.com/drimpossible/GDumb
+    """
+
+    def __init__(self, model, device, nepochs=100, lr=0.05, lr_min=0.0005, lr_factor=3, lr_patience=5, clipgrad=10.0,
+                 momentum=0.9, wd=1e-6, multi_softmax=False, wu_nepochs=1, wu_lr_factor=1, fix_bn=False,
+                 eval_on_train=False, logger=None, exemplars_dataset=None, regularization='cutmix'):
+        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.regularization = regularization
+        self.init_model = None
+
+        have_exemplars = self.exemplars_dataset.max_num_exemplars + self.exemplars_dataset.max_num_exemplars_per_class
+        assert (have_exemplars > 0), 'Error: GDumb needs exemplars.'
+
+    @staticmethod
+    def exemplars_dataset_class():
+        return ExemplarsDataset
+
+    @staticmethod
+    def extra_parser(args):
+        """Returns a parser containing the approach specific parameters"""
+        parser = ArgumentParser()
+        parser.add_argument('--regularization', default='cutmix', required=False,
+                            help='Use regularization (default=%(default)s)')
+        return parser.parse_known_args(args)
+
+    def train_loop(self, t, trn_loader, val_loader):
+        """Contains the epochs loop"""
+        # 1. GDumb resets the network before learning a new task, relying only on the exemplars stored so far
+        if t == 0:
+            # Keep the randomly initialized model from time step 0
+            self.init_model = deepcopy(self.model)
+        else: 
+            # Reinitialize the model (backbone) for very task from time step 1
+            self.model.model = deepcopy(self.init_model.model)
+            for layer in self.model.heads.children():
+                layer.reset_parameters()
+
+        # EXEMPLAR MANAGEMENT -- select training subset from current task and exemplar memory
+        aux_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)
+        self.exemplars_dataset.collect_exemplars(self.model, aux_loader, val_loader.dataset.transform)
+
+        # Set new set of exemplars as the only data to train on
+        trn_loader = torch.utils.data.DataLoader(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)
+
+    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:
+            # Get exemplars
+            if len(self.exemplars_dataset) > 0:
+                # 2. Balanced batches
+                exemplar_indices = torch.randperm(len(self.exemplars_dataset))[:trn_loader.batch_size]
+                images_exemplars, targets_exemplars = default_collate([self.exemplars_dataset[i]
+                                                                       for i in exemplar_indices])
+                images = torch.cat((images, images_exemplars), dim=0)
+                targets = torch.cat((targets, targets_exemplars), dim=0) 
+
+            # 3. Apply cutmix as regularization
+            do_cutmix = self.regularization == 'cutmix' and np.random.rand(1) < 0.5  # cutmix_prob (Sec.4)
+            if do_cutmix: 
+                images, targets_a, targets_b, lamb = cutmix_data(x=images, y=targets, alpha=1.0)  # cutmix_alpha (Sec.4)
+                # Forward current model
+                outputs = self.model(images.to(self.device))
+                loss = lamb * self.criterion(t, outputs, targets_a.to(self.device))
+                loss += (1.0 - lamb) * self.criterion(t, outputs, targets_b.to(self.device))
+            else:
+                outputs = self.model(images.to(self.device))
+                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()
+
+    def criterion(self, t, outputs, targets):
+        """Returns the loss value"""
+        return torch.nn.functional.cross_entropy(torch.cat(outputs, dim=1), targets)

src/approach/lwm.py

@@ -160,9 +160,11 @@ def eval(self, t, val_loader):
 
     def attention_distillation_loss(self, attention_map1, attention_map2):
         """Calculates the attention distillation loss"""
-        attention_map1 = torch.norm(attention_map1, p=2, dim=1)
-        attention_map2 = torch.norm(attention_map2, p=2, dim=1)
-        return torch.norm(attention_map2 - attention_map1, p=1, dim=1).sum(dim=1).mean()
+        attention_map1 = torch.nn.functional.normalize(attention_map1.view(attention_map1.size(0), -1),
+                                                       p=2, dim=1, eps=1e-12, out=None)
+        attention_map2 = torch.nn.functional.normalize(attention_map2.view(attention_map1.size(0), -1),
+                                                       p=2, dim=1, eps=1e-12, out=None)
+        return torch.norm(attention_map2 - attention_map1, p=1, dim=1).mean()
 
     def cross_entropy(self, outputs, targets, exp=1.0, size_average=True, eps=1e-5):
         """Calculates cross-entropy with temperature scaling"""

src/datasets/dataset_config.py

@@ -26,6 +26,20 @@
         'flip': True,
         'normalize': ((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023))
     },
+    'cifar100_fixed': {
+        'path': join(_BASE_DATA_PATH, 'cifar100'),
+        'resize': None,
+        'pad': 4,
+        'crop': 32,
+        'flip': True,
+        'normalize': ((0.5071, 0.4866, 0.4409), (0.2009, 0.1984, 0.2023)),
+        'class_order': [
+            0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28,
+            29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55,
+            56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82,
+            83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99
+        ]
+    },
     'cifar100_icarl': {
         'path': join(_BASE_DATA_PATH, 'cifar100'),
         'resize': None,

src/gridsearch_config.py

@@ -65,6 +65,8 @@ def __init__(self):
             'r_walk': {
                 'lamb': [20],
             },
+            'gdumb': {
+            },
         }
         self.current_lr = self.params['general']['lr'][0]
         self.current_tradeoff = 0

src/main_incremental.py

@@ -4,6 +4,7 @@
 import argparse
 import importlib
 import numpy as np
+import torch.multiprocessing
 from functools import reduce
 
 import utils
@@ -91,8 +92,8 @@ def main(argv=None):
     parser.add_argument('--eval-on-train', action='store_true',
                         help='Show train loss and accuracy (default=%(default)s)')
     # gridsearch args
-    parser.add_argument('--gridsearch-tasks', default=-1, type=int,
-                        help='Number of tasks to apply GridSearch (-1: all tasks) (default=%(default)s)')
+    parser.add_argument('--gridsearch-tasks', default=0, type=int,
+                        help='Number of tasks to apply GridSearch (default=%(default)s)')
 
     # Args -- Incremental Learning Framework
     args, extra_args = parser.parse_known_args(argv)
@@ -120,6 +121,8 @@ def main(argv=None):
     else:
         print('WARNING: [CUDA unavailable] Using CPU instead!')
         device = 'cpu'
+    # In case the dataset is too large
+    torch.multiprocessing.set_sharing_strategy('file_system')
     # Multiple gpus
     # if torch.cuda.device_count() > 1:
     #     self.C = torch.nn.DataParallel(C)