Merge branch 'fix-agem' of https://github.com/JonasFrey96/avalanche i…

…nto agem

avalanche/training/plugins/agem.py

@@ -53,22 +53,24 @@ def after_backward(self, strategy, **kwargs):
         """
         Project gradient based on reference gradients
         """
-
         if self.memory_x is not None:
-            for (n1, p1), (n2, refg) in zip(strategy.model.named_parameters(),
-                                            self.reference_gradients):
-
-                assert n1 == n2, "Different model parameters in AGEM projection"
-                assert (p1.grad is not None and refg is not None) or \
-                       (p1.grad is None and refg is None)
-
-                if refg is None:
-                    continue
-
-                dotg = torch.dot(p1.grad.view(-1), refg.view(-1))
-                dotref = torch.dot(refg.view(-1), refg.view(-1))
-                if dotg < 0:
-                    p1.grad -= (dotg / dotref) * refg
+            current_gradients = [p.grad.view(-1)
+                for n, p in strategy.model.named_parameters() if p.requires_grad]
+            current_gradients = torch.cat(current_gradients)
+
+            assert current_gradients.shape == self.reference_gradients.shape , "Different model parameters in AGEM projection"
+
+            dotg = torch.dot( current_gradients, self.reference_gradients)
+            if dotg < 0:
+                alpha2 = dotg / torch.dot(self.reference_gradients, self.reference_gradients)
+                grad_proj = current_gradients - self.reference_gradients * alpha2
+
+                count = 0 
+                for n, p in strategy.model.named_parameters():
+                    if p.requires_grad:
+                        n_param = p.numel()      
+                        p.grad.copy_( grad_proj[count:count+n_param].view_as(p) )
+                        count += n_param
 
     def after_training_exp(self, strategy, **kwargs):
         """
@@ -97,26 +99,30 @@ def update_memory(self, dataloader):
         """
         Update replay memory with patterns from current experience.
         """
-
         tot = 0
-        for mbatch in dataloader:
-            x, y, _ = mbatch
-            if tot + x.size(0) <= self.patterns_per_experience:
-                if self.memory_x is None:
-                    self.memory_x = x.clone()
-                    self.memory_y = y.clone()
-                else:
-                    self.memory_x = torch.cat((self.memory_x, x), dim=0)
-                    self.memory_y = torch.cat((self.memory_y, y), dim=0)
-            else:
-                diff = self.patterns_per_experience - tot
-                if self.memory_x is None:
-                    self.memory_x = x[:diff].clone()
-                    self.memory_y = y[:diff].clone()
+        done = False
+        for batches in dataloader:
+            for _, (x, y) in batches.items():
+                if tot + x.size(0) <= self.patterns_per_experience:
+                    if self.memory_x is None:
+                        self.memory_x = x.clone()
+                        self.memory_y = y.clone()
+                    else:
+                        self.memory_x = torch.cat((self.memory_x, x), dim=0)
+                        self.memory_y = torch.cat((self.memory_y, y), dim=0)
+                    tot += x.size(0)
                 else:
-                    self.memory_x = torch.cat((self.memory_x,
-                                               x[:diff]), dim=0)
-                    self.memory_y = torch.cat((self.memory_y,
-                                               y[:diff]), dim=0)
-                break
-            tot += x.size(0)
+                    diff = self.patterns_per_experience - tot
+                    if self.memory_x is None:
+                        self.memory_x = x[:diff].clone()
+                        self.memory_y = y[:diff].clone()
+                    else:
+                        self.memory_x = torch.cat((self.memory_x,
+                                                   x[:diff]), dim=0)
+                        self.memory_y = torch.cat((self.memory_y,
+                                                   y[:diff]), dim=0)
+                    tot += diff
+                    done = True
+
+                if done: break
+            if done: break

examples/gem_agem_mnist.py

@@ -32,7 +32,7 @@
 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: 51.4%
+experience: 83.5%
 
 AGEM-SMNIST:
 Patterns per experience = sample size: 256, 512, 1024. Performance on previous 
@@ -41,7 +41,7 @@
 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: 23.5%
+experience: 67.0%
 
 """