Merge branch 'master' of https://github.com/ContinualAI/avalanche int…

…o newdocs

avalanche/training/plugins/early_stopping.py

@@ -115,18 +115,20 @@ def _update_best(self, strategy):
                 f"Metric {self.metric_name} used by the EarlyStopping plugin "
                 f"is not computed yet. EarlyStopping will not be triggered."
             )
-        if self.best_val is None or self.operator(val_acc, self.best_val):
+
+        if self.best_val is None:
+            self.best_state = deepcopy(strategy.model.state_dict())
+            self.best_val = val_acc
+            self.best_step = self._get_strategy_counter(strategy)
+            return None
+
+        delta_val = float(val_acc - self.best_val)
+        if self.operator(delta_val, 0) and abs(delta_val) >= self.margin:
             self.best_state = deepcopy(strategy.model.state_dict())
-            if self.best_val is None:
-                self.best_val = val_acc
-                self.best_step = 0
-                return None
-
-            if self.operator(float(val_acc - self.best_val), self.margin):
-                self.best_step = self._get_strategy_counter(strategy)
-                self.best_val = val_acc
-                if self.verbose:
-                    print("EarlyStopping: new best value:", val_acc)
+            self.best_val = val_acc
+            self.best_step = self._get_strategy_counter(strategy)
+            if self.verbose:
+                print("EarlyStopping: new best value:", val_acc)
 
         return self.best_val
 

avalanche/training/supervised/lamaml_v2.py

@@ -99,7 +99,6 @@ def __init__(
             buffer_mb_size=buffer_mb_size,
             device=device,
         )
-
         self.model.apply(init_kaiming_normal)
 
     def _before_training_exp(self, **kwargs):
@@ -305,16 +304,15 @@ def __len__(self):
 
     def get_buffer_batch(self):
         rnd_ind = torch.randperm(len(self))[: self.buffer_mb_size]
-        buff_x = torch.cat(
-            [self.storage_policy.buffer[i][0].unsqueeze(0) for i in rnd_ind]
-        ).to(self.device)
-        buff_y = torch.LongTensor(
-            [self.storage_policy.buffer[i][1] for i in rnd_ind]
-        ).to(self.device)
-        buff_t = torch.LongTensor(
-            [self.storage_policy.buffer[i][2] for i in rnd_ind]
-        ).to(self.device)
-
+        buff = self.storage_policy.buffer.subset(rnd_ind)
+        buff_x, buff_y, buff_t = [], [], []
+        for bx, by, bt in buff:
+            buff_x.append(bx)
+            buff_y.append(by)
+            buff_t.append(bt)
+        buff_x = torch.stack(buff_x, dim=0).to(self.device)
+        buff_y = torch.tensor(buff_y).to(self.device).long()
+        buff_t = torch.tensor(buff_t).to(self.device).long()
         return buff_x, buff_y, buff_t
 
 

avalanche/training/supervised/supervised_contrastive_replay.py

@@ -34,6 +34,7 @@ class SCR(SupervisedTemplate):
 
     def __init__(
         self,
+        *,
         model: SCRModel,
         optimizer: Optimizer,
         augmentations=Compose([Lambda(lambda el: el)]),

docs/benchmarks.rst

@@ -37,6 +37,7 @@ Scenarios
 
     NCScenario
     NIScenario
+    benchmark_with_validation_stream
 
 Streams
 """""""""
@@ -279,6 +280,7 @@ Train/Validation splits for streams:
 
 .. currentmodule:: avalanche.benchmarks.utils
 
+
 Utils (Data Loading and AvalancheDataset)
 --------------------------------------------------------------------------------
 

extra_dependencies.txt

@@ -4,8 +4,8 @@
 # package[version_required]: tag1, tag2, ...
 
 higher:               extra
-ctrl-benchmark:       extra
 torchaudio:           extra
+ctrl-benchmark:       ctrl
 gym:                  rl
 pycocotools:          detection
 lvis:                 detection