Merge pull request #65 from rctn/64-bug-learn_dictionary-uses-only-on…

…e-batch-update-per-epoch

Iterate through all of dataloader each epoch.

sparsecoding/models.py

@@ -104,29 +104,18 @@ def learn_dictionary(self, dataset, n_epoch, batch_size):
         losses = []
 
         dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=True)
-        iterloader = iter(dataloader)
-        for i in range(n_epoch):
-            try:
-                batch = next(iterloader)
-            except StopIteration:
-                dataloader = DataLoader(dataset, batch_size=batch_size,
-                                        shuffle=True)
-                iterloader = iter(dataloader)
-                batch = next(iterloader)
-
-            # infer coefficients
-            a = self.inference_method.infer(batch, self.dictionary)
-
-            # update dictionary
-            self.update_dictionary(batch, a)
-
-            # normalize dictionary
-            self.normalize_dictionary()
-
-            # compute current loss
-            loss = self.compute_loss(batch, a)
-
-            losses.append(loss)
+        for _ in range(n_epoch):
+            loss = 0.0
+            for batch in dataloader:
+                # infer coefficients
+                a = self.inference_method.infer(batch, self.dictionary)
+                # update dictionary
+                self.update_dictionary(batch, a)
+                # normalize dictionary
+                self.normalize_dictionary()
+                # compute current loss
+                loss += self.compute_loss(batch, a)
+            losses.append(loss/len(dataloader))
         return np.asarray(losses)
 
     def compute_loss(self, data, a):