fix bugs in original code

hoedt · Jan 14, 2021 · c96c2ea · c96c2ea
1 parent 97a930c
commit c96c2ea
Show file tree

Hide file tree

Showing 5 changed files with 27 additions and 33 deletions.
diff --git a/experiments/sequential_mnist.py b/experiments/sequential_mnist.py
@@ -257,7 +257,8 @@
     operation=args.operation,
     use_cuda=args.cuda,
     seed=args.seed,
-    mnist_digits=args.mnist_digits
+    mnist_digits=args.mnist_digits,
+    num_workers=2  # debugging
 )
 dataset_train = dataset.fork(seq_length=args.interpolation_length, subset='train').dataloader(shuffle=True)
 # Seeds are from random.org

diff --git a/experiments/simple_function_recurrent.py b/experiments/simple_function_recurrent.py
@@ -56,7 +56,7 @@
 print(f'  - max_iterations: {args.max_iterations}')
 
 # Prepear logging
-results_writer = stable_nalu.writer.ResultsWriter('simple_function_recurrent')
+# results_writer = stable_nalu.writer.ResultsWriter('simple_function_recurrent')
 summary_writer = stable_nalu.writer.SummaryWriter(
     f'simple_function_recurrent/{args.layer_type.lower()}_{args.operation.lower()}_{args.seed}'
 )
@@ -138,15 +138,15 @@ def test_model(dataloader):
 print(f'  - loss_valid_extra: {loss_valid_extra}')
 
 # save results
-results_writer.add({
-    'seed': args.seed,
-    'operation': args.operation,
-    'layer_type': args.layer_type,
-    'simple': args.simple,
-    'cuda': args.cuda,
-    'verbose': args.verbose,
-    'max_iterations': args.max_iterations,
-    'loss_train': loss_train,
-    'loss_valid_inter': loss_valid_inter,
-    'loss_valid_extra': loss_valid_extra
-})
+#results_writer.add({
+#    'seed': args.seed,
+#    'operation': args.operation,
+#    'layer_type': args.layer_type,
+#    'simple': args.simple,
+#    'cuda': args.cuda,
+#    'verbose': args.verbose,
+#    'max_iterations': args.max_iterations,
+#    'loss_train': loss_train,
+#    'loss_valid_inter': loss_valid_inter,
+#    'loss_valid_extra': loss_valid_extra
+#})
diff --git a/stable_nalu/dataset/sequential_mnist.py b/stable_nalu/dataset/sequential_mnist.py
@@ -16,7 +16,7 @@ class ItemShape(NamedTuple):
 class OPERATIONS:
     @staticmethod
     def sum(seq):
-        return OPERATIONS.sum(seq)
+        return np.sum(seq, keepdims=True).reshape(-1, 1)
 
     @staticmethod
     def cumsum(seq):
@@ -55,25 +55,9 @@ def __init__(self, operation,
         self._rng = np.random.RandomState(seed)
         self._mnist_digits = set(mnist_digits)
 
-    def is_cum_task():
-        if self._operation == OPERATIONS.sum:
-            return False
-        elif self._operation == OPERATIONS.cumsum:
-            return True
-        elif self._operation == OPERATIONS.prod:
-            return False
-        elif self._operation == OPERATIONS.cumprod:
-            return True
-        elif self._operation == OPERATIONS.div:
-            return False
-        elif self._operation == OPERATIONS.cumdiv:
-            return True
-        else:
-            raise ValueError('bad operation')
-
     def get_item_shape(self):
         if self._operation == OPERATIONS.sum:
-            return ItemShape((None, 28, 28), (None, 1))
+            return ItemShape((None, 28, 28), (1, 1))
         elif self._operation == OPERATIONS.cumsum:
             return ItemShape((None, 28, 28), (None, 1))
         elif self._operation == OPERATIONS.prod:

diff --git a/stable_nalu/writer/save_model.py b/stable_nalu/writer/save_model.py
@@ -13,4 +13,13 @@
 def save_model(name, model):
     save_file = path.join(SAVE_DIR, name) + '.pth'
     os.makedirs(path.dirname(save_file), exist_ok=True)
+
+    def remove_writer(m):
+        # remove writers from ExtendedTorchModules
+        try:
+            del m.writer
+        except AttributeError:
+            pass
+
+    model.apply(remove_writer)
     torch.save(model, save_file)
diff --git a/stable_nalu/writer/summary_writer.py b/stable_nalu/writer/summary_writer.py
@@ -86,7 +86,7 @@ def add_tensor(self, name, matrix, verbose_only=True):
     def add_histogram(self, name, tensor, verbose_only=True):
         if torch.isnan(tensor).any():
             print(f'nan detected in {self._namespace}/{name}')
-            tensor = torch.where(torch.isnan(tensor), torch.tensor(0, dtype=tensor.dtype), tensor)
+#            tensor = torch.where(torch.isnan(tensor), torch.tensor(0, dtype=tensor.dtype), tensor)
             raise ValueError('nan detected')
 
         if self.is_log_iteration() and self.is_logging_enabled() and self.is_verbose(verbose_only):