Update

diff --git a/beaver.py b/beaver.py
index bd17365..f5f092b 100755 (executable)
--- a/beaver.py
+++ b/beaver.py
@@ -134,7 +134,7 @@ for n in vars(args):
 def generation_order(x, fixed_len):
     if args.random_regression_order:
         order = torch.rand(x.size(), device=x.device)
-        order[:, :fixed_len] = torch.linspace(-2, -1, fixed_len, device=order.device)
+        order[:, :fixed_len] = torch.linspace(-2, -1, fixed_len, device=x.device)
         order = order.sort(1).indices
     else:
         order = (
@@ -167,7 +167,9 @@ def shuffle(x, fixed_len):
 
 
 def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
-    for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
+    for input, ar_mask, order in zip(
+        input.split(batch_size), ar_mask.split(batch_size), order.split(batch_size)
+    ):
         i = (ar_mask.sum(0) > 0).nonzero()
         if i.min() > 0:
             # Needed to initialize the model's cache
@@ -186,7 +188,7 @@ def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None)
 ######################################################################
 
 
-def compute_perplexity(model, fixed_len, split="train"):
+def compute_perplexity(model, task, fixed_len, split="train"):
     with torch.autograd.no_grad():
         t = model.training
         model.eval()
@@ -469,7 +471,11 @@ class TaskMaze(Task):
             ar_mask = result.new_zeros(result.size())
             ar_mask[:, self.height * self.width :] = 1
             result *= 1 - ar_mask
-            masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
+            x, order = shuffle(result, self.height * self.width)
+            masked_inplace_autoregression(
+                model, self.batch_size, x, ar_mask, order=order
+            )
+            result = reorder(x, order, back=True)
 
             mazes, paths = self.seq2map(input)
             _, predicted_paths = self.seq2map(result)
@@ -588,10 +594,10 @@ log_string(f"learning_rate_schedule {learning_rate_schedule}")
 if nb_epochs_finished >= args.nb_epochs:
     n_epoch = nb_epochs_finished
     train_perplexity = compute_perplexity(
-        model, fixed_len=task.height * task.width, split="train"
+        model, task, fixed_len=task.height * task.width, split="train"
     )
     test_perplexity = compute_perplexity(
-        model, fixed_len=task.height * task.width, split="test"
+        model, task, fixed_len=task.height * task.width, split="test"
     )
 
     log_string(
@@ -635,7 +641,7 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 
     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
     test_perplexity = compute_perplexity(
-        model, fixed_len=task.height * task.width, split="test"
+        model, task, fixed_len=task.height * task.width, split="test"
     )
 
     log_string(