Update

diff --git a/beaver.py b/beaver.py
index 065cda0..5407859 100755 (executable)
--- a/beaver.py
+++ b/beaver.py
@@ -133,10 +133,10 @@ for n in vars(args):
 ######################################################################
 
 
-def generation_order(x, fixed_len=0):
+def generation_order(x, prompt_len=0):
     if args.random_regression_order:
         order = torch.rand(x.size(), device=x.device)
-        order[:, :fixed_len] = torch.arange(-fixed_len, 0, device=x.device)
+        order[:, :prompt_len] = torch.arange(-prompt_len, 0, device=x.device)
         order = order.sort(1).indices
     else:
         order = (
@@ -156,13 +156,13 @@ def reorder(x, order, reverse=False):  # x is NxTxD1x...xDk, order is NxT'
     return v
 
 
-def shuffle(x, fixed_len):
-    order = generation_order(x, fixed_len)
+def shuffle(x, prompt_len):
+    order = generation_order(x, prompt_len)
     return reorder(x, order), order
 
 
-def eval_mygpt(model, input, mode="standard", fixed_len=0):
-    x, order = shuffle(input, fixed_len)
+def eval_mygpt(model, input, mode="standard", prompt_len=0):
+    x, order = shuffle(input, prompt_len)
     x = model(mygpt.BracketedSequence(x), mode=mode, order=order).x
     return reorder(x, order, reverse=True)
 
@@ -195,7 +195,7 @@ def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None)
 ######################################################################
 
 
-def compute_perplexity(model, task, fixed_len, split="train"):
+def compute_perplexity(model, task, prompt_len, split="train"):
     with torch.autograd.no_grad():
         t = model.training
         model.eval()
@@ -204,7 +204,7 @@ def compute_perplexity(model, task, fixed_len, split="train"):
 
         for input in task.batches(split=split):
             input = input.to(device)
-            output = eval_mygpt(model, input, fixed_len=fixed_len)
+            output = eval_mygpt(model, input, prompt_len=prompt_len)
             if args.noncausal_prompt:
                 d = input.size(1) // 2
                 loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
@@ -273,7 +273,7 @@ def oneshot(gpt, task):
         acc_train_loss, nb_train_samples = 0, 0
         for mazes, policies in task.policy_batches(split="train"):
             output_gpt = eval_mygpt(
-                gpt, mazes, mode=args.oneshot_input, fixed_len=task.height * task.width
+                gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
             )
             output = model(output_gpt)
 
@@ -288,7 +288,7 @@ def oneshot(gpt, task):
         acc_test_loss, nb_test_samples = 0, 0
         for mazes, policies in task.policy_batches(split="test"):
             output_gpt = eval_mygpt(
-                gpt, mazes, mode=args.oneshot_input, fixed_len=task.height * task.width
+                gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
             )
             output = model(output_gpt)
             loss = compute_loss(mazes, output, policies, task.height, task.width)
@@ -303,7 +303,7 @@ def oneshot(gpt, task):
         mazes = task.test_input[:32, : task.height * task.width]
         policies = task.test_policies[:32]
         output_gpt = eval_mygpt(
-            gpt, mazes, mode=args.oneshot_input, fixed_len=task.height * task.width
+            gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
         )
         output = model(output_gpt)
         if args.oneshot_output == "policy":
@@ -523,13 +523,15 @@ log_string(f"vocabulary_size {vocabulary_size}")
 
 ##############################
 
+
 def noncausal_prompt_amm_generator(d):
     q = torch.arange(d)[:, None]
     k = torch.arange(d)[None, :]
     s = args.maze_height * args.maze_width
-#    return torch.logical_and(q < k, torch.logical_or(q >= s, k >= s))
+    #    return torch.logical_and(q < k, torch.logical_or(q >= s, k >= s))
     return q < k
 
+
 amm_generator = None
 
 if args.noncausal_prompt:
@@ -616,10 +618,10 @@ 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, task, fixed_len=task.height * task.width, split="train"
+        model, task, prompt_len=task.height * task.width, split="train"
     )
     test_perplexity = compute_perplexity(
-        model, task, fixed_len=task.height * task.width, split="test"
+        model, task, prompt_len=task.height * task.width, split="test"
     )
 
     log_string(
@@ -650,9 +652,7 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 
     for input in task.batches(split="train"):
         input = input.to(device)
-        output = eval_mygpt(
-            model, input, fixed_len=task.height * task.width
-        )
+        output = eval_mygpt(model, input, prompt_len=task.height * task.width)
         if args.noncausal_prompt:
             d = input.size(1) // 2
             loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
@@ -667,7 +667,7 @@ 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, task, fixed_len=task.height * task.width, split="test"
+        model, task, prompt_len=task.height * task.width, split="test"
     )
 
     log_string(

diff --git a/mygpt.py b/mygpt.py
index 06b56df..4555b1e 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -148,8 +148,7 @@ class QKVAttention(nn.Module):
 
         if amm_generator is None:
             self.amm_generator = (
-                lambda d: torch.arange(d)[:, None]
-                < torch.arange(d)[None, :]
+                lambda d: torch.arange(d)[:, None] < torch.arange(d)[None, :]
             )
         else:
             self.amm_generator = amm_generator
@@ -190,7 +189,9 @@ class QKVAttention(nn.Module):
 
         if self.causal:
             if bs_q.first == 0:
-                self.cache_attzero = self.amm_generator(x_q.size(1)).to(q.device)[None, None,:,:]
+                self.cache_attzero = self.amm_generator(x_q.size(1)).to(q.device)[
+                    None, None, :, :
+                ]
             a = a.masked_fill(
                 self.cache_attzero[
                     :, :, bs_q.first : bs_q.first + bs_q.nb, : bs_q.first + bs_q.nb