Update.

diff --git a/README.txt b/README.txt
index d4cb93d..a4cd46b 100644 (file)
--- a/README.txt
+++ b/README.txt
@@ -1,10 +1,12 @@
 
+======================================================================
 For the stack experiment:
 
 ./main.py --task=stack
 
 Takes ~1h10min on a 4090.
 
+======================================================================
 For the arithmetic expressions experiments
 
 # 38M parameters / 250k samples
@@ -15,3 +17,4 @@ For the arithmetic expressions experiments
   learning rate schedule is obviously terrible
 
 ./main.py --task=expr --nb_blocks=48 --dim_model=1024 --nb_train_samples=2500000 --result_dir=results_expr_48b_d1024_2.5M
+======================================================================

diff --git a/main.py b/main.py
index 71026c5..ba8843b 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -79,7 +79,18 @@ parser.add_argument("--overwrite_results", action="store_true", default=False)
 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 
 ##############################
-# picoclvr options
+# rpl options
+
+parser.add_argument("--rpl-nb_starting_values", type=int, default=5)
+
+parser.add_argument("--rpl-max_input", type=int, default=9)
+
+parser.add_argument("--rpl-prog_len", type=int, default=10)
+
+parser.add_argument("--rpl-nb_runs", type=int, default=8)
+
+##############################
+# sandbox options
 
 parser.add_argument("--sandbox_level", type=int, default=0)
 
@@ -427,6 +438,10 @@ elif args.task == "rpl":
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
         batch_size=args.batch_size,
+        nb_starting_values=args.rpl_nb_starting_values,
+        max_input=args.rpl_max_input,
+        prog_len=args.rpl_prog_len,
+        nb_runs=args.rpl_nb_runs,
         logger=log_string,
         device=device,
     )

diff --git a/tasks.py b/tasks.py
index bad4536..0fac0a7 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -1097,7 +1097,10 @@ class RPL(Task):
         self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
         self.id2token = dict([(n, c) for c, n in self.token2id.items()])
 
-        self.t_nul, self.t_prog = self.token2id["<nul>"], self.token2id["<prog>"]
+        self.t_nul = self.token2id["<nul>"]
+        self.t_prog = self.token2id["<prog>"]
+        self.t_input = self.token2id["<input>"]
+        self.t_output = self.token2id["<output>"]
 
         self.train_input = self.tensorize(train_sequences)
         self.test_input = self.tensorize(test_sequences)
@@ -1133,7 +1136,7 @@ class RPL(Task):
         self, n_epoch, model, result_dir, logger, deterministic_synthesis
     ):
         # --------------------------------------------------------------------
-        def compute_nb_errors(input, nb_to_log=0):
+        def compute_nb_errors_prog(input, nb_to_log=0):
             result = input.clone()
             s = (result == self.t_prog).long()
             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
@@ -1173,9 +1176,51 @@ class RPL(Task):
 
             return sum_nb_total, sum_nb_errors
 
+        # --------------------------------------------------------------------
+        def compute_nb_errors_output(input, nb_to_log=0):
+            result = input.clone()
+            k = torch.arange(result.size(1), device=result.device)[None,:]
+            last_output_idx = ((result == self.t_output) * k).max(dim=1,keep_dim=True)
+            first_prog_idx = ((result == self.t_prog) * k).min(dim=1,keep_dim=True)
+            ar_mask = (k > last_output_idx).long() * (k < first_prog_idx)
+            result = (1 - ar_mask) * result + ar_mask * self.t_nul
+
+            masked_inplace_autoregression(
+                model,
+                self.batch_size,
+                result,
+                ar_mask,
+                deterministic_synthesis,
+                device=self.device,
+            )
+
+            sum_nb_total, sum_nb_errors = 0, 0
+            for x, y in zip(input, result):
+                seq = [self.id2token[i.item()] for i in y]
+                sum_nb_total += 1
+                sum_nb_errors += 0 if (x-y).abs().max() == 0 else 1
+                if nb_to_log > 0:
+                    gt_seq = [self.id2token[i.item()] for i in x]
+                    _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
+                    gt_prog = " ".join([str(x) for x in gt_prog])
+                    prog = " ".join([str(x) for x in prog])
+                    comment = "*" if nb_errors == 0 else "-"
+                    logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
+                    for start_stack, target_stack, result_stack, correct in stacks:
+                        comment = "*" if correct else "-"
+                        start_stack = " ".join([str(x) for x in start_stack])
+                        target_stack = " ".join([str(x) for x in target_stack])
+                        result_stack = " ".join([str(x) for x in result_stack])
+                        logger(
+                            f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
+                        )
+                    nb_to_log -= 1
+
+            return sum_nb_total, sum_nb_errors
+
         # --------------------------------------------------------------------
 
-        test_nb_total, test_nb_errors = compute_nb_errors(
+        test_nb_total, test_nb_errors = compute_nb_errors_prog(
             self.test_input[:1000].to(self.device), nb_to_log=10
         )