X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=e14ceb76e18ba38283acb99a5f214d42b92cc1b1;hb=5703df4c32a0856c8fa4b1ff97810cdc1fb76253;hp=a3d47f54451464470bd9d37710bda4f42c1ab4ce;hpb=c9dbc3abf436df8af1379d04ab51159e821496f1;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index a3d47f5..e14ceb7 100755
--- a/tasks.py
+++ b/tasks.py
@@ -1044,11 +1044,18 @@ class RPL(Task):
             0,
         ).to(self.device)
 
+    def seq2str(self, seq):
+        return " ".join([self.id2token[i] for i in seq])
+
     def __init__(
         self,
         nb_train_samples,
         nb_test_samples,
         batch_size,
+        nb_starting_values=3,
+        max_input=9,
+        prog_len=6,
+        nb_runs=5,
         device=torch.device("cpu"),
     ):
         super().__init__()
@@ -1057,11 +1064,23 @@ class RPL(Task):
         self.device = device
 
         train_sequences = [
-            rpl.generate()
+            rpl.generate(
+                nb_starting_values=nb_starting_values,
+                max_input=max_input,
+                prog_len=prog_len,
+                nb_runs=nb_runs,
+            )
             for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
         ]
+
         test_sequences = [
-            rpl.generate() for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
+            rpl.generate(
+                nb_starting_values=nb_starting_values,
+                max_input=max_input,
+                prog_len=prog_len,
+                nb_runs=nb_runs,
+            )
+            for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
         ]
 
         symbols = list(
@@ -1117,22 +1136,33 @@ class RPL(Task):
                 device=self.device,
             )
 
-            if nb_to_log > 0:
-                for x in result[:nb_to_log]:
-                    s = " ".join([self.id2token[i.item()] for i in x])
-                    logger(f"check {n_epoch} {s}")
-                nb_to_log -= min(nb_to_log, result.size(0))
-
             sum_nb_total, sum_nb_errors = 0, 0
-            for x in result:
-                seq = [self.id2token[i.item()] for i in x]
-                nb_total, nb_errors = rpl.check(seq)
-                sum_nb_total += nb_total
-                sum_nb_errors += nb_errors
+            for x, y in zip(input, result):
+                seq = [self.id2token[i.item()] for i in y]
+                nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
+                sum_nb_total += 1
+                sum_nb_errors += 0 if nb_errors == 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])
+                    logger(f"PROG [{gt_prog}] PREDICTED [{prog}]")
+                    for start_stack, target_stack, result_stack, correct in stacks:
+                        comment = " CORRECT" 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"  [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]{comment}"
+                        )
+                    nb_to_log -= 1
 
             return sum_nb_total, sum_nb_errors
 
-        test_nb_total, test_nb_errors = compute_nb_errors(self.test_input, nb_to_log=10)
+        test_nb_total, test_nb_errors = compute_nb_errors(
+            self.test_input[:1000], nb_to_log=10
+        )
 
         logger(
             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"