Update.

[picoclvr.git] / main.py

diff --git a/main.py b/main.py
index 0f1fbb5..9c28e47 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -42,6 +42,10 @@ parser.add_argument("--result_dir", type=str, default=None)
 
 parser.add_argument("--seed", type=int, default=0)
 
+parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
+
+########################################
+
 parser.add_argument("--nb_epochs", type=int, default=None)
 
 parser.add_argument("--batch_size", type=int, default=None)
@@ -56,6 +60,8 @@ parser.add_argument("--learning_rate", type=float, default=1e-4)
 
 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
 
+########################################
+
 parser.add_argument("--model", type=str, default="37M")
 
 parser.add_argument("--dim_model", type=int, default=None)
@@ -70,6 +76,8 @@ parser.add_argument("--nb_blocks", type=int, default=None)
 
 parser.add_argument("--dropout", type=float, default=0.1)
 
+########################################
+
 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 
 parser.add_argument("--no_checkpoint", action="store_true", default=False)
@@ -355,8 +363,10 @@ if args.task == "sandbox":
         raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
 
     task = tasks.SandBox(
-        problem,
+        # problem,
         # problems.ProblemAddition(zero_padded=False, inverted_result=False),
+        # problems.ProblemLenId(len_max=args.sandbox_levels_len_source),
+        problems.ProblemTwoTargets(len_total=12, len_targets=4),
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
         batch_size=args.batch_size,
@@ -541,34 +551,37 @@ token_probas = token_count / token_count.sum()
 entropy = -torch.xlogy(token_probas, token_probas).sum()
 train_set_perplexity = math.exp(entropy)
 
-##############################
-
+######################################################################
 # A bit of paranoia never hurts
 
-train_examples = {}
 
+def subsets_as_tuples(batches, cs):
+    s = set()
+    for batch in batches:
+        for x in batch:
+            s.add(tuple([v.item() for v in x]))
+            if len(s) == cs:
+                yield s
+                s = set()
+    yield s
 
-for input in task.batches(split="train"):
-    assert input.dim() == 2 and input.dtype == torch.int64
-    for x in input:
-        train_examples[x.sum().item()] = x
-
-nb_total, nb_collisions = 0, 0
-for input in task.batches(split="test"):
-    assert input.dim() == 2 and input.dtype == torch.int64
-    for x in input:
-        nb_total += 1
-        y = train_examples.get(x.sum().item())
-        if y is not None:
-            if x.size() == y.size() and (x - y).abs().sum() == 0:
-                nb_collisions += 1
-
-del train_examples
+
+nb_test, nb_in_train = 0, 0
+for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
+    in_train = set()
+    for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
+        in_train.update(test_subset.intersection(train_subset))
+    nb_in_train += len(in_train)
+    nb_test += len(test_subset)
 
 log_string(
-    f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
+    f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 )
 
+assert (
+    nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
+), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
+
 ##############################
 
 if args.learning_rate_schedule == "cos":