Update.

[culture.git] / main.py

diff --git a/main.py b/main.py
index 634363f..d400ab1 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -29,7 +29,6 @@ else:
 ######################################################################
 
 parser = argparse.ArgumentParser(
-    description="An implementation of GPT with cache.",
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 
@@ -274,53 +273,6 @@ vocabulary_size = quiz_machine.vocabulary_size()
 log_string(f"vocabulary_size {vocabulary_size}")
 
 ######################################################################
-
-# Compute the entropy of the training tokens
-
-token_count = 0
-for input in quiz_machine.batches(split="train", desc="train-entropy"):
-    token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
-        (0, 1)
-    )
-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
-
-if args.max_percents_of_test_in_train >= 0:
-
-    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
-
-    nb_test, nb_in_train = 0, 0
-    for test_subset in subsets_as_tuples(
-        quiz_machine.batches(split="test", desc="test-check"), 25000
-    ):
-        in_train = set()
-        for train_subset in subsets_as_tuples(
-            quiz_machine.batches(split="train", desc="train-check"), 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_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"
-
 ##############################
 
 
@@ -331,7 +283,7 @@ def one_epoch(model, quiz_machine):
 
     nb_train_samples, acc_train_loss = 0, 0.0
 
-    for input in quiz_machine.batches(split="train"):
+    for input in quiz_machine.batches(model, split="train"):
         input = input.to(device)
 
         if nb_train_samples % args.batch_size == 0:
@@ -363,7 +315,7 @@ def run_tests(model, quiz_machine, deterministic_synthesis):
         nb_test_samples, acc_test_loss = 0, 0.0
         nb_samples_accumulated = 0
 
-        for input in quiz_machine.batches(split="test"):
+        for input in quiz_machine.batches(model, split="test"):
             input = input.to(device)
 
             bs = model(mygpt.BracketedSequence(input))
@@ -596,6 +548,15 @@ for k in range(args.nb_gpts):
     model.main_test_accuracy = 0.0
     model.id = k
 
+    model.train_w_quizzes = quiz_machine.generate_token_sequences(
+        args.nb_train_samples
+    ).to(device)
+    quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
+    model.test_w_quizzes = quiz_machine.generate_token_sequences(
+        args.nb_test_samples
+    ).to(device)
+    quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
+
     models.append(model)
 
 
@@ -604,6 +565,54 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 
 ######################################################################
 
+# Compute the entropy of the training tokens
+
+token_count = 0
+for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
+    token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
+        (0, 1)
+    )
+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
+
+if args.max_percents_of_test_in_train >= 0:
+
+    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
+
+    nb_test, nb_in_train = 0, 0
+    for test_subset in subsets_as_tuples(
+        quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
+    ):
+        in_train = set()
+        for train_subset in subsets_as_tuples(
+            quiz_machine.batches(models[0], split="train", desc="train-check"), 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_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"
+
+######################################################################
+
 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 
@@ -654,7 +663,7 @@ for n_epoch in range(args.nb_epochs):
     log_string(
         f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
     )
-    quiz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+    quiz_machine.renew_w_quizzes(model, args.nb_train_samples // args.nb_gpts)
 
     ##################################################
     # If all the models are good enough, generate new quizzes and