From 2186d96fccfc525884f1b3fb722c40642891ab0a Mon Sep 17 00:00:00 2001
From: =?utf8?q?Fran=C3=A7ois=20Fleuret?= <francois@fleuret.org>
Date: Wed, 26 Jun 2024 08:41:45 +0200
Subject: [PATCH] Update.

---
 main.py | 60 +++++++++++++++++++++++++++++++++++++--------------------
 1 file changed, 39 insertions(+), 21 deletions(-)

diff --git a/main.py b/main.py
index c5acea7..d0de5af 100755
--- a/main.py
+++ b/main.py
@@ -355,12 +355,19 @@ def create_c_quizzes(
     nb_for_test=100,
     min_ave_seq_logproba=None,
 ):
-    kept = []
+    # We will store the generated quizzes for each number of
+    # correct prediction
+    recorded = dict([(n, []) for n in range(len(models) + 1)])
+
     model_indexes = []
     sum_logits, sum_nb_c_quizzes = 0, 0
+    nb_correct_to_validate = len(models) - 1
 
-    while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
-        nb_to_generate = nb_for_train + nb_for_test
+    while (
+        sum([x.size(0) for x in recorded[nb_correct_to_validate]])
+        < nb_for_train + nb_for_test
+    ):
+        nb_to_validate = nb_for_train + nb_for_test
 
         if len(model_indexes) == 0:
             model_indexes = [i.item() for i in torch.randperm(len(models))]
@@ -368,7 +375,7 @@ def create_c_quizzes(
         model = models[model_indexes.pop()]
 
         new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes(
-            nb=nb_to_generate,
+            nb=nb_to_validate,
             model_for_generation=model,
             models_for_validation=models,
             min_ave_seq_logproba=min_ave_seq_logproba,
@@ -380,33 +387,44 @@ def create_c_quizzes(
         sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
         sum_nb_c_quizzes += new_c_quizzes.size(0)
 
-        to_keep = new_c_quizzes[nb_correct == len(models) - 1]
-
         if args.dirty_debug:
-            to_keep = new_c_quizzes[
-                torch.randint(3, (new_c_quizzes.size(0),), device=new_c_quizzes.device)
-                == 0
-            ]
+            nb_correct = torch.randint(
+                len(models) + 1, nb_correct.size(), device=new_c_quizzes.device
+            )
+
+        for n in range(nb_correct.max() + 1):
+            recorded[n].append(new_c_quizzes[nb_correct == n].clone())
 
-        kept.append(to_keep)
+        nb_validated = sum([x.size(0) for x in recorded[nb_correct_to_validate]])
+        nb_generated = sum(
+            [sum([x.size(0) for x in recorded[n]]) for n in recorded.keys()]
+        )
 
         log_string(
-            f"keep c_quizzes {to_keep.size(0)}/{new_c_quizzes.size(0)} ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%) total {sum([ x.size(0) for x in kept])}/{nb_to_generate}"
+            f"keep c_quizzes {nb_validated*100/nb_generated:.02f}% kept total {nb_validated}/{nb_to_validate}"
         )
 
-    new_c_quizzes = torch.cat(kept, dim=0)
-    new_c_quizzes = new_c_quizzes[
-        torch.randperm(new_c_quizzes.size(0))[: nb_for_train + nb_for_test]
-    ]
+    # concatenate and shuffle
+    for n in recorded.keys():
+        if len(recorded[n]) > 0:
+            q = torch.cat(recorded[n], dim=0)
+            q = q[torch.randperm(q.size(0), device=q.device)]
+            recorded[n] = q
+        else:
+            del recorded[n]
+
+    new_c_quizzes = recorded[nb_correct_to_validate][: nb_for_train + nb_for_test]
 
     quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
     quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 
-    quizz_machine.problem.save_quizzes(
-        new_c_quizzes[:72],
-        args.result_dir,
-        f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",
-    )
+    for n in recorded.keys():
+        s = "_validated" if n == nb_correct_to_validate else ""
+        quizz_machine.problem.save_quizzes(
+            recorded[n][:72],
+            args.result_dir,
+            f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
+        )
 
     return sum_logits / sum_nb_c_quizzes
 
-- 
2.39.5