X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=ee4e9e5b3aea5fb5f6dcd426bbecbd7687bfbb35;hb=b3392c295bdb75140916e2db70efc6fa50962f63;hp=3b5a7c6e47d3b813d675b5c5cfdf82a99b452316;hpb=ac0d58ae608b02b00226e8c08c98fca7295a2b3b;p=culture.git

diff --git a/main.py b/main.py
index 3b5a7c6..ee4e9e5 100755
--- a/main.py
+++ b/main.py
@@ -183,8 +183,8 @@ for n in vars(args):
 ######################################################################
 
 if args.check:
-    args.nb_train_samples = 2500
-    args.nb_test_samples = 100
+    args.nb_train_samples = 25000
+    args.nb_test_samples = 1000
 
 if args.physical_batch_size is None:
     args.physical_batch_size = args.batch_size
@@ -338,11 +338,13 @@ def create_quizzes(
     desired_average_logits=None,
 ):
     kept = []
-    nb_generated_tokens, sum_logits = 0, 0
+
+    sum_logits = 0
 
     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
         nb_to_generate = 4 * (nb_for_train + nb_for_test)
-        new_quizzes, nb_correct, average_logits = task.create_new_quizzes(
+
+        new_quizzes, nb_correct, _sum_logits = task.create_new_quizzes(
             n_epoch=n_epoch,
             result_dir=args.result_dir,
             logger=log_string,
@@ -352,8 +354,7 @@ def create_quizzes(
             desired_average_logits=desired_average_logits,
         )
 
-        nb_generated_tokens += new_quizzes.numel()
-        sum_logits += average_logits * new_quizzes.numel()
+        sum_logits += _sum_logits
 
         to_keep = new_quizzes[nb_correct == len(other_models) - 1]
         log_string(
@@ -373,7 +374,7 @@ def create_quizzes(
         log_string,
     )
 
-    return sum_logits / nb_generated_tokens
+    return sum_logits / new_quizzes.size(0)
 
 
 ######################################################################
@@ -409,13 +410,13 @@ nb_new_quizzes_for_test = 100
 
 if args.check:
     accuracy_to_make_quizzes = 0.0
-    nb_new_quizzes_for_train = 10
+    nb_new_quizzes_for_train = 100
     nb_new_quizzes_for_test = 10
 
 desired_average_logits = None
 
 for n_epoch in range(args.nb_epochs):
-    log_string(f"--- epoch {n_epoch+1} ----------------------------------------")
+    log_string(f"--- epoch {n_epoch} ----------------------------------------")
 
     a = [(model.id, float(model.main_test_accuracy)) for model in models]
     a.sort(key=lambda p: p[0])