X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=inline;f=main.py;h=803383626669647984f3a542421f5b2c57d461c2;hb=779e3675414e061ad294c6b5599a7843d9e887bc;hp=a021a71c76ed7525c40db2b01ef11b6e9b1e9059;hpb=17c63771f2ca82ce39d8406e377ace2015fe69fc;p=culture.git

diff --git a/main.py b/main.py
index a021a71..8033836 100755
--- a/main.py
+++ b/main.py
@@ -73,7 +73,7 @@ parser.add_argument("--deterministic_synthesis", action="store_true", default=Fa
 
 parser.add_argument("--nb_gpts", type=int, default=5)
 
-parser.add_argument("--check", action="store_true", default=False)
+parser.add_argument("--dirty_debug", action="store_true", default=False)
 
 ######################################################################
 
@@ -182,7 +182,7 @@ for n in vars(args):
 
 ######################################################################
 
-if args.check:
+if args.dirty_debug:
     args.nb_train_samples = 2500
     args.nb_test_samples = 100
 
@@ -338,10 +338,12 @@ def create_quizzes(
     desired_average_logits=None,
 ):
     kept = []
-    nb_generated_tokens, sum_logits = 0, 0
+
+    sum_logits, sum_nb_quizzes = 0, 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(
             n_epoch=n_epoch,
             result_dir=args.result_dir,
@@ -352,13 +354,18 @@ def create_quizzes(
             desired_average_logits=desired_average_logits,
         )
 
-        nb_generated_tokens += new_quizzes.numel()
-        sum_logits += average_logits * new_quizzes.numel()
+        sum_logits += new_quizzes.size(0) * average_logits
+        sum_nb_quizzes += new_quizzes.size(0)
 
         to_keep = new_quizzes[nb_correct == len(other_models) - 1]
+
+        if args.dirty_debug:
+            to_keep = new_quizzes
+
         log_string(
             f"keep {to_keep.size(0)}/{new_quizzes.size(0)} quizzes ({to_keep.size(0)*100/new_quizzes.size(0):.02f}%)"
         )
+
         kept.append(to_keep)
 
     new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
@@ -373,7 +380,7 @@ def create_quizzes(
         log_string,
     )
 
-    return sum_logits / nb_generated_tokens
+    return sum_logits / sum_nb_quizzes
 
 
 ######################################################################
@@ -407,14 +414,16 @@ accuracy_to_make_quizzes = 0.975
 nb_new_quizzes_for_train = 1000
 nb_new_quizzes_for_test = 100
 
-if args.check:
+if args.dirty_debug:
     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} ----------------------------------------")
+
     a = [(model.id, float(model.main_test_accuracy)) for model in models]
     a.sort(key=lambda p: p[0])
     log_string(f"current accuracies {a}")