Update.

[culture.git] / main.py

diff --git a/main.py b/main.py
index d063423..b88847e 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -13,13 +13,12 @@ from torch.nn import functional as F
 
 import ffutils
 import mygpt
-import sky, quizz_machine
+import sky, wireworld, quizz_machine
 
 # world quizzes vs. culture quizzes
 
 ######################################################################
 
-accuracy_to_make_c_quizzes = 0.975
 nb_new_c_quizzes_for_train = 1000
 nb_new_c_quizzes_for_test = 100
 
@@ -38,7 +37,7 @@ parser = argparse.ArgumentParser(
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 
-parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
+parser.add_argument("--log_filename", type=str, default="train.log")
 
 parser.add_argument("--result_dir", type=str, default=None)
 
@@ -80,9 +79,19 @@ parser.add_argument("--dropout", type=float, default=0.1)
 
 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 
+parser.add_argument("--problem", type=str, default="sky")
+
 parser.add_argument("--nb_gpts", type=int, default=5)
 
-parser.add_argument("--nb_correct_to_validate", type=int, default=4)
+parser.add_argument("--nb_models_for_generation", type=int, default=1)
+
+parser.add_argument("--generation_mode", type=str, default="groupthink")
+
+parser.add_argument("--min_to_validate", type=int, default=4)
+
+parser.add_argument("--max_to_validate", type=int, default=4)
+
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
 
 parser.add_argument("--dirty_debug", action="store_true", default=False)
 
@@ -96,7 +105,7 @@ if args.result_dir is None:
 ######################################################################
 
 if args.dirty_debug:
-    accuracy_to_make_c_quizzes = 0.0
+    args.accuracy_to_make_c_quizzes = 0.0
     nb_new_c_quizzes_for_train = 100
     nb_new_c_quizzes_for_test = 10
 
@@ -212,8 +221,15 @@ else:
 assert args.nb_train_samples % args.batch_size == 0
 assert args.nb_test_samples % args.batch_size == 0
 
+if args.problem=="sky":
+    problem=sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2, speed=2),
+elif args.problem="wireworld":
+    problem=wireworld.Wireworld(height=10, width=15, nb_iterations=4)
+else:
+    raise ValueError
+
 quizz_machine = quizz_machine.QuizzMachine(
-    problem=sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2),
+    problem=problem,
     nb_train_samples=args.nb_train_samples,
     nb_test_samples=args.nb_test_samples,
     batch_size=args.physical_batch_size,
@@ -336,7 +352,6 @@ def run_tests(model, quizz_machine, deterministic_synthesis):
             n_epoch=n_epoch,
             model=model,
             result_dir=args.result_dir,
-            logger=log_string,
             deterministic_synthesis=deterministic_synthesis,
         )
 
@@ -371,26 +386,25 @@ def create_c_quizzes(
         return sum(
             [
                 sum([x.size(0) for x in recorded[n]])
-                for n in range(args.nb_correct_to_validate, len(models))
+                for n in range(args.min_to_validate, args.max_to_validate + 1)
             ]
         )
 
-    while nb_validated() < 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))]
-
-        model = models[model_indexes.pop()]
-
-        new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes(
-            nb=nb_to_validate,
-            model_for_generation=model,
-            models_for_validation=models,
+    nb_to_create = nb_for_train + nb_for_test
+
+    while nb_validated() < nb_to_create:
+        (
+            new_c_quizzes,
+            nb_correct,
+            ave_seq_logproba,
+        ) = quizz_machine.gang_create_c_quizzes(
+            nb=nb_to_create,
+            nb_models_for_generation=args.nb_models_for_generation,
+            models=models,
+            mode=args.generation_mode,
             min_ave_seq_logproba=min_ave_seq_logproba,
             n_epoch=n_epoch,
             result_dir=args.result_dir,
-            logger=log_string,
         )
 
         sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
@@ -405,7 +419,7 @@ def create_c_quizzes(
             recorded[n].append(new_c_quizzes[nb_correct == n].clone())
 
         log_string(
-            f"keep c_quizzes {nb_validated()*100/nb_generated():.02f}% kept total {nb_validated()} / {nb_to_validate}"
+            f"keep c_quizzes {nb_validated()*100/nb_generated():.02f}% kept total {nb_validated()} / {nb_to_create}"
         )
 
     # concatenate and shuffle
@@ -418,7 +432,8 @@ def create_c_quizzes(
             del recorded[n]
 
     new_c_quizzes = torch.cat(
-        [recorded[n] for n in range(args.nb_correct_to_validate, len(models))], dim=0
+        [recorded[n] for n in range(args.min_to_validate, args.max_to_validate + 1)],
+        dim=0,
     )
 
     new_c_quizzes = new_c_quizzes[
@@ -431,7 +446,11 @@ def create_c_quizzes(
     quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 
     for n in recorded.keys():
-        s = "_validated" if n >= args.nb_correct_to_validate and n < len(models) else ""
+        s = (
+            "_validated"
+            if n >= args.min_to_validate and n <= args.max_to_validate
+            else ""
+        )
         quizz_machine.problem.save_quizzes(
             recorded[n][:72],
             args.result_dir,
@@ -475,7 +494,8 @@ for n_epoch in range(args.nb_epochs):
 
     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}")
+    s = " ".join([f"{p[1]*100:.02f}%" for p in a])
+    log_string(f"current accuracies {s}")
 
     # select the model with lowest accuracy
     models.sort(key=lambda model: model.main_test_accuracy)
@@ -501,7 +521,7 @@ for n_epoch in range(args.nb_epochs):
         f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
     )
 
-    if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
+    if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
         ave_seq_logproba = create_c_quizzes(
             models,
             quizz_machine,