Update.

[culture.git] / main.py

diff --git a/main.py b/main.py
index d194a8d..d63398c 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -79,9 +79,7 @@ parser.add_argument("--dropout", type=float, default=0.1)
 
 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 
-parser.add_argument("--reverse_cleanup", action="store_true", default=True)
-
-parser.add_argument("--validation_forward_only", action="store_true", default=False)
+parser.add_argument("--both_directions", action="store_true", default=False)
 
 parser.add_argument("--problem", type=str, default="sky")
 
@@ -409,11 +407,10 @@ def create_c_quizzes(
         c_quizzes, ave_seq_logproba = quizz_machine.generate_quizzes(
             nb_to_create,
             model_for_generation=model_for_generation,
-            reverse_cleanup=args.reverse_cleanup,
         )
 
         nb_correct = quizz_machine.compute_correctness(
-            c_quizzes, models, both_directions=not args.validation_forward_only
+            c_quizzes, models, both_directions=args.both_directions
         )
 
         if args.dirty_debug:
@@ -429,7 +426,7 @@ def create_c_quizzes(
         nb_validated = valid_c_quizzes(recorded, standard_validity).size(0)
 
         log_string(
-            f"keep c_quizzes kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
+            f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
         )
 
     # store the new c_quizzes which have been validated
@@ -487,20 +484,20 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 for n_epoch in range(args.nb_epochs):
     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 
+    # Select, improve, and eval the worst model
+
     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
 
     log_string(
         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
     )
 
-    # improve it
     one_epoch(weakest_model, quizz_machine)
 
     log_string(
         f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
     )
 
-    # test it
     run_tests(weakest_model, quizz_machine, deterministic_synthesis=False)
 
     log_string(
@@ -510,9 +507,13 @@ for n_epoch in range(args.nb_epochs):
     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
     log_string(f"current_test_accuracies {cta}")
 
-    # replace a fraction of the w_quizzes with fresh ones
+    # Replace a fraction of the w_quizzes with fresh ones
+
     quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 
+    # If all the models are good enough, generate new quizzes and
+    # re-compute the test errors
+
     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
         create_c_quizzes(
             models,
@@ -521,7 +522,6 @@ for n_epoch in range(args.nb_epochs):
             nb_for_test=nb_new_c_quizzes_for_test,
         )
 
-        # We update everyone
         for model in models:
             run_tests(model, quizz_machine, deterministic_synthesis=False)