Update.

[culture.git] / main.py

diff --git a/main.py b/main.py
index 585cbdf..9c3d7f1 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -12,18 +12,15 @@ from torch import nn
 from torch.nn import functional as F
 
 import ffutils
+
 import mygpt
-import sky, reasoning, quiz_machine
+import sky, grids, quiz_machine
+from problem import MultiThreadProblem
 
 # world quizzes vs. culture quizzes
 
 ######################################################################
 
-nb_new_c_quizzes_for_train = 1000
-nb_new_c_quizzes_for_test = 100
-
-######################################################################
-
 if torch.cuda.is_available():
     device = torch.device("cuda")
     torch.backends.cuda.matmul.allow_tf32 = True
@@ -57,7 +54,7 @@ parser.add_argument("--nb_train_samples", type=int, default=None)
 
 parser.add_argument("--nb_test_samples", type=int, default=None)
 
-parser.add_argument("--learning_rate", type=float, default=1e-3)
+parser.add_argument("--learning_rate", type=float, default=5e-4)
 
 ########################################
 
@@ -79,7 +76,9 @@ 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("--problem", type=str, default="grids")
+
+parser.add_argument("--multi_thread_problem", action="store_true", default=False)
 
 parser.add_argument("--nb_gpts", type=int, default=5)
 
@@ -124,13 +123,6 @@ if args.result_dir is None:
 
 ######################################################################
 
-if args.dirty_debug:
-    args.accuracy_to_make_c_quizzes = 0.0
-    nb_new_c_quizzes_for_train = 100
-    nb_new_c_quizzes_for_test = 10
-
-######################################################################
-
 default_args = {
     "model": "37M",
     "batch_size": 100,
@@ -250,12 +242,15 @@ if args.problem == "sky":
         speed=args.sky_speed,
     )
     back_accuracy = False
-elif args.problem == "reasoning":
-    problem = reasoning.Reasoning(device=device)
+elif args.problem == "grids":
+    problem = grids.Grids(device=device)
     back_accuracy = True
 else:
     raise ValueError
 
+if args.multi_thread_problem:
+    problem = MultiThreadProblem(problem, args.nb_train_samples, chunk_size=1000)
+
 quiz_machine = quiz_machine.QuizMachine(
     problem=problem,
     nb_train_samples=args.nb_train_samples,
@@ -417,6 +412,7 @@ def create_c_quizzes(
     )
 
     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
+
     with open(file_name, "w") as logp_file:
         while (
             valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0)
@@ -432,23 +428,26 @@ def create_c_quizzes(
                 temperature=args.generation_temperature,
             )
 
-            nb_correct, seq_logproba = quiz_machine.compute_correctness(
-                c_quizzes,
-                models,
-                bidirectional_validation=args.bidirectional_validation,
-                deterministic_validation=args.deterministic_validation,
-            )
-
-            for n, l in zip(nb_correct, seq_logproba):
-                s = " ".join([str(x.item()) for x in l])
-                logp_file.write(f"{n} {s}\n")
+            c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 
-            if args.dirty_debug:
-                nb_correct = torch.randint(
-                    len(models) + 1, nb_correct.size(), device=c_quizzes.device
+            if c_quizzes.size(0) > 0:
+                nb_correct, seq_logproba = quiz_machine.compute_correctness(
+                    c_quizzes,
+                    models,
+                    bidirectional_validation=args.bidirectional_validation,
+                    deterministic_validation=args.deterministic_validation,
                 )
 
-            quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
+                for n, l in zip(nb_correct, seq_logproba):
+                    s = " ".join([str(x.item()) for x in l])
+                    logp_file.write(f"{n} {s}\n")
+
+                if args.dirty_debug:
+                    nb_correct = torch.randint(
+                        len(models) + 1, nb_correct.size(), device=c_quizzes.device
+                    )
+
+                quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
 
             nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
             nv = " ".join([str(x.item()) for x in nv])
@@ -484,6 +483,8 @@ def create_c_quizzes(
             quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
         )[:72]
 
+        quiz_machine.reverse_random_half_in_place(q)
+
         if q.size(0) > 0:
             quiz_machine.save_quizzes(
                 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
@@ -517,12 +518,30 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 
 ######################################################################
 
+nb_new_c_quizzes_for_train = args.nb_train_samples // 50
+nb_new_c_quizzes_for_test = args.nb_test_samples // 50
+
+log_string(
+    f"nb_new_c_quizzes_for_train {nb_new_c_quizzes_for_train} nb_new_c_quizzes_for_test {nb_new_c_quizzes_for_test}"
+)
+
+######################################################################
+
+if args.dirty_debug:
+    args.accuracy_to_make_c_quizzes = 0.0
+    args.nb_gpts = 2
+    nb_new_c_quizzes_for_train = 100
+    nb_new_c_quizzes_for_test = 10
+
+######################################################################
+
 for n_epoch in range(args.nb_epochs):
     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 
     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
     log_string(f"current_test_accuracies {cta}")
 
+    ##################################################
     # Select, improve, and eval the worst model
 
     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
@@ -543,10 +562,12 @@ for n_epoch in range(args.nb_epochs):
         f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
     )
 
+    ##################################################
     # Replace a fraction of the w_quizzes with fresh ones
 
     quiz_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