Update.

[culture.git] / main.py

diff --git a/main.py b/main.py
index 918f75d..02e1a8d 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -13,7 +13,7 @@ from torch.nn import functional as F
 
 import ffutils
 import mygpt
 
 import ffutils
 import mygpt

-import sky, wireworld, quizz_machine
+import sky, reasoning, quiz_machine

 
 # world quizzes vs. culture quizzes
 
 
 # world quizzes vs. culture quizzes
 


@@ -57,7 +57,7 @@ parser.add_argument("--nb_train_samples", type=int, default=None)
 
 parser.add_argument("--nb_test_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)

 
 ########################################
 
 
 ########################################
 


@@ -79,23 +79,23 @@ parser.add_argument("--dropout", type=float, default=0.1)
 
 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 
 
 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 

-parser.add_argument("--both_directions", 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("--problem", type=str, default="sky")
 
 parser.add_argument("--nb_gpts", type=int, default=5)
 

-parser.add_argument("--min_to_validate", type=int, default=4)
+parser.add_argument("--min_to_validate", type=int, default=None)

 
 

-parser.add_argument("--max_to_validate", type=int, default=4)
+parser.add_argument("--max_to_validate", type=int, default=None)

 
 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
 
 
 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
 

-parser.add_argument("--dirty_debug", action="store_true", default=False)
+parser.add_argument("--generation_temperature", type=float, default=2.0)

 
 

-parser.add_argument("--generation_temperature", type=float, default=1.0)
+parser.add_argument("--deterministic_validation", action="store_true", default=False)

 
 

-parser.add_argument("--stochastic_validation", action="store_true", default=False)
+parser.add_argument("--bidirectional_validation", action="store_true", default=False)
+
+parser.add_argument("--dirty_debug", action="store_true", default=False)

 
 ######################################################################
 
 
 ######################################################################
 


@@ -113,6 +113,12 @@ parser.add_argument("--sky_speed", type=int, default=3)
 
 args = parser.parse_args()
 
 
 args = parser.parse_args()
 

+if args.min_to_validate is None:
+    args.min_to_validate = args.nb_gpts - 1
+
+if args.max_to_validate is None:
+    args.max_to_validate = args.nb_gpts - 1
+

 if args.result_dir is None:
     args.result_dir = f"results_culture"
 
 if args.result_dir is None:
     args.result_dir = f"results_culture"
 


@@ -120,6 +126,7 @@ if args.result_dir is None:
 
 if args.dirty_debug:
     args.accuracy_to_make_c_quizzes = 0.0
 
 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
 
     nb_new_c_quizzes_for_train = 100
     nb_new_c_quizzes_for_test = 10
 


@@ -243,15 +250,18 @@ if args.problem == "sky":
         nb_iterations=args.sky_nb_iterations,
         speed=args.sky_speed,
     )
         nb_iterations=args.sky_nb_iterations,
         speed=args.sky_speed,
     )

-elif args.problem == "wireworld":
-    problem = wireworld.Wireworld(height=8, width=10, nb_iterations=2, speed=5)
+    back_accuracy = False
+elif args.problem == "reasoning":
+    problem = reasoning.Reasoning(device=device)
+    back_accuracy = True

 else:
     raise ValueError
 
 else:
     raise ValueError
 

-quizz_machine = quizz_machine.QuizzMachine(
+quiz_machine = quiz_machine.QuizMachine(

     problem=problem,
     nb_train_samples=args.nb_train_samples,
     nb_test_samples=args.nb_test_samples,
     problem=problem,
     nb_train_samples=args.nb_train_samples,
     nb_test_samples=args.nb_test_samples,

+    back_accuracy=back_accuracy,

     batch_size=args.physical_batch_size,
     result_dir=args.result_dir,
     logger=log_string,
     batch_size=args.physical_batch_size,
     result_dir=args.result_dir,
     logger=log_string,


@@ -262,7 +272,7 @@ quizz_machine = quizz_machine.QuizzMachine(
 
 log_string(f"device {device}")
 
 
 log_string(f"device {device}")
 

-vocabulary_size = quizz_machine.vocabulary_size()
+vocabulary_size = quiz_machine.vocabulary_size()

 
 log_string(f"vocabulary_size {vocabulary_size}")
 
 
 log_string(f"vocabulary_size {vocabulary_size}")
 


@@ -271,8 +281,8 @@ log_string(f"vocabulary_size {vocabulary_size}")
 # Compute the entropy of the training tokens
 
 token_count = 0
 # Compute the entropy of the training tokens
 
 token_count = 0

-for input in quizz_machine.batches(split="train", desc="train-entropy"):
-    token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
+for input in quiz_machine.batches(split="train", desc="train-entropy"):
+    token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(

         (0, 1)
     )
 token_probas = token_count / token_count.sum()
         (0, 1)
     )
 token_probas = token_count / token_count.sum()


@@ -296,11 +306,11 @@ if args.max_percents_of_test_in_train >= 0:
 
     nb_test, nb_in_train = 0, 0
     for test_subset in subsets_as_tuples(
 
     nb_test, nb_in_train = 0, 0
     for test_subset in subsets_as_tuples(

-        quizz_machine.batches(split="test", desc="test-check"), 25000
+        quiz_machine.batches(split="test", desc="test-check"), 25000

     ):
         in_train = set()
         for train_subset in subsets_as_tuples(
     ):
         in_train = set()
         for train_subset in subsets_as_tuples(

-            quizz_machine.batches(split="train", desc="train-check"), 25000
+            quiz_machine.batches(split="train", desc="train-check"), 25000

         ):
             in_train.update(test_subset.intersection(train_subset))
         nb_in_train += len(in_train)
         ):
             in_train.update(test_subset.intersection(train_subset))
         nb_in_train += len(in_train)


@@ -317,14 +327,14 @@ if args.max_percents_of_test_in_train >= 0:
 ##############################
 
 
 ##############################
 
 

-def one_epoch(model, quizz_machine):
+def one_epoch(model, quiz_machine):

     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 
     model.train()
 
     nb_train_samples, acc_train_loss = 0, 0.0
 
     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 
     model.train()
 
     nb_train_samples, acc_train_loss = 0, 0.0
 

-    for input in quizz_machine.batches(split="train"):
+    for input in quiz_machine.batches(split="train"):

         input = input.to(device)
 
         if nb_train_samples % args.batch_size == 0:
         input = input.to(device)
 
         if nb_train_samples % args.batch_size == 0:


@@ -349,14 +359,14 @@ def one_epoch(model, quizz_machine):
 ######################################################################
 
 
 ######################################################################
 
 

-def run_tests(model, quizz_machine, deterministic_synthesis):
+def run_tests(model, quiz_machine, deterministic_synthesis):

     with torch.autograd.no_grad():
         model.eval()
 
         nb_test_samples, acc_test_loss = 0, 0.0
         nb_samples_accumulated = 0
 
     with torch.autograd.no_grad():
         model.eval()
 
         nb_test_samples, acc_test_loss = 0, 0.0
         nb_samples_accumulated = 0
 

-        for input in quizz_machine.batches(split="test"):
+        for input in quiz_machine.batches(split="test"):

             input = input.to(device)
 
             bs = model(mygpt.BracketedSequence(input))
             input = input.to(device)
 
             bs = model(mygpt.BracketedSequence(input))


@@ -372,7 +382,7 @@ def run_tests(model, quizz_machine, deterministic_synthesis):
 
         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 
 
         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 

-        model.main_test_accuracy = quizz_machine.produce_results(
+        model.main_test_accuracy = quiz_machine.produce_results(

             n_epoch=n_epoch,
             model=model,
             result_dir=args.result_dir,
             n_epoch=n_epoch,
             model=model,
             result_dir=args.result_dir,


@@ -393,11 +403,11 @@ def valid_c_quizzes(recorded, criteria):
 
 def create_c_quizzes(
     models,
 
 def create_c_quizzes(
     models,

-    quizz_machine,
+    quiz_machine,

     nb_for_train=1000,
     nb_for_test=100,
 ):
     nb_for_train=1000,
     nb_for_test=100,
 ):

-    recorded = []
+    quizzes_and_nb_correct_records = []

 
     nb_to_create = nb_for_train + nb_for_test
 
 
     nb_to_create = nb_for_train + nb_for_test
 


@@ -409,22 +419,25 @@ 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:
 
     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(recorded, standard_validity).size(0) < nb_to_create:
+        while (
+            valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0)
+            < nb_to_create
+        ):

             # Select a model at random to generate the new quizzes
 
             model_for_generation = models[torch.randint(len(models), (1,))]
 
             # Select a model at random to generate the new quizzes
 
             model_for_generation = models[torch.randint(len(models), (1,))]
 

-            c_quizzes = quizz_machine.generate_quizzes(
+            c_quizzes = quiz_machine.generate_quizzes(

                 nb_to_create,
                 model_for_generation=model_for_generation,
                 temperature=args.generation_temperature,
             )
 
                 nb_to_create,
                 model_for_generation=model_for_generation,
                 temperature=args.generation_temperature,
             )
 

-            nb_correct, seq_logproba = quizz_machine.compute_correctness(
+            nb_correct, seq_logproba = quiz_machine.compute_correctness(

                 c_quizzes,
                 models,
                 c_quizzes,
                 models,

-                both_directions=args.both_directions,
-                deterministic_validation=not args.stochastic_validation,
+                bidirectional_validation=args.bidirectional_validation,
+                deterministic_validation=args.deterministic_validation,

             )
 
             for n, l in zip(nb_correct, seq_logproba):
             )
 
             for n, l in zip(nb_correct, seq_logproba):


@@ -436,12 +449,14 @@ def create_c_quizzes(
                     len(models) + 1, nb_correct.size(), device=c_quizzes.device
                 )
 
                     len(models) + 1, nb_correct.size(), device=c_quizzes.device
                 )
 

-            recorded.append((c_quizzes, nb_correct))
+            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])
 
 
             nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
             nv = " ".join([str(x.item()) for x in nv])
 

-            nb_validated = valid_c_quizzes(recorded, standard_validity).size(0)
+            nb_validated = valid_c_quizzes(
+                quizzes_and_nb_correct_records, standard_validity
+            ).size(0)

 
             log_string(
                 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
 
             log_string(
                 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"


@@ -449,10 +464,12 @@ def create_c_quizzes(
 
     # store the new c_quizzes which have been validated
 
 
     # store the new c_quizzes which have been validated
 

-    new_c_quizzes = valid_c_quizzes(recorded, standard_validity)
+    new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
+
+    quiz_machine.reverse_random_half_in_place(new_c_quizzes)

 
 

-    quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
-    quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
+    quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
+    quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)

 
     # save a bunch of images to investigate what quizzes with a
     # certain nb of correct predictions look like
 
     # save a bunch of images to investigate what quizzes with a
     # certain nb of correct predictions look like


@@ -464,10 +481,14 @@ def create_c_quizzes(
             else ""
         )
 
             else ""
         )
 

-        q = valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72]
+        q = valid_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:
 
         if q.size(0) > 0:

-            quizz_machine.save_quizzes(
+            quiz_machine.save_quizzes(

                 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
             )
 
                 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
             )
 


@@ -502,6 +523,9 @@ 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} ----------------------------------------")
 
 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))
     # Select, improve, and eval the worst model
 
     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))


@@ -510,24 +534,21 @@ for n_epoch in range(args.nb_epochs):
         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
     )
 
         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
     )
 

-    one_epoch(weakest_model, quizz_machine)
+    one_epoch(weakest_model, quiz_machine)

 
     log_string(
 
     log_string(

-        f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
+        f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"

     )
 
     )
 

-    run_tests(weakest_model, quizz_machine, deterministic_synthesis=False)
+    run_tests(weakest_model, quiz_machine, deterministic_synthesis=False)

 
     log_string(
 
     log_string(

-        f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
+        f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"

     )
 
     )
 

-    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)
+    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
 
     # If all the models are good enough, generate new quizzes and
     # re-compute the test errors


@@ -535,13 +556,13 @@ for n_epoch in range(args.nb_epochs):
     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
         create_c_quizzes(
             models,
     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
         create_c_quizzes(
             models,

-            quizz_machine,
+            quiz_machine,

             nb_for_train=nb_new_c_quizzes_for_train,
             nb_for_test=nb_new_c_quizzes_for_test,
         )
 
         for model in models:
             nb_for_train=nb_new_c_quizzes_for_train,
             nb_for_test=nb_new_c_quizzes_for_test,
         )
 
         for model in models:

-            run_tests(model, quizz_machine, deterministic_synthesis=False)
+            run_tests(model, quiz_machine, deterministic_synthesis=False)

 
 
 ######################################################################
 
 
 ######################################################################