Update.

[culture.git] / main.py

diff --git a/main.py b/main.py
index ebecad8..a6c482f 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -12,13 +12,13 @@ from torch import nn
 from torch.nn import functional as F
 
 import ffutils
 from torch.nn import functional as F
 
 import ffutils

-import mygpt, tasks
+import mygpt
+import sky, wireworld, quizz_machine

 
 # world quizzes vs. culture quizzes
 
 ######################################################################
 
 
 # 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
 
 nb_new_c_quizzes_for_train = 1000
 nb_new_c_quizzes_for_test = 100
 


@@ -37,7 +37,7 @@ parser = argparse.ArgumentParser(
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 
     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)
 
 
 parser.add_argument("--result_dir", type=str, default=None)
 


@@ -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-4)
+parser.add_argument("--learning_rate", type=float, default=1e-3)

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


@@ -79,8 +79,20 @@ 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("--problem", type=str, default="sky")
+

 parser.add_argument("--nb_gpts", type=int, default=5)
 
 parser.add_argument("--nb_gpts", type=int, default=5)
 

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


@@ -93,7 +105,7 @@ if args.result_dir is None:
 ######################################################################
 
 if args.dirty_debug:
 ######################################################################
 
 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
 
     nb_new_c_quizzes_for_train = 100
     nb_new_c_quizzes_for_test = 10
 


@@ -102,7 +114,7 @@ if args.dirty_debug:
 default_args = {
     "model": "37M",
     "batch_size": 100,
 default_args = {
     "model": "37M",
     "batch_size": 100,

-    "nb_train_samples": 250000,
+    "nb_train_samples": 100000,

     "nb_test_samples": 10000,
 }
 
     "nb_test_samples": 10000,
 }
 


@@ -209,7 +221,15 @@ else:
 assert args.nb_train_samples % args.batch_size == 0
 assert args.nb_test_samples % args.batch_size == 0
 
 assert args.nb_train_samples % args.batch_size == 0
 assert args.nb_test_samples % args.batch_size == 0
 

-task = tasks.World(
+if args.problem == "sky":
+    problem = sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2, speed=3)
+elif args.problem == "wireworld":
+    problem = wireworld.Wireworld(height=8, width=10, nb_iterations=2, speed=5)
+else:
+    raise ValueError
+
+quizz_machine = quizz_machine.QuizzMachine(
+    problem=problem,

     nb_train_samples=args.nb_train_samples,
     nb_test_samples=args.nb_test_samples,
     batch_size=args.physical_batch_size,
     nb_train_samples=args.nb_train_samples,
     nb_test_samples=args.nb_test_samples,
     batch_size=args.physical_batch_size,


@@ -222,7 +242,7 @@ task = tasks.World(
 
 log_string(f"device {device}")
 
 
 log_string(f"device {device}")
 

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

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


@@ -231,8 +251,10 @@ 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 task.batches(split="train", desc="train-entropy"):
-    token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
+for input in quizz_machine.batches(split="train", desc="train-entropy"):
+    token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
+        (0, 1)
+    )

 token_probas = token_count / token_count.sum()
 entropy = -torch.xlogy(token_probas, token_probas).sum()
 train_set_perplexity = math.exp(entropy)
 token_probas = token_count / token_count.sum()
 entropy = -torch.xlogy(token_probas, token_probas).sum()
 train_set_perplexity = math.exp(entropy)


@@ -254,11 +276,11 @@ 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(

-        task.batches(split="test", desc="test-check"), 25000
+        quizz_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(

-            task.batches(split="train", desc="train-check"), 25000
+            quizz_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)


@@ -275,14 +297,14 @@ if args.max_percents_of_test_in_train >= 0:
 ##############################
 
 
 ##############################
 
 

-def one_epoch(model, task):
+def one_epoch(model, quizz_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 task.batches(split="train"):
+    for input in quizz_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:


@@ -307,14 +329,14 @@ def one_epoch(model, task):
 ######################################################################
 
 
 ######################################################################
 
 

-def run_tests(model, task, deterministic_synthesis):
+def run_tests(model, quizz_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 task.batches(split="test"):
+        for input in quizz_machine.batches(split="test"):

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


@@ -326,11 +348,10 @@ def run_tests(model, task, deterministic_synthesis):
 
             nb_test_samples += input.size(0)
 
 
             nb_test_samples += input.size(0)
 

-        main_test_accuracy = task.produce_results(
+        main_test_accuracy = quizz_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,

-            logger=log_string,

             deterministic_synthesis=deterministic_synthesis,
         )
 
             deterministic_synthesis=deterministic_synthesis,
         )
 


@@ -345,55 +366,96 @@ def run_tests(model, task, deterministic_synthesis):
 
 
 def create_c_quizzes(
 
 
 def create_c_quizzes(

-    model,
-    other_models,
-    task,
+    models,
+    quizz_machine,

     nb_for_train=1000,
     nb_for_test=100,
     nb_for_train=1000,
     nb_for_test=100,

-    desired_average_logits=None,
+    min_ave_seq_logproba=None,

 ):
 ):

-    kept = []
+    # We will store the generated quizzes for each number of
+    # correct prediction
+    recorded = dict([(n, []) for n in range(len(models) + 1)])

 
 

+    model_indexes = []

     sum_logits, sum_nb_c_quizzes = 0, 0
 
     sum_logits, sum_nb_c_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)
+    def nb_generated():
+        return sum([sum([x.size(0) for x in recorded[n]]) for n in recorded.keys()])

 
 

-        new_c_quizzes, nb_correct, average_logits = task.create_c_quizzes(
+    def nb_validated():
+        return sum(
+            [
+                sum([x.size(0) for x in recorded[n]])
+                for n in range(args.min_to_validate, args.max_to_validate + 1)
+            ]
+        )
+
+    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,
             n_epoch=n_epoch,
             result_dir=args.result_dir,

-            logger=log_string,
-            nb=nb_to_generate,
-            model=model,
-            other_models=other_models,
-            desired_average_logits=desired_average_logits,

         )
 
         )
 

-        sum_logits += new_c_quizzes.size(0) * average_logits
+        sum_logits += new_c_quizzes.size(0) * ave_seq_logproba

         sum_nb_c_quizzes += new_c_quizzes.size(0)
 
         sum_nb_c_quizzes += new_c_quizzes.size(0)
 

-        to_keep = new_c_quizzes[nb_correct == len(other_models) - 1]
-

         if args.dirty_debug:
         if args.dirty_debug:

-            to_keep = new_c_quizzes
+            nb_correct = torch.randint(
+                len(models) + 1, nb_correct.size(), device=new_c_quizzes.device
+            )
+
+        for n in range(nb_correct.max() + 1):
+            recorded[n].append(new_c_quizzes[nb_correct == n].clone())

 
         log_string(
 
         log_string(

-            f"keep {to_keep.size(0)}/{new_c_quizzes.size(0)} c_quizzes ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%)"
+            f"keep c_quizzes {nb_validated()*100/nb_generated():.02f}% kept total {nb_validated()} / {nb_to_create}"

         )
 
         )
 

-        kept.append(to_keep)
+    # concatenate and shuffle
+    for n in recorded.keys():
+        if len(recorded[n]) > 0:
+            q = torch.cat(recorded[n], dim=0)
+            q = q[torch.randperm(q.size(0), device=q.device)]
+            recorded[n] = q
+        else:
+            del recorded[n]

 
 

-    new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
+    new_c_quizzes = torch.cat(
+        [recorded[n] for n in range(args.min_to_validate, args.max_to_validate + 1)],
+        dim=0,
+    )

 
 

-    task.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
-    task.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
+    new_c_quizzes = new_c_quizzes[
+        torch.randperm(new_c_quizzes.size(0), device=new_c_quizzes.device)[
+            : nb_for_train + nb_for_test
+        ]
+    ]

 
 

-    task.save_quizzes(
-        new_c_quizzes[:72],
-        args.result_dir,
-        f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",
-        log_string,
-    )
+    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)
+
+    for n in recorded.keys():
+        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,
+            f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
+        )

 
     return sum_logits / sum_nb_c_quizzes
 
 
     return sum_logits / sum_nb_c_quizzes
 


@@ -425,14 +487,15 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 
 ######################################################################
 
 
 ######################################################################
 

-desired_average_logits = None
+min_ave_seq_logproba = 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])
 
 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}")
+    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)
 
     # select the model with lowest accuracy
     models.sort(key=lambda model: model.main_test_accuracy)


@@ -443,45 +506,41 @@ for n_epoch in range(args.nb_epochs):
     )
 
     # improve it
     )
 
     # improve it

-    one_epoch(model, task)
+    one_epoch(model, quizz_machine)

 
 

-    task.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+    quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)

 
     log_string(
 
     log_string(

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

     )
 
     # test it
     )
 
     # test it

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

 
     log_string(
 
     log_string(

-        f"test_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}"
+        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:
-        other_models = models.copy()
-        other_models.remove(model)
-
-        average_logits = create_c_quizzes(
-            model,
-            other_models,
-            task,
+    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,

             nb_for_train=nb_new_c_quizzes_for_train,
             nb_for_test=nb_new_c_quizzes_for_test,
             nb_for_train=nb_new_c_quizzes_for_train,
             nb_for_test=nb_new_c_quizzes_for_test,

-            desired_average_logits=desired_average_logits,
+            min_ave_seq_logproba=min_ave_seq_logproba,

         )
 
         # We keep the first average logits as a reference
         )
 
         # We keep the first average logits as a reference

-        if desired_average_logits is None:
-            desired_average_logits = average_logits
-        else:
-            log_string(
-                f"desired_average_logits {desired_average_logits} average_logits {average_logits}"
-            )
+        # if min_ave_seq_logproba is None:
+        # min_ave_seq_logproba = ave_seq_logproba
+        # else:
+        # log_string(
+        # f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
+        # )

 
         # We update everyone
         for model in models:
 
         # We update everyone
         for model in models:

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

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