X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=sidebyside;f=main.py;h=a6c482f746e2b15dbe750dd964d2edddf0266b44;hb=db8c21397d370ae16fd6078858c649e2ab14fe4e;hp=e0588224f07ecafde8105b00c2b004e0b195e249;hpb=5c751aa1bbfbcf42654f4626f81905acfa946c15;p=culture.git

diff --git a/main.py b/main.py
index e058822..a6c482f 100755
--- a/main.py
+++ b/main.py
@@ -12,7 +12,15 @@ from torch import nn
 from torch.nn import functional as F
 
 import ffutils
-import mygpt, tasks, problems
+import mygpt
+import sky, wireworld, quizz_machine
+
+# world quizzes vs. culture quizzes
+
+######################################################################
+
+nb_new_c_quizzes_for_train = 1000
+nb_new_c_quizzes_for_test = 100
 
 ######################################################################
 
@@ -29,9 +37,7 @@ parser = argparse.ArgumentParser(
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 
-parser.add_argument("--task", type=str, default="world", help="world")
-
-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)
 
@@ -51,7 +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("--learning_rate", type=float, default=1e-4)
+parser.add_argument("--learning_rate", type=float, default=1e-3)
 
 ########################################
 
@@ -73,32 +79,48 @@ 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("--check", action="store_true", default=False)
+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)
 
 ######################################################################
 
 args = parser.parse_args()
 
 if args.result_dir is None:
-    args.result_dir = f"results_{args.task}"
+    args.result_dir = f"results_culture"
 
 ######################################################################
 
-default_task_args = {
-    "world": {
-        "model": "37M",
-        "batch_size": 100,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
+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,
+    "nb_train_samples": 100000,
+    "nb_test_samples": 10000,
 }
 
-if args.task in default_task_args:
-    for k, v in default_task_args[args.task].items():
-        if getattr(args, k) is None:
-            setattr(args, k, v)
+for k, v in default_args.items():
+    if getattr(args, k) is None:
+        setattr(args, k, v)
 
 ######################################################################
 
@@ -187,8 +209,8 @@ for n in vars(args):
 
 ######################################################################
 
-if args.check:
-    args.nb_train_samples = 500
+if args.dirty_debug:
+    args.nb_train_samples = 2500
     args.nb_test_samples = 100
 
 if args.physical_batch_size is None:
@@ -199,235 +221,28 @@ else:
 assert args.nb_train_samples % args.batch_size == 0
 assert args.nb_test_samples % args.batch_size == 0
 
-if args.task == "file":
-    assert (
-        args.filetask_train_file is not None and args.filetask_test_file is not None
-    ), "You have to specify the task train and test files"
-    task = tasks.TaskFromFile(
-        args.filetask_train_file,
-        args.filetask_test_file,
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        shuffle=True,
-        device=device,
-    )
-    args.max_percents_of_test_in_train = 0
-
-elif args.task == "byheart":
-    task = tasks.SandBox(
-        problem=problems.ProblemByHeart(separation=args.byheart_separation),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        device=device,
-    )
-    args.max_percents_of_test_in_train = -1
-
-elif args.task == "world":
-    task = tasks.World(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        result_dir=args.result_dir,
-        logger=log_string,
-        device=device,
-    )
-    args.max_percents_of_test_in_train = -1
-
-elif args.task == "learnop":
-    task = tasks.SandBox(
-        problem=problems.ProblemLearnOperator(),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        device=device,
-    )
-
-
-elif args.task == "guessop":
-    task = tasks.SandBox(
-        problem=problems.ProblemGuessOperator(),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        device=device,
-    )
-
-
-elif args.task == "twotargets":
-    task = tasks.SandBox(
-        problem=problems.ProblemTwoTargets(),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        device=device,
-    )
-
-elif args.task == "memory":
-    task = tasks.SandBox(
-        problem=problems.ProblemMemory(),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        device=device,
-    )
-
-elif args.task == "mixing":
-    task = tasks.SandBox(
-        problem=problems.ProblemMixing(
-            hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
-        ),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        device=device,
-    )
-
-elif args.task == "addition":
-    task = tasks.SandBox(
-        problem=problems.ProblemAddition(),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        device=device,
-    )
-
-elif args.task == "picoclvr":
-    task = tasks.PicoCLVR(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        height=args.picoclvr_height,
-        width=args.picoclvr_width,
-        nb_colors=args.picoclvr_nb_colors,
-        logger=log_string,
-        device=device,
-        pruner_train=picoclvr_pruner_train,
-        pruner_eval=picoclvr_pruner_eval,
-    )
-
-elif args.task == "mnist":
-    task = tasks.MNIST(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        device=device,
-    )
-
-elif args.task == "maze":
-    task = tasks.Maze(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        height=args.maze_height,
-        width=args.maze_width,
-        nb_walls=args.maze_nb_walls,
-        device="cpu",
-    )
-
-elif args.task == "snake":
-    task = tasks.Snake(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        height=args.snake_height,
-        width=args.snake_width,
-        nb_colors=args.snake_nb_colors,
-        length=args.snake_length,
-        prompt_length=args.snake_length // 2,
-        device=device,
-    )
-
-elif args.task == "stack":
-    task = tasks.Stack(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        logger=log_string,
-        nb_steps=args.stack_nb_steps,
-        nb_stacks=args.stack_nb_stacks,
-        nb_digits=args.stack_nb_digits,
-        fraction_values_for_train=args.stack_fraction_values_for_train,
-        device=device,
-    )
-
-elif args.task == "expr":
-    task = tasks.Expr(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        nb_variables=args.expr_nb_variables,
-        sequence_length=args.expr_sequence_length,
-        operand_max=args.expr_operand_max,
-        result_max=args.expr_result_max,
-        batch_size=args.physical_batch_size,
-        device=device,
-    )
-
-elif args.task == "rpl":
-    task = tasks.RPL(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        nb_starting_values=args.rpl_nb_starting_values,
-        max_input=args.rpl_max_input,
-        prog_len=args.rpl_prog_len,
-        nb_runs=args.rpl_nb_runs,
-        no_prog=args.rpl_no_prog,
-        logger=log_string,
-        device=device,
-    )
-
-elif args.task == "grid":
-    task = tasks.Grid(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        size=args.grid_size,
-        fraction_play=args.grid_fraction_play,
-        logger=log_string,
-        device=device,
-    )
-
-elif args.task == "qmlp":
-    task = tasks.QMLP(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        result_dir=args.result_dir,
-        logger=log_string,
-        device=device,
-    )
-
-elif args.task == "greed":
-    task = tasks.Greed(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.physical_batch_size,
-        height=args.greed_height,
-        width=args.greed_width,
-        T=args.greed_T,
-        nb_walls=args.greed_nb_walls,
-        nb_coins=args.greed_nb_coins,
-        logger=log_string,
-        device=device,
-    )
-
+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(f"Unknown task {args.task}")
+    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,
+    result_dir=args.result_dir,
+    logger=log_string,
+    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}")
 
@@ -436,8 +251,10 @@ log_string(f"vocabulary_size {vocabulary_size}")
 # 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)
@@ -459,11 +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(
-        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(
-            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)
@@ -480,14 +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
 
-    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:
@@ -512,14 +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
 
-        for input in task.batches(split="test"):
+        for input in quizz_machine.batches(split="test"):
             input = input.to(device)
 
             bs = model(mygpt.BracketedSequence(input))
@@ -531,11 +348,10 @@ def run_tests(model, task, deterministic_synthesis):
 
             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,
-            logger=log_string,
             deterministic_synthesis=deterministic_synthesis,
         )
 
@@ -549,41 +365,100 @@ def run_tests(model, task, deterministic_synthesis):
 ######################################################################
 
 
-def create_quizzes(
-    model,
-    other_models,
-    task,
+def create_c_quizzes(
+    models,
+    quizz_machine,
     nb_for_train=1000,
     nb_for_test=100,
+    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
+
+    def nb_generated():
+        return sum([sum([x.size(0) for x in recorded[n]]) for n in recorded.keys()])
+
+    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)
+            ]
+        )
 
-    while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
-        new_quizzes, nb_correct = task.create_new_quizzes(
+    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,
-            nb=4 * (nb_for_train + nb_for_test),
-            model=model,
-            other_models=other_models,
         )
 
-        to_keep = new_quizzes[nb_correct == len(other_models) - 1]
-        log_string(f"keep {to_keep.size(0)} quizzes")
-        kept.append(to_keep)
+        sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
+        sum_nb_c_quizzes += new_c_quizzes.size(0)
 
-    new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
+        if args.dirty_debug:
+            nb_correct = torch.randint(
+                len(models) + 1, nb_correct.size(), device=new_c_quizzes.device
+            )
 
-    task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
-    task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
+        for n in range(nb_correct.max() + 1):
+            recorded[n].append(new_c_quizzes[nb_correct == n].clone())
 
-    task.save_image(
-        new_quizzes[:96],
-        args.result_dir,
-        f"world_quiz_{n_epoch:04d}_{model.id:02d}.png",
-        log_string,
+        log_string(
+            f"keep c_quizzes {nb_validated()*100/nb_generated():.02f}% kept total {nb_validated()} / {nb_to_create}"
+        )
+
+    # 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(
+        [recorded[n] for n in range(args.min_to_validate, args.max_to_validate + 1)],
+        dim=0,
     )
 
+    new_c_quizzes = new_c_quizzes[
+        torch.randperm(new_c_quizzes.size(0), device=new_c_quizzes.device)[
+            : nb_for_train + nb_for_test
+        ]
+    ]
+
+    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
+
 
 ######################################################################
 
@@ -612,16 +487,16 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 
 ######################################################################
 
-accuracy_to_make_quizzes = 0.975
-nb_new_quizzes_for_train = 1000
-nb_new_quizzes_for_test = 100
-
-if args.check:
-    accuracy_to_make_quizzes = 0.0
-    nb_new_quizzes_for_train = 10
-    nb_new_quizzes_for_test = 10
+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])
+    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)
     model = models[0]
@@ -631,26 +506,41 @@ for n_epoch in range(args.nb_epochs):
     )
 
     # improve it
-    one_epoch(model, task)
+    one_epoch(model, quizz_machine)
+
+    quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 
     log_string(
-        f"train_set_composition world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_quizzes}"
+        f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
     )
 
     # test it
-    run_tests(model, task, deterministic_synthesis=False)
-
-    if model.main_test_accuracy >= accuracy_to_make_quizzes:
-        other_models = models.copy()
-        other_models.remove(model)
-
-        create_quizzes(
-            model,
-            other_models,
-            task,
-            nb_for_train=nb_new_quizzes_for_train,
-            nb_for_test=nb_new_quizzes_for_test,
+    run_tests(model, quizz_machine, deterministic_synthesis=False)
+
+    log_string(
+        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]) >= 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,
+            min_ave_seq_logproba=min_ave_seq_logproba,
         )
 
+        # We keep the first average logits as a reference
+        # 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:
+            run_tests(model, quizz_machine, deterministic_synthesis=False)
+
 
 ######################################################################