Update.

[culture.git] / main.py

diff --git a/main.py b/main.py
index 22edf7b..524715a 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -12,7 +12,16 @@ 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, problems
+import mygpt
+import sky, quizz_machine
+
+# 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

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


@@ -29,13 +38,6 @@ parser = argparse.ArgumentParser(
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 

-parser.add_argument(
-    "--task",
-    type=str,
-    default="twotargets",
-    help="file, byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp, greed",
-)
-

 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
 
 parser.add_argument("--result_dir", type=str, default=None)
 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
 
 parser.add_argument("--result_dir", type=str, default=None)


@@ -46,7 +48,7 @@ parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
 
 ########################################
 
 
 ########################################
 

-parser.add_argument("--nb_epochs", type=int, default=50)
+parser.add_argument("--nb_epochs", type=int, default=10000)

 
 parser.add_argument("--batch_size", type=int, default=None)
 
 
 parser.add_argument("--batch_size", type=int, default=None)
 


@@ -56,12 +58,8 @@ 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("--optim", type=str, default="adam")
-

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

-parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
-

 ########################################
 
 parser.add_argument("--model", type=str, default=None)
 ########################################
 
 parser.add_argument("--model", type=str, default=None)


@@ -82,245 +80,36 @@ 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)
 

-##############################
-# filetask
-
-parser.add_argument("--filetask_train_file", type=str, default=None)
-
-parser.add_argument("--filetask_test_file", type=str, default=None)
-
-##############################
-# rpl options
-
-parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
-
-parser.add_argument("--rpl_max_input", type=int, default=9)
-
-parser.add_argument("--rpl_prog_len", type=int, default=8)
-
-parser.add_argument("--rpl_nb_runs", type=int, default=5)
-
-parser.add_argument("--rpl_no_prog", action="store_true", default=False)
-
-##############################
-# grid options
-
-parser.add_argument("--grid_size", type=int, default=6)
-
-parser.add_argument("--grid_fraction_play", type=float, default=0)
-
-##############################
-# picoclvr options
-
-parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
-
-parser.add_argument("--picoclvr_height", type=int, default=12)
-
-parser.add_argument("--picoclvr_width", type=int, default=16)
-
-parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
-
-##############################
-# Maze options
-
-parser.add_argument("--maze_height", type=int, default=13)
-
-parser.add_argument("--maze_width", type=int, default=21)
-
-parser.add_argument("--maze_nb_walls", type=int, default=15)
-
-##############################
-# Snake options
-
-parser.add_argument("--snake_height", type=int, default=9)
-
-parser.add_argument("--snake_width", type=int, default=12)
-
-parser.add_argument("--snake_nb_colors", type=int, default=5)
-
-parser.add_argument("--snake_length", type=int, default=200)
-
-##############################
-# ByHeart options
-
-parser.add_argument("--byheart_separation", type=int, default=1)
-
-##############################
-# Stack options
-
-parser.add_argument("--stack_nb_steps", type=int, default=100)
-
-parser.add_argument("--stack_nb_stacks", type=int, default=3)
-
-parser.add_argument("--stack_nb_digits", type=int, default=3)
-
-parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
-
-##############################
-# Expr options
+parser.add_argument("--nb_gpts", type=int, default=5)

 
 

-parser.add_argument("--expr_nb_variables", type=int, default=5)
-
-parser.add_argument("--expr_sequence_length", type=int, default=40)
-
-parser.add_argument("--expr_operand_max", type=int, default=9)
-
-parser.add_argument("--expr_result_max", type=int, default=99)
-
-parser.add_argument("--expr_input_file", type=str, default=None)
-
-##############################
-# Mixing
-
-parser.add_argument("--mixing_hard", action="store_true", default=False)
-
-parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
-
-##############################
-# greed options
-
-parser.add_argument("--greed_height", type=int, default=5)
-
-parser.add_argument("--greed_width", type=int, default=7)
-
-parser.add_argument("--greed_T", type=int, default=25)
-
-parser.add_argument("--greed_nb_walls", type=int, default=5)
-
-parser.add_argument("--greed_nb_coins", type=int, default=2)
+parser.add_argument("--dirty_debug", action="store_true", default=False)

 
 ######################################################################
 
 args = parser.parse_args()
 
 
 ######################################################################
 
 args = parser.parse_args()
 

-assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
-

 if args.result_dir is None:
 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,
-    },
-    "file": {
-        "model": "37M",
-        "batch_size": 25,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "addition": {
-        "model": "352M",
-        "batch_size": 25,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "byheart": {
-        "model": "37M",
-        "batch_size": 25,
-        "nb_train_samples": 50000,
-        "nb_test_samples": 10000,
-    },
-    "expr": {
-        "model": "352M",
-        "batch_size": 25,
-        "nb_train_samples": 2500000,
-        "nb_test_samples": 10000,
-    },
-    "grid": {
-        "model": "37M",
-        "batch_size": 25,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "qmlp": {
-        "model": "37M",
-        "batch_size": 10,
-        "nb_train_samples": 100000,
-        "nb_test_samples": 1000,
-    },
-    "guessop": {
-        "model": "352M",
-        "batch_size": 25,
-        "nb_train_samples": 1000000,
-        "nb_test_samples": 10000,
-    },
-    "learnop": {
-        "model": "37M",
-        "batch_size": 25,
-        "nb_train_samples": 50000,
-        "nb_test_samples": 10000,
-    },
-    "maze": {
-        "model": "37M",
-        "batch_size": 5,
-        "nb_train_samples": 100000,
-        "nb_test_samples": 10000,
-    },
-    "picoclvr": {
-        "model": "37M",
-        "batch_size": 25,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "rpl": {
-        "model": "352M",
-        "batch_size": 5,
-        "nb_train_samples": 2500000,
-        "nb_test_samples": 10000,
-    },
-    "snake": {
-        "model": "37M",
-        "batch_size": 25,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "stack": {
-        "model": "37M",
-        "batch_size": 25,
-        "nb_train_samples": 100000,
-        "nb_test_samples": 1000,
-    },
-    "twotargets": {
-        "model": "37M",
-        "batch_size": 25,
-        "nb_train_samples": 50000,
-        "nb_test_samples": 10000,
-    },
-    "memory": {
-        "model": "37M",
-        "batch_size": 100,
-        "nb_train_samples": 25000,
-        "nb_test_samples": 1000,
-    },
-    "mixing": {
-        "model": "37M",
-        "batch_size": 25,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "mnist": {
-        "model": "37M",
-        "batch_size": 10,
-        "nb_train_samples": 60000,
-        "nb_test_samples": 10000,
-    },
-    "greed": {
-        "model": "37M",
-        "batch_size": 25,
-        "nb_train_samples": 25000,
-        "nb_test_samples": 10000,
-    },
+if args.dirty_debug:
+    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": 250000,
+    "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)

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


@@ -374,9 +163,8 @@ else:
 try:
     os.mkdir(args.result_dir)
 except FileExistsError:
 try:
     os.mkdir(args.result_dir)
 except FileExistsError:

-    if not args.resume:
-        print(f"result directory {args.result_dir} already exists")
-        exit(1)
+    print(f"result directory {args.result_dir} already exists")
+    exit(1)

 
 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 
 
 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 


@@ -410,24 +198,9 @@ for n in vars(args):
 
 ######################################################################
 
 
 ######################################################################
 

-
-def picoclvr_pruner_horizontal_green(p):
-    return not ("green" in p and ("left" in p or "right" in p))
-
-
-picoclvr_pruner_train = (
-    picoclvr_pruner_horizontal_green
-    if args.picocvlr_prune_properties in {"train+eval"}
-    else None
-)
-
-picoclvr_pruner_eval = (
-    (lambda p: not picoclvr_pruner_horizontal_green(p))
-    if args.picocvlr_prune_properties in {"train+eval", "eval"}
-    else None
-)
-
-######################################################################
+if args.dirty_debug:
+    args.nb_train_samples = 2500
+    args.nb_test_samples = 100

 
 if args.physical_batch_size is None:
     args.physical_batch_size = args.batch_size
 
 if args.physical_batch_size is None:
     args.physical_batch_size = args.batch_size


@@ -437,267 +210,33 @@ 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
 

-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,
-    )
-
-else:
-    raise ValueError(f"Unknown task {args.task}")
+quizz_machine = quizz_machine.QuizzMachine(
+    sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2),
+    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}")
 
 
 ######################################################################
 
 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}")
 

-##############################
-
-models = []
-
-for k in range(2):
-    models.append(
-        mygpt.MyGPT(
-            vocabulary_size=vocabulary_size,
-            dim_model=args.dim_model,
-            dim_keys=args.dim_keys,
-            dim_hidden=args.dim_hidden,
-            nb_heads=args.nb_heads,
-            nb_blocks=args.nb_blocks,
-            causal=True,
-            dropout=args.dropout,
-        ).to(device)
-    )
-
-
-nb_parameters = sum(p.numel() for p in models[0].parameters())
-log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
-

 ######################################################################
 
 # 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)


@@ -719,11 +258,11 @@ 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)


@@ -739,50 +278,15 @@ if args.max_percents_of_test_in_train >= 0:
 
 ##############################
 
 
 ##############################
 

-if args.learning_rate_schedule == "cos":
-    learning_rate_schedule = {}
-    for n_epoch in range(args.nb_epochs):
-        u = n_epoch / args.nb_epochs * math.pi
-        learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
-else:
-    u = {
-        int(k): float(v)
-        for k, v in [
-            tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
-        ]
-    }
-
-    learning_rate_schedule = {}
-    learning_rate = args.learning_rate
-    for n_epoch in range(args.nb_epochs):
-        if n_epoch in u:
-            learning_rate = u[n_epoch]
-        learning_rate_schedule[n_epoch] = learning_rate
-
-log_string(f"learning_rate_schedule {learning_rate_schedule}")
-
-time_pred_result = None
-
-######################################################################
-

 
 

-def one_epoch(model, task, learning_rate):
-    log_string(f"learning_rate {learning_rate}")
-
-    if args.optim == "sgd":
-        optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
-    elif args.optim == "adam":
-        optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
-    elif args.optim == "adamw":
-        optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
-    else:
-        raise ValueError(f"Unknown optimizer {args.optim}.")
+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
 
 
     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:


@@ -807,14 +311,14 @@ def one_epoch(model, task, learning_rate):
 ######################################################################
 
 
 ######################################################################
 
 

-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))


@@ -826,7 +330,7 @@ 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,


@@ -838,77 +342,150 @@ def run_tests(model, task, deterministic_synthesis):
 
         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 
 
         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 

-    return main_test_accuracy
+    model.main_test_accuracy = main_test_accuracy

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

-def create_quizzes(
+def create_c_quizzes(
+    model,

     other_models,
     other_models,

-    task,
+    quizz_machine,

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

-    nb_runs=10,
-    nb_min_correct=9,
-    nb_max_correct=9,
+    min_ave_seq_logproba=None,

 ):
     kept = []
 
 ):
     kept = []
 

+    sum_logits, sum_nb_c_quizzes = 0, 0
+

     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
     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_generate = 4 * (nb_for_train + nb_for_test)
+
+        new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes(

             n_epoch=n_epoch,
             result_dir=args.result_dir,
             logger=log_string,
             n_epoch=n_epoch,
             result_dir=args.result_dir,
             logger=log_string,

-            nb=4 * (nb_for_train + nb_for_test),
-            models=other_models,
-            nb_runs=nb_runs,
+            nb=nb_to_generate,
+            model=model,
+            other_models=other_models,
+            min_ave_seq_logproba=min_ave_seq_logproba,
+        )
+
+        sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
+        sum_nb_c_quizzes += new_c_quizzes.size(0)
+
+        to_keep = new_c_quizzes[nb_correct == len(other_models) - 1]
+
+        if args.dirty_debug:
+            to_keep = new_c_quizzes
+
+        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}%)"

         )
 
         )
 

-        to_keep = new_quizzes[
-            torch.logical_and(
-                nb_correct >= nb_min_correct, nb_correct <= nb_max_correct
-            )
-        ]
-        log_string(f"keep {to_keep.size(0)} quizzes")

         kept.append(to_keep)
 
         kept.append(to_keep)
 

-    new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
+    new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]

 
 

-    task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
-    task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
+    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)

 
 

-    task.save_image(
-        new_quizzes[:96],
+    quizz_machine.problem.save_quizzes(
+        new_c_quizzes[:72],

         args.result_dir,
         args.result_dir,

-        f"world_new_{n_epoch:04d}.png",
+        f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",

         log_string,
     )
 
         log_string,
     )
 

+    return sum_logits / sum_nb_c_quizzes
+

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

-accuracy_to_make_quizzes = 0.95
+models = []
+
+for k in range(args.nb_gpts):
+    model = mygpt.MyGPT(
+        vocabulary_size=vocabulary_size,
+        dim_model=args.dim_model,
+        dim_keys=args.dim_keys,
+        dim_hidden=args.dim_hidden,
+        nb_heads=args.nb_heads,
+        nb_blocks=args.nb_blocks,
+        causal=True,
+        dropout=args.dropout,
+    ).to(device)
+
+    model.main_test_accuracy = 0.0
+    model.id = k

 
 

-for n_epoch in range(nb_epochs_finished, args.nb_epochs):
-    learning_rate = learning_rate_schedule[n_epoch]
+    models.append(model)

 
 

-    for m in models:
-        one_epoch(m, task, learning_rate)
-        test_accuracy = run_tests(m, task, deterministic_synthesis=False)

 
 

-    if test_accuracy >= accuracy_to_make_quizzes:
+nb_parameters = sum(p.numel() for p in models[0].parameters())
+log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
+
+######################################################################
+
+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])
+    log_string(f"current accuracies {a}")
+
+    # select the model with lowest accuracy
+    models.sort(key=lambda model: model.main_test_accuracy)
+    model = models[0]
+
+    log_string(
+        f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
+    )
+
+    # improve it
+    one_epoch(model, quizz_machine)
+
+    quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+
+    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(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]) >= accuracy_to_make_c_quizzes:

         other_models = models.copy()
         other_models.remove(model)
         other_models = models.copy()
         other_models.remove(model)

-        create_quizzes(other_models, task)

 
 

-    # --------------------------------------------
-
-    time_current_result = datetime.datetime.now()
-    if time_pred_result is not None:
-        log_string(
-            f"next_result {time_current_result + (time_current_result - time_pred_result)}"
+        ave_seq_logproba = create_c_quizzes(
+            model,
+            other_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,

         )
         )

-    time_pred_result = time_current_result
+
+        # 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)
+

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