X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=6b00bbfd991178841468b75006f94121668c2b4f;hb=HEAD;hp=63e66680fe37126b2b9ead3847a0b44cf179eb40;hpb=c9dbc3abf436df8af1379d04ab51159e821496f1;p=culture.git

diff --git a/main.py b/main.py
index 63e6668..40772c2 100755
--- a/main.py
+++ b/main.py
@@ -3,63 +3,70 @@
 # Any copyright is dedicated to the Public Domain.
 # https://creativecommons.org/publicdomain/zero/1.0/
 
-# Written by Francois Fleuret <francois@fleuret.org>
+# > A > f(A) > B ; > f(B)
+# < f(B) ; < B < f(A) < A
 
-# torch.backends.cuda.matmul.allow_tf23
-# torch.autocast(torch.bfloat16)
+# Written by Francois Fleuret <francois@fleuret.org>
 
-import math, sys, argparse, time, tqdm, os
+import math, sys, argparse, time, tqdm, os, datetime, warnings
 
 import torch, torchvision
 from torch import nn
 from torch.nn import functional as F
 
 import ffutils
-import mygpt, tasks
 
-######################################################################
+import mygpt
+import sky, grids, quiz_machine
 
-if torch.cuda.is_available():
-    device = torch.device("cuda")
-    torch.backends.cuda.matmul.allow_tf32 = True
-else:
-    device = torch.device("cpu")
+from quiz_machine import one_batch_masked_inplace_autoregression
+
+import threading, subprocess
+
+import torch.multiprocessing as mp
 
 ######################################################################
 
 parser = argparse.ArgumentParser(
-    description="An implementation of GPT with cache.",
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 
-parser.add_argument(
-    "--task",
-    type=str,
-    default="sandbox",
-    help="sandbox, picoclvr, mnist, maze, snake, stack, expr, rpl, 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)
 
 parser.add_argument("--seed", type=int, default=0)
 
-parser.add_argument("--nb_epochs", type=int, default=None)
+parser.add_argument("--resume", action="store_true", default=False)
+
+parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
+
+parser.add_argument("--log_command", type=str, default=None)
+
+# ----------------------------------
+
+parser.add_argument("--nb_epochs", type=int, default=10000)
 
 parser.add_argument("--batch_size", type=int, default=None)
 
+parser.add_argument("--physical_batch_size", type=int, default=None)
+
+parser.add_argument("--inference_batch_size", type=int, default=None)
+
 parser.add_argument("--nb_train_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("--nb_new_c_quizzes_for_train", type=int, default=None)
+
+parser.add_argument("--nb_new_c_quizzes_for_test", type=int, default=None)
 
-parser.add_argument("--learning_rate", type=float, default=1e-4)
+parser.add_argument("--learning_rate", type=float, default=5e-4)
 
-parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
+parser.add_argument("--schedule_free", action="store_true", default=False)
 
-parser.add_argument("--model", type=str, default="37M")
+# ----------------------------------
+parser.add_argument("--model", type=str, default=None)
 
 parser.add_argument("--dim_model", type=int, default=None)
 
@@ -73,157 +80,97 @@ parser.add_argument("--nb_blocks", type=int, default=None)
 
 parser.add_argument("--dropout", type=float, default=0.1)
 
+# ----------------------------------
 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 
-parser.add_argument("--no_checkpoint", action="store_true", default=False)
-
-parser.add_argument("--overwrite_results", action="store_true", default=False)
-
-parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
-
-##############################
-# picoclvr options
-
-parser.add_argument("--sandbox_level", type=int, default=0)
-
-parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
-
-parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
+parser.add_argument("--problem", type=str, default="grids")
 
-parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
+parser.add_argument("--nb_threads", type=int, default=1)
 
-##############################
-# picoclvr options
+parser.add_argument("--gpus", type=str, default="all")
 
-parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
+# ----------------------------------
 
-parser.add_argument("--picoclvr_height", type=int, default=12)
+parser.add_argument("--nb_gpts", type=int, default=5)
 
-parser.add_argument("--picoclvr_width", type=int, default=16)
+parser.add_argument("--max_fail_to_validate", type=int, default=3)
 
-parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.95)
 
-##############################
-# Maze options
+parser.add_argument("--proba_understands", type=float, default=0.95)
 
-parser.add_argument("--maze_height", type=int, default=23)
+parser.add_argument("--proba_not_understands", type=float, default=0.1)
 
-parser.add_argument("--maze_width", type=int, default=39)
+parser.add_argument("--temperature_hot", type=float, default=1.5)
 
-parser.add_argument("--maze_nb_walls", type=int, default=45)
+parser.add_argument("--temperature_cold", type=float, default=1)
 
-##############################
-# Snake options
+parser.add_argument("--prompt_noise", type=float, default=0.05)
 
-parser.add_argument("--snake_height", type=int, default=6)
+parser.add_argument("--dirty_debug", action="store_true", default=False)
 
-parser.add_argument("--snake_width", type=int, default=8)
+parser.add_argument("--test", type=str, default=None)
 
-parser.add_argument("--snake_nb_colors", type=int, default=5)
-
-parser.add_argument("--snake_length", type=int, default=200)
-
-##############################
-# 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=0.75)
+grids_tasks = ", ".join(
+    [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
+)
 
-##############################
-# Expr options
+parser.add_argument(
+    "--grids_world_tasks",
+    type=str,
+    default="replace_color,translate,grow,frame",
+    help="A comma-separated subset of: " + grids_tasks + ".",
+)
 
-parser.add_argument("--expr_nb_variables", type=int, default=5)
+parser.add_argument(
+    "--grids_science_tasks",
+    type=str,
+    default=None,
+    help="A comma-separated subset of: " + grids_tasks + ", or None.",
+)
 
-parser.add_argument("--expr_sequence_length", type=int, default=40)
+######################################################################
 
-parser.add_argument("--expr_operand_max", type=int, default=9)
+parser.add_argument("--sky_height", type=int, default=6)
 
-parser.add_argument("--expr_result_max", type=int, default=99)
+parser.add_argument("--sky_width", type=int, default=8)
 
-parser.add_argument("--expr_input_file", type=str, default=None)
+parser.add_argument("--sky_nb_birds", type=int, default=3)
 
-##############################
-# World options
+parser.add_argument("--sky_nb_iterations", type=int, default=2)
 
-parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
+parser.add_argument("--sky_speed", type=int, default=3)
 
 ######################################################################
 
 args = parser.parse_args()
 
-assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
-
 if args.result_dir is None:
-    args.result_dir = f"results_{args.task}"
+    args.result_dir = f"results_culture"
+
+assert not args.grids_science_tasks or (
+    len(
+        set(args.grids_world_tasks.split(","))
+        & set(args.grids_science_tasks.split(","))
+    )
+    == 0
+), "World and science tasks have to be disjoint"
 
 ######################################################################
 
-default_task_args = {
-    "sandbox": {
-        "nb_epochs": 50,
-        "batch_size": 25,
-        "nb_train_samples": 100000,
-        "nb_test_samples": 10000,
-    },
-    "picoclvr": {
-        "nb_epochs": 25,
-        "batch_size": 25,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "mnist": {
-        "nb_epochs": 25,
-        "batch_size": 10,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "maze": {
-        "nb_epochs": 25,
-        "batch_size": 5,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "snake": {
-        "nb_epochs": 5,
-        "batch_size": 25,
-        "nb_train_samples": 250000,
-        "nb_test_samples": 10000,
-    },
-    "stack": {
-        "nb_epochs": 5,
-        "batch_size": 25,
-        "nb_train_samples": 100000,
-        "nb_test_samples": 1000,
-    },
-    "expr": {
-        "nb_epochs": 40,
-        "batch_size": 25,
-        "nb_train_samples": 1000000,
-        "nb_test_samples": 10000,
-    },
-    "rpl": {
-        "nb_epochs": 40,
-        "batch_size": 25,
-        "nb_train_samples": 1000000,
-        "nb_test_samples": 10000,
-    },
-    "world": {
-        "nb_epochs": 10,
-        "batch_size": 25,
-        "nb_train_samples": 25000,
-        "nb_test_samples": 1000,
-    },
+default_args = {
+    "model": "37M",
+    "batch_size": 25,
+    "inference_batch_size": 50,
+    "nb_train_samples": 40000,
+    "nb_test_samples": 1000,
 }
 
-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)
 
 ######################################################################
 
@@ -235,6 +182,13 @@ default_model_args = {
         "nb_heads": 2,
         "nb_blocks": 2,
     },
+    "4M": {
+        "dim_model": 256,
+        "dim_keys": 32,
+        "dim_hidden": 1024,
+        "nb_heads": 4,
+        "nb_blocks": 6,
+    },
     "37M": {
         "dim_model": 512,
         "dim_keys": 64,
@@ -267,10 +221,13 @@ else:
 
 ######################################################################
 
-try:
-    os.mkdir(args.result_dir)
-except FileExistsError:
-    if not args.overwrite_results:
+if args.resume:
+    assert os.path.isdir(args.result_dir)
+
+else:
+    try:
+        os.mkdir(args.result_dir)
+    except FileExistsError:
         print(f"result directory {args.result_dir} already exists")
         exit(1)
 
@@ -298,363 +255,1122 @@ def log_string(s):
     sys.stdout.flush()
 
 
+######################################################################
+# Create a time-stamped archive of the source code
+
+with open("this_run.sh", "w") as f:
+    f.write(f"{' '.join(sys.argv)}\n")
+
+now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
+
+os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
+
+######################################################################
+
+log_string(f"argv {' '.join(sys.argv)}")
+
 for n in vars(args):
     log_string(f"args.{n} {getattr(args, n)}")
 
 
 ######################################################################
 
+if args.gpus == "all":
+    gpus_idx = range(torch.cuda.device_count())
+else:
+    gpus_idx = [int(k) for k in args.gpus.split(",")]
 
-def picoclvr_pruner_horizontal_green(p):
-    return not ("green" in p and ("left" in p or "right" in p))
+gpus = [torch.device(f"cuda:{n}") for n in gpus_idx]
 
+if torch.cuda.is_available():
+    main_device = gpus[0]
+else:
+    assert len(gpus) == 0
+    main_device = torch.device("cpu")
 
-picoclvr_pruner_train = (
-    picoclvr_pruner_horizontal_green
-    if args.picocvlr_prune_properties in {"train+eval"}
-    else None
-)
+if args.dirty_debug:
+    args.nb_train_samples = 2500
+    args.nb_test_samples = 100
 
-picoclvr_pruner_eval = (
-    (lambda p: not picoclvr_pruner_horizontal_green(p))
-    if args.picocvlr_prune_properties in {"train+eval", "eval"}
-    else None
+if args.physical_batch_size is None:
+    args.physical_batch_size = args.batch_size
+else:
+    assert args.batch_size % args.physical_batch_size == 0
+
+assert args.nb_train_samples % args.batch_size == 0
+assert args.nb_test_samples % args.batch_size == 0
+
+if args.problem == "sky":
+    problem = sky.Sky(
+        height=args.sky_height,
+        width=args.sky_width,
+        nb_birds=args.sky_nb_birds,
+        nb_iterations=args.sky_nb_iterations,
+        speed=args.sky_speed,
+        max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
+        chunk_size=100,
+        nb_threads=args.nb_threads,
+    )
+
+elif args.problem == "grids":
+    problem = grids.Grids(
+        max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
+        chunk_size=100,
+        nb_threads=args.nb_threads,
+        tasks=args.grids_world_tasks,
+    )
+
+    if args.grids_science_tasks is None:
+        science_w_quizzes = None
+    else:
+        science_problem = grids.Grids(
+            max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
+            chunk_size=100,
+            nb_threads=args.nb_threads,
+            tasks=args.grids_science_tasks,
+        )
+        science_w_quizzes = science_problem.generate_w_quizzes(100)
+
+        if not args.resume:
+            science_problem.save_some_examples(args.result_dir, "science_")
+
+
+else:
+    raise ValueError
+
+if not args.resume:
+    problem.save_some_examples(args.result_dir)
+
+quiz_machine = quiz_machine.QuizMachine(
+    problem=problem,
+    batch_size=args.inference_batch_size,
+    result_dir=args.result_dir,
+    prompt_noise=args.prompt_noise,
+    logger=log_string,
+    device=main_device,
 )
 
 ######################################################################
 
-if args.task == "sandbox":
-    if args.sandbox_level == 0:
-        problem = tasks.ProblemLevel0(
-            nb_sentences=args.sandbox_levels_nb_items,
-            len_prompt=args.sandbox_levels_len_source,
-            len_result=args.sandbox_levels_len_result,
+log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}")
+
+vocabulary_size = quiz_machine.vocabulary_size()
+
+log_string(f"vocabulary_size {vocabulary_size}")
+
+######################################################################
+
+
+def optimizer_to(optim, device):
+    for param in optim.state.values():
+        # Not sure there are any global tensors in the state dict
+        if isinstance(param, torch.Tensor):
+            param.data = param.data.to(device)
+            if param._grad is not None:
+                param._grad.data = param._grad.data.to(device)
+        elif isinstance(param, dict):
+            for subparam in param.values():
+                if isinstance(subparam, torch.Tensor):
+                    subparam.data = subparam.data.to(device)
+                    if subparam._grad is not None:
+                        subparam._grad.data = subparam._grad.data.to(device)
+
+
+######################################################################
+
+
+def run_tests(model, quiz_machine, local_device=main_device):
+    with torch.autograd.no_grad():
+        model.to(local_device).eval()
+        if args.schedule_free:
+            model.optimizer.eval()
+
+        nb_test_samples, acc_test_loss = 0, 0.0
+        nb_samples_accumulated = 0
+
+        full_input, full_mask_loss = quiz_machine.data_input(model, split="test")
+        src = zip(
+            full_input.split(args.batch_size), full_mask_loss.split(args.batch_size)
         )
-    elif args.sandbox_level == 1:
-        problem = tasks.ProblemLevel1(
-            nb_operators=args.sandbox_levels_nb_items,
-            len_source=args.sandbox_levels_len_source,
-            len_result=args.sandbox_levels_len_result,
+
+        for input, mask_loss in tqdm.tqdm(
+            src,
+            dynamic_ncols=True,
+            desc="test",
+            total=full_input.size(0) // args.batch_size,
+        ):
+            input = input.to(local_device)
+            mask_loss = mask_loss.to(local_device)
+            targets = input
+
+            output = model(mygpt.BracketedSequence(input)).x
+            loss_per_token = F.cross_entropy(
+                output.transpose(1, 2), targets, reduction="none"
+            )
+            loss = (loss_per_token * mask_loss).mean()
+            acc_test_loss += loss.item() * input.size(0)
+            nb_test_samples += input.size(0)
+
+        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+
+        log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
+
+        model.main_test_accuracy = quiz_machine.produce_results(
+            n_epoch=n_epoch,
+            model=model,
+            input=full_input[:2000],
+            result_dir=args.result_dir,
         )
-    elif args.sandbox_level == 2:
-        problem = tasks.ProblemLevel2(
-            len_source=args.sandbox_levels_len_source,
-            len_result=args.sandbox_levels_len_result,
+
+
+######################################################################
+
+
+def one_epoch(model, quiz_machine, local_device=main_device):
+    model.to(local_device).train()
+    optimizer_to(model.optimizer, local_device)
+
+    if args.schedule_free:
+        model.optimizer.train()
+
+    nb_train_samples, acc_train_loss = 0, 0.0
+
+    hard_w_quizzes = []
+
+    full_input, full_mask_loss = quiz_machine.data_input(model, split="train")
+    src = zip(full_input.split(args.batch_size), full_mask_loss.split(args.batch_size))
+
+    for input, mask_loss in tqdm.tqdm(
+        src,
+        dynamic_ncols=True,
+        desc="training",
+        total=full_input.size(0) // args.batch_size,
+    ):
+        input = input.to(local_device)
+        mask_loss = mask_loss.to(local_device)
+
+        if nb_train_samples % args.batch_size == 0:
+            model.optimizer.zero_grad()
+
+        targets = input
+
+        output = model(mygpt.BracketedSequence(input)).x
+        loss_per_token = F.cross_entropy(
+            output.transpose(1, 2), targets, reduction="none"
         )
-    else:
-        raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
-
-    task = tasks.SandBox(
-        problem,
-        # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
-        logger=log_string,
-        device=device,
-    )
+        loss = (loss_per_token * mask_loss).mean() + model.loss
+        acc_train_loss += loss.item() * input.size(0)
+
+        loss_per_samples = loss_per_token.detach().flatten(1).mean(dim=1)
+
+        nb_train_samples += input.size(0)
+
+        loss.backward()
+
+        if nb_train_samples % args.batch_size == 0:
+            model.optimizer.step()
+
+    train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+
+    log_string(f"train_perplexity {n_epoch} model {model.id} {train_perplexity}")
+
+    run_tests(model, quiz_machine)
+
+    # threshold = torch.cat([l for _, l in hard_w_quizzes], dim=0).sort().values
+    # threshold = threshold[threshold.size(0) // 2]
+
+    # model.hard_w_quizzes = torch.cat(
+    # [x[l >= threshold] for x, l in hard_w_quizzes], dim=0
+    # )
+
+    model.to(main_device)
+    optimizer_to(model.optimizer, main_device)
+
+
+######################################################################
+
+
+def model_transformer_hot(model):
+    model.temperature = args.temperature_hot
+    # model.set_noise_injection(1.0, ("ffw", args.nb_blocks // 2))
+
 
-elif args.task == "picoclvr":
-    task = tasks.PicoCLVR(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.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,
+def model_transformer_cold(model):
+    model.temperature = args.temperature_cold
+    # pass
+
+
+c_quizzes_procedure = [
+    (("f_B", "f_A", "A", "B"), (1, 0, 0, 0), model_transformer_hot),
+    (("f_B", "f_A", "A", "B"), (0, 1, 1, 1), model_transformer_cold),
+    (("A", "f_A", "B", "f_B"), (0, 0, 0, 1), model_transformer_cold),
+    (("f_A", "A", "f_B", "B"), (0, 0, 0, 1), model_transformer_cold),
+]
+
+######################################################################
+
+
+def save_additional_results(model, models, science_w_quizzes):
+    # Save generated quizzes with the successive steps
+
+    recorder = []
+
+    c_quizzes = quiz_machine.generate_c_quizzes(
+        64,
+        model_for_generation=model,
+        procedure=c_quizzes_procedure,
+        recorder=recorder,
     )
 
-elif args.task == "mnist":
-    task = tasks.MNIST(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
-        device=device,
+    # This is nb_quizzes x nb_models
+
+    seq_logproba = quiz_machine.models_logprobas(
+        models, c_quizzes, ("A", "f_A", "B", "f_B"), (0, 0, 0, 1), (0, 0, 1, 0)
+    ) + quiz_machine.models_logprobas(
+        models, c_quizzes, ("f_A", "A", "f_B", "B"), (0, 0, 0, 1), (0, 0, 1, 0)
     )
 
-elif args.task == "maze":
-    task = tasks.Maze(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
-        height=args.maze_height,
-        width=args.maze_width,
-        nb_walls=args.maze_nb_walls,
-        device=device,
+    probas = seq_logproba.exp()
+
+    comments = []
+
+    for l in seq_logproba:
+        comments.append("proba " + " ".join([f"{x.exp().item():.02f}" for x in l]))
+
+    ##
+
+    c_quizzes = torch.cat([c[:, None, :] for c, _, in recorder], dim=1)
+    predicted_parts = torch.cat([t[:, None, :] for _, t in recorder], dim=1)
+    nb_steps = c_quizzes.size(1)
+    c_quizzes = c_quizzes.reshape(-1, c_quizzes.size(-1))
+    predicted_parts = predicted_parts.reshape(-1, predicted_parts.size(-1))
+
+    # We have comments only for the final quiz, not the successive
+    # steps, so we have to add nb_steps-1 empty comments
+
+    steps_comments = []
+    for c in comments:
+        steps_comments += [""] * (nb_steps - 1) + [c]
+
+    filename = f"non_validated_{n_epoch:04d}_{model.id:02d}.png"
+
+    quiz_machine.problem.save_quizzes_as_image(
+        args.result_dir,
+        filename,
+        quizzes=c_quizzes,
+        predicted_parts=predicted_parts,
+        comments=steps_comments,
+        nrow=nb_steps * 2,  # two quiz per row
     )
 
-elif args.task == "snake":
-    task = tasks.Snake(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.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,
+    log_string(f"wrote {filename}")
+
+    ######################################################################
+
+    if science_w_quizzes is not None:
+        struct = ("A", "f_A", "B", "f_B")
+        mask = (0, 0, 0, 1)
+        result, correct = quiz_machine.predict(
+            model=model,
+            quizzes=science_w_quizzes.to(main_device),
+            struct=struct,
+            mask=mask,
+        )
+
+        predicted_parts = torch.tensor(mask, device=correct.device)[None, :].expand(
+            correct.size(0), -1
+        )
+        correct = (2 * correct - 1) * (predicted_parts.sum(dim=-1) == 1).long()
+
+        nb_correct = (correct == 1).long().sum()
+        nb_total = (correct != 0).long().sum()
+
+        log_string(
+            f"science_accuracy {n_epoch} model {model.id} val {nb_correct} / {nb_total}"
+        )
+
+        i = correct == 1
+        j = correct != 1
+
+        result = torch.cat([result[i], result[j]], dim=0)
+        correct = torch.cat([correct[i], correct[j]], dim=0)
+        correct_parts = predicted_parts * correct[:, None]
+
+        result = result[:128]
+        predicted_parts = predicted_parts[:128]
+        correct_parts = correct_parts[:128]
+
+        quiz_machine.problem.save_quizzes_as_image(
+            args.result_dir,
+            f"culture_science_{n_epoch:04d}_{model.id:02d}.png",
+            quizzes=result,
+            predicted_parts=predicted_parts,
+            correct_parts=correct_parts,
+        )
+
+
+######################################################################
+
+
+def record_new_c_quizzes(models, quiz_machine, nb_for_train=1000, nb_for_test=100):
+    nb_to_validate = nb_for_train + nb_for_test
+    nb_to_generate_per_iteration = max(args.physical_batch_size, nb_to_validate)
+    nb_validated = 0
+
+    recorded_validated = []
+
+    start_time = time.perf_counter()
+
+    nb_validated_per_model = torch.zeros(len(models), dtype=torch.int64)
+
+    while nb_validated_per_model.sum() < nb_to_validate:
+        # We use the model that has generated the fewest quizzes to
+        # balance the number of quizzes per model overall
+
+        # model_for_generation = sorted(
+        # models, key=lambda m: nb_validated_per_model[m.id]
+        # )[0]
+
+        model_for_generation = models[torch.randint(len(models), (1,)).item()]
+
+        # We generate quizzes with a procedure that injects some
+        # structured noise
+
+        c_quizzes = quiz_machine.generate_c_quizzes(
+            nb_to_generate_per_iteration,
+            model_for_generation=model,
+            procedure=c_quizzes_procedure,
+        )
+
+        # We discard the trivial ones, according to a criterion
+        # specific to the world quizzes (e.g. B=f(B))
+
+        to_keep = quiz_machine.problem.trivial(c_quizzes) == False
+
+        c_quizzes = c_quizzes[to_keep]
+
+        # This is nb_quizzes x nb_models
+
+        seq_logproba = quiz_machine.models_logprobas(
+            models, c_quizzes, ("A", "f_A", "B", "f_B"), (0, 0, 0, 1), (0, 0, 1, 0)
+        ) + quiz_machine.models_logprobas(
+            models, c_quizzes, ("f_A", "A", "f_B", "B"), (0, 0, 0, 1), (0, 0, 1, 0)
+        )
+
+        probas = seq_logproba.exp()
+
+        nb_succeed = (probas >= args.proba_understands).long().sum(dim=1)
+        nb_fail = (probas <= args.proba_not_understands).long().sum(dim=1)
+
+        to_keep = (
+            (nb_succeed + nb_fail == probas.size(1))
+            & (nb_fail >= 1)
+            & (nb_fail <= args.max_fail_to_validate)
+        )
+
+        c_quizzes = c_quizzes[to_keep]
+
+        if c_quizzes.size(0) > 0:
+            nb_validated_per_model[model_for_generation.id] += c_quizzes.size(0)
+            recorded_validated.append(c_quizzes)
+            nb_validated = c_quizzes.size(0)
+        else:
+            nb_validated = 0
+
+        total_nb_validated = nb_validated_per_model.sum().item()
+
+        duration = time.perf_counter() - start_time
+
+        if total_nb_validated > 0:
+            if total_nb_validated < nb_to_validate:
+                d = (
+                    (nb_to_validate - total_nb_validated)
+                    * duration
+                    / total_nb_validated
+                )
+                e = (datetime.datetime.now() + datetime.timedelta(seconds=d)).strftime(
+                    "%a %H:%M"
+                )
+            else:
+                e = "now!"
+        else:
+            e = "???"
+
+        log_string(
+            f"keep c_quizzes model {model_for_generation.id} validated {nb_validated} / {nb_to_generate_per_iteration} ({100*nb_validated/nb_to_generate_per_iteration:.02f}%) nb_accumulated {total_nb_validated} / {nb_to_validate} (finishes {e} -- {int((total_nb_validated * 3600)/duration)}/h)"
+        )
+
+    validated_quizzes = torch.cat(recorded_validated, dim=0)
+
+    ######################################################################
+    # store the new c_quizzes which have been validated
+
+    v_train = validated_quizzes[:nb_for_train]
+    quiz_machine.store_c_quizzes(v_train, for_train=True)
+
+    v_test = validated_quizzes[nb_for_train:nb_to_validate]
+    quiz_machine.store_c_quizzes(v_test, for_train=False)
+
+    ######################################################################
+    # save images
+
+    vq = validated_quizzes[torch.randperm(validated_quizzes.size(0))[:128]]
+
+    if vq.size(0) > 0:
+        seq_logproba = quiz_machine.models_logprobas(
+            models, vq, ("A", "f_A", "B", "f_B"), (0, 0, 0, 1), (0, 0, 1, 0)
+        ) + quiz_machine.models_logprobas(
+            models, vq, ("f_A", "A", "f_B", "B"), (0, 0, 0, 1), (0, 0, 1, 0)
+        )
+
+        probas = seq_logproba.exp()
+
+        comments = []
+
+        for l in seq_logproba:
+            comments.append("proba " + " ".join([f"{x.exp().item():.02f}" for x in l]))
+
+        filename = f"culture_c_quiz_{n_epoch:04d}.png"
+        quiz_machine.problem.save_quizzes_as_image(
+            args.result_dir, filename, vq, comments=comments
+        )
+
+
+######################################################################
+
+# The generator is very similar to a "solving GPT" except that it
+# deals with quizzes prologued with one token per solving GPT that
+# indicates if the said model solves it or not.
+#
+# There are three levels of solving 0->proba<=proba_not_understands,
+# 2->proba>=proba_understands and 1 otherwise.
+
+
+def generate_c_quizzes_with_generator(generator, quiz_machine, nb):
+    generator.to(main_device)
+
+    struct = ("A", "f_A", "B", "f_B")
+
+    c_quizzes = quiz_machine.problem.create_empty_quizzes(nb, struct=struct)
+    ar_mask = quiz_machine.make_quiz_mask(c_quizzes, struct, (1, 1, 1, 1))
+
+    i = F.one_hot(
+        torch.randint(args.nb_gpts, (c_quizzes.size(0),)),
+        num_classes=args.nb_gpts,
     )
 
-elif args.task == "stack":
-    task = tasks.Stack(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.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,
+    prologs_c_quizzes = token_prolog_0 * i + token_prolog_2 * (1 - i)
+    prologs_ar_mask = ar_mask.new_zeros(ar_mask.size(0), prologs_c_quizzes.size(1))
+
+    prologued_c_quizzes = torch.cat([prologs_c_quizzes, c_quizzes], dim=1).to(
+        main_device
     )
+    prologued_ar_mask = torch.cat([prologs_ar_mask, ar_mask], dim=1).to(main_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.batch_size,
-        device=device,
+    seq_logproba = torch.zeros(
+        prologued_c_quizzes.size(0), device=prologued_c_quizzes.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.batch_size,
-        device=device,
+    generator.temperature = args.temperature_hot
+
+    with torch.autograd.no_grad():
+        t = generator.training
+        generator.eval()
+
+        one_batch_masked_inplace_autoregression(
+            generator,
+            prologued_c_quizzes,
+            prologued_ar_mask,
+            seq_logproba,
+            deterministic_synthesis=False,
+        )
+
+        generator.train(t)
+
+    generator.reset_transformations()
+
+    prologued_c_quizzes = (
+        prologued_c_quizzes * (prologued_c_quizzes < vocabulary_size).long()
     )
 
-elif args.task == "world":
-    task = tasks.World(
-        nb_train_samples=args.nb_train_samples,
-        nb_test_samples=args.nb_test_samples,
-        batch_size=args.batch_size,
-        vqae_nb_epochs=args.world_vqae_nb_epochs,
-        logger=log_string,
-        device=device,
+    c_quizzes = prologued_c_quizzes[:, prologs_c_quizzes.size(1) :]
+
+    return c_quizzes.to("cpu"), prologs_c_quizzes.to("cpu")
+
+
+def batches_for_generator(generator, quiz_machine, models, fraction_w_quizzes=1.0):
+    samples = []
+
+    for _ in range(args.nb_train_samples // args.batch_size):
+        while sum([x.size(0) for x in samples]) < args.batch_size:
+            # Generate a bunch of quizzes
+
+            if torch.rand(1).item() <= fraction_w_quizzes:
+                # Either we start with the world quizzes
+                c_quizzes = quiz_machine.problem.generate_w_quizzes(
+                    args.batch_size, progress_bar=False
+                )
+            else:
+                # Or we use the generator itself to generate them
+                c_quizzes, _ = generate_c_quizzes_with_generator(
+                    generator, quiz_machine, args.batch_size
+                )
+
+            # We remove the trivial ones
+            to_keep = quiz_machine.problem.trivial(c_quizzes) == False
+            c_quizzes = c_quizzes[to_keep]
+
+            # If there are remaining ones, we compute the true prolog
+            # that indicates how the GPTs solve it
+
+            if c_quizzes.size(0) > 0:
+                seq_logproba = quiz_machine.models_logprobas(
+                    models,
+                    c_quizzes,
+                    ("A", "f_A", "B", "f_B"),
+                    (0, 0, 0, 1),
+                    (0, 0, 1, 0),
+                ) + quiz_machine.models_logprobas(
+                    models,
+                    c_quizzes,
+                    ("f_A", "A", "f_B", "B"),
+                    (0, 0, 0, 1),
+                    (0, 0, 1, 0),
+                )
+
+                probas = seq_logproba.exp()
+
+                u0 = probas <= args.proba_not_understands
+                u2 = probas >= args.proba_understands
+                u1 = (u0 | u2) == False
+
+                prologs = (
+                    (u0.long() * token_prolog_0)
+                    + (u1.long() * token_prolog_1)
+                    + (u2.long() * token_prolog_2)
+                )
+
+                prologued_c_quizzes = torch.cat([prologs, c_quizzes], dim=1)
+
+                # nb_u2 = u2.long().sum(dim=1)
+                # nb_u0 = u0.long().sum(dim=1)
+                # prologued_c_quizzes = prologued_c_quizzes[(nb_u2 >= 1) & (nb_u0 >= 1)]
+
+                if prologued_c_quizzes.size(0) > 0:
+                    samples.append(prologued_c_quizzes)
+
+        # Now we yield a batch
+
+        x = torch.cat(samples, dim=0)
+        samples = [x[args.batch_size :]]
+
+        yield x[: args.batch_size]
+
+
+def one_generator_epoch(
+    generator, quiz_machine, models, fraction_w_quizzes, local_device=main_device
+):
+    model.to(local_device).train()
+
+    optimizer = torch.optim.Adam(generator.parameters(), lr=args.learning_rate)
+
+    nb_train_samples, acc_train_loss = 0, 0.0
+
+    src = batches_for_generator(
+        generator=generator,
+        quiz_machine=quiz_machine,
+        models=models,
+        fraction_w_quizzes=fraction_w_quizzes,
     )
 
-else:
-    raise ValueError(f"Unknown task {args.task}")
+    for input in tqdm.tqdm(
+        src,
+        dynamic_ncols=True,
+        desc="training",
+        total=args.nb_train_samples // args.batch_size,
+    ):
+        input = input.to(local_device)
 
-######################################################################
+        if nb_train_samples % args.batch_size == 0:
+            optimizer.zero_grad()
 
-log_string(f"device {device}")
+        targets = input
 
-vocabulary_size = task.vocabulary_size()
+        output = generator(mygpt.BracketedSequence(input)).x
+        loss = F.cross_entropy(output.transpose(1, 2), targets)
+        acc_train_loss += loss.item() * input.size(0)
+        nb_train_samples += input.size(0)
 
-log_string(f"vocabulary_size {vocabulary_size}")
+        loss.backward()
 
-##############################
-
-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,
-)
+        if nb_train_samples % args.batch_size == 0:
+            optimizer.step()
 
-model.to(device)
+    train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+
+    log_string(f"train_perplexity {n_epoch} generator - {train_perplexity}")
+
+    generator.to(main_device)
 
-nb_parameters = sum(p.numel() for p in model.parameters())
-log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 
 ######################################################################
 
-nb_epochs_finished = 0
 
-if args.no_checkpoint:
-    log_string(f"not trying to load checkpoint.")
+def train_complexifier(model_gen, model_pred1, model_pred2):
+    samples = []
+    perf = []
 
-else:
-    try:
-        checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
-        checkpoint = torch.load(checkpoint_name)
-        nb_epochs_finished = checkpoint["nb_epochs_finished"]
-        model.load_state_dict(checkpoint["model_state"])
-        torch.set_rng_state(checkpoint["rng_state"])
-        if torch.cuda.is_available():
-            torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
+    optimizer = torch.optim.Adam(model_gen.parameters(), lr=args.learning_rate)
 
-        log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
+    nb_train_samples, acc_train_loss = 0, 0.0
 
-    except FileNotFoundError:
-        log_string("starting from scratch.")
+    for n_epoch in range(args.nb_epochs):
+        for b in range(args.nb_train_samples // args.batch_size):
+            while sum([x.size(0) for x in samples]) < args.batch_size:
+                c_quizzes = quiz_machine.generate_c_quizzes(
+                    args.inference_batch_size,
+                    model_for_generation=model_gen,
+                    procedure=c_quizzes_procedure,
+                )
+                to_keep = quiz_machine.problem.trivial(c_quizzes) == False
+                c_quizzes = c_quizzes[to_keep]
+                if c_quizzes.size(0) > 0:
+                    seq_logproba = quiz_machine.models_logprobas(
+                        [model_pred1, model_pred2],
+                        c_quizzes,
+                        ("A", "f_A", "B", "f_B"),
+                        (0, 0, 0, 1),
+                    ) + quiz_machine.models_logprobas(
+                        [model_pred1, model_pred2],
+                        c_quizzes,
+                        ("f_A", "A", "f_B", "B"),
+                        (0, 0, 0, 1),
+                    )
+                    probas = seq_logproba.exp()
+                    to_keep = (probas[:, model_pred1.id] >= args.proba_understands) & (
+                        probas[:, model_pred2.id] <= args.proba_not_understands
+                    )
+                    log_string(
+                        f"generating {to_keep.long().sum()} / {c_quizzes.size(0)}"
+                    )
+                    c_quizzes = c_quizzes[to_keep]
+                    if c_quizzes.size(0):
+                        samples.append(c_quizzes)
+
+            log_string(f"full batch {sum([x.size(0) for x in samples])}")
+
+            x = torch.cat(samples, dim=0)
+
+            input = x[: args.batch_size]
+            samples = [x[args.batch_size :]]
+
+            # -------------------
+
+            seq_logproba = quiz_machine.models_logprobas(
+                [model_pred1, model_pred2],
+                input,
+                ("A", "f_A", "B", "f_B"),
+                (0, 0, 0, 1),
+            ) + quiz_machine.models_logprobas(
+                [model_pred1, model_pred2],
+                input,
+                ("f_A", "A", "f_B", "B"),
+                (0, 0, 0, 1),
+            )
+
+            comments = []
+
+            for l in seq_logproba:
+                comments.append(
+                    f"proba {l[model_pred1.id].exp().item():.02f} {l[model_pred2.id].exp().item():.02f}"
+                )
+
+            filename = f"batch_{n_epoch:04d}_{b:04d}.png"
+            quiz_machine.problem.save_quizzes_as_image(
+                args.result_dir, filename, input, comments=comments
+            )
+            log_string(f"wrote {filename}")
+
+            # ------------------------
+
+            input = input.to(main_device)
+
+            if nb_train_samples % args.batch_size == 0:
+                optimizer.zero_grad()
+
+            output = model_gen(mygpt.BracketedSequence(input)).x
+            loss = F.cross_entropy(output.transpose(1, 2), input)
+            acc_train_loss += loss.item() * input.size(0)
+            nb_train_samples += input.size(0)
+
+            loss.backward()
+
+            if nb_train_samples % args.batch_size == 0:
+                optimizer.step()
+
+        train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+
+        log_string(f"train_perplexity {n_epoch} model ae {train_perplexity}")
 
-    except:
-        log_string("error when loading the checkpoint.")
-        exit(1)
 
 ######################################################################
 
-if args.task == "expr" and args.expr_input_file is not None:
-    task.produce_results(
-        nb_epochs_finished,
-        model,
-        args.result_dir,
-        log_string,
-        args.deterministic_synthesis,
-        args.expr_input_file,
+models = []
+
+
+def compute_causal_attzero(t_q, t_k):
+    return t_q < t_k
+
+
+if args.schedule_free:
+    import schedulefree
+
+for k in range(args.nb_gpts):
+    log_string(f"creating model {k} and its w_quizzes")
+
+    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,
+        compute_attzero=compute_causal_attzero,
+        dropout=args.dropout,
+    ).to(main_device)
+
+    model.id = k
+
+    if args.schedule_free:
+        model.optimizer = schedulefree.AdamWScheduleFree(
+            model.parameters(), lr=args.learning_rate
+        )
+    else:
+        model.optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
+
+    model.main_test_accuracy = 0.0
+
+    model.train_w_quizzes = quiz_machine.problem.generate_w_quizzes(
+        args.nb_train_samples
     )
 
-    exit(0)
+    model.test_w_quizzes = quiz_machine.problem.generate_w_quizzes(args.nb_test_samples)
+
+    models.append(model)
+
+######################################################################
+
+if args.test == "quant":
+    nb_bits = 8
+    for model in models:
+        model.trunk.insert(
+            12,
+            mygpt.CacheWrapper(
+                mygpt.RandomBypass(
+                    nn.Sequential(
+                        nn.Linear(args.dim_model, nb_bits),
+                        mygpt.BSQ(nb_bits),
+                        nn.Linear(nb_bits, args.dim_model),
+                    ),
+                    0.1,
+                )
+            ),
+        )
+
+        print(model)
+        exit(0)
+
 
 ######################################################################
 
-nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
+current_epoch = 0
 
-# Compute the entropy of the training tokens
+if args.resume:
+    for model in models:
+        filename = f"gpt_{model.id:03d}.pth"
 
-token_count = 0
-for input in task.batches(split="train"):
-    token_count += F.one_hot(input, num_classes=task.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)
+        try:
+            d = torch.load(os.path.join(args.result_dir, filename))
+            model.load_state_dict(d["state_dict"])
+            model.optimizer.load_state_dict(d["optimizer_state_dict"])
+            model.main_test_accuracy = d["main_test_accuracy"]
+            log_string(f"successfully loaded {filename}")
+        except FileNotFoundError:
+            log_string(f"cannot find {filename}")
+            pass
 
-##############################
+    try:
+        filename = "c_quizzes.pth"
+        quiz_machine.load_c_quizzes(os.path.join(args.result_dir, filename))
+        log_string(f"successfully loaded {filename}")
+    except FileNotFoundError:
+        log_string(f"cannot find {filename}")
+        pass
 
-# A bit of paranoia never hurts
+    try:
+        filename = "state.pth"
+        state = torch.load(os.path.join(args.result_dir, filename))
+        log_string(f"successfully loaded {filename}")
+        current_epoch = state["current_epoch"]
+    except FileNotFoundError:
+        log_string(f"cannot find {filename}")
+        pass
 
-train_examples = {}
+######################################################################
 
+nb_parameters = sum(p.numel() for p in models[0].parameters())
+log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 
-for input in task.batches(split="train"):
-    assert input.dim() == 2 and input.dtype == torch.int64
-    for x in input:
-        train_examples[x.sum().item()] = x
+######################################################################
 
-nb_total, nb_collisions = 0, 0
-for input in task.batches(split="test"):
-    assert input.dim() == 2 and input.dtype == torch.int64
-    for x in input:
-        nb_total += 1
-        y = train_examples.get(x.sum().item())
-        if y is not None:
-            if x.size() == y.size() and (x - y).abs().sum() == 0:
-                nb_collisions += 1
+if args.nb_new_c_quizzes_for_train is None:
+    args.nb_new_c_quizzes_for_train = args.nb_train_samples // 100
 
-del train_examples
+if args.nb_new_c_quizzes_for_test is None:
+    args.nb_new_c_quizzes_for_test = args.nb_test_samples // 100
 
 log_string(
-    f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
+    f"nb_new_c_quizzes_for_train {args.nb_new_c_quizzes_for_train} nb_new_c_quizzes_for_test {args.nb_new_c_quizzes_for_test}"
 )
 
-##############################
+######################################################################
+
+if args.dirty_debug:
+    args.accuracy_to_make_c_quizzes = 0.0
+    args.nb_gpts = 2
+    args.nb_new_c_quizzes_for_train = 100
+    args.nb_new_c_quizzes_for_test = 10
+
+######################################################################
+
+if args.test == "tsne":
+    model = models[0]
+
+    quizzes = []
+    labels = []
+    nb_samples_per_task = 1000
+
+    for n, t in enumerate(args.grids_world_tasks.split(",")):
+        quizzes.append(
+            quiz_machine.problem.generate_w_quizzes(nb_samples_per_task, [t])
+        )
+        labels.append(torch.full((quizzes[-1].size(0),), n))
+
+    quizzes = torch.cat(quizzes, dim=0)
+    labels = torch.cat(labels, dim=0)
+
+    with torch.autograd.no_grad():
+        model.eval().to(main_device)
+        record = []
+        for input, targets in zip(
+            quizzes.split(args.batch_size), labels.split(args.batch_size)
+        ):
+            input = input.to(main_device)
+            bs = mygpt.BracketedSequence(input)
+            bs = mygpt.BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb)
+            bs = model.embedding(bs)
+            bs = model.trunk[args.nb_blocks // 2](bs)
+            record.append((bs.x.to("cpu"), targets))
+
+    x = torch.cat([x for x, y in record], dim=0).flatten(1)
+    y = torch.cat([y for x, y in record], dim=0)
+
+    print(f"{x.size()=} {y.size()=}")
+    # torch.save((x,y), "/tmp/embed.pth")
+    # exit(0)
+
+    from sklearn.manifold import TSNE
+
+    x_np = x.numpy()
+    z_np = TSNE(n_components=2, perplexity=50).fit_transform(x_np)
+    z = torch.from_numpy(z_np)
+
+    print(f"{z.size()=}")
+
+    with open("/tmp/result.dat", "w") as f:
+        for k in range(z.size(0)):
+            f.write(f"{y[k]} {z[k,0]} {z[k,1]}\n")
+
+    exit(0)
+
+######################################################################
+
+if args.test == "generator":
+    token_prolog_0 = vocabulary_size + 0
+    token_prolog_1 = vocabulary_size + 1
+    token_prolog_2 = vocabulary_size + 2
+    generator_vocabulary_size = vocabulary_size + 3
+
+    generator = mygpt.MyGPT(
+        vocabulary_size=generator_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,
+        compute_attzero=compute_causal_attzero,
+        dropout=args.dropout,
+    ).to(main_device)
+
+    generator.main_test_accuracy = 0.0
+
+    filename = f"generator.pth"
+
+    try:
+        d = torch.load(os.path.join(args.result_dir, filename))
+        generator.load_state_dict(d[0])
+        generator.main_test_accuracy = d[1]
+        log_string(f"successfully loaded {filename}")
+    except FileNotFoundError:
+        log_string(f"cannot find {filename}")
+        pass
 
-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
+        one_generator_epoch(
+            generator,
+            quiz_machine=quiz_machine,
+            models=models,
+            fraction_w_quizzes=1 if n_epoch < 25 else 0.5,
+            local_device=main_device,
+        )
 
-log_string(f"learning_rate_schedule {learning_rate_schedule}")
+        filename = f"generator.pth"
+        torch.save(
+            (generator.state_dict(), generator.main_test_accuracy),
+            os.path.join(args.result_dir, filename),
+        )
+        log_string(f"wrote {filename}")
 
-##############################
+        c_quizzes, prologs = generate_c_quizzes_with_generator(
+            generator, quiz_machine, args.batch_size
+        )
 
-nb_samples_seen = 0
+        seq_logproba = quiz_machine.models_logprobas(
+            models, c_quizzes, ("A", "f_A", "B", "f_B"), (0, 0, 0, 1), (0, 0, 1, 0)
+        ) + quiz_machine.models_logprobas(
+            models, c_quizzes, ("f_A", "A", "f_B", "B"), (0, 0, 0, 1), (0, 0, 1, 0)
+        )
 
-if nb_epochs_finished >= nb_epochs:
-    task.produce_results(
-        nb_epochs_finished,
-        model,
-        args.result_dir,
-        log_string,
-        args.deterministic_synthesis,
-    )
+        probas = seq_logproba.exp()
 
-for n_epoch in range(nb_epochs_finished, nb_epochs):
-    learning_rate = learning_rate_schedule[n_epoch]
+        u0 = probas <= args.proba_not_understands
+        u2 = probas >= args.proba_understands
+        u1 = (u0 | u2) == False
 
-    log_string(f"learning_rate {learning_rate}")
+        predicted_prologs = (
+            (u0.long() * token_prolog_0)
+            + (u1.long() * token_prolog_1)
+            + (u2.long() * token_prolog_2)
+        )
 
-    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}.")
+        comments = []
 
-    model.train()
+        nb_errors = (predicted_prologs != prologs).long().sum()
+        nb_total = prologs.numel()
 
-    nb_train_samples, acc_train_loss = 0, 0.0
+        log_string(f"generator_error {nb_errors} / {nb_total}")
 
-    for input in task.batches(split="train"):
-        input = input.to(device)
-        output = model(mygpt.BracketedSequence(input)).x
-        loss = F.cross_entropy(output.transpose(1, 2), input)
-        acc_train_loss += loss.item() * input.size(0)
-        nb_train_samples += input.size(0)
-        nb_samples_seen += input.size(0)
+        def readable(prologs):
+            return (prologs == token_prolog_1) + 2 * (prologs == token_prolog_2)
 
-        optimizer.zero_grad()
-        loss.backward()
-        optimizer.step()
+        for aa, ee, ff in zip(probas, readable(predicted_prologs), readable(prologs)):
+            sa = "prolog " + " ".join(
+                [f"{e.item()}/{f.item()}" for e, f in zip(ee, ff)]
+            )
+            sp = "proba " + " ".join([f"{p.item():.02f}" for p in aa])
+            comments.append(sa + "\n" + sp)
 
-    with torch.autograd.no_grad():
-        model.eval()
+        filename = f"generator_batch_{n_epoch:04d}.png"
+        quiz_machine.problem.save_quizzes_as_image(
+            args.result_dir, filename, c_quizzes, comments=comments
+        )
+        log_string(f"wrote {filename}")
 
-        nb_test_samples, acc_test_loss = 0, 0.0
+    exit(0)
 
-        for input in task.batches(split="test"):
-            input = input.to(device)
+######################################################################
 
-            output = model(mygpt.BracketedSequence(input)).x
-            loss = F.cross_entropy(output.transpose(1, 2), input)
-            acc_test_loss += loss.item() * input.size(0)
-            nb_test_samples += input.size(0)
+for n_epoch in range(current_epoch, args.nb_epochs):
+    state = {"current_epoch": n_epoch}
+    filename = "state.pth"
+    torch.save(state, os.path.join(args.result_dir, filename))
+    log_string(f"wrote {filename}")
 
-        train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
-        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+    log_string(f"--- epoch {n_epoch} ----------------------------------------")
 
-        log_string(
-            f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
+    cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
+    log_string(f"current_test_accuracies {cta}")
+
+    ##################################################
+    # If all the models are good enough, generate new quizzes and
+    # re-compute the test errors
+
+    if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
+        record_new_c_quizzes(
+            models,
+            quiz_machine,
+            nb_for_train=args.nb_new_c_quizzes_for_train,
+            nb_for_test=args.nb_new_c_quizzes_for_test,
         )
 
-        task.produce_results(
-            n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
+        filename = "c_quizzes.pth"
+        quiz_machine.save_c_quizzes(os.path.join(args.result_dir, filename))
+        log_string(f"wrote {filename}")
+
+        # Force one epoch of training
+        for model in models:
+            model.main_test_accuracy = 0.0
+
+    ##################################################
+    # Select, improve, and eval the worst model(s)
+
+    ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
+
+    weakest_models = ranked_models[: len(gpus)]
+
+    threads = []
+
+    for gpu, model in zip(gpus, weakest_models):
+        log_string(f"training model {model.id}")
+
+        t = threading.Thread(
+            target=one_epoch, daemon=True, args=(model, quiz_machine, gpu)
         )
 
-    checkpoint = {
-        "nb_epochs_finished": n_epoch + 1,
-        "model_state": model.state_dict(),
-        "rng_state": torch.get_rng_state(),
-    }
+        threads.append(t)
 
-    if torch.cuda.is_available():
-        checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
+        t.start()
+
+    for t in threads:
+        t.join()
+
+    # Save the models to disk
+
+    for model in weakest_models:
+        filename = f"gpt_{model.id:03d}.pth"
+        torch.save(
+            {
+                "state_dict": model.state_dict(),
+                "optimizer_state_dict": model.optimizer.state_dict(),
+                "main_test_accuracy": model.main_test_accuracy,
+            },
+            os.path.join(args.result_dir, filename),
+        )
+        log_string(f"wrote {filename}")
+
+    for model in weakest_models:
+        save_additional_results(model, models, science_w_quizzes)
+
+    ######################################################################
+
+    # Renew the training samples
+
+    for model in weakest_models:
+        quiz_machine.renew_train_w_quizzes(model=model)
 
-    checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
-    torch.save(checkpoint, checkpoint_name)
-    log_string(f"saved checkpoint {checkpoint_name}")
+    if args.log_command is not None:
+        s = args.log_command.split()
+        s.insert(1, args.result_dir)
+        subprocess.run(s)
 
 ######################################################################