X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=sidebyside;f=main.py;h=6b00bbfd991178841468b75006f94121668c2b4f;hb=refs%2Fheads%2Fmaster;hp=ebecad8a471353400fc7e6e472bdb95f594b48ab;hpb=15192743a5dee8d88650319d64610f1603d21472;p=culture.git

diff --git a/main.py b/main.py
index ebecad8..5dceefc 100755
--- a/main.py
+++ b/main.py
@@ -5,63 +5,67 @@
 
 # Written by Francois Fleuret <francois@fleuret.org>
 
-import math, sys, argparse, time, tqdm, os, datetime, warnings
+import math, sys, argparse, time, tqdm, os, datetime, warnings, copy
 
 import torch, torchvision
 from torch import nn
 from torch.nn import functional as F
 
-import ffutils
-import mygpt, tasks
+import ffutils, grids, attae
 
-# world quizzes vs. culture quizzes
+import threading, subprocess
 
-######################################################################
+# import torch.multiprocessing as mp
 
-accuracy_to_make_c_quizzes = 0.975
-nb_new_c_quizzes_for_train = 1000
-nb_new_c_quizzes_for_test = 100
+torch.set_float32_matmul_precision("high")
 
-######################################################################
-
-if torch.cuda.is_available():
-    device = torch.device("cuda")
-    torch.backends.cuda.matmul.allow_tf32 = True
-else:
-    device = torch.device("cpu")
+# torch.set_default_dtype(torch.bfloat16)
 
 ######################################################################
 
 parser = argparse.ArgumentParser(
-    description="An implementation of GPT with cache.",
     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 )
 
-parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
+parser.add_argument("--log_filename", type=str, default="train.log")
 
 parser.add_argument("--result_dir", type=str, default=None)
 
 parser.add_argument("--seed", type=int, default=0)
 
-parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
+parser.add_argument("--resume", action="store_true", default=False)
 
-########################################
+# ----------------------------------
 
 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=25)
+
+parser.add_argument("--train_batch_size", type=int, default=None)
+
+parser.add_argument("--eval_batch_size", type=int, default=25)
 
-parser.add_argument("--physical_batch_size", type=int, default=None)
+parser.add_argument("--nb_train_samples", type=int, default=50000)
 
-parser.add_argument("--nb_train_samples", type=int, default=None)
+parser.add_argument("--nb_test_samples", type=int, default=2500)
 
-parser.add_argument("--nb_test_samples", type=int, default=None)
+parser.add_argument("--nb_c_quizzes", type=int, default=5000)
 
-parser.add_argument("--learning_rate", type=float, default=1e-4)
+parser.add_argument("--c_quiz_multiplier", type=int, default=1)
 
-########################################
+parser.add_argument("--learning_rate", type=float, default=5e-4)
 
-parser.add_argument("--model", type=str, default=None)
+parser.add_argument("--nb_have_to_be_correct", type=int, default=3)
+
+parser.add_argument("--nb_have_to_be_wrong", type=int, default=1)
+
+parser.add_argument("--nb_mistakes_to_be_wrong", type=int, default=5)
+
+# ----------------------------------
+
+parser.add_argument("--model_type", type=str, default="standard")
+
+parser.add_argument("--model", type=str, default="37M")
 
 parser.add_argument("--dim_model", type=int, default=None)
 
@@ -73,42 +77,49 @@ parser.add_argument("--nb_heads", type=int, default=None)
 
 parser.add_argument("--nb_blocks", type=int, default=None)
 
-parser.add_argument("--dropout", type=float, default=0.1)
+parser.add_argument("--dropout", type=float, default=0.5)
 
-########################################
+# ----------------------------------
 
-parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
+parser.add_argument("--nb_threads", type=int, default=1)
 
-parser.add_argument("--nb_gpts", type=int, default=5)
+parser.add_argument("--gpus", type=str, default="all")
 
-parser.add_argument("--dirty_debug", action="store_true", default=False)
+# ----------------------------------
 
-######################################################################
+parser.add_argument("--nb_models", type=int, default=5)
 
-args = parser.parse_args()
+parser.add_argument("--diffusion_nb_iterations", type=int, default=25)
 
-if args.result_dir is None:
-    args.result_dir = f"results_culture"
+parser.add_argument("--diffusion_proba_corruption", type=float, default=0.05)
+
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.95)
+
+parser.add_argument("--proba_prompt_noise", type=float, default=0.05)
+
+parser.add_argument("--proba_hint", type=float, default=0.25)
+
+parser.add_argument("--quizzes", type=str, default=None)
 
 ######################################################################
 
-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
+grids_tasks = ", ".join(
+    [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
+)
+
+parser.add_argument(
+    "--grids_world_tasks",
+    type=str,
+    default="replace_color,translate,grow,frame",
+    help="A comma-separated subset of: " + grids_tasks + ".",
+)
 
 ######################################################################
 
-default_args = {
-    "model": "37M",
-    "batch_size": 100,
-    "nb_train_samples": 250000,
-    "nb_test_samples": 10000,
-}
+args = parser.parse_args()
 
-for k, v in default_args.items():
-    if getattr(args, k) is None:
-        setattr(args, k, v)
+if args.result_dir is None:
+    args.result_dir = f"results_culture"
 
 ######################################################################
 
@@ -159,11 +170,16 @@ else:
 
 ######################################################################
 
-try:
-    os.mkdir(args.result_dir)
-except FileExistsError:
-    print(f"result directory {args.result_dir} already exists")
-    exit(1)
+if args.resume:
+    if not os.path.isdir(args.result_dir):
+        print(f"Trying to resume from a non-existing result dir {args.result_dir}.")
+        exit(1)
+else:
+    try:
+        os.mkdir(args.result_dir)
+    except FileExistsError:
+        print(f"result directory {args.result_dir} already exists")
+        exit(1)
 
 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 
@@ -179,6 +195,9 @@ if args.seed >= 0:
 
 
 def log_string(s):
+    """print the given string prefixed with a time stamps, and log it
+    into log_file is not None"""
+
     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 
     if log_file is not None:
@@ -189,6 +208,18 @@ 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):
@@ -197,291 +228,773 @@ for n in vars(args):
 
 ######################################################################
 
-if args.dirty_debug:
-    args.nb_train_samples = 2500
-    args.nb_test_samples = 100
+if args.gpus == "all":
+    gpus_idx = range(torch.cuda.device_count())
+else:
+    gpus_idx = [int(k) for k in args.gpus.split(",")]
+
+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")
 
-if args.physical_batch_size is None:
-    args.physical_batch_size = args.batch_size
+if args.train_batch_size is None:
+    args.train_batch_size = args.batch_size
 else:
-    assert args.batch_size % args.physical_batch_size == 0
+    assert args.batch_size % args.train_batch_size == 0
 
 assert args.nb_train_samples % args.batch_size == 0
 assert args.nb_test_samples % args.batch_size == 0
 
-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,
-)
+######################################################################
+
+
+def optimizer_to(optim, device):
+    """Move the optimizer optim to the 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)
+
 
 ######################################################################
 
-log_string(f"device {device}")
 
-vocabulary_size = task.vocabulary_size()
+def generate_quiz_set(nb_samples, c_quizzes, c_quiz_multiplier=1):
+    if c_quizzes is None:
+        quizzes = problem.generate_w_quizzes(nb_samples)
+        nb_w_quizzes = quizzes.size(0)
+        nb_c_quizzes = 0
+    else:
+        if c_quiz_multiplier > 1:
+            n = min(c_quiz_multiplier, (nb_samples // 2) // c_quizzes.size(0))
+            body = c_quizzes.repeat(n, 1)
+            if n < c_quiz_multiplier:
+                tail = c_quizzes[
+                    torch.randperm(c_quizzes.size(0))[: nb_samples // 2 - body.size(0)]
+                ]
+                c_quizzes = torch.cat([body, tail], dim=0)
+            else:
+                c_quizzes = body
+
+        if c_quizzes.size(0) > nb_samples // 2:
+            i = torch.randperm(c_quizzes.size(0))[: nb_samples // 2]
+            c_quizzes = c_quizzes[i]
+
+        w_quizzes = problem.generate_w_quizzes(nb_samples - c_quizzes.size(0))
+
+        quizzes = torch.cat([w_quizzes, c_quizzes], dim=0)
+        nb_w_quizzes = w_quizzes.size(0)
+        nb_c_quizzes = c_quizzes.size(0)
+
+    i = torch.randperm(quizzes.size(0), device=quizzes.device)
+    quizzes = quizzes[i].contiguous()
+
+    log_string(f"quiz_set nb_w_quizzes {nb_w_quizzes} nb_c_quizzes {nb_c_quizzes}")
+
+    return quizzes
 
-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))
-token_probas = token_count / token_count.sum()
-entropy = -torch.xlogy(token_probas, token_probas).sum()
-train_set_perplexity = math.exp(entropy)
+def add_hints_imt(imt_set):
+    """Set every component of the mask to zero with probability
+    args.proba_hint, and for each component set to zero, copy the
+    corresponding value from the target into the input
+
+    """
+    input, masks, targets = imt_set.unbind(dim=1)
+    # h = torch.rand(masks.size(), device=masks.device) - masks
+    # t = h.sort(dim=1).values[:, args.nb_hints, None]
+    # mask_hints = (h < t).long()
+    mask_hints = (
+        torch.rand(input.size(), device=input.device) < args.proba_hint
+    ).long() * masks
+    masks = (1 - mask_hints) * masks
+    input = (1 - mask_hints) * input + mask_hints * targets
+    return torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1)
+
+
+def add_noise_imt(imt_set):
+    """Replace every component of the input by a random value with
+    probability args.proba_prompt_noise."""
+    input, masks, targets = imt_set.unbind(dim=1)
+    noise = problem.pure_noise(input.size(0), input.device)
+    change = (1 - masks) * (
+        torch.rand(input.size(), device=input.device) < args.proba_prompt_noise
+    ).long()
+    input = (1 - change) * input + change * noise
+    return torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1)
+
 
 ######################################################################
-# A bit of paranoia never hurts
-
-if args.max_percents_of_test_in_train >= 0:
-
-    def subsets_as_tuples(batches, cs):
-        s = set()
-        for batch in batches:
-            for x in batch:
-                s.add(tuple([v.item() for v in x]))
-                if len(s) == cs:
-                    yield s
-                    s = set()
-        yield s
-
-    nb_test, nb_in_train = 0, 0
-    for test_subset in subsets_as_tuples(
-        task.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
-        ):
-            in_train.update(test_subset.intersection(train_subset))
-        nb_in_train += len(in_train)
-        nb_test += len(test_subset)
+# Prediction
 
-    log_string(
-        f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
-    )
 
-    assert (
-        nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
-    ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
+def samples_for_prediction_imt(input):
+    nb = input.size(0)
+    masks = input.new_zeros(input.size())
+    u = F.one_hot(torch.randint(4, (nb,), device=masks.device), num_classes=4)
+    masks.view(nb, 4, -1)[...] = u[:, :, None]
+    targets = input
+    input = (1 - masks) * targets
 
-##############################
+    return torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1)
 
 
-def one_epoch(model, task):
-    optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
+def ae_predict(model, imt_set, local_device=main_device):
+    model.eval().to(local_device)
 
-    model.train()
+    record = []
 
-    nb_train_samples, acc_train_loss = 0, 0.0
+    src = tqdm.tqdm(
+        imt_set.split(args.eval_batch_size),
+        dynamic_ncols=True,
+        desc="predict",
+        total=imt_set.size(0) // args.eval_batch_size,
+        delay=10,
+    )
+
+    for imt in src:
+        # some paranoia
+        imt = imt.clone()
+        imt[:, 0] = imt[:, 0] * (1 - imt[:, 1])
 
-    for input in task.batches(split="train"):
-        input = input.to(device)
+        with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
+            logits = model(imt[:, 0] * 2 + imt[:, 1])
+        dist = torch.distributions.categorical.Categorical(logits=logits)
+        result = (1 - imt[:, 1]) * imt[:, 0] + imt[:, 1] * dist.sample()
+        record.append(result)
 
-        if nb_train_samples % args.batch_size == 0:
-            optimizer.zero_grad()
+    return torch.cat(record)
 
-        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)
+def predict_the_four_grids(
+    model, input, with_noise=False, with_hints=False, local_device=main_device
+):
+    input = input[:, None, :].expand(-1, 4, -1).reshape(-1, input.size(1))
+    nb = input.size(0)
+    masks = input.new_zeros(input.size())
+    u = F.one_hot(torch.arange(nb, device=masks.device) % 4, num_classes=4)
+    masks.view(nb, 4, -1)[...] = u[:, :, None]
+    targets = input
+    input = (1 - masks) * targets
+    imt_set = torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1)
 
-        loss.backward()
+    if with_hints:
+        imt_set = add_hints_imt(imt_set)
 
-        if nb_train_samples % args.batch_size == 0:
-            optimizer.step()
+    if with_noise:
+        imt_set = add_noise_imt(imt_set)
 
-    train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+    result = ae_predict(model, imt_set, local_device=local_device)
+    result = (result * masks).reshape(-1, 4, result.size(1)).sum(dim=1)
 
-    log_string(f"train_perplexity {n_epoch} {train_perplexity}")
+    return result
 
 
 ######################################################################
 
 
-def run_tests(model, task, deterministic_synthesis):
-    with torch.autograd.no_grad():
-        model.eval()
+def samples_for_generation_imt(input):
+    nb = input.size(0)
+    probs_iterations = 0.1 ** torch.linspace(
+        0, 1, args.diffusion_nb_iterations, device=input.device
+    )
+    probs_iterations = probs_iterations[None, :] / probs_iterations.sum()
+    probs_iterations = probs_iterations.expand(nb, -1)
+    dist = torch.distributions.categorical.Categorical(probs=probs_iterations)
+    t = dist.sample() + 1
+    r = torch.rand(input.size(), device=input.device)
+    proba_erased = 1 - (1 - args.diffusion_proba_corruption) ** t
+    mask_erased = (r <= proba_erased[:, None]).long()
 
-        nb_test_samples, acc_test_loss = 0, 0.0
-        nb_samples_accumulated = 0
+    noise = problem.pure_noise(nb, input.device)
+    targets = input
+    input = (1 - mask_erased) * input + mask_erased * noise
+    masks = input.new_full(input.size(), 1)
 
-        for input in task.batches(split="test"):
-            input = input.to(device)
+    return torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1)
 
-            bs = model(mygpt.BracketedSequence(input))
-            output = bs.x
 
-            loss = F.cross_entropy(output.transpose(1, 2), input)
+def prioritized_rand(low):
+    x = torch.rand(low.size(), device=low.device).sort(dim=1, descending=True).values
+    k = torch.rand(low.size(), device=low.device) + low.long()
+    k = k.sort(dim=1).indices
+    y = x.new(x.size())
+    y.scatter_(dim=1, index=k, src=x)
+    return y
 
-            acc_test_loss += loss.item() * input.size(0)
 
-            nb_test_samples += input.size(0)
+def ae_generate(model, nb, local_device=main_device):
+    model.eval().to(local_device)
 
-        main_test_accuracy = task.produce_results(
-            n_epoch=n_epoch,
-            model=model,
-            result_dir=args.result_dir,
-            logger=log_string,
-            deterministic_synthesis=deterministic_synthesis,
+    # We loop through the iterations first and through the
+    # mini-batches second so that we keep only the samples that have
+    # not stabilized
+
+    all_input = problem.pure_noise(nb, local_device)
+    all_masks = all_input.new_full(all_input.size(), 1)
+    all_changed = torch.full((all_input.size(0),), True, device=all_input.device)
+
+    for it in range(args.diffusion_nb_iterations):
+        # log_string(f"nb_changed {all_changed.long().sum().item()}")
+
+        if not all_changed.any():
+            break
+
+        sub_input = all_input[all_changed].clone()
+        sub_masks = all_masks[all_changed].clone()
+        sub_changed = all_changed[all_changed].clone()
+
+        src = zip(
+            sub_input.split(args.eval_batch_size),
+            sub_masks.split(args.eval_batch_size),
+            sub_changed.split(args.eval_batch_size),
         )
 
-        test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+        for input, masks, changed in src:
+            with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
+                logits = model(input * 2 + masks)
+            dist = torch.distributions.categorical.Categorical(logits=logits)
+            output = dist.sample()
+            r = prioritized_rand(input != output)
+            mask_changes = (r <= args.diffusion_proba_corruption).long() * masks
+            update = (1 - mask_changes) * input + mask_changes * output
+            changed[...] = changed & (update != input).max(dim=1).values
+            input[...] = update
 
-        log_string(f"test_perplexity {n_epoch} {test_perplexity}")
+        a = all_changed.clone()
+        all_input[a] = sub_input
+        all_masks[a] = sub_masks
+        all_changed[a] = sub_changed
 
-    model.main_test_accuracy = main_test_accuracy
+    return all_input
 
 
 ######################################################################
 
 
-def create_c_quizzes(
-    model,
-    other_models,
-    task,
-    nb_for_train=1000,
-    nb_for_test=100,
-    desired_average_logits=None,
+def one_epoch(model, n_epoch, c_quizzes, train=True, local_device=main_device):
+    quizzes = generate_quiz_set(
+        args.nb_train_samples if train else args.nb_test_samples,
+        c_quizzes,
+        args.c_quiz_multiplier,
+    )
+
+    q_p, q_g = quizzes.to(local_device).chunk(2)
+
+    # Half of the samples train the prediction, and we inject noise in
+    # all, and hints in half
+    b_p = samples_for_prediction_imt(q_p)
+    b_p = add_noise_imt(b_p)
+    half = torch.rand(b_p.size(0)) < 0.5
+    b_p[half] = add_hints_imt(b_p[half])
+
+    # The other half are denoising examples for the generation
+    b_g = samples_for_generation_imt(q_g)
+
+    imt_set = torch.cat([b_p, b_g])
+    imt_set = imt_set[torch.randperm(imt_set.size(0), device=imt_set.device)]
+
+    if train:
+        label = "train"
+        model.train().to(local_device)
+        optimizer_to(model.optimizer, local_device)
+        batch_size = args.train_batch_size
+    else:
+        label = "test"
+        model.eval().to(local_device)
+        batch_size = args.eval_batch_size
+
+    nb_samples, acc_loss = 0, 0.0
+
+    for imt in tqdm.tqdm(
+        imt_set.split(batch_size),
+        dynamic_ncols=True,
+        desc=label,
+        total=quizzes.size(0) // batch_size,
+        delay=10,
+    ):
+        input, masks, targets = imt.unbind(dim=1)
+        if train and nb_samples % args.batch_size == 0:
+            model.optimizer.zero_grad()
+
+        with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
+            logits = model(input * 2 + masks)
+
+        loss_per_token = F.cross_entropy(
+            logits.transpose(1, 2), targets, reduction="none"
+        )
+        loss = (loss_per_token * masks).mean()
+        acc_loss += loss.item() * imt.size(0)
+        nb_samples += imt.size(0)
+
+        if train:
+            loss.backward()
+
+            if nb_samples % args.batch_size == 0:
+                model.optimizer.step()
+
+    log_string(f"{label}_loss {n_epoch} model {model.id} {acc_loss/nb_samples}")
+
+
+######################################################################
+
+
+def save_inference_images(model, n_epoch, c_quizzes, c_quiz_multiplier, local_device):
+    # Save some images of the prediction results
+
+    quizzes = generate_quiz_set(150, c_quizzes, args.c_quiz_multiplier)
+    imt_set = samples_for_prediction_imt(quizzes.to(local_device))
+    result = ae_predict(model, imt_set, local_device=local_device).to("cpu")
+    masks = imt_set[:, 1].to("cpu")
+
+    correct = (quizzes == result).min(dim=1).values.long()
+    correct_parts = (2 * correct - 1)[:, None] * masks.reshape(masks.size(0), 4, -1)[
+        :, :, 1
+    ]
+    predicted_parts = correct_parts.abs()
+
+    problem.save_quizzes_as_image(
+        args.result_dir,
+        f"culture_prediction_{n_epoch}_{model.id}.png",
+        quizzes=result[:128],
+        predicted_parts=predicted_parts[:128],
+        correct_parts=correct_parts[:128],
+    )
+
+    # Save some images of the ex nihilo generation of the four grids
+
+    result = ae_generate(model, 150, local_device=local_device).to("cpu")
+    problem.save_quizzes_as_image(
+        args.result_dir,
+        f"culture_generation_{n_epoch}_{model.id}.png",
+        quizzes=result[:128],
+    )
+
+
+######################################################################
+
+
+def one_complete_epoch(
+    model, n_epoch, train_c_quizzes, test_c_quizzes, local_device=main_device
 ):
-    kept = []
+    one_epoch(model, n_epoch, train_c_quizzes, train=True, local_device=local_device)
+
+    one_epoch(model, n_epoch, test_c_quizzes, train=False, local_device=local_device)
+
+    # Compute the test accuracy
+
+    quizzes = generate_quiz_set(args.nb_test_samples, c_quizzes, args.c_quiz_multiplier)
+    imt_set = samples_for_prediction_imt(quizzes.to(local_device))
+    result = ae_predict(model, imt_set, local_device=local_device).to("cpu")
+    correct = (quizzes == result).min(dim=1).values.long()
+
+    nb_correct, nb_total = correct.sum().item(), quizzes.size(0)
+    model.test_accuracy = nb_correct / nb_total
+
+    log_string(
+        f"test_accuracy {n_epoch} model {model.id} nb_correct {nb_correct} / {nb_total} ({model.test_accuracy*100:.02f}%)"
+    )
+
+    save_inference_images(
+        model, n_epoch, c_quizzes, args.c_quiz_multiplier, local_device=local_device
+    )
+
+
+######################################################################
+
+
+def max_nb_mistakes_on_one_grid(quizzes, prediction):
+    return (
+        (prediction != quizzes)
+        .long()
+        .reshape(quizzes.size(0), 4, -1)
+        .sum(dim=2)
+        .max(dim=1)
+        .values
+    )
 
-    sum_logits, sum_nb_c_quizzes = 0, 0
 
-    while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
-        nb_to_generate = 4 * (nb_for_train + nb_for_test)
+def evaluate_quizzes(quizzes, models, with_hints, local_device):
+    nb_correct, nb_wrong = 0, 0
 
-        new_c_quizzes, nb_correct, average_logits = task.create_c_quizzes(
-            n_epoch=n_epoch,
-            result_dir=args.result_dir,
-            logger=log_string,
-            nb=nb_to_generate,
+    for model in models:
+        model = copy.deepcopy(model).to(local_device).eval()
+        predicted = predict_the_four_grids(
             model=model,
-            other_models=other_models,
-            desired_average_logits=desired_average_logits,
+            input=quizzes,
+            with_noise=False,
+            with_hints=with_hints,
+            local_device=local_device,
         )
+        nb_mistakes = max_nb_mistakes_on_one_grid(quizzes, predicted)
+        nb_correct += (nb_mistakes == 0).long()
+        nb_wrong += (nb_mistakes >= args.nb_mistakes_to_be_wrong).long()
+
+    # print("\n\n", nb_correct, nb_wrong)
+
+    return nb_correct, nb_wrong
+
+
+######################################################################
+
+
+def identity_quizzes(quizzes):
+    quizzes = quizzes.reshape(quizzes.size(0), 4, -1)
+    return (quizzes[:, 0] == quizzes[:, 1]).min(dim=1).values | (
+        quizzes[:, 2] == quizzes[:, 3]
+    ).min(dim=1).values
 
-        sum_logits += new_c_quizzes.size(0) * average_logits
-        sum_nb_c_quizzes += new_c_quizzes.size(0)
 
-        to_keep = new_c_quizzes[nb_correct == len(other_models) - 1]
+def generate_c_quizzes(models, nb_to_generate, local_device=main_device):
+    record = []
+    nb_validated = 0
 
-        if args.dirty_debug:
-            to_keep = new_c_quizzes
+    start_time = time.perf_counter()
+    last_log = -1
 
-        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}%)"
+    while nb_validated < nb_to_generate:
+        # Generate new quizzes
+
+        model = models[torch.randint(len(models), (1,)).item()]
+        model = copy.deepcopy(model).to(local_device).eval()
+        generator_id = model.id
+
+        c_quizzes = ae_generate(
+            model=model, nb=args.eval_batch_size * 10, local_device=local_device
         )
 
-        kept.append(to_keep)
+        c_quizzes = c_quizzes[identity_quizzes(c_quizzes) == False]
+
+        if c_quizzes.size(0) > 0:
+            # Select the ones that are solved properly by some models and
+            # not understood by others
+
+            nb_correct, nb_wrong = evaluate_quizzes(
+                quizzes=c_quizzes,
+                models=models,
+                with_hints=True,
+                local_device=local_device,
+            )
+
+            to_keep = (nb_correct >= args.nb_have_to_be_correct) & (
+                nb_wrong >= args.nb_have_to_be_wrong
+            )
+
+            nb_validated += to_keep.long().sum().item()
+            record.append(c_quizzes[to_keep])
+
+        #####################
+
+        duration = time.perf_counter() - start_time
+
+        if last_log < 0 or duration > last_log + 10:
+            last_log = duration
+            if nb_validated > 0:
+                if nb_validated < nb_to_generate:
+                    d = (nb_to_generate - nb_validated) * duration / nb_validated
+                    e = (
+                        datetime.datetime.now() + datetime.timedelta(seconds=d)
+                    ).strftime("%a %H:%M")
+                else:
+                    e = "now!"
+            else:
+                e = "???"
+
+            log_string(
+                f"nb_validated {nb_validated} model {generator_id} (finishes {e} -- {int((nb_validated * 3600)/duration)}/h)"
+            )
+
+        #####################
+
+    duration = time.perf_counter() - start_time
 
-    new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
+    log_string(f"generate_c_quizz_speed {int(3600 * nb_validated / duration)}/h")
+
+    return torch.cat(record).to("cpu")
+
+
+######################################################################
+
+
+def multithread_execution(fun, arguments):
+    # Single instance, no thread
+    if len(arguments) == 1:
+        return fun(*(arguments[0]))
+
+    records, threads = [], []
+
+    def threadable_fun(*args):
+        r = fun(*args)
+        if type(r) is not tuple:
+            r = (r,)
+        records.append(r)
+
+    for args in arguments:
+        # To get a different sequence between threads
+        log_string(f"dummy_rand {torch.rand(1)}")
+        # torch.rand(1)
+        t = threading.Thread(target=threadable_fun, daemon=True, args=args)
+        threads.append(t)
+        t.start()
+
+    for t in threads:
+        t.join()
+
+    if records[0] == (None,):
+        return
+    else:
+        return [
+            torch.cat([x[k] for x in records], dim=0) for k in range(len(records[0]))
+        ]
+
+
+######################################################################
+
+
+def save_models(models, suffix=""):
+    if suffix != "":
+        suffix = "_" + suffix
+
+    for model in models:
+        filename = f"ae_{model.id:03d}{suffix}.pth"
+        torch.save(
+            {
+                "state_dict": model.state_dict(),
+                "optimizer_state_dict": model.optimizer.state_dict(),
+                "test_accuracy": model.test_accuracy,
+            },
+            os.path.join(args.result_dir, filename),
+        )
+
+    log_string(f"wrote ae_*{suffix}.pth")
+
+
+######################################################################
+
+
+def save_quiz_image(models, c_quizzes, filename, local_device=main_device):
+    c_quizzes = c_quizzes.to(local_device)
+
+    nb_correct, nb_wrong = evaluate_quizzes(
+        quizzes=c_quizzes,
+        models=models,
+        with_hints=False,
+        local_device=local_device,
+    )
 
-    task.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
-    task.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
+    comments = [f"nb_correct {c} nb_wrong {w}" for c, w in zip(nb_correct, nb_wrong)]
 
-    task.save_quizzes(
-        new_c_quizzes[:72],
+    problem.save_quizzes_as_image(
         args.result_dir,
-        f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",
-        log_string,
+        filename,
+        quizzes=c_quizzes,
+        comments=comments,
+        delta=True,
+        nrow=8,
     )
 
-    return sum_logits / sum_nb_c_quizzes
+    log_string(f"wrote {filename}")
 
 
+######################################################################
+
+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 not args.resume:
+    problem.save_some_examples(args.result_dir)
+
+
+log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}")
+
+vocabulary_size = problem.vocabulary_size()
+
+log_string(f"vocabulary_size {vocabulary_size}")
+
 ######################################################################
 
 models = []
 
-for k in range(args.nb_gpts):
-    model = mygpt.MyGPT(
-        vocabulary_size=vocabulary_size,
+if args.model_type == "standard":
+    model_constructor = attae.AttentionAE
+elif args.model_type == "functional":
+    model_constructor = attae.FunctionalAttentionAE
+else:
+    raise ValueError(f"Unknown model type {args.model_type}")
+
+
+for i in range(args.nb_models):
+    model = model_constructor(
+        vocabulary_size=vocabulary_size * 2,
         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
+    # model = torch.compile(model)
+
+    model.id = i
+    model.test_accuracy = 0.0
+    model.optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 
     models.append(model)
 
+######################################################################
+
+current_epoch = 0
+
+if args.resume:
+    for model in models:
+        filename = f"ae_{model.id:03d}.pth"
+
+        d = torch.load(
+            os.path.join(args.result_dir, filename),
+            map_location="cpu",
+            weights_only=False,
+        )
+        model.load_state_dict(d["state_dict"])
+        model.optimizer.load_state_dict(d["optimizer_state_dict"])
+        model.test_accuracy = d["test_accuracy"]
+        log_string(f"successfully loaded {filename}")
+
+    filename = "state.pth"
+    state = torch.load(
+        os.path.join(args.result_dir, filename),
+        map_location="cpu",
+        weights_only=False,
+    )
+
+    log_string(f"successfully loaded {filename}")
+
+    current_epoch = state["current_epoch"]
+    train_c_quizzes = state["train_c_quizzes"]
+    test_c_quizzes = state["test_c_quizzes"]
+
+######################################################################
 
 nb_parameters = sum(p.numel() for p in models[0].parameters())
 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 
+
 ######################################################################
 
-desired_average_logits = None
+train_c_quizzes, test_c_quizzes = None, None
+
+######################################################################
+
+for n_epoch in range(current_epoch, args.nb_epochs):
+    start_time = time.perf_counter()
+
+    state = {
+        "current_epoch": n_epoch,
+        "train_c_quizzes": train_c_quizzes,
+        "test_c_quizzes": test_c_quizzes,
+    }
+
+    filename = "state.pth"
+    torch.save(state, os.path.join(args.result_dir, filename))
+    log_string(f"wrote {filename}")
 
-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}")
+    cta = " ".join([f"{float(m.test_accuracy):.04f}" for m in models])
+    log_string(f"current_test_accuracies {cta}")
 
-    # 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}"
-    )
+    lowest_test_accuracy = min([float(m.test_accuracy) for m in models])
 
-    # improve it
-    one_epoch(model, task)
+    if lowest_test_accuracy >= args.accuracy_to_make_c_quizzes:
+        if train_c_quizzes is None:
+            save_models(models, "naive")
 
-    task.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+        nb_gpus = len(gpus)
+        nb_c_quizzes_to_generate = (args.nb_c_quizzes + nb_gpus - 1) // nb_gpus
 
-    log_string(
-        f"train_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}"
-    )
+        (new_c_quizzes,) = multithread_execution(
+            generate_c_quizzes,
+            [(models, nb_c_quizzes_to_generate, gpu) for gpu in gpus],
+        )
 
-    # test it
-    run_tests(model, task, deterministic_synthesis=False)
+        save_quiz_image(
+            models, new_c_quizzes[:256], f"culture_c_quiz_{n_epoch:04d}.png"
+        )
 
-    log_string(
-        f"test_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}"
-    )
+        log_string(f"generated_c_quizzes {new_c_quizzes.size()}")
 
-    if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
-        other_models = models.copy()
-        other_models.remove(model)
-
-        average_logits = create_c_quizzes(
-            model,
-            other_models,
-            task,
-            nb_for_train=nb_new_c_quizzes_for_train,
-            nb_for_test=nb_new_c_quizzes_for_test,
-            desired_average_logits=desired_average_logits,
+        train_c_quizzes = (
+            new_c_quizzes
+            if train_c_quizzes is None
+            else torch.cat([train_c_quizzes, new_c_quizzes])
         )
+        train_c_quizzes = train_c_quizzes[-args.nb_train_samples :]
 
-        # We keep the first average logits as a reference
-        if desired_average_logits is None:
-            desired_average_logits = average_logits
-        else:
-            log_string(
-                f"desired_average_logits {desired_average_logits} average_logits {average_logits}"
-            )
+        nb_correct, _ = evaluate_quizzes(
+            quizzes=train_c_quizzes,
+            models=models,
+            with_hints=False,
+            local_device=local_device,
+        )
+
+        test_c_quizzes = train_c_quizzes[nb_correct >= args.nb_have_to_be_correct]
 
-        # We update everyone
         for model in models:
-            run_tests(model, task, deterministic_synthesis=False)
+            model.test_accuracy = 0
 
+    if train_c_quizzes is None:
+        log_string("no_c_quiz")
+    else:
+        log_string(f"nb_c_quizzes {train_c_quizzes.size(0)}")
 
-######################################################################
+    # --------------------------------------------------------------------
+
+    ranked_models = sorted(models, key=lambda m: float(m.test_accuracy))
+    weakest_models = ranked_models[: len(gpus)]
+
+    log_string(
+        f"weakest_accuracies {[model.test_accuracy for model in weakest_models]}"
+    )
+
+    multithread_execution(
+        one_complete_epoch,
+        [
+            (model, n_epoch, train_c_quizzes, test_c_quizzes, gpu)
+            for model, gpu in zip(weakest_models, gpus)
+        ],
+    )
+
+    save_models(models)
+
+    # --------------------------------------------------------------------
+
+    duration = time.perf_counter() - start_time
+    str_duration = ""
+    if duration >= 60:
+        str_duration += f"{int(duration)//60}min"
+    str_duration += f"{int(duration)%60}s"
+    str_next = (
+        datetime.datetime.now() + datetime.timedelta(seconds=duration)
+    ).strftime("%H:%M:%S")
+    log_string(f"epoch_duration {str_duration} next_finish {str_next}")